<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: Mindutil</title>
    <description>The latest articles on DEV Community by Mindutil (@li_yi_dcec96bf2647dec1b7e).</description>
    <link>https://dev.to/li_yi_dcec96bf2647dec1b7e</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4174928%2F40eeeb0e-a4ba-4e3c-9fe5-aed702539161.png</url>
      <title>DEV Community: Mindutil</title>
      <link>https://dev.to/li_yi_dcec96bf2647dec1b7e</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/li_yi_dcec96bf2647dec1b7e"/>
    <language>en</language>
    <item>
      <title>I rewrote my 20 highest-traffic pages. The other 130 still show three sentences I never wrote.</title>
      <dc:creator>Mindutil</dc:creator>
      <pubDate>Fri, 20 Nov 2026 00:36:33 +0000</pubDate>
      <link>https://dev.to/li_yi_dcec96bf2647dec1b7e/i-rewrote-my-20-highest-traffic-pages-the-other-130-still-show-three-sentences-i-never-wrote-bea</link>
      <guid>https://dev.to/li_yi_dcec96bf2647dec1b7e/i-rewrote-my-20-highest-traffic-pages-the-other-130-still-show-three-sentences-i-never-wrote-bea</guid>
      <description>&lt;p&gt;My directory lists 4,323 AI tools. Every day I get one hour to make the highest-traffic pages better, and I have been spending that hour wrong. Here is the number that finally showed me.&lt;/p&gt;

&lt;h2&gt;
  
  
  What "written" means on my site
&lt;/h2&gt;

&lt;p&gt;Each listing has a description body. Most of them were filled in by a crawler, and the crawler copy is a template with the product name slotted in. My enrichment queue (&lt;code&gt;scripts/seo-queue.mjs&lt;/code&gt;) strips four known boilerplate sentences, a templated opening line, affiliate-tracking URLs, and the tagline; whatever word count is left is the only part I'm willing to call "written." Target is 80+ unique words.&lt;/p&gt;

&lt;p&gt;I ranked all 4,323 rows by the mirrored read count, took the top 150, and ran that filter. The result:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;20 pages pass. 130 don't.&lt;/li&gt;
&lt;li&gt;Those 130 templated pages are 86.7% of the top 150 by count.&lt;/li&gt;
&lt;li&gt;They are 64.0% of the top 150 by traffic — 14,103,474 of 22,025,504 mirrored reads.&lt;/li&gt;
&lt;li&gt;My 20 hand-written pages, the ones I've been grinding on, hold 36.0%.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;So the effort curve and the traffic curve point in opposite directions. I write a page, feel good, and it's one of the lower-read ones in the set. The bulk of the people landing on my "best of" rows are reading text a script wrote.&lt;/p&gt;

&lt;h2&gt;
  
  
  What a reader actually sees on an unwritten page
&lt;/h2&gt;

&lt;p&gt;The single most-read page I haven't rewritten is "Muse Wonder" — 211,026 mirrored reads. The raw description in my database runs longer than the reader sees, because my renderer (the same filter) drops the boilerplate sentences at render time. After filtering, the entire "About" paragraph a visitor reads is three sentences:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;Muse Wonder is a freemium AI tool in the LLM models category. It currently ranks #4 in the LLM models category. Common tags for this tool include Latest AI.&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;I pulled that off the live page with curl, no cookies, to make sure I wasn't describing the database instead of the site. Those three sentences contain: one price label, one category name repeated twice, one rank, and one tag. There is no fact in there about what the product does that the vendor would recognize as theirs. And the next most-read unwritten pages — "Gemini 4 Argon" at 210,792 reads, "GPT-6 Sol and Luna" at 207,352 — render the identical shape with a different proper noun.&lt;/p&gt;

&lt;h2&gt;
  
  
  The part that actually stings
&lt;/h2&gt;

&lt;p&gt;81 of those 130 templated pages carry a "claimed" badge — the vendor came to my site, verified they own the listing, and said "yes, this is ours." Some of those show "Verified Gold."&lt;/p&gt;

&lt;p&gt;Think about the sequence of events that implies. A vendor was enough of a human to find their listing and claim it. Then they left. Nobody — not them, not me — put a single real sentence about their product on the page the claimer's own customers would land on. The claim badge and the empty body coexist on 62% of my unwritten top pages. The overall claim rate across the whole directory is only 14.2% (614 of 4,323), so this isn't random noise: high-traffic rows are far more likely to be claimed and far more likely to still be a template.&lt;/p&gt;

&lt;h2&gt;
  
  
  It isn't a free-vs-paid story
&lt;/h2&gt;

&lt;p&gt;I expected the gap to sit in one pricing bucket, which would at least make it a clean bug. It doesn't. The 130 templated pages split free 67 / freemium 44 / paid 19. The free bucket is biggest simply because free is the biggest bucket overall — about half my directory is labeled free. So the /free-ai-tools list people filter on is, for its highest-traffic rows, a list of pages whose only body is those three template sentences. The filter is honest about the price label (that comes from the claim, where it can be); it says nothing about whether the description is worth reading, because for 67 of them there isn't one.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'm changing
&lt;/h2&gt;

&lt;p&gt;The queue already sorts by mirrored reads, so it has been handing me the right order — I was just treating it as a "nice to have" backlog instead of the top of a funnel. Two concrete changes:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Stop counting "written" per page and count it per read. My daily hour should be judged by how much of the 14.1M-read templated block I convert, not by how many pages leave the todo list. Converting the top 10 by reads is worth more than converting 40 near the tail.&lt;/li&gt;
&lt;li&gt;Prioritize the 81 claimed-but-empty pages specifically. That's a list I can email, because the claim already gave me a contact and a demonstrated willingness to engage.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;None of the fix is clever. The number that was missing is the traffic-weight one. A page with zero unique words and 211,026 reads is not a minor content task — it's the most-read thing on my site that I refuse to describe.&lt;/p&gt;

&lt;p&gt;I'm keeping the whole thing reproducible rather than asking you to trust my arithmetic: the same filter lives in one file (&lt;code&gt;scripts/seo-queue.mjs&lt;/code&gt;), and every figure above is &lt;code&gt;findMany&lt;/code&gt; over one SQLite table plus a cookie-less curl against a live URL. The full directory, including the free filter that 67 of these pages feed, is at &lt;a href="https://aiverse.mindutil.com/free-ai-tools?utm_source=devto&amp;amp;utm_medium=social&amp;amp;utm_campaign=topic" rel="noopener noreferrer"&gt;aiverse.mindutil.com/free-ai-tools&lt;/a&gt; if you want to go find a three-sentence page yourself. The experiment I'd rather run is the boring one: rewrite the top ten, then check whether the reads move.&lt;/p&gt;

</description>
      <category>webdev</category>
      <category>ai</category>
      <category>seo</category>
      <category>datascience</category>
    </item>
    <item>
      <title>Five of my most-read catalog pages 404 because of a hyphen I never chose</title>
      <dc:creator>Mindutil</dc:creator>
      <pubDate>Thu, 19 Nov 2026 00:50:55 +0000</pubDate>
      <link>https://dev.to/li_yi_dcec96bf2647dec1b7e/five-of-my-most-read-catalog-pages-404-because-of-a-hyphen-i-never-chose-4jjk</link>
      <guid>https://dev.to/li_yi_dcec96bf2647dec1b7e/five-of-my-most-read-catalog-pages-404-because-of-a-hyphen-i-never-chose-4jjk</guid>
      <description>&lt;p&gt;On the first page of the category that carries more mirrored reads than any other in my catalog — LLM models, 208 tools, 24 cards rendered per page — three of those cards link to pages that return 404. Cards #5, #15 and #18. Not a redirect, not a slow render: a 404. All three are among the most-read rows in the whole catalog.&lt;/p&gt;

&lt;p&gt;I want to write up exactly what happened, because the cause is a single character, and the character is not mine.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the URLs look like
&lt;/h2&gt;

&lt;p&gt;My catalog has 4,323 rows. Five of them have a slug that contains a percent sign:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;gpt%e2%80%916-1-sol-openai
gpt%e2%80%915-6-openai
gpt%e2%80%915-5
gpt%e2%80%915-5-instant-openai
gpt%e2%80%91rosalind
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;%e2%80%91&lt;/code&gt; is the UTF-8 percent-encoding of U+2011, the NON-BREAKING HYPHEN. So the stored slug is not &lt;code&gt;gpt-6-1-sol-openai&lt;/code&gt; with a normal hyphen — it is the string &lt;code&gt;gpt%e2%80%916-1-sol-openai&lt;/code&gt;, percent signs included, sitting in a &lt;code&gt;@unique&lt;/code&gt; column.&lt;/p&gt;

&lt;p&gt;Here is what those five rows are worth to me:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;name&lt;/th&gt;
&lt;th&gt;rank by mirrored reads&lt;/th&gt;
&lt;th&gt;reads&lt;/th&gt;
&lt;th&gt;price label&lt;/th&gt;
&lt;th&gt;vendor-claimed badge&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;GPT‑6.1 Sol&lt;/td&gt;
&lt;td&gt;22&lt;/td&gt;
&lt;td&gt;210,984&lt;/td&gt;
&lt;td&gt;freemium&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT‑5.6&lt;/td&gt;
&lt;td&gt;34&lt;/td&gt;
&lt;td&gt;200,819&lt;/td&gt;
&lt;td&gt;paid&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT‑5.5&lt;/td&gt;
&lt;td&gt;37&lt;/td&gt;
&lt;td&gt;198,748&lt;/td&gt;
&lt;td&gt;paid&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT‑5.5 Instant&lt;/td&gt;
&lt;td&gt;51&lt;/td&gt;
&lt;td&gt;180,819&lt;/td&gt;
&lt;td&gt;freemium&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;GPT‑Rosalind&lt;/td&gt;
&lt;td&gt;521&lt;/td&gt;
&lt;td&gt;15,405&lt;/td&gt;
&lt;td&gt;paid&lt;/td&gt;
&lt;td&gt;yes&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;806,775 mirrored reads between them — 1.62% of every read counted across the site. ("Mirrored": that counter is copied from the source directory, not measured on my own traffic.) Four of the five sit inside the Top 150 list I use to decide which pages to write properly. All five carry my "certified, vendor claimed" badge. All five have &lt;code&gt;clickCount = 0&lt;/code&gt;, which is the least surprising number in this table.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why it 404s
&lt;/h2&gt;

&lt;p&gt;The link is built with a template string:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;href&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="sb"&gt;`/tool/${tool.slug}`&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;So the HTML contains &lt;code&gt;/tool/gpt%e2%80%915-5&lt;/code&gt;. That is a &lt;em&gt;valid&lt;/em&gt; URL — it is the correct encoding of the path segment &lt;code&gt;gpt‑5-5&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;The server receives the request, decodes it, and now holds the decoded string &lt;code&gt;gpt‑5-5&lt;/code&gt;. It looks that up in the slug column, which contains the encoded string &lt;code&gt;gpt%e2%80%915-5&lt;/code&gt;. No match. 404.&lt;/p&gt;

&lt;p&gt;The value was stored already-encoded, and every consumer decodes once. Two representations of the same name, permanently one decode apart. Nothing in the request path is wrong; the database is holding the escaped form of its own identifier.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the character came from
&lt;/h2&gt;

&lt;p&gt;I did not type it. The crawler builds a tool's slug by taking the last path segment of the source directory's permalink, verbatim:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;slug&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="no"&gt;URL&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;sourceUrl&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;pathname&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;replace&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;/^&lt;/span&gt;\&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="n"&gt;en&lt;/span&gt;\&lt;span class="c1"&gt;//, "").replace(/\/+$/, "");&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;And the source URLs in my database are, literally:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;https://www.aixploria.com/en/gpt%e2%80%916-1-sol-openai/
https://www.aixploria.com/en/gpt%e2%80%915-5/
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Upstream had the same bug one level higher, and published percent-encoded permalinks. I inherited them as primary keys.&lt;/p&gt;

&lt;p&gt;The hyphen itself starts one level higher still. OpenAI writes its model names with U+2011 in some places. Evidence inside my own table: I hold both &lt;code&gt;GPT-5.3 Instant&lt;/code&gt; (plain U+002D) and &lt;code&gt;GPT‑5.5 Instant&lt;/code&gt; (U+2011) — same vendor, same naming scheme, two different characters, because different releases were scraped from differently-typed pages.&lt;/p&gt;

&lt;p&gt;It only took five rows through, though. My catalog has 13 names containing non-ASCII characters: the five above, plus &lt;code&gt;Darwin Gödel Machine&lt;/code&gt;, &lt;code&gt;XoMotion™&lt;/code&gt;, &lt;code&gt;Traverse AI™&lt;/code&gt;, &lt;code&gt;All in One Accessibility®&lt;/code&gt;, &lt;code&gt;PIXART-α&lt;/code&gt;, &lt;code&gt;DALL·E 3 Avatar Generator&lt;/code&gt;, &lt;code&gt;Night Café&lt;/code&gt; and &lt;code&gt;Møbel&lt;/code&gt;. Every one of those eight has a clean ASCII slug — &lt;code&gt;pixart-a&lt;/code&gt;, &lt;code&gt;night-cafe&lt;/code&gt;, &lt;code&gt;mobel&lt;/code&gt;, &lt;code&gt;dall-e-3-avatar-generator&lt;/code&gt; — because upstream normalized them and failed only on the hyphen. A slugifier that handles &lt;code&gt;ö&lt;/code&gt; and &lt;code&gt;™&lt;/code&gt; but not &lt;code&gt;‑&lt;/code&gt; is a slugifier with a gap shaped exactly like this bug.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I checked, and what that missed
&lt;/h2&gt;

