DEV Community: Bulti

There's a formula for content AI quotes — 5 GEO moves you can apply today

Bulti — Sun, 16 Aug 2026 14:32:33 +0000

⏰ The 30-second version

Generative engine optimization means structuring content so an AI can retrieve, understand, and quote the answer to a customer question.
The formula for content AI quotes has five parts — conclusion first · self-contained paragraphs · question-style subheads · FAQs · a readable surface.
All five require no code. You can do them today, just by editing content you already have.
The best bang for your buck is conclusion first. Moving your conclusion paragraph to the top alone changes the structure.
The sneaky blocker is the surface. Image-based product detail pages — on your own store or on marketplaces — mean AI can't read the text at all.

1. Conclusion first: intros don't get quoted

When AI assembles an answer, it grabs the paragraph that directly answers the question first. You know the kind of opener — "As the weather turns cooler and everyone starts thinking about immunity…" That warm-up is exactly why the piece never gets quoted.

If the customer's question is "Can a protein shake replace a meal?", give the answer in your first two sentences: "To use one as a meal replacement, look for at least 20g of protein plus a filling ingredient." Like that. And as you may have noticed, the summary block at the top of this very article works the same way.

2. Self-contained paragraphs: AI doesn't read your whole article

AI quotes in paragraph-sized chunks. But phrases like "as mentioned above" or "that product" lose their meaning the moment a paragraph is cut out of context. Write so that any paragraph stands on its own — swap pronouns for the actual product name and concrete criteria. It may feel slightly repetitive to a human reader, but to AI, that's what makes a paragraph quotable.

3. Subheads = your customer's questions: don't write "Product Features"

Subheads like "Product Features" or "Brand Story" signal nothing to AI. If you use the questions customers actually ask — "Can I use this on sensitive skin?" — as your H2 subheads, AI preferentially pulls the section that matches the question.

Where do you find those questions? You already have them. Reviews, customer inquiries, and search autocomplete are all warehouses full of questions.

4. Three FAQs: AI's favorite structure

A follow-up question plus a two-or-three-sentence answer — an FAQ is question-and-answer by design, which makes it the easiest thing for AI to quote. Questions that come up right before purchase — shipping, how to use, comparisons — work especially well.

💡 FAQ schema
Marking up your page's FAQ in a format search engines understand (structured data). If you have the bandwidth, it's worth adding — but a plain-text FAQ delivers GEO value on its own. Don't let perfectionism keep you from starting.

5. A readable surface: it doesn't matter how well you write if it can't be read

This is many sellers' blind spot. Most marketplace listings (Amazon, Etsy, and the like) are a single flat image, so AI can't read the text. No matter how good the content you pack into that image, AI sees a blank page.

GEO's home turf is surfaces where the text actually exists in the HTML — your own store's product pages and your blog. One blog post that moves your listing's key content into real text is far more visible to AI than ten image-based detail pages.

Your 5-item to-do for today

[ ] Put the conclusion in the first paragraph
[ ] Rewrite paragraphs to stand alone
[ ] Turn subheads into questions
[ ] Add 3 FAQs
[ ] Secure a text-based surface

FAQ

Q. Of the five, which should I do first?
Number one. Just moving an existing piece's conclusion to the top changes the structure, and it's the cheapest edit you'll make.

Q. How often should I publish?
Coverage matters more than frequency. It's more efficient to first build a list of the questions customers ask AI, then fill in the unanswered ones one at a time.

Q. I already have dozens of posts. Do I have to fix them all?
No. Start with the posts closest to a buying decision — comparisons, recommendations, how-tos. For everything else, just write new posts in this structure going forward.

Originally published at en.bultibulti.com. We run a free GEO audit there — it checks one of your pages against 30 buyer questions, no signup.

What is GEO, exactly? It means your customers have started asking AI instead of Google.

Bulti — Sun, 16 Aug 2026 14:32:09 +0000

⏰ The 30-second version

GEO is the work of getting your brand cited inside an AI's answer. If SEO is about ranking near the top of a list of links, GEO is about getting inside the answer itself.
Search volume shows the direction. Over the last three months, searches for "SEO optimization" fell -42%, while GEO-related searches rose +29% to +66% (Google Keyword Planner data).
An AI answer usually names just two or three brands. The seats are scarce — and whoever gets in first tends to stay in for a while.
Right now, most sellers have never heard of GEO. This is the window where the people who know can claim the ground.

The shift has already happened — your customers moved first

Picture a customer a few years ago. They searched "best probiotics," opened ten blue links, filtered out the ads, and compared three or four blog reviews. Now picture a customer today. They ask ChatGPT, "Recommend a probiotic — which one should I pick?" and choose from the two or three brands the AI hands back. Even Google now leads with an AI Overview above the results.

Sounds like an early-adopter thing? The search-volume data says otherwise.

Searches for the old playbook are sliding; searches for the new one are climbing. The absolute volumes are still small — a few hundred a month — but that's exactly the point. If you wait until the volume gets big, you're already late. By then you're sprinting alongside every competitor.