&lt;p&gt;The reader-facing consequences, all re-verified today with plain &lt;code&gt;curl&lt;/code&gt;:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;/tool/gpt%e2%80%915-5&lt;/code&gt; → 404 (19,223 bytes of my not-found page). Same for all five.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;/tool/gpt-5-5&lt;/code&gt; (the URL a human would guess) → 404. The row does not exist under the ASCII spelling either.&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;/badge/gpt%e2%80%915-5&lt;/code&gt; → 404, while &lt;code&gt;/badge/chatgpt&lt;/code&gt; → 200. That badge route is what the "embed us" snippet tells vendors to paste.&lt;/li&gt;
&lt;li&gt;Five of the 4,443 URLs in my live &lt;code&gt;sitemap.xml&lt;/code&gt; contain a percent sign. Those five are the ones I submitted to Google and Bing.&lt;/li&gt;
&lt;li&gt;Site search for &lt;code&gt;GPT-5.5&lt;/code&gt; answers "No results". Search with the non-breaking hyphen returns exactly two hits — and both result cards link to a 404. So the reader who somehow types the right invisible character still gets nowhere.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Two days ago I published a link audit of 300 outbound tool URLs. That job measured whether &lt;em&gt;other companies&lt;/em&gt; answer. Not one line of it fetched my own internal paths. My status-code habit pointed outward, so five broken pages sat in my sitemap, in the category that carries the most mirrored reads, undetected by the one process that was supposed to catch broken pages.&lt;/p&gt;

&lt;h2&gt;
  
  
  The fix, and why I had not done it
&lt;/h2&gt;

&lt;p&gt;The obvious objection — "just re-slug them, that will collide with something" — is measurable. Running my own slugifier over all 4,323 names yields 4,320 distinct keys and exactly 3 collisions, involving 6 rows (&lt;code&gt;Bardeen AI&lt;/code&gt; / &lt;code&gt;bardeen AI&lt;/code&gt;, &lt;code&gt;sync.labs&lt;/code&gt; / &lt;code&gt;Sync labs&lt;/code&gt;, &lt;code&gt;Liner AI&lt;/code&gt; / &lt;code&gt;Liner.ai&lt;/code&gt;). None of the five OpenAI rows collides with anything. And the crawler matches an existing row by its &lt;code&gt;sourceUrl&lt;/code&gt; before it looks at slugs, so rewriting the slug while leaving &lt;code&gt;sourceUrl&lt;/code&gt; untouched does not create a duplicate entry next crawl.&lt;/p&gt;

&lt;p&gt;So the whole repair is: decode the stored value, resolve the three genuine name collisions, and keep the raw upstream permalink for matching. What I actually want to add is the check that would have caught this on day one:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Assert my identifier is not an encoding of itself — &lt;code&gt;slug LIKE '%!%%' ESCAPE '!'&lt;/code&gt; on every crawl, plus "the slug must be the value that survives &lt;code&gt;decodeURIComponent&lt;/code&gt;".&lt;/li&gt;
&lt;li&gt;Crawl my own internal links the way I crawl other people's: fetch a sample of rendered pages, follow every &lt;code&gt;/tool/&lt;/code&gt; href, compare against the row it should resolve to.&lt;/li&gt;
&lt;li&gt;Normalize at the boundary. If I copy an identifier from another site, decode it first, then run it through my own slug rule. A borrowed permalink is data about the source site, not a key.&lt;/li&gt;
&lt;li&gt;Never treat &lt;code&gt;verified&lt;/code&gt; as a health signal. These five are certified listings whose page does not load.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;I have not touched the data or the code in this pass — today's job was the audit, and I'd rather publish the measurement before the fix so that the "before" numbers can't drift. The live pages I'd point a reader at for the ranking these rows occupy are here: &lt;a href="https://aiverse.mindutil.com/best?utm_source=devto&amp;amp;utm_medium=social&amp;amp;utm_campaign=topic" rel="noopener noreferrer"&gt;the tools I rank by real signals&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;One character, two hops upstream, five dead pages, and a link audit that only ever looked outward. That last part is the one worth keeping: whatever you check, you check at one boundary, and the boundary you chose is usually the one you chose last month.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>webdev</category>
      <category>seo</category>
      <category>database</category>
    </item>
    <item>
      <title>My directory's "New" page sorts by a date I don't own. Its freshest row is 31 days old.</title>
      <dc:creator>Mindutil</dc:creator>
      <pubDate>Wed, 18 Nov 2026 00:36:45 +0000</pubDate>
      <link>https://dev.to/li_yi_dcec96bf2647dec1b7e/my-directorys-new-page-sorts-by-a-date-i-dont-own-its-freshest-row-is-31-days-old-1hae</link>
      <guid>https://dev.to/li_yi_dcec96bf2647dec1b7e/my-directorys-new-page-sorts-by-a-date-i-dont-own-its-freshest-row-is-31-days-old-1hae</guid>
      <description>&lt;p&gt;Every directory has a &lt;code&gt;/new&lt;/code&gt; page, and mine has a sentence on it that I never re-read: &lt;em&gt;"Check back often — new tools are added every week."&lt;/em&gt; I went to look at what my newest page actually contains today, and the top row was published &lt;strong&gt;October 18, 2026&lt;/strong&gt;. That is 31 days ago, on a page whose whole job is freshness.&lt;/p&gt;

&lt;p&gt;Here's the part that made me laugh: the page is not broken. It is sorting exactly as I told it to. I just told it to sort by a date I don't own.&lt;/p&gt;

&lt;h2&gt;
  
  
  What "newest" means in my code
&lt;/h2&gt;

&lt;p&gt;The directory is 4,323 AI tool listings. The sort behind the newest view is three lines:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="k"&gt;case&lt;/span&gt; &lt;span class="s2"&gt;"newest"&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;publishedAt&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"desc"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="n"&gt;createdAt&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s2"&gt;"desc"&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;createdAt&lt;/code&gt; is the row entering &lt;em&gt;my&lt;/em&gt; database. &lt;code&gt;publishedAt&lt;/code&gt; is — per the comment I wrote next to the column in the schema — "Original publication date &lt;strong&gt;on the source site&lt;/strong&gt;". It comes from the upstream directory I ingest from, and it is the primary key of the sort. So &lt;code&gt;/new-ai-tools&lt;/code&gt; doesn't mean "new to me". It means "new at the place I copy from", and the reader can't see that field anywhere: the cards render name, logo, rating, pricing badge. No date.&lt;/p&gt;

&lt;p&gt;One page promise, two possible meanings, and the sort picks the one I control least.&lt;/p&gt;

&lt;h2&gt;
  
  
  How old is "new"?
&lt;/h2&gt;

&lt;p&gt;I ran the numbers over the whole database this morning (read-only, &lt;code&gt;date&lt;/code&gt; computed from the same &lt;code&gt;publishedAt&lt;/code&gt; column):&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Listings whose source publish date is within&lt;/th&gt;
&lt;th&gt;count&lt;/th&gt;
&lt;th&gt;share&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;30 days&lt;/td&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;60 days&lt;/td&gt;
&lt;td&gt;188&lt;/td&gt;
&lt;td&gt;4.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;90 days&lt;/td&gt;
&lt;td&gt;257&lt;/td&gt;
&lt;td&gt;5.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;1 year&lt;/td&gt;
&lt;td&gt;827&lt;/td&gt;
&lt;td&gt;19.1%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;2 years&lt;/td&gt;
&lt;td&gt;1,555&lt;/td&gt;
&lt;td&gt;36.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;older than 2 years&lt;/td&gt;
&lt;td&gt;2,768&lt;/td&gt;
&lt;td&gt;64.0%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Median &lt;code&gt;publishedAt&lt;/code&gt; is &lt;strong&gt;March 15, 2024&lt;/strong&gt; — 2.7 years before today. Mean age 2.4 years. So the collection is a backlog with a "New" tab on top of it, which is normal for a directory built by ingestion and worth saying out loud instead of papering over with a page that implies a cadence.&lt;/p&gt;

&lt;h2&gt;
  
  
  The lag is the real number
&lt;/h2&gt;

&lt;p&gt;Comparing &lt;code&gt;publishedAt&lt;/code&gt; to &lt;code&gt;createdAt&lt;/code&gt; for each row tells me how stale a listing was at the moment I ingested it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;median lag: &lt;strong&gt;957 days&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;90th percentile: &lt;strong&gt;1,345 days&lt;/strong&gt;
&lt;/li&gt;
&lt;li&gt;rows already more than a year old at ingest: &lt;strong&gt;3,445 of 4,323 (79.7%)&lt;/strong&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;When I "added" something, it had usually been public for two and a half years. That's fine as an archive. It is not fine as a promise, and the promise is what's in the H2.&lt;/p&gt;

&lt;h2&gt;
  
  
  The newest page is a staircase, not a stream
&lt;/h2&gt;

&lt;p&gt;The 24 rows on page one of &lt;code&gt;/new-ai-tools&lt;/code&gt; today carry only five distinct dates: 2026-10-18 (5 rows), 10-13 (6), 10-07 (6), 10-02 (4), 09-30 (3). The source publishes in weekly-ish batches, so a sort on its publish date gives you a staircase — a jump of six "new" tools, then silence for four days, then another six. Any reader who subscribes to that page learns the upstream's editorial calendar, not mine.&lt;/p&gt;

&lt;h2&gt;
  
  
  Two more dates on the page that don't mean what they look like
&lt;/h2&gt;

&lt;p&gt;Each tool page renders "Added" and "Last updated". A few live pages, fetched today without cookies:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;chatgpt&lt;/code&gt;: Added Sep 30, 2026 / Last updated Nov 16, 2026&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;capcut&lt;/code&gt;: Added Sep 30, 2026 / Last updated Nov 17, 2026&lt;/li&gt;
&lt;li&gt;
&lt;code&gt;midjourney&lt;/code&gt;: Added Sep 30, 2026 / Last updated Nov 14, 2026&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Those "last updated" values are real, and they do not mean the description changed. &lt;code&gt;updatedAt&lt;/code&gt; is bumped whenever the crawler touches the row — 4,177 of 4,323 rows have &lt;code&gt;updatedAt &amp;gt; publishedAt&lt;/code&gt;, and &lt;code&gt;createdAt&lt;/code&gt; in my local copy spans exactly three distinct dates, because the directory was built in three crawl runs. A timestamp that moves when nobody edited anything is not a content signal; it's pipeline noise wearing a label a reader will interpret as editorial attention.&lt;/p&gt;

&lt;p&gt;The same lag shows up in death detection. The dead-tool bucket holds 405 listings, and the newest &lt;code&gt;publishedAt&lt;/code&gt; among them is &lt;strong&gt;January 24, 2026&lt;/strong&gt; — so a tool has to have been around for months before it can even be flagged as gone. Zero of the 24 newest rows are flagged, which is either good news or the same blind spot pointed the other way; last week's link audit found 4 supposedly-live listings whose domains no longer resolve, so I'd bet on blind spot.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'm changing
&lt;/h2&gt;

&lt;p&gt;None of this is shipped yet, so I'm writing it as a list rather than a changelog:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Split the two meanings.&lt;/strong&gt; "New to AIVerse" sorts by &lt;code&gt;createdAt&lt;/code&gt;; "Recently released" sorts by &lt;code&gt;publishedAt&lt;/code&gt;. Different pages, different promises, so the sort key matches the headline.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Put the date on the card.&lt;/strong&gt; A "new" listing without a visible date is a claim; with one it's a measurement.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Delete the cadence sentence&lt;/strong&gt; until the crawl genuinely adds something weekly. Today it added zero in seven days.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Rename "Last updated" to something that survives an audit&lt;/strong&gt; — "record touched by the pipeline" is ugly and true, and I'd rather be ugly and true than have a page claim editorial attention it isn't spending.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The honest summary: my directory is 64% older than two years, ingests at a two-and-a-half-year median lag, and had a page titled "New" whose newest row was 31 days stale. All four problems are copy-and-sort-key problems, not data problems — which is why I keep writing these audits up on the site itself, &lt;a href="https://aiverse.mindutil.com/new-ai-tools?utm_source=devto&amp;amp;utm_medium=social&amp;amp;utm_campaign=topic" rel="noopener noreferrer"&gt;the newest-tools view is here if you want to check my dates&lt;/a&gt; instead of trusting my table.&lt;/p&gt;

&lt;p&gt;If you run anything with an "updated" or "new" surface fed by ingestion, the question worth asking is which of your timestamps a reader can actually defend — and whether the field driving your sort is one you own.&lt;/p&gt;

</description>
      <category>seo</category>
      <category>webdev</category>
      <category>database</category>
      <category>ai</category>
    </item>
    <item>
      <title>I link-checked 300 of my 4,323 AI tool listings. My own network wrote the dead-link list.</title>
      <dc:creator>Mindutil</dc:creator>
      <pubDate>Tue, 17 Nov 2026 00:45:18 +0000</pubDate>
      <link>https://dev.to/li_yi_dcec96bf2647dec1b7e/i-link-checked-300-of-my-4323-ai-tool-listings-84-failed-including-notion-3o4i</link>
      <guid>https://dev.to/li_yi_dcec96bf2647dec1b7e/i-link-checked-300-of-my-4323-ai-tool-listings-84-failed-including-notion-3o4i</guid>
      <description>&lt;p&gt;Every so often a directory like mine stops adding tools and starts shipping broken links instead. I run one: about 4,323 AI tool listings. I had been assuming "the crawler refreshes the descriptions" was close enough to "the links still work". So I spent a morning checking whether the stored URLs actually load. The method was simple, and it collapsed on the first control group.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I ran
&lt;/h2&gt;

&lt;p&gt;Sample: 300 of 4,323 listings (6.9%) — the top 150 by the traffic number I display, plus 150 random from the rest. One machine, one pass, Node &lt;code&gt;fetch&lt;/code&gt;, concurrency 16, follow redirects, 15-second timeout.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;outcome&lt;/th&gt;
&lt;th&gt;count&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;2xx / 3xx&lt;/td&gt;
&lt;td&gt;193&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;4xx&lt;/td&gt;
&lt;td&gt;21&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;5xx&lt;/td&gt;
&lt;td&gt;2&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;no TCP or TLS answer at all&lt;/td&gt;
&lt;td&gt;84&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;First read of that table: "28% of my catalog is dead." Wrong, and I can prove it's wrong.&lt;/p&gt;

&lt;h2&gt;
  
  
  The control group that killed the method
&lt;/h2&gt;

&lt;p&gt;Then I ran a control: the same script against &lt;code&gt;notion.so&lt;/code&gt; and &lt;code&gt;writesonic.com&lt;/code&gt; - two companies I know are alive. Both timed out from this machine, exactly like the 84 that landed in my failure bucket.&lt;/p&gt;

&lt;p&gt;So that bucket isn't "dead tools", it's "unreachable from where I sit", and those are not the same set. The 4xx column is misleading in the other direction too: all 21 of them are 403, and they include Canva, Midjourney, Lovable and DataRobot — sites that answer a browser and refuse my fetch. Not dead either.&lt;/p&gt;

&lt;p&gt;Now count the part of the catalog I structurally cannot check from here: &lt;strong&gt;370 of 4,323 listings (8.6%)&lt;/strong&gt; sit on domains this network does not let me reach — &lt;code&gt;*.google.com&lt;/code&gt; and &lt;code&gt;withgoogle.com&lt;/code&gt; (152), &lt;code&gt;huggingface.co&lt;/code&gt; (78), OpenAI and ChatGPT (55), &lt;code&gt;x.ai&lt;/code&gt; / &lt;code&gt;grok.com&lt;/code&gt; (46), Anthropic and Claude (23), Meta AI (16).&lt;/p&gt;

&lt;p&gt;A single-vantage link check on a global catalog is mostly a measurement of the network path, not of the product. My dead-link list had an 8.6% floor of "unknowable" in it before any server said anything.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I could actually prove
&lt;/h2&gt;

&lt;p&gt;Throw away the ambiguous failures and keep only ones that are self-evidently about the domain rather than the path. That's 8 listings:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;4 domains that no longer resolve.&lt;/strong&gt; &lt;code&gt;sharktank-ai.com&lt;/code&gt;, &lt;code&gt;superprompts.com&lt;/code&gt;, &lt;code&gt;aicoloringpage.org&lt;/code&gt;, &lt;code&gt;postai.pro&lt;/code&gt;. I checked each against three independent resolvers (system, AliDNS, DNSPod); all three returned nothing. That isn't a firewall artifact, that's an unregistered domain.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;1 with a lapsed certificate.&lt;/strong&gt; &lt;code&gt;moviepong.app&lt;/code&gt; still serves a Let's Encrypt cert with &lt;code&gt;notAfter = Jun 17 2026&lt;/code&gt; — expired about five months ago — and the page now returns 404. Nobody renewed it because nobody is running it.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;1 Cloudflare 530, error code 1016.&lt;/strong&gt; &lt;code&gt;summarist.ai&lt;/code&gt; reached the edge, then Cloudflare could not resolve the origin. The tunnel is gone; the DNS record never got the memo.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;1 answering 404 while DNS points at a free host's default name.&lt;/strong&gt; &lt;code&gt;naturallanguageplaylist.com&lt;/code&gt; resolves to a &lt;code&gt;pythonanywhere.com&lt;/code&gt; host. The app left, the A record stayed.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;1 TLS handshake failure at an AWS Global Accelerator address.&lt;/strong&gt; &lt;code&gt;promptmixer.dev&lt;/code&gt; opens port 443 and then nothing speaks TLS. No listener behind the accelerator.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Four of those eight are already filed under my "RIP AI" category. The other four — ColorPage Lab, MoviePong, Summarist, Natural Language Playlist — are still listed as ordinary, live tools. Three of those four carry a &lt;strong&gt;free&lt;/strong&gt; pricing label, and pricing is the first filter on the site. A reader filtering to free tools has no way to know the thing they clicked has had an expired certificate since June.&lt;/p&gt;

&lt;h2&gt;
  
  
  The label also errs the other way
&lt;/h2&gt;

&lt;p&gt;Of the 16 sampled listings tagged "RIP", &lt;strong&gt;7 returned HTTP 200&lt;/strong&gt;. Most are stale tags on things that got bought or moved — harmless. Two aren't:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;Pandora Avatars is stored with &lt;code&gt;socialbook.io&lt;/code&gt;. The &lt;code&gt;&amp;lt;title&amp;gt;&lt;/code&gt; of that page today is "2-Row Infinite Testimonial Slider". Whatever that domain sells now, it isn't that product.&lt;/li&gt;
&lt;li&gt;Stocknews AI at &lt;code&gt;stocknewsai.com&lt;/code&gt; redirects to &lt;code&gt;click-v4.junclikrmedi.com/click?i=…&lt;/code&gt; — a parked-domain click script.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;So "dead" in my catalog is a category name, and the category comes from a source I mirror. It is not an observation with a timestamp attached. Neither is "live".&lt;/p&gt;

&lt;h2&gt;
  
  
  29 links that quietly changed owner
&lt;/h2&gt;

&lt;p&gt;Of the 216 listings that answered, 29 finished on a different registrable domain than the one I store. 18 of those start from a tracking short link, expected. The remaining 11 are domain moves I have not applied: &lt;code&gt;runwayml.com → runway.com&lt;/code&gt;, &lt;code&gt;moonvalley.ai → moonvalley.com&lt;/code&gt;, &lt;code&gt;preview.reve.art → reve.com&lt;/code&gt;, &lt;code&gt;freepik.com → magnific.com&lt;/code&gt;, &lt;code&gt;textero.ai → textero.io&lt;/code&gt;, &lt;code&gt;relume.io → relume.ai&lt;/code&gt;, &lt;code&gt;plusdocs.com → plusai.com&lt;/code&gt;, &lt;code&gt;www.anthropic.com → claude.com&lt;/code&gt;, &lt;code&gt;hollyhires.ai → hiringagents.ai&lt;/code&gt;, &lt;code&gt;aimyads.ai → brandnetworks.com&lt;/code&gt; (a different company entirely), plus the parked-domain case above.&lt;/p&gt;

&lt;p&gt;Each of those is a listing whose visible name and actual destination now disagree.&lt;/p&gt;

&lt;h2&gt;
  
  
  The part that is my own fault
&lt;/h2&gt;

&lt;p&gt;76 listings — 17 of them inside the top 150 — store a tracking short link (&lt;code&gt;prf.hn&lt;/code&gt;, &lt;code&gt;*.pxf.io&lt;/code&gt;, &lt;code&gt;*.sjv.io&lt;/code&gt;, &lt;code&gt;partnerlinks.io&lt;/code&gt;) as the tool's website. That's deliberate; it's how the directory pays for itself. But it costs me diagnostic power. Probing &lt;code&gt;prf.hn&lt;/code&gt; tells me whether the affiliate network is up, never whether the product is. And when the vendor moves, the tracker still reports 200, so the change is invisible to me by construction.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I would change — none of it shipped
&lt;/h2&gt;

&lt;p&gt;This is a dev log, not a roadmap, so let me be exact: &lt;strong&gt;I have changed nothing yet.&lt;/strong&gt; There is no HTTP status column, no check timestamp, no second vantage point.&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;Store the vendor's canonical domain as its own field and keep the monetized redirect separate, so link hygiene is measurable at all.&lt;/li&gt;
&lt;li&gt;Make a dead link a state with evidence attached — status code, error class, check date, resolver set — and give "unreachable from one machine" its own third value instead of folding it into dead.&lt;/li&gt;
&lt;li&gt;Refuse to render a pricing label for a listing whose domain fails a multi-resolver check. The free filter is the loudest page on the site; it shouldn't be able to surface an unregistered domain.&lt;/li&gt;
&lt;li&gt;Hold "RIP" pages out of the sitemap until something external confirms the death, because 7 of 16 sampled dead tags still serve 200.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Those 300 checks were one pass from one place. If you run a crawler or a directory, point the same script at a couple of domains you &lt;em&gt;know&lt;/em&gt; are alive before you trust your dead list. Mine reported 84 dead and could defend 8.&lt;/p&gt;

&lt;p&gt;The free-tools index — 2,198 listings with their pricing labels, all of the above included — is at &lt;a href="https://aiverse.mindutil.com/free-ai-tools?utm_source=devto&amp;amp;utm_medium=social&amp;amp;utm_campaign=topic" rel="noopener noreferrer"&gt;AIVerse free AI tools&lt;/a&gt;. Now you know how little I can promise about the links in it.&lt;/p&gt;

</description>
      <category>testing</category>
      <category>showdev</category>
      <category>ai</category>
      <category>webdev</category>
    </item>
    <item>
      <title>My most-read page has 1,322,170 reads. My site has had 4 outbound clicks.</title>
      <dc:creator>Mindutil</dc:creator>
      <pubDate>Mon, 16 Nov 2026 00:35:56 +0000</pubDate>
      <link>https://dev.to/li_yi_dcec96bf2647dec1b7e/my-most-read-page-has-1322170-reads-my-site-has-had-4-outbound-clicks-1960</link>
      <guid>https://dev.to/li_yi_dcec96bf2647dec1b7e/my-most-read-page-has-1322170-reads-my-site-has-had-4-outbound-clicks-1960</guid>
      <description>&lt;p&gt;My daily writing queue is a list of the "most-read" pages on my AI tools directory. Today it put &lt;code&gt;Deevid AI&lt;/code&gt; at #1 with 1,322,170 reads, &lt;code&gt;Text to Video AI: Deevid&lt;/code&gt; at #2 with 983,454, and &lt;code&gt;Deevid AI Image&lt;/code&gt; at #3 with 637,094.&lt;/p&gt;

&lt;p&gt;Those are three landing pages for one product, and my site has produced &lt;strong&gt;4 outbound clicks&lt;/strong&gt; in its entire history. Here is how I found out, and what the number actually is.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the number comes from
&lt;/h2&gt;

&lt;p&gt;The directory has ~4,300 tool listings, imported by a crawler that reads another AI tool directory's public listing pages. In &lt;code&gt;scripts/crawler.mjs&lt;/code&gt; the field is extracted like this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="n"&gt;views&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nf"&gt;text&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="s2"&gt;".aix_meta-views"&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That selector is a popularity counter on &lt;em&gt;someone else's&lt;/em&gt; page. It is stored in a column whose own schema comment says "Popularity counter mirrored from the source directory". Nothing in the pipeline ever mixed it with a request log from my domain.&lt;/p&gt;

&lt;p&gt;So the "reads" number in my queue is not my traffic. It is an inherited signal about how much attention a product gets on the site I scraped — and I treated it as if it were mine.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I print it as
&lt;/h2&gt;

&lt;p&gt;This is the part I underestimated. The number is not just an internal ranking input; it is rendered to visitors as if it described my page:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;On every tool page, in a stats block labelled &lt;code&gt;Popularity&lt;/code&gt;, the value reads &lt;code&gt;1,322,170 reads&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;On every "best of" category page, the comparison table has a column literally headed &lt;code&gt;Directory reads&lt;/code&gt;.&lt;/li&gt;
&lt;li&gt;On the A-vs-B pages, the winner of the "reach" row is chosen by comparing the two scraped counters.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;A visitor has no way to know the difference between "this product is popular on the site we imported data from" and "this page was read 1.3 million times on ours". The second one is a claim about my site. I have never made it deliberately, and I could not support it if asked — the field that &lt;em&gt;would&lt;/em&gt; support a claim about my site, an outbound click counter, sums to 4 across all 4,323 rows.&lt;/p&gt;

&lt;h2&gt;
  
  
  The top three are one product
&lt;/h2&gt;

&lt;p&gt;Three rows, three URLs, same vendor: &lt;code&gt;deevid.ai/image-to-video&lt;/code&gt;, &lt;code&gt;deevid.ai/text-to-video&lt;/code&gt;, &lt;code&gt;deevid.ai/ai-image-generator&lt;/code&gt;. Between them they hold 2,942,718 of the counter's units, which is 13.4% of everything the top 150 rows carry. They also carry inconsistent pricing labels — two rows say &lt;code&gt;free&lt;/code&gt;, one says &lt;code&gt;freemium&lt;/code&gt; — because each was scraped independently and nothing reconciled them.&lt;/p&gt;

&lt;p&gt;If I spend my hour writing body copy in queue order, my first three days of work are three variations of the same page for the same company, ranked #1 partly because that vendor's parent directory counts each landing page separately.&lt;/p&gt;

&lt;h2&gt;
  
  
  Dedup by hostname doesn't rescue it
&lt;/h2&gt;