SEO vs GEO: what's the same, and what's completely different

	SEO	GEO
Goal	Rank your link higher in a results list	Get your brand/content cited in an AI answer
Judge	Search-engine ranking algorithm	Language model ("does this directly answer the question?")
Weapon	Keywords, backlinks, domain authority	Sentences and structure that answer the question head-on
Slots available	10+	2–3

The core difference is the judge. SEO is scored by an algorithm; GEO is scored by a language model asking, "Is this clear enough to quote?" That's why answering the question in your very first paragraph beats repeating a keyword ten times.

💡 GEO (Generative Engine Optimization)
The practice of designing content so that "generative" engines — ChatGPT, Perplexity, Google AI Overviews — use it as raw material for their answers. The concept was first proposed in a 2023 academic paper, and it has since crossed over into e-commerce.

Are SEO and GEO enemies? No. Well-structured content is good for both. What changes is the standard you design to — "keyword rank" versus "question coverage."

Why this matters for sellers: the seats are scarce

An AI answer mentions only two or three brands. Compared to a ten-link results page, the shelf just shrank to a third of its size. Scarcity cuts both ways: miss the cut and you effectively don't exist to the customer; make it and you all but monopolize the exposure.

And right now, almost no sellers are competing for those seats. In most categories, a large share of real customer questions sit unanswered — no one has claimed them. If the game is grabbing empty chairs, then when you start is the whole strategy.

What to do next

Start with five practical GEO moves if you already have pages to improve.
Run a GEO audit to find the highest-priority gap.
Use a fixed routine to track your brand in AI search after you publish.

FAQ

Q. Is GEO only for big brands?
No. AI doesn't look at brand size — it looks for content that answers the question. If the content exists, a small DTC store gets cited. And large brands with no useful content get left out all the time.

Q. If I already do SEO, does GEO happen automatically?
There's overlap, but it's not automatic. Keyword-stuffed ranking pages are hard for AI to quote. You have to redesign them into a question-and-answer structure.

Q. When do results show up?
It depends on how often the AI re-reads the web, but usually within a few weeks of publishing. Because there's more open space than in SEO — which carries 15 years of accumulated competition — it tends to reflect faster.

Originally published at en.bultibulti.com. We run a free GEO audit there — it checks one of your pages against 30 buyer questions, no signup.

How ChatGPT picks the products it shows — and how a small brand gets in

Bulti — Sun, 16 Aug 2026 14:30:09 +0000

⏰ The 30-second version

ChatGPT shopping results have two layers: a product carousel fed by structured product data, and the written recommendation around it, which is assembled from ordinary web pages.
Semrush ran the same product query 100 times and found ChatGPT's top product appeared in Google Shopping's first three results 75% of the time. Your Google Shopping health leaks straight into the carousel.
Shopify merchants are already in the catalog integration by default. Non-Shopify merchants apply for a direct feed at chatgpt.com/merchants.
The feed gets you into the carousel. It does not write the sentence that says why you're the right choice — that comes from content, reviews and third-party pages.
Small brands lose in the written layer more often than in the feed. That's also the layer you can fix this week without an integration.

Two different machines

When someone asks ChatGPT "what's a good fragrance-free sunscreen for sensitive skin", two things happen at once, and they run on different fuel.

The carousel. Products with images and prices. This is powered by structured product data — a feed pushed to OpenAI, plus third-party product metadata. OpenAI's help documentation says merchants are ranked on factors like availability, price, quality and whether they're the maker or primary seller, and that results are organic rather than paid placement.

The answer. The paragraphs explaining what matters for sensitive skin, which ingredients to avoid, and which brands people rate. This is retrieved from the open web: guides, reviews, forums, and your own pages if they happen to answer the question in extractable text.

Most DTC brands obsess over the first machine and ignore the second. The second is where "and X is the one usually recommended for this" gets decided.

What the evidence says about the carousel

Semrush's experiment is the most useful public data point. Running a product query 100 times, they found ChatGPT issues two sets of background searches: contextual queries that build the written answer, and Google Shopping queries that populate the products. The top ChatGPT product sat in Google Shopping's top three results 75% of the time.

The practical read: for the carousel, your Google Merchant Center feed quality, pricing, availability and review volume are doing more work than anything you write.

Meanwhile the official path keeps getting more formal:

Shopify merchants are integrated through Shopify Catalog. OpenAI states no additional work is required from individual merchants.
Everyone else applies at chatgpt.com/merchants and, once approved, pushes a feed (CSV, TSV, XML or JSON) to an allow-listed HTTPS endpoint, refreshable as often as every 15 minutes.
Feed fields include explicit flags — is_eligible_search controls whether a product can appear in ChatGPT search results at all.
Shopping is live for ChatGPT users in the U.S. first, expanding from there.