&lt;p&gt;The obvious fix is to group listings by domain and only let the highest-scoring one occupy a slot. I tried it in SQL against the current data, and it fails in both directions:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;code&gt;openai.com&lt;/code&gt; legitimately appears 50 times, because those are different products and model pages. Collapsing it would delete real entries.&lt;/li&gt;
&lt;li&gt;11 of the top 150 rows don't have a product hostname at all. Their URL is an affiliate redirect — &lt;code&gt;prf.hn&lt;/code&gt;, &lt;code&gt;*.pxf.io&lt;/code&gt;, &lt;code&gt;partnerlinks.io&lt;/code&gt; — 45 rows site-wide. You cannot cluster those by hostname, because the hostname belongs to the affiliate network, not the vendor.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The identifier I'd need — "which product is this" — is exactly the thing the import never captured. It's the same reason my rebrand cleanup is manual: the rows have unique URLs, so URL uniqueness proves nothing about product uniqueness.&lt;/p&gt;

&lt;h2&gt;
  
  
  The queue is a cliff, not a curve
&lt;/h2&gt;

&lt;p&gt;Summed over the 4,176 listings that have any value, the counter totals 49,736,374. The top 150 rows — 3.5% of the directory — account for 22,025,504, i.e. 44.3%. Inside that top band the differences are noise: row 150 sits at 44,554 and row 151 at 44,250. A gap of 304 counter-units, out of 44,554, decides whether a page gets written proper copy or stays a template sentence.&lt;/p&gt;

&lt;p&gt;Worse, 147 listings have the counter at exactly 0. They are last in every ordering forever, no matter what they are. And because this field is copied from a third-party page, I have no way to tell whether a value is current, stale, or a scraping artifact — when I tried to re-check a handful of source pages today I got an HTTP 403, so I could not even confirm a single number externally.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I'm actually changing
&lt;/h2&gt;

&lt;p&gt;Being precise about what is done versus planned, because "here's my fix" sections usually lie:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Not shipped yet:&lt;/strong&gt; the public labels. &lt;code&gt;Popularity: N reads&lt;/code&gt; and the &lt;code&gt;Directory reads&lt;/code&gt; column are honest only if the number is relabeled as what it is — attention on the source directory — or hidden. Until that edit ships, the pages overstate my site.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Not shipped yet:&lt;/strong&gt; a product-level identity on each listing, so near-duplicate landing pages of one vendor compete for one slot instead of three.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;In progress today:&lt;/strong&gt; stop treating the queue as a traffic ranking. It's a demand-prioritization heuristic borrowed from someone else's site, and it's fine for deciding &lt;em&gt;what readers probably want to read about&lt;/em&gt; and useless for deciding &lt;em&gt;what my pages performed&lt;/em&gt;.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Already true:&lt;/strong&gt; only 20 of the top 150 pages have hand-written body copy; the other 130 are still templated. The real bottleneck is writing, not ranking — and 98 of those 150 rows do have a vendor claim badge, versus 14.2% of the whole directory, so the demand band at least correlates with listings someone accountable has touched.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;The lesson generalizes past my directory: if a metric is inherited, label it as inherited at the point of display, not just in the schema comment. Nobody reads the comment. Everybody reads the number.&lt;/p&gt;

&lt;p&gt;If you want to see what the ranking currently produces, the category leaderboards are at &lt;a href="https://aiverse.mindutil.com/best?utm_source=devto&amp;amp;utm_medium=social&amp;amp;utm_campaign=sideproject" rel="noopener noreferrer"&gt;aiverse.mindutil.com/best&lt;/a&gt; — the ordering there is this same counter, which is now the honest way to describe it.&lt;/p&gt;

</description>
      <category>seo</category>
      <category>datascience</category>
      <category>ai</category>
      <category>webdev</category>
    </item>
    <item>
      <title>Entropy is not strength: the arithmetic behind a password meter</title>
      <dc:creator>Mindutil</dc:creator>
      <pubDate>Wed, 11 Nov 2026 02:34:50 +0000</pubDate>
      <link>https://dev.to/li_yi_dcec96bf2647dec1b7e/entropy-is-not-strength-the-arithmetic-behind-a-password-meter-2ie6</link>
      <guid>https://dev.to/li_yi_dcec96bf2647dec1b7e/entropy-is-not-strength-the-arithmetic-behind-a-password-meter-2ie6</guid>
      <description>&lt;p&gt;A password meter is six lines of arithmetic. Almost every one you have used runs a variant of this:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;let&lt;/span&gt; &lt;span class="n"&gt;poolSize&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;z&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;test&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;pw&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="n"&gt;poolSize&lt;/span&gt; &lt;span class="o"&gt;+=&lt;/span&gt; &lt;span class="mi"&gt;26&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="no"&gt;A&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="no"&gt;Z&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;test&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;pw&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="n"&gt;poolSize&lt;/span&gt; &lt;span class="o"&gt;+=&lt;/span&gt; &lt;span class="mi"&gt;26&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="mi"&gt;9&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;test&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;pw&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="n"&gt;poolSize&lt;/span&gt; &lt;span class="o"&gt;+=&lt;/span&gt; &lt;span class="mi"&gt;10&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="k"&gt;if&lt;/span&gt; &lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="p"&gt;[&lt;/span&gt;&lt;span class="o"&gt;^&lt;/span&gt;&lt;span class="n"&gt;a&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="n"&gt;zA&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="no"&gt;Z0&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="mi"&gt;9&lt;/span&gt;&lt;span class="p"&gt;]&lt;/span&gt;&lt;span class="o"&gt;/&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;test&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;pw&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="n"&gt;poolSize&lt;/span&gt; &lt;span class="o"&gt;+=&lt;/span&gt; &lt;span class="mi"&gt;32&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="n"&gt;bits&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="n"&gt;pw&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="n"&gt;length&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="nb"&gt;Math&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;log2&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="n"&gt;poolSize&lt;/span&gt; &lt;span class="o"&gt;||&lt;/span&gt; &lt;span class="mi"&gt;1&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Then a band — under 28 bits is "weak", under 50 "fair", under 70 "good", above "strong".&lt;/p&gt;

&lt;p&gt;There is nothing dishonest in that code. The problem is what the number means: it is a statement about a &lt;em&gt;process&lt;/em&gt; (draw each character uniformly from a &lt;code&gt;poolSize&lt;/code&gt;-symbol alphabet) applied to an &lt;em&gt;output&lt;/em&gt; (a string). When the string really was produced that way, the arithmetic is exact. When a human typed it, the same arithmetic produces numbers that are wrong in a specific direction, and the direction is always "too optimistic".&lt;/p&gt;

&lt;h2&gt;
  
  
  Four outputs from the same function
&lt;/h2&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Password&lt;/th&gt;
&lt;th&gt;Chars&lt;/th&gt;
&lt;th&gt;Pool&lt;/th&gt;
&lt;th&gt;Bits&lt;/th&gt;
&lt;th&gt;Band&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;hunter2&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;7&lt;/td&gt;
&lt;td&gt;36&lt;/td&gt;
&lt;td&gt;36.19&lt;/td&gt;
&lt;td&gt;fair&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;P@ssw0rd1&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;9&lt;/td&gt;
&lt;td&gt;94&lt;/td&gt;
&lt;td&gt;58.99&lt;/td&gt;
&lt;td&gt;good&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;Tr0ub4dor&amp;amp;3&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;11&lt;/td&gt;
&lt;td&gt;94&lt;/td&gt;
&lt;td&gt;72.10&lt;/td&gt;
&lt;td&gt;strong&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;&lt;code&gt;aaaaaaaaaaaaaaaaaaaaaaaa&lt;/code&gt;&lt;/td&gt;
&lt;td&gt;24&lt;/td&gt;
&lt;td&gt;26&lt;/td&gt;
&lt;td&gt;112.81&lt;/td&gt;
&lt;td&gt;strong&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;The last row is the whole problem in one line. Twenty-four copies of the same letter outrank an eleven-character password that has upper case, digits and a symbol — because the formula has no term for repetition, and no term for anything a human does.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. The meter's pool is bigger than the generator's alphabet
&lt;/h2&gt;

&lt;p&gt;The generator on the same page defines its pools literally:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="no"&gt;POOLS&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="n"&gt;uppercase&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'ABCDEFGHIJKLMNOPQRSTUVWXYZ'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;   &lt;span class="c1"&gt;// 26&lt;/span&gt;
  &lt;span class="n"&gt;lowercase&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="s1"&gt;'abcdefghijklmnopqrstuvwxyz'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;   &lt;span class="c1"&gt;// 26&lt;/span&gt;
  &lt;span class="n"&gt;digits&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;    &lt;span class="s1"&gt;'0123456789'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;                   &lt;span class="c1"&gt;// 10&lt;/span&gt;
  &lt;span class="n"&gt;symbols&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt;   &lt;span class="s1"&gt;'!@#$%^&amp;amp;*()-_=+[]{}|;:,.&amp;lt;&amp;gt;?'&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt;   &lt;span class="c1"&gt;// 26&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;All four together are 88 symbols, so a generated character carries &lt;code&gt;log2(88) = 6.45943&lt;/code&gt; bits. The &lt;em&gt;meter&lt;/em&gt;, reading the same string, awards 32 for "any non-alphanumeric" — which makes the pool 94 and the per-character value &lt;code&gt;log2(94) = 6.55459&lt;/code&gt;.&lt;/p&gt;

&lt;p&gt;A 16-character password drawn from all four pools therefore contains 103.35 bits and is displayed as 104.87. It is one and a half bits of flattery, from the same function that decides whether to call it "strong". The direction is harmless here; the lesson is that "pool size" is defined twice, in two places, and the two definitions are not the same number.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Uniformity is the assumption, and it is the part humans break
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;P@ssw0rd1&lt;/code&gt; really does have 58.99 bits &lt;em&gt;if&lt;/em&gt; each of its nine characters was chosen uniformly from 94 symbols. At 10^10 guesses per second, half of &lt;code&gt;2^58.99&lt;/code&gt; is about 2.9e17 guesses — just under a year.&lt;/p&gt;

&lt;p&gt;But nobody guesses uniformly. A rule-based cracker tries dictionary words first, then capitalisation, then leetspeak (&lt;code&gt;a&lt;/code&gt;→&lt;code&gt;@&lt;/code&gt;, &lt;code&gt;o&lt;/code&gt;→&lt;code&gt;0&lt;/code&gt;, &lt;code&gt;s&lt;/code&gt;→&lt;code&gt;$&lt;/code&gt;), then a short digit suffix, and &lt;code&gt;P@ssw0rd1&lt;/code&gt; falls out of that schedule in the first few minutes. The entropy number was never wrong; the sample space it describes was never the one being searched.&lt;/p&gt;

&lt;p&gt;This is why &lt;code&gt;xxxxxxxxxxxxxxxxxxxxxxxx&lt;/code&gt; and &lt;code&gt;P@ssw0rd1&lt;/code&gt; both show a green bar, and only one of them survives contact with a real attacker.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Even "random" leaks, in two more places
&lt;/h2&gt;

&lt;p&gt;Entropy is not the only quantity that gets overstated. Two examples that are measured rather than remembered:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;A UUID is not 128 random bits.&lt;/strong&gt; A v4 UUID starts from 16 random bytes and then overwrites six bits — four for the version (&lt;code&gt;0100&lt;/code&gt;) and two for the variant (&lt;code&gt;10&lt;/code&gt;). What is left is 122 bits of entropy, about 5.3e36 values. Still far past guessing, but if you are budgeting bits, you are over by six.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Modulo bias depends on the size of the source.&lt;/strong&gt; Reducing a random number with &lt;code&gt;buf[0] % n&lt;/code&gt; is exact only when &lt;code&gt;2^32&lt;/code&gt; is a multiple of &lt;code&gt;n&lt;/code&gt;; otherwise the first &lt;code&gt;2^32 mod n&lt;/code&gt; values get one extra slot. At 32 bits the leftover is tiny: for &lt;code&gt;n = 94&lt;/code&gt;, &lt;code&gt;2^32 mod 94 = 42&lt;/code&gt;, a relative bias of about 1e-8 — irrelevant. The same line with a single byte (0–255) and &lt;code&gt;n = 100&lt;/code&gt; is not irrelevant: 56 of the 100 values appear 1.5 times as often, because 256 = 2×100 + 56. The fix is rejection sampling — draw, discard anything at or above &lt;code&gt;floor(max/n) * n&lt;/code&gt;, draw again — which is what the random number tool does.&lt;/p&gt;

&lt;h2&gt;
  
  
  4. What the number is actually good for
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;If you generated it&lt;/strong&gt;, &lt;code&gt;length × log2(alphabet)&lt;/code&gt; is the correct entropy. Pick length, not symbols: from an 88-symbol alphabet, 12 characters is 77.51 bits, 16 is 103.35, 20 is 129.19. Adding two characters buys more than adding a symbol class.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;If a human typed it&lt;/strong&gt;, the number is an upper bound that is never reached. Use a dictionary-and-pattern model (zxcvbn-style) or do not score at all — instruct people to generate instead.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;If it is a passphrase&lt;/strong&gt;, count words, not letters. A 7,776-word diceware list gives &lt;code&gt;log2(7776) = 12.92&lt;/code&gt; bits per word, so six words is 77.55 bits. The letter-based meter scores that same 28-character string at 164.02 — roughly twice its real entropy, because it treats the letters as if they had been drawn independently.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The useful rule is blunt: &lt;strong&gt;entropy describes how a value was produced&lt;/strong&gt;. If you know the process, the formula is exact and you should trust it. If you only have the string, any meter can be fooled by a pattern, and the one thing you must not do is build a policy around the bar turning green.&lt;/p&gt;




&lt;p&gt;If you want to watch the numbers above come out of the actual code, the &lt;a href="https://tools.mindutil.com/en/tools/password-generator" rel="noopener noreferrer"&gt;password generator&lt;/a&gt; shows the character pools and the entropy it assigns, the &lt;a href="https://tools.mindutil.com/en/tools/uuid-generator" rel="noopener noreferrer"&gt;UUID generator&lt;/a&gt; produces v4 values (version and variant bits included), and the &lt;a href="https://tools.mindutil.com/en/tools/random-number-generator" rel="noopener noreferrer"&gt;random number generator&lt;/a&gt; uses rejection sampling rather than a modulo. All three run in the page — the values are generated by your browser's &lt;code&gt;crypto&lt;/code&gt; API and nothing you type or generate is sent anywhere.&lt;/p&gt;

</description>
      <category>javascript</category>
      <category>security</category>
      <category>beginners</category>
      <category>webdev</category>
    </item>
    <item>
      <title>A Unix timestamp has no unit: the four ways JavaScript reads your date wrong</title>
      <dc:creator>Mindutil</dc:creator>
      <pubDate>Mon, 09 Nov 2026 02:34:06 +0000</pubDate>
      <link>https://dev.to/li_yi_dcec96bf2647dec1b7e/a-unix-timestamp-has-no-unit-the-four-ways-javascript-reads-your-date-wrong-20m1</link>
      <guid>https://dev.to/li_yi_dcec96bf2647dec1b7e/a-unix-timestamp-has-no-unit-the-four-ways-javascript-reads-your-date-wrong-20m1</guid>
      <description>&lt;p&gt;&lt;code&gt;1700000000&lt;/code&gt; is a moment in time: November 2023. Read as milliseconds, the identical number is three weeks into 1970. Both readings are "correct" — which is exactly the problem.&lt;/p&gt;

&lt;p&gt;A Unix timestamp is just a count. It carries no unit, no timezone, no calendar. Every system that stores one is relying on an agreement that was never written into the number. Here are the four places where that agreement breaks, with the output of each run measured rather than remembered.&lt;/p&gt;

&lt;h2&gt;
  
  
  Trap 1: seconds or milliseconds — and the digit-count heuristic
&lt;/h2&gt;

&lt;p&gt;&lt;code&gt;Date&lt;/code&gt; in JavaScript always takes &lt;strong&gt;milliseconds&lt;/strong&gt;. Most APIs, logs and database columns give you &lt;strong&gt;seconds&lt;/strong&gt;. Feed one to the other:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nc"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;1700000000&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;toISOString&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;      &lt;span class="c1"&gt;// 1970-01-20T16:13:20.000Z&lt;/span&gt;
&lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nc"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;1700000000&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;1000&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;toISOString&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt; &lt;span class="c1"&gt;// 2023-11-14T22:13:20.000Z&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No exception, no warning. Just a date 53 years early.&lt;/p&gt;

&lt;p&gt;The standard trick is to count digits, because today a seconds value has 10 digits and a milliseconds value has 13:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;ms&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nc"&gt;String&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;value&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nx"&gt;length&lt;/span&gt; &lt;span class="o"&gt;&amp;gt;=&lt;/span&gt; &lt;span class="mi"&gt;12&lt;/span&gt; &lt;span class="p"&gt;?&lt;/span&gt; &lt;span class="nx"&gt;value&lt;/span&gt; &lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="nx"&gt;value&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;1000&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That is the rule this site's own converter uses, and it holds for everything in living memory. Here is why, measured rather than asserted: milliseconds is seconds × 1000, so the two representations of the &lt;em&gt;same instant&lt;/em&gt; always differ by exactly three digits. They therefore cross every decimal boundary together:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Instant&lt;/th&gt;
&lt;th&gt;as seconds&lt;/th&gt;
&lt;th&gt;as milliseconds&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;1973-03-03T09:46:40Z&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;100000000&lt;/code&gt; (9)&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;100000000000&lt;/code&gt; (12)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;2001-09-09T01:46:40Z&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;1000000000&lt;/code&gt; (10)&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;1000000000000&lt;/code&gt; (13)&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;2286-11-20T17:46:40Z&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;10000000000&lt;/code&gt; (11)&lt;/td&gt;
&lt;td&gt;
&lt;code&gt;10000000000000&lt;/code&gt; (14)&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;So a "≥ 12 digits" cut is simultaneously "every seconds value up to year 5138" and "every milliseconds value from 1973-03-03 onward". That is an enormous window, and it is why the heuristic feels unbreakable.&lt;/p&gt;

&lt;p&gt;It does break, in the one direction nobody expects. Milliseconds values &lt;strong&gt;before 1973-03-03T09:46:40Z&lt;/strong&gt; have 11 digits or fewer, so they get multiplied by 1000 a second time:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// 1971-09-27T15:43:20Z stored in milliseconds&lt;/span&gt;
&lt;span class="mi"&gt;54834200000&lt;/span&gt;            &lt;span class="c1"&gt;// 11 digits → heuristic multiplies again&lt;/span&gt;
&lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nc"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;54834200000&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;1000&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;toISOString&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;  &lt;span class="c1"&gt;// 3707-08-18T02:13:20.000Z&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;Anything with pre-1973 dates in millisecond form — a 1971 log archive, a 1969 master tape, birth dates on old records — silently lands in the 38th century. The mirror case (seconds from year 5138 onward) is theoretical; this one is data.&lt;/p&gt;

&lt;p&gt;The honest fix is not a smarter heuristic — it is to know which unit the field is in and say so in the column name (&lt;code&gt;created_at_ms&lt;/code&gt;), or to never accept a bare number from a human without asking which unit they pasted.&lt;/p&gt;

&lt;p&gt;One more consequence of counting digits at all: a validator written as &lt;code&gt;^\d{1,16}$&lt;/code&gt; rejects the minus sign, so &lt;strong&gt;every date before 1970 is unparsable&lt;/strong&gt;, not just mis-parsed. If your dataset has pre-epoch rows, a tool built this way will tell you the input is invalid — which is at least an honest failure rather than a wrong answer.&lt;/p&gt;

&lt;h2&gt;
  
  
  Trap 2: the same string means different instants on different machines
&lt;/h2&gt;

&lt;p&gt;This one is in the language spec, and most people meet it only once:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="nb"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;parse&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;2024-01-15&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;          &lt;span class="c1"&gt;// 2024-01-15T00:00:00.000Z  (UTC)&lt;/span&gt;
&lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nc"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;2024-01-15T00:00&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;      &lt;span class="c1"&gt;// local midnight&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A &lt;strong&gt;date-only&lt;/strong&gt; string is parsed as UTC midnight. A &lt;strong&gt;date-time without an offset&lt;/strong&gt; is parsed as local time. So the second line's result depends on the machine's timezone:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;in New York (UTC−5): &lt;code&gt;2024-01-15T05:00:00.000Z&lt;/code&gt;
&lt;/li&gt;
&lt;li&gt;in Shanghai (UTC+8): &lt;code&gt;2024-01-14T16:00:00.000Z&lt;/code&gt;
&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;The first line is identical everywhere. Run both lines on one machine and they are five hours apart — not because either is wrong, but because they answer different questions. A test that passes locally and fails in CI, a "daily report" that silently drops or double-counts the boundary day, a booking that shifts by eight hours when the server moves region: this is almost always it.&lt;/p&gt;

&lt;p&gt;Rules that survive contact with production:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;always write the offset (&lt;code&gt;2024-01-15T00:00:00Z&lt;/code&gt;, or &lt;code&gt;+08:00&lt;/code&gt; if you mean a wall clock);&lt;/li&gt;
&lt;li&gt;store UTC, format for a zone only at the last moment;&lt;/li&gt;
&lt;li&gt;if you want a &lt;em&gt;date&lt;/em&gt; with no time, keep it a string (&lt;code&gt;YYYY-MM-DD&lt;/code&gt;) and don't put it through &lt;code&gt;Date&lt;/code&gt; at all.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Note the direction of the damage: &lt;code&gt;toISOString()&lt;/code&gt; is always UTC and never uses the host zone, so it is the safe one. &lt;code&gt;toLocaleString()&lt;/code&gt;, &lt;code&gt;toString()&lt;/code&gt;, and &lt;code&gt;getHours()&lt;/code&gt; all read the host zone, so the same instant prints as &lt;code&gt;10:31:41&lt;/code&gt; on one laptop and &lt;code&gt;02:31:41&lt;/code&gt; on the CI box.&lt;/p&gt;

&lt;h2&gt;
  
  
  Trap 3: wall clocks skip and repeat
&lt;/h2&gt;

&lt;p&gt;Local time is not a number line. On 2024-03-10 at 02:00, New York jumped to 03:00. So &lt;code&gt;02:30&lt;/code&gt; that morning does not exist:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nc"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt;2024-03-10T02:30&lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;toString&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;
&lt;span class="c1"&gt;// Sun Mar 10 2024 03:30:00 GMT-0400 (EDT)&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;No error — it slides an hour forward. Twice a year, in the other direction, the opposite happens: 01:30 on 2024-11-03 occurs twice, and the parser has to guess. It picks the first occurrence (&lt;code&gt;05:30Z&lt;/code&gt;, EDT), so a 25-hour "day" of data collapses to 23 hours of distinct timestamps.&lt;/p&gt;

&lt;p&gt;Anything that schedules, bills, or timestamps in local wall-clock terms needs either an explicit offset or a real timezone library — the browser's &lt;code&gt;Intl&lt;/code&gt; API is enough for most formatting and is built in.&lt;/p&gt;

&lt;h2&gt;
  
  
  Trap 4: the ceiling you can see coming
&lt;/h2&gt;

&lt;p&gt;Two hard limits worth knowing before a customer finds them:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nc"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt; &lt;span class="o"&gt;**&lt;/span&gt; &lt;span class="mi"&gt;31&lt;/span&gt; &lt;span class="o"&gt;*&lt;/span&gt; &lt;span class="mi"&gt;1000&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;toISOString&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt; &lt;span class="c1"&gt;// 2038-01-19T03:14:08.000Z&lt;/span&gt;
&lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nc"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mf"&gt;8.64e15&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;toISOString&lt;/span&gt;&lt;span class="p"&gt;()&lt;/span&gt;        &lt;span class="c1"&gt;// +275760-09-13T00:00:00.000Z&lt;/span&gt;
&lt;span class="k"&gt;new&lt;/span&gt; &lt;span class="nc"&gt;Date&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="mi"&gt;2&lt;/span&gt; &lt;span class="o"&gt;**&lt;/span&gt; &lt;span class="mi"&gt;53&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt;                      &lt;span class="c1"&gt;// RangeError&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;&lt;code&gt;2**31 − 1&lt;/code&gt; seconds is 2038-01-19T03:14:07Z — the point where anything that stores epoch seconds in a signed 32-bit integer rolls over to 1901. JavaScript itself is immune (its &lt;code&gt;Date&lt;/code&gt; is 64-bit ms), but the columns, file formats and embedded firmware it talks to are not, and migrating those is a multi-year job that has to start before 2038 rather than after.&lt;/p&gt;

&lt;p&gt;And &lt;code&gt;Date&lt;/code&gt; only spans ±100,000,000 days: roughly year −271,821 to +275,760. Beyond that you get &lt;code&gt;Invalid Date&lt;/code&gt;, and past &lt;code&gt;2**53&lt;/code&gt; ms the constructor throws instead. So a stray unit multiplication doesn't just give you a wrong century — it can give you &lt;code&gt;NaN&lt;/code&gt; two layers away from the code that caused it.&lt;/p&gt;

&lt;p&gt;Pre-epoch has one more trap of the same family. Converting ms → seconds with &lt;code&gt;| 0&lt;/code&gt; truncates toward zero, so it disagrees with &lt;code&gt;Math.floor&lt;/code&gt; for negative values:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="nb"&gt;Math&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;floor&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="mi"&gt;5500&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;1000&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="c1"&gt;// -6&lt;/span&gt;
&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="o"&gt;-&lt;/span&gt;&lt;span class="mi"&gt;5500&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;1000&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="o"&gt;|&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;       &lt;span class="c1"&gt;// -5&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A one-second error on 1969 data, in exactly the code path that a unit heuristic already made hard to read.&lt;/p&gt;

&lt;h2&gt;
  
  
  What to do with all this
&lt;/h2&gt;

&lt;p&gt;The order that keeps every one of these traps away: keep instants as UTC milliseconds (&lt;code&gt;Date.now()&lt;/code&gt;), keep calendar dates as &lt;code&gt;YYYY-MM-DD&lt;/code&gt; strings, carry the zone separately when you mean a wall clock, and name the unit of every numeric timestamp column.&lt;/p&gt;

&lt;p&gt;If you just need to look at a value and know which of the four traps you're in, &lt;a href="https://tools.mindutil.com/en/tools/timestamp-converter" rel="noopener noreferrer"&gt;the timestamp converter&lt;/a&gt; does the digit-count detection in the open: it shows the unit it guessed, then prints the same instant as a bare local string and as &lt;code&gt;… UTC&lt;/code&gt;, so an eight-hour surprise is visible instead of embedded in your output. Paste &lt;code&gt;1700000000&lt;/code&gt; and &lt;code&gt;1700000000000&lt;/code&gt; and watch both lines land on the same moment.&lt;/p&gt;

&lt;p&gt;For the second and third traps — where the instant is fine but the zone is the question — &lt;a href="https://tools.mindutil.com/en/tools/timezone-converter" rel="noopener noreferrer"&gt;the timezone converter&lt;/a&gt; pins one UTC instant against several zones at once, which is the only way to sanity-check a meeting time across a DST boundary. And when the bug shows up as a duration instead of a date — a countdown a day short, an "age" off by one — &lt;a href="https://tools.mindutil.com/en/tools/date-difference" rel="noopener noreferrer"&gt;date difference&lt;/a&gt; computes it in both calendar days and elapsed 24-hour periods, which differ exactly when a DST skip or a month boundary is in the range.&lt;/p&gt;

&lt;p&gt;All of it runs in the page: the parsing is done by the same &lt;code&gt;Date&lt;/code&gt; calls quoted above, in your browser, and the value you paste is not sent anywhere. Which is the nice thing about timestamps — the bugs are purely arithmetic, so once you've seen the four failure modes above, you can spot them in any tool, including this one.&lt;/p&gt;

</description>
      <category>javascript</category>
      <category>beginners</category>
      <category>webdev</category>
      <category>todayilearned</category>
    </item>
    <item>
      <title>“Double it and add 30” is close enough to fool you — until you need the real Celsius to Fahrenheit number</title>
      <dc:creator>Mindutil</dc:creator>
      <pubDate>Thu, 05 Nov 2026 06:31:41 +0000</pubDate>
      <link>https://dev.to/li_yi_dcec96bf2647dec1b7e/double-it-and-add-30-is-close-enough-to-fool-you-until-you-need-the-real-celsius-to-fahrenheit-18n</link>
      <guid>https://dev.to/li_yi_dcec96bf2647dec1b7e/double-it-and-add-30-is-close-enough-to-fool-you-until-you-need-the-real-celsius-to-fahrenheit-18n</guid>
      <description>&lt;p&gt;There's a trick everyone half-remembers for Celsius to Fahrenheit: &lt;strong&gt;double it and add 30&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;This is the second post in a short series about conversion factors people half-remember. The first one covered &lt;a href="https://dev.to/li_yi_dcec96bf2647dec1b7e/1-mile-is-not-16-km-the-exact-factors-and-where-the-shortcut-breaks-1d3f"&gt;why 1 mile is not 1.6 km&lt;/a&gt; — same pattern: a rounded rule that is right at one point and quietly wrong everywhere else.&lt;/p&gt;

&lt;p&gt;It's not wrong. It's not right either. It's a straight line that happens to cross the real one at exactly 10°C, and then drifts away from it in both directions.&lt;/p&gt;

&lt;h2&gt;
  
  
  The exact formula
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;°F = °C × 9/5 + 32
°C = (°F − 32) × 5/9
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's it. The Fahrenheit scale puts freezing at 32 and boiling at 212 — 180 degrees apart — while Celsius uses 100. The ratio 180/100 is 9/5, and the 32 is the offset. Everything else is arithmetic.&lt;/p&gt;

&lt;h2&gt;
  
  
  How far the shortcut drifts
&lt;/h2&gt;

&lt;p&gt;The shortcut is &lt;code&gt;2C + 30&lt;/code&gt;. Subtract the exact value and you get a clean expression for the error:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;error = (2C + 30) − (1.8C + 32) = 0.2C − 2
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;So the error is zero at 10°C, and grows by &lt;strong&gt;2°F for every 10°C&lt;/strong&gt; you move away from it:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;°C&lt;/th&gt;
&lt;th&gt;Exact °F&lt;/th&gt;
&lt;th&gt;×2 + 30&lt;/th&gt;
&lt;th&gt;Error&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;−10&lt;/td&gt;
&lt;td&gt;14.0&lt;/td&gt;
&lt;td&gt;10&lt;/td&gt;
&lt;td&gt;−4.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;0&lt;/td&gt;
&lt;td&gt;32.0&lt;/td&gt;
&lt;td&gt;30&lt;/td&gt;
&lt;td&gt;−2.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;10&lt;/td&gt;
&lt;td&gt;50.0&lt;/td&gt;
&lt;td&gt;50&lt;/td&gt;
&lt;td&gt;0.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;20&lt;/td&gt;
&lt;td&gt;68.0&lt;/td&gt;
&lt;td&gt;70&lt;/td&gt;
&lt;td&gt;+2.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;30&lt;/td&gt;
&lt;td&gt;86.0&lt;/td&gt;
&lt;td&gt;90&lt;/td&gt;
&lt;td&gt;+4.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;37&lt;/td&gt;
&lt;td&gt;98.6&lt;/td&gt;
&lt;td&gt;104&lt;/td&gt;
&lt;td&gt;+5.4&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;40&lt;/td&gt;
&lt;td&gt;104.0&lt;/td&gt;
&lt;td&gt;110&lt;/td&gt;
&lt;td&gt;+6.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;100&lt;/td&gt;
&lt;td&gt;212.0&lt;/td&gt;
&lt;td&gt;230&lt;/td&gt;
&lt;td&gt;+18.0&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;200&lt;/td&gt;
&lt;td&gt;392.0&lt;/td&gt;
&lt;td&gt;430&lt;/td&gt;
&lt;td&gt;+38.0&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;For weather, that's fine. "It's 30 outside" doesn't need to be 86 rather than 90. For anything you'll act on precisely, it stops being fine quickly.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where it actually bites: body temperature
&lt;/h2&gt;

&lt;p&gt;Normal body temperature is about &lt;strong&gt;37°C = 98.6°F&lt;/strong&gt;. A fever of 38°C is &lt;strong&gt;100.4°F&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;Run the reverse shortcut — subtract 30, divide by 2 — on 100.4 and you get &lt;strong&gt;35.2°C&lt;/strong&gt;. That's not a fever. That's hypothermia territory, from a reading that means the opposite. This is the one place where a "close enough" conversion is worse than no conversion at all.&lt;/p&gt;

&lt;p&gt;The exact reverse formula gives 38.0°C, as it should.&lt;/p&gt;

&lt;h2&gt;
  
  
  The mental model worth keeping
&lt;/h2&gt;

&lt;p&gt;If you want a shortcut that behaves predictably, use this one instead: &lt;strong&gt;double it, subtract 10%, add 32.&lt;/strong&gt;&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight plaintext"&gt;&lt;code&gt;°F ≈ C × 2 × 0.9 + 32 = 1.8C + 32
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;That's algebraically identical to the exact formula, so the error is zero everywhere. It's one more step than "double and add 30", and it removes the entire error table above.&lt;/p&gt;

&lt;p&gt;Two other anchors worth memorising:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;−40 is the same in both scales.&lt;/strong&gt; It's the only point where Celsius and Fahrenheit agree.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Every 5°C is 9°F.&lt;/strong&gt; So 0 → 32, 5 → 41, 10 → 50, 15 → 59, 20 → 68. If you know one, you can walk up and down the scale in steps of 9.&lt;/li&gt;
&lt;/ul&gt;




&lt;p&gt;When I need the number rather than the trick, I use a converter with a pre-computed table: &lt;a href="https://tools.mindutil.com/en/tools/celsius-to-fahrenheit" rel="noopener noreferrer"&gt;celsius to fahrenheit&lt;/a&gt; and the reverse direction at &lt;a href="https://tools.mindutil.com/en/tools/fahrenheit-to-celsius" rel="noopener noreferrer"&gt;fahrenheit to celsius&lt;/a&gt;. Both run in the browser, no upload, no account.&lt;/p&gt;

&lt;p&gt;The shortcut isn't a lie. It's a rounded line that crosses the truth once, at 10°C — which is exactly why it feels trustworthy.&lt;/p&gt;

</description>
      <category>math</category>
      <category>beginners</category>
      <category>todayilearned</category>
      <category>productivity</category>
    </item>
    <item>
      <title>Count words correctly in JavaScript — including Chinese, Japanese, and Korean</title>
      <dc:creator>Mindutil</dc:creator>
      <pubDate>Tue, 03 Nov 2026 02:40:17 +0000</pubDate>
      <link>https://dev.to/li_yi_dcec96bf2647dec1b7e/count-words-correctly-in-javascript-including-chinese-japanese-and-korean-1pmp</link>
      <guid>https://dev.to/li_yi_dcec96bf2647dec1b7e/count-words-correctly-in-javascript-including-chinese-japanese-and-korean-1pmp</guid>
      <description>&lt;h2&gt;
  
  
  The problem with &lt;code&gt;text.split(' ')&lt;/code&gt;
&lt;/h2&gt;

&lt;p&gt;Almost every "word counter" tool you find online does the same thing: split the text on spaces, count the pieces. That works for English — until it doesn't.&lt;/p&gt;

&lt;p&gt;Try it on a sentence with an apostrophe: &lt;code&gt;"don't worry"&lt;/code&gt; becomes two words instead of one. Try it on Chinese: &lt;code&gt;"你好世界"&lt;/code&gt; has no spaces at all, so you get one word instead of four. Try it on text with multiple spaces or line breaks, and your count inflates with empty strings.&lt;/p&gt;

&lt;p&gt;Here is how to count words correctly across languages, and what the edge cases actually are.&lt;/p&gt;

&lt;h2&gt;
  
  
  The two-script problem
&lt;/h2&gt;

&lt;p&gt;Most of the world's text falls into one of two categories:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Space-separated scripts&lt;/strong&gt; (English, French, German, Spanish, etc.): Words are delimited by whitespace. A word is a run of letters and digits, possibly with inner apostrophes or hyphens (&lt;code&gt;don't&lt;/code&gt;, &lt;code&gt;well-known&lt;/code&gt;).&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Character-based scripts&lt;/strong&gt; (Chinese, Japanese, Korean): There are no word boundaries. Each character — each Han ideograph, each kana, each Hangul syllable — counts as one word.&lt;/p&gt;

&lt;p&gt;A correct word counter needs to handle both, and it needs to handle them in the same text. Mix English and Chinese in one paragraph (as technical documentation often does), and you need both algorithms firing at once.&lt;/p&gt;

&lt;h2&gt;
  
  
  The implementation
&lt;/h2&gt;



&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="c1"&gt;// CJK scripts: Han ideographs, Japanese kana, Korean Hangul&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;CJK&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="sr"&gt;/&lt;/span&gt;&lt;span class="se"&gt;[\u&lt;/span&gt;&lt;span class="sr"&gt;1100-&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;11FF&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;3040-&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;30FF&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;3130-&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;318F&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;3400-&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;4DBF&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;4E00-&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;9FFF&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;AC00-&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;D7AF&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;F900-&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;FAFF&lt;/span&gt;&lt;span class="se"&gt;]&lt;/span&gt;&lt;span class="sr"&gt;/g&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="c1"&gt;// Space-separated words: letters/digits with inner apostrophes and hyphens&lt;/span&gt;
&lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;WORD&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="sr"&gt;/&lt;/span&gt;&lt;span class="se"&gt;[\p&lt;/span&gt;&lt;span class="sr"&gt;{L}&lt;/span&gt;&lt;span class="se"&gt;\p&lt;/span&gt;&lt;span class="sr"&gt;{N}&lt;/span&gt;&lt;span class="se"&gt;]&lt;/span&gt;&lt;span class="sr"&gt;+&lt;/span&gt;&lt;span class="se"&gt;(?:[&lt;/span&gt;&lt;span class="sr"&gt;'&lt;/span&gt;&lt;span class="se"&gt;\u&lt;/span&gt;&lt;span class="sr"&gt;2019-&lt;/span&gt;&lt;span class="se"&gt;][\p&lt;/span&gt;&lt;span class="sr"&gt;{L}&lt;/span&gt;&lt;span class="se"&gt;\p&lt;/span&gt;&lt;span class="sr"&gt;{N}&lt;/span&gt;&lt;span class="se"&gt;]&lt;/span&gt;&lt;span class="sr"&gt;+&lt;/span&gt;&lt;span class="se"&gt;)&lt;/span&gt;&lt;span class="sr"&gt;*/gu&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;

&lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;countWords&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;text&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;cjk&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;text&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;match&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;CJK&lt;/span&gt;&lt;span class="p"&gt;)?.&lt;/span&gt;&lt;span class="nx"&gt;length&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="kd"&gt;const&lt;/span&gt; &lt;span class="nx"&gt;other&lt;/span&gt; &lt;span class="o"&gt;=&lt;/span&gt; &lt;span class="nx"&gt;text&lt;/span&gt;&lt;span class="p"&gt;.&lt;/span&gt;&lt;span class="nf"&gt;replace&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;CJK&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="s1"&gt; &lt;/span&gt;&lt;span class="dl"&gt;'&lt;/span&gt;&lt;span class="p"&gt;).&lt;/span&gt;&lt;span class="nf"&gt;match&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;WORD&lt;/span&gt;&lt;span class="p"&gt;)?.&lt;/span&gt;&lt;span class="nx"&gt;length&lt;/span&gt; &lt;span class="o"&gt;??&lt;/span&gt; &lt;span class="mi"&gt;0&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nx"&gt;cjk&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="nx"&gt;other&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;The trick: strip out all CJK characters first (replacing them with spaces so we don't accidentally merge adjacent Latin words), then count what's left with the word pattern. Add the two counts together.&lt;/p&gt;

&lt;p&gt;Two details worth noting:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;1. The Unicode property escapes.&lt;/strong&gt; &lt;code&gt;\p{L}&lt;/code&gt; matches any letter in any script; &lt;code&gt;\p{N}&lt;/code&gt; matches any numeric character. This handles accented Latin (&lt;code&gt;café&lt;/code&gt;), Cyrillic, Greek, Arabic, Devanagari — everything. A regex like &lt;code&gt;[a-zA-Z]+&lt;/code&gt; would miss most of the world's text.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;2. The apostrophe handling.&lt;/strong&gt; &lt;code&gt;don't&lt;/code&gt; should be one word, not two. The pattern &lt;code&gt;(?:['\u2019-][\p{L}\p{N}]+)*&lt;/code&gt; allows inner apostrophes (both straight &lt;code&gt;'&lt;/code&gt; and curly &lt;code&gt;\u2019&lt;/code&gt;) and hyphens, but only when surrounded by letters. A leading or trailing apostrophe doesn't count.&lt;/p&gt;

&lt;h2&gt;
  
  
  Reading time: it's not just words
&lt;/h2&gt;

&lt;p&gt;Once you have a word count, the next question is usually "how long will this take to read?" The standard answer is 200 words per minute — but that's only accurate for space-separated scripts.&lt;/p&gt;

&lt;p&gt;CJK text is read character by character, at roughly 400 characters per minute for an average reader. So the correct formula is:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight javascript"&gt;&lt;code&gt;&lt;span class="kd"&gt;function&lt;/span&gt; &lt;span class="nf"&gt;estimateReadingMinutes&lt;/span&gt;&lt;span class="p"&gt;(&lt;/span&gt;&lt;span class="nx"&gt;cjkCharacters&lt;/span&gt;&lt;span class="p"&gt;,&lt;/span&gt; &lt;span class="nx"&gt;words&lt;/span&gt;&lt;span class="p"&gt;)&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt;
  &lt;span class="k"&gt;return&lt;/span&gt; &lt;span class="nx"&gt;cjkCharacters&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;400&lt;/span&gt; &lt;span class="o"&gt;+&lt;/span&gt; &lt;span class="nx"&gt;words&lt;/span&gt; &lt;span class="o"&gt;/&lt;/span&gt; &lt;span class="mi"&gt;200&lt;/span&gt;&lt;span class="p"&gt;;&lt;/span&gt;
&lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;A 1000-character Chinese article takes about 2.5 minutes to read. A 500-word English article takes about 2.5 minutes. Mix them, and you add the two times together.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three tests for any word counter
&lt;/h2&gt;

&lt;p&gt;Before you trust a tool:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;"你好世界"&lt;/code&gt; should give 4 words.&lt;/strong&gt; If you get 1, the tool doesn't handle CJK.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;"don't worry"&lt;/code&gt; should give 2 words.&lt;/strong&gt; If you get 3, the tool splits on apostrophes.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;&lt;code&gt;"  multiple   spaces  "&lt;/code&gt; should give 2 words.&lt;/strong&gt; If you get 5 or more, the tool counts empty strings from splitting on whitespace.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;If a tool fails any of these, its word count is wrong.&lt;/p&gt;

&lt;h2&gt;
  
  
  Words are not tokg��
&lt;/h2&gt;

&lt;p&gt;If the text is going into an LLM rather than a human, stop using the word count. A word is a linguistic unit; a token is a BPE vocabulary entry, and they drift apart by script: English lands around 1.3 words per token, CJK is closer to one token per character or worse. The estimate and the exact count are different problems, and a counter that claims both is guessing.&lt;/p&gt;

&lt;p&gt;I put the honest version of that behind &lt;a href="https://tools.mindutil.com/en/tools/token-counter" rel="noopener noreferrer"&gt;ToolStack Token Counter&lt;/a&gt;: real &lt;code&gt;o200k_base&lt;/code&gt; / &lt;code&gt;cl100k_base&lt;/code&gt; byte-pair encodings when the vendor maps to one we can load, and a clearly labelled statistical estimate where the vocabulary is private.&lt;/p&gt;

&lt;h2&gt;
  
  
  What I shipped
&lt;/h2&gt;

&lt;p&gt;I put this exact pipeline behind a page that counts words, characters, sentences, paragraphs, lines, and estimated reading time — with live updates as you type: &lt;a href="https://tools.mindutil.com/en/tools/word-counter" rel="noopener noreferrer"&gt;ToolStack Word Counter&lt;/a&gt;.&lt;/p&gt;

&lt;p&gt;It handles CJK and Latin scripts correctly, ignores leading/trailing whitespace, and runs entirely in the browser — your text never leaves the tab. Part of a set of 125 browser-local tools in 7 languages, no sign-up.&lt;/p&gt;

&lt;p&gt;Disclosure, since I would check it on someone else's page: GA4 page-level analytics is enabled; text input never leaves the tab, and there is no servfWr-side storage of anything you paste or type.&lt;/p&gt;

</description>
      <category>javascript</category>
      <category>webdev</category>
      <category>i18n</category>
      <category>productivity</category>
    </item>
    <item>
      <title>My AI directory has 92 categories. Ten of them hold 37% of the listings.</title>
      <dc:creator>Mindutil</dc:creator>
      <pubDate>Sun, 01 Nov 2026 00:32:24 +0000</pubDate>
      <link>https://dev.to/li_yi_dcec96bf2647dec1b7e/my-ai-directory-has-92-categories-ten-of-them-hold-37-of-the-listings-3ei8</link>
      <guid>https://dev.to/li_yi_dcec96bf2647dec1b7e/my-ai-directory-has-92-categories-ten-of-them-hold-37-of-the-listings-3ei8</guid>
      <description>&lt;p&gt;I ran one query against my AI tools directory last night, and the result explained why some of its pages get crawled while others sit ignored.&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Category&lt;/th&gt;
&lt;th&gt;Tools&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;RIP AI&lt;/td&gt;
&lt;td&gt;405&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Video Generators&lt;/td&gt;
&lt;td&gt;294&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Image Generators&lt;/td&gt;
&lt;td&gt;269&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Productivity&lt;/td&gt;
&lt;td&gt;253&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Developer Tools&lt;/td&gt;
&lt;td&gt;247&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Education / Studies&lt;/td&gt;
&lt;td&gt;245&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Business&lt;/td&gt;
&lt;td&gt;233&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Image editing&lt;/td&gt;
&lt;td&gt;230&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;LLM models&lt;/td&gt;
&lt;td&gt;208&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;Life Assistants&lt;/td&gt;
&lt;td&gt;192&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Those ten categories hold 2,576 of 7,006 category assignments — 36.8% of the whole taxonomy. The directory has 4,323 tools across 92 categories.&lt;/p&gt;

&lt;p&gt;That last number is not the problem. The tail is.&lt;/p&gt;

&lt;h2&gt;
  
  
  14 categories have ten tools or fewer
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;New AI Tools: 0&lt;/li&gt;
&lt;li&gt;AIxploria Selection: 0&lt;/li&gt;
&lt;li&gt;Programming: 0&lt;/li&gt;
&lt;li&gt;AI Presentations: 2&lt;/li&gt;
&lt;li&gt;AI Music: 4&lt;/li&gt;
&lt;li&gt;AI Education: 4&lt;/li&gt;
&lt;li&gt;AI 3D: 4&lt;/li&gt;
&lt;li&gt;AI Research: 4&lt;/li&gt;
&lt;li&gt;AI SEO: 5&lt;/li&gt;
&lt;li&gt;AI Translation: 7&lt;/li&gt;
&lt;li&gt;AI Chatbots: 8&lt;/li&gt;
&lt;li&gt;AI Image Generators: 8&lt;/li&gt;
&lt;li&gt;AI Social Media: 8&lt;/li&gt;
&lt;li&gt;AI Customer Support: 10&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;Three of those are empty, and they are also the most tempting pages to publish, because the names sound valuable: "New AI Tools" and "Programming" are exactly what a searcher types. So I checked what the site actually returns for them:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight shell"&gt;&lt;code&gt;&lt;span class="nv"&gt;$ BASE&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;https://aiverse.mindutil.com
&lt;span class="nv"&gt;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-o&lt;/span&gt; /dev/null &lt;span class="nt"&gt;-w&lt;/span&gt; &lt;span class="s2"&gt;"%{http_code}"&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$BASE&lt;/span&gt;&lt;span class="s2"&gt;/category/last-ai-en"&lt;/span&gt;
404
&lt;span class="nv"&gt;$ &lt;/span&gt;curl &lt;span class="nt"&gt;-s&lt;/span&gt; &lt;span class="s2"&gt;"&lt;/span&gt;&lt;span class="nv"&gt;$BASE&lt;/span&gt;&lt;span class="s2"&gt;/category/last-ai-en"&lt;/span&gt; | &lt;span class="nb"&gt;grep &lt;/span&gt;robots
&amp;lt;meta &lt;span class="nv"&gt;name&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"robots"&lt;/span&gt; &lt;span class="nv"&gt;content&lt;/span&gt;&lt;span class="o"&gt;=&lt;/span&gt;&lt;span class="s2"&gt;"noindex, nofollow"&lt;/span&gt;/&amp;gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;And the sitemap filter is one line:&lt;br&gt;
&lt;/p&gt;

&lt;div class="highlight js-code-highlight"&gt;
&lt;pre class="highlight typescript"&gt;&lt;code&gt;&lt;span class="nx"&gt;where&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="nl"&gt;tools&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{&lt;/span&gt; &lt;span class="na"&gt;some&lt;/span&gt;&lt;span class="p"&gt;:&lt;/span&gt; &lt;span class="p"&gt;{}&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt; &lt;span class="p"&gt;}&lt;/span&gt;
&lt;/code&gt;&lt;/pre&gt;

&lt;/div&gt;



&lt;p&gt;So 89 of 92 categories ship in the sitemap. The three empty ones do not.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why a 404 beats an empty page
&lt;/h2&gt;

&lt;p&gt;An empty category that returns 200 with "0 tools found" is a page a crawler can index. Do that twenty times and you have told a search engine your site has twenty useless URLs before it has read a single tool page. The 404 does the opposite: it says this hub does not exist yet.&lt;/p&gt;

&lt;p&gt;None of this is free. A user who bookmarked the page gets a dead end, and the internal link from the categories index has to be removed or the link graph starts leaking 404s. The trade is simple: a little maintenance for the absence of thin pages. On a directory where the median tool page is still fighting for its first indexing, that trade is worth taking.&lt;/p&gt;

&lt;h2&gt;
  
  
  Category is not a tag
&lt;/h2&gt;

&lt;p&gt;4,323 tools produce 7,006 category assignments, an average of 1.62 per tool. Exactly 2,100 tools (48.6%) sit in two or more categories. That overlap is real — a video editor can be "Video Editing", "Productivity", and "Business" without stretching any of those labels.&lt;/p&gt;

&lt;p&gt;Overlap is also what keeps category pages from being template clones: each one is a different basket of results. Directory sites get flagged as doorway pages when the only unique word on 92 pages is the category name. The fix is not fewer categories. It is making each category page answer something the tool pages cannot: which five of these should I try first, and why.&lt;/p&gt;

&lt;h2&gt;
  
  
  The 405-tool graveyard
&lt;/h2&gt;

&lt;p&gt;RIP AI is the largest bucket in the whole directory: 405 tools, 9.4% of everything, 236 of them marked free. These are discontinued or dead products. Publishing that number is useful — a directory that never deletes anything is lying by omission — but RIP AI is not a category. It is a status.&lt;/p&gt;

&lt;p&gt;Mixing the two is how you end up with a tool that is both "RIP AI" and "Video Generators", an empty "Programming" bucket, and all three looking equally important in a sitemap.&lt;/p&gt;

&lt;h2&gt;
  
  
  Four rules I am adopting
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;A category ships only if it has at least one tool. Zero-tool hub gets a 404 and noindex, never a 200 with an empty state.&lt;/li&gt;
&lt;li&gt;A category with fewer than 10 tools does not get its own indexable template until it reaches 10. Until then it is a filter on the parent listing.&lt;/li&gt;
&lt;li&gt;Statuses — dead, free, new, verified — stay filters, not categories.&lt;/li&gt;
&lt;li&gt;Every category page needs a hand-written shortlist paragraph. If I cannot write one, the category is not ready to be indexed.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;Only the first rule is already live. The other three are the work for the next few weeks.&lt;/p&gt;

&lt;h2&gt;
  
  
  The part that surprised me
&lt;/h2&gt;

&lt;p&gt;The top ten are not the categories I would have guessed. I expected "AI Chatbots" to lead; it has 8 tools. Video Generators has 294. The taxonomy drifted toward whatever the crawler found in bulk rather than toward what people actually search for. That is a data problem wearing a design problem's clothes, and it is the next thing on the list.&lt;/p&gt;

&lt;p&gt;If you want to see the current state of the catalog, including the free and paid filters, it is here: &lt;a href="https://aiverse.mindutil.com/free-ai-tools?utm_source=devto&amp;amp;utm_medium=social&amp;amp;utm_campaign=topic" rel="noopener noreferrer"&gt;https://aiverse.mindutil.com/free-ai-tools?utm_source=devto&amp;amp;utm_medium=social&amp;amp;utm_campaign=topic&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>webdev</category>
      <category>database</category>
      <category>seo</category>
    </item>
    <item>
      <title>Half of my AI directory is labeled "free". 88% of those listings were never claimed.</title>
      <dc:creator>Mindutil</dc:creator>
      <pubDate>Sat, 31 Oct 2026 00:36:14 +0000</pubDate>
      <link>https://dev.to/li_yi_dcec96bf2647dec1b7e/half-of-my-ai-directory-is-labeled-free-88-of-those-listings-were-never-claimed-53e4</link>
      <guid>https://dev.to/li_yi_dcec96bf2647dec1b7e/half-of-my-ai-directory-is-labeled-free-88-of-those-listings-were-never-claimed-53e4</guid>
      <description>&lt;p&gt;I run a directory with 4,323 AI tool listings. Every row has exactly one pricing label: free, freemium, or paid. On paper the split looks healthy:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;free: 2,198 (50.8%)&lt;/li&gt;
&lt;li&gt;freemium: 1,425 (33.0%)&lt;/li&gt;
&lt;li&gt;paid: 700 (16.2%)&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;For a site whose main feature is "filter by price", those are the numbers I care about most. Then I asked a different question: how many of these rows has anyone actually confirmed?&lt;/p&gt;

&lt;p&gt;Each listing can carry a verified badge, which means the vendor went through a claim flow and said "yes, this is us". 614 of 4,323 do. That is 14.2%. The split by pricing tier is where it gets interesting:&lt;/p&gt;

&lt;div class="table-wrapper-paragraph"&gt;&lt;table&gt;
&lt;thead&gt;
&lt;tr&gt;
&lt;th&gt;Label&lt;/th&gt;
&lt;th&gt;Listings&lt;/th&gt;
&lt;th&gt;Verified&lt;/th&gt;
&lt;th&gt;Verified rate&lt;/th&gt;
&lt;/tr&gt;
&lt;/thead&gt;
&lt;tbody&gt;
&lt;tr&gt;
&lt;td&gt;free&lt;/td&gt;
&lt;td&gt;2,198&lt;/td&gt;
&lt;td&gt;261&lt;/td&gt;
&lt;td&gt;11.9%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;freemium&lt;/td&gt;
&lt;td&gt;1,425&lt;/td&gt;
&lt;td&gt;225&lt;/td&gt;
&lt;td&gt;15.8%&lt;/td&gt;
&lt;/tr&gt;
&lt;tr&gt;
&lt;td&gt;paid&lt;/td&gt;
&lt;td&gt;700&lt;/td&gt;
&lt;td&gt;128&lt;/td&gt;
&lt;td&gt;18.3%&lt;/td&gt;
&lt;/tr&gt;
&lt;/tbody&gt;
&lt;/table&gt;&lt;/div&gt;

&lt;p&gt;Free listings are the least verified segment, paid listings the most. I am not going to pretend this is causal. This directory has produced 4 outbound clicks in total, so nothing here is statistically impressive. But the direction makes sense: a paid product has revenue, a support inbox, and someone whose job includes checking listings. A free tool often has one person, no budget, and no reason to care that a directory in another timezone has the wrong price attached to it.&lt;/p&gt;

&lt;p&gt;The uncomfortable part is that "free" is the label readers trust most. It is the first filter people apply, and it is the one applied to the listings that nobody has checked.&lt;/p&gt;

&lt;h2&gt;
  
  
  Where the label actually comes from
&lt;/h2&gt;

&lt;p&gt;The pricing label is inferred, not declared. It comes from the text that arrived with the listing: vendor marketing copy, aggregator feeds, and the crawl itself. So "free" means "the page said free when it was crawled", not "there is a usable free tier with a published limit".&lt;/p&gt;

&lt;p&gt;That gap is easy to underestimate. While hand-writing 20 tool pages last week, I hit three failure modes that no automated pass had caught:&lt;/p&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Version drift.&lt;/strong&gt; The Adobe Firefly entry still named Firefly 3. The current flagship is Firefly Image Model 5, two generations later.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Product merging.&lt;/strong&gt; One row had collapsed Nano Banana 2 (&lt;code&gt;gemini-3.1-flash-image&lt;/code&gt;) and Nano Banana Pro (&lt;code&gt;gemini-3-pro-image&lt;/code&gt;) into a single tool, with one set of limits.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Conflicting prices.&lt;/strong&gt; Google AI Plus appeared at both $4.99 and $7.99 across 2026 sources. I wrote "unconfirmed" instead of picking the friendlier number.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;All three are the kind of error that only surfaces when a human opens the vendor's pricing page and reads it.&lt;/p&gt;

&lt;h2&gt;
  
  
  Unverified rows rot in predictable ways
&lt;/h2&gt;

&lt;p&gt;The 85.8% of rows without a verified badge are not all wrong. Most are just fine. But the drift follows patterns:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;
&lt;strong&gt;Stale names.&lt;/strong&gt; Products rename themselves faster than feeds update.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Merged or split products.&lt;/strong&gt; Two tools become one, or one tool becomes a tier of another, and the old row keeps collecting traffic.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Dead tools.&lt;/strong&gt; 405 rows (9.4% of the directory) sit in a discontinued category. Some are genuinely shut down; others are dormant and might come back.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;There is also a category-level version of the same problem. Of 92 category pages, 3 currently have zero tools: &lt;em&gt;New AI Tools&lt;/em&gt;, &lt;em&gt;Aixploria Selection&lt;/em&gt;, and &lt;em&gt;Programming&lt;/em&gt;. They still render, and they are still in the sitemap. That is a thin page I created by accident, and it is the sort of thing a crawler is right to ignore.&lt;/p&gt;

&lt;p&gt;Related: the average tool sits in 1.62 categories, and 2,100 of 4,323 rows (48.6%) are in more than one. Overlaps are normal, but they also mean a single bad row can produce several bad pages.&lt;/p&gt;

&lt;h2&gt;
  
  
  Freshness is a number, not a vibe
&lt;/h2&gt;

&lt;p&gt;The "new tools" page is the one most likely to lie by omission, so I started measuring it:&lt;/p&gt;

&lt;ul&gt;
&lt;li&gt;22 listings were published in the last 30 days.&lt;/li&gt;
&lt;li&gt;0 were published in the last 7 days.&lt;/li&gt;
&lt;li&gt;The newest row is dated 2026-10-18, which is 13 days old today.&lt;/li&gt;
&lt;/ul&gt;

&lt;p&gt;For a directory fed by crawling, that is not a disaster, but it does mean the page cannot honestly present a daily "fresh arrivals" feed. The useful metric is not "did the crawl succeed" but "how many rows did this run add or change" — a crawl that succeeds with zero additions is a silent outage.&lt;/p&gt;

&lt;h2&gt;
  
  
  Three changes that followed
&lt;/h2&gt;

&lt;ol&gt;
&lt;li&gt;
&lt;strong&gt;Treat the pricing label as a claim.&lt;/strong&gt; Store the source and the date it was read next to the value. "Free, checked 2026-10-30" is a different statement from "free".&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Measure verification per segment, not overall.&lt;/strong&gt; The headline 14.2% hides a real 11.9% versus 18.3% gap, and that gap tells you where to look first.&lt;/li&gt;
&lt;li&gt;
&lt;strong&gt;Prune and mark.&lt;/strong&gt; A discontinued category that is visible is better than dead rows scattered through live categories, and empty category pages should be noindexed until they have something on them.&lt;/li&gt;
&lt;/ol&gt;

&lt;p&gt;None of this fixes the underlying economics: free listings will always have the weakest data, because nobody is paid to keep them accurate. The best I can do is stop presenting an inferred label as a verified fact, and show readers which rows have been checked.&lt;/p&gt;

&lt;p&gt;The free list, such as it is: &lt;a href="https://aiverse.mindutil.com/free-ai-tools?utm_source=devto&amp;amp;utm_medium=social&amp;amp;utm_campaign=sideproject" rel="noopener noreferrer"&gt;https://aiverse.mindutil.com/free-ai-tools?utm_source=devto&amp;amp;utm_medium=social&amp;amp;utm_campaign=sideproject&lt;/a&gt;&lt;/p&gt;

</description>
      <category>ai</category>
      <category>webdev</category>
      <category>sideprojects</category>
      <category>database</category>
    </item>
    <item>
      <title>I fact-checked the free tiers of 20 AI tools. Most listings were wrong.</title>
      <dc:creator>Mindutil</dc:creator>
      <pubDate>Fri, 30 Oct 2026 00:32:11 +0000</pubDate>
      <link>https://dev.to/li_yi_dcec96bf2647dec1b7e/i-fact-checked-the-free-tiers-of-20-ai-tools-most-listings-were-wrong-3l19</link>
      <guid>https://dev.to/li_yi_dcec96bf2647dec1b7e/i-fact-checked-the-free-tiers-of-20-ai-tools-most-listings-were-wrong-3l19</guid>
      <description>&lt;p&gt;I run a small AI tools directory - 4,323 entries, filterable by pricing. For the first month, almost every tool page was built from a single sentence of upstream text plus a template paragraph. That is the standard scaled-content setup, and it is wrong in ways that are easy to measure.&lt;/p&gt;

&lt;p&gt;So I picked the 20 highest-traffic pages (211K to 414K reads each in the source data) and rewrote them from primary sources: vendor pricing pages, docs changelogs, model cards, and - where the vendor site was unreachable from my network - two or three dated 2026 secondary sources cross-checked against each other. Every claim now carries a source URL and a fetch date.&lt;/p&gt;

&lt;p&gt;Three error classes came up over and over. None of them were typos.&lt;/p&gt;

&lt;h2&gt;
  
  
  1. Version drift: the listing describes a product two generations old
&lt;/h2&gt;

&lt;p&gt;The Adobe Firefly entry still said "Firefly 3". The current flagship image model is Firefly Image Model 5, announced at Adobe MAX in October 2025, with native 4-megapixel output and broad availability since early 2026. The page also described Firefly as an image tool, while the web app now covers video, audio, vector and Boards.&lt;/p&gt;

&lt;p&gt;This is not cosmetic. Firefly's paid tiers are published (Standard $9.99/mo with 2,000 credits, Pro $19.99 with 4,000, Pro Plus $49.99 with 10,000, Premium $199.99 with 50,000), and standard image and vector generations are unlimited on any paid plan - credits get spent on premium models at 10 per Image 5 generation and 100 per second of 1080p video. A reader who budgets against the old listing buys the wrong tier.&lt;/p&gt;

&lt;h2&gt;
  
  
  2. Model merging: two products collapsed into one row
&lt;/h2&gt;

&lt;p&gt;One entry mixed Nano Banana 2 (gemini-3.1-flash-image) with Nano Banana Pro (gemini-3-pro-image). They are different tiers with different prices. Per-image API pricing on the 2 line runs $0.067 at 1K, $0.101 at 2K and $0.151 at 4K, and Nano Banana 2.1 - GA on 2026-10-06 - roughly halves that to $0.0336, $0.0504 and $0.113.&lt;/p&gt;

&lt;p&gt;Merging them is how a directory ends up telling people a model costs twice what it does.&lt;/p&gt;

&lt;h2&gt;
  
  
  3. Confident numbers the vendor never published
&lt;/h2&gt;

&lt;p&gt;Google's free image-generation allowance is published as roughly 20 images a day on the basic tier, 50 on Plus, 100 on Pro and 1,000 on Ultra - with Google itself warning the caps change often. Fine. But the price of Google AI Plus showed up as both $4.99 and $7.99 in two independent 2026 sources, and I could not settle it. That page now says the price is unconfirmed instead of picking the nicer-sounding one.&lt;/p&gt;

&lt;p&gt;Same with Anthropic: Claude Opus 5.5's usage is governed by a rolling 5-hour cap plus a weekly cap, published as multiples of Pro, not absolute message counts. Anthropic does not print the numbers. So we do not either - and the page says why, instead of inventing "200 messages per 5 hours".&lt;/p&gt;

&lt;h2&gt;
  
  
  The rules I now apply to every tool page
&lt;/h2&gt;

&lt;ul&gt;
&lt;li&gt;The free tier is described as &lt;em&gt;what you can do&lt;/em&gt;, not as a credit count. "Unlimited text messages since 2026-08-10, metered images, deep research and voice" is useful; "free plan available" is not.&lt;/li&gt;
&lt;li&gt;Every number has a source URL and a fetch date, stored next to the text.&lt;/li&gt;
&lt;li&gt;Conflicts are written down as conflicts, with the date of each figure. A directory that resolves a disagreement by guessing is just a blog post with a sort button.&lt;/li&gt;
&lt;li&gt;Two pros and two cons, taken from the vendor's own failure notes where they exist. Google's model card, for example, is candid that small text blurs at 1K and long paragraphs go wrong.&lt;/li&gt;
&lt;/ul&gt;

&lt;h2&gt;
  
  
  Why bother, for a directory
&lt;/h2&gt;

&lt;p&gt;Because the only thing that separates a directory from a search result is the answer to "can I afford this, and will it do the thing". The pricing split across those 4,323 entries is 2,198 free, 1,425 freemium, 700 paid - which means most of the catalogue &lt;em&gt;is&lt;/em&gt; affordable, and the useful work is telling people where the wall actually is.&lt;/p&gt;

&lt;p&gt;I have done 20 of 150 top pages. The remaining 130 are templated, and I would rather ship them slowly than ship them wrong.&lt;/p&gt;

&lt;p&gt;If you want to check the argument against real listings, the free-tier filter is here - every row shows its pricing type, and the pages I rewrote carry the facts above:&lt;/p&gt;

&lt;p&gt;&lt;a href="https://aiverse.mindutil.com/free-ai-tools?utm_source=devto&amp;amp;utm_medium=social&amp;amp;utm_campaign=topic" rel="noopener noreferrer"&gt;https://aiverse.mindutil.com/free-ai-tools?utm_source=devto&amp;amp;utm_medium=social&amp;amp;utm_campaign=topic&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;Corrections welcome, especially from people who maintain these tools - if a listing is out of date, send the pricing page that beats mine.&lt;/p&gt;

</description>
      <category>ai</category>
      <category>webdev</category>
      <category>database</category>
      <category>productivity</category>
    </item>
  </channel>
</rss>