If you sell on Shopify in the U.S., you are probably already eligible and don't know it. Check by asking ChatGPT a category question and seeing whether you appear.

Where small brands actually lose

We scan a lot of small ecommerce sites. The pattern repeats:

The answer layer never had a chance. The brand's product page describes the product in marketing language and puts the useful detail — ingredients, who it suits, what it doesn't do — in an image. A model extracting text finds a slogan and moves on.

No comparison content exists. Buyers ask comparative questions ("X vs Y for oily skin"). The only pages answering those are affiliate roundups, so the affiliate roundup decides who gets named.

Third-party evidence is thin. Reviews concentrated on your own site help conversion but travel poorly. Mentions on independent sites, forums and retailer pages are what the answer layer reads.

The buyer question was never written down. Teams optimise for the phrase they use internally, not the sentence a customer types. "Barrier-repair moisturiser" versus "moisturizer for oily but dehydrated skin" — one of those gets asked.

An order of work that fits a small team

Feed hygiene first, because it's mechanical. Google Merchant Center clean, titles descriptive, availability accurate, Product schema on product pages. If you're not on Shopify, apply for the OpenAI feed.
Move the buying detail out of images into text. Ingredients, sizing, who it's for, who it isn't for. Same page, readable format.
Write the comparison you're currently outsourcing. One honest page comparing your product to the obvious alternative, including where the alternative wins. Models quote balanced pages more readily than one-sided ones.
Answer the top five questions your support inbox already gets. In the customer's words, in an FAQ block, conclusion first.
Re-ask the same questions 30 days later. Same questions, same models. Otherwise you're measuring noise.

This week, in order

[ ] Merchant Center feed clean and Product schema live
[ ] One product page rewritten so key detail is text not image
[ ] One honest comparison page published
[ ] Five real customer questions answered in an FAQ
[ ] Baseline recorded so Day 30 means something

What about Perplexity and the others

Perplexity's shopping surface leans on its own retrieval plus merchant data, and Google's AI Mode leans on the Shopping Graph it already owns. The mechanics differ, but the two-layer shape holds everywhere: a structured product layer you feed, and a written recommendation layer assembled from pages anyone can publish.

That second layer is the one a small brand can win, because it's decided by whoever answered the question best, not by whoever has the biggest catalog.

The free scan shows which of 30 real buyer questions your page currently answers, and what people actually search around them. It takes about 30 seconds and doesn't ask for a card. For the full picture of what an audit covers, see what a GEO audit measures.

FAQ

Q. How do I get my products into ChatGPT shopping results?
If you're on Shopify, product data already flows through Shopify Catalog. Otherwise apply at chatgpt.com/merchants for direct feed access, then push a compliant feed with is_eligible_search set to true.

Q. Can I pay to appear in ChatGPT shopping results?
OpenAI states shopping results are organic and not influenced by paid placement. Feed quality, relevance and user context determine what surfaces.

Q. Does ChatGPT rank products by price?
Not directly. OpenAI's documentation says merchants are ranked on factors including availability, price, quality and whether they are the maker or primary seller, and price shown may not be the lowest available.

Q. Why does ChatGPT recommend my competitor and not me?
Usually because the written layer, not the product layer, named them: a review site, a comparison page or a forum thread answered the buyer's question and cited them. That's content you can compete with.

Originally published at en.bultibulti.com. We run a free GEO audit there — it checks one of your pages against 30 buyer questions, no signup.

AI visibility tools: what they actually cost, and what they don't do

Bulti — Sun, 16 Aug 2026 14:29:40 +0000

⏰ The 30-second version

Every tool in this category does roughly the same thing: run a list of prompts against AI models on a schedule and count how often your brand shows up.
Published entry prices run from $29/mo to $295/mo, and the tiers a real brand ends up on sit closer to $199–$495/mo. That's $2,400–$5,900 a year.
Free plans are rare. Most give a trial, and several hide their real price behind a demo.
Every one of them tells you that you're invisible. None of them writes the page that fixes it, and none of them knows what your customers actually ask.
Measurement is the cheap part. Deciding what to publish is the part that costs you a quarter if you get it wrong.

What this category actually sells

An AI visibility tool does three things:

Sends a fixed set of prompts to ChatGPT, Gemini, Perplexity, Claude and others, usually daily.
Records whether you were named, which sources were cited, and how competitors placed.
Draws the trend on a dashboard.

That's a monitoring product. It's genuinely useful once you have something to monitor. It's also, in most cases, the second thing a brand needs, not the first.

Published pricing, checked 2026-08-15

All figures come from each vendor's own public pricing page on the date checked. Prices move, and "Enterprise" is custom everywhere, so treat this as a starting map and click through before you budget.

Tool	Entry price	Highest published tier	Free access
OtterlyAI	$29/mo (Lite)	$489/mo (Premium); Enterprise from $1,000/mo	Free trial, no card
Peec AI	$95/mo (Brand Starter)	$495/mo (Brand Advanced); agency plans to $795/mo	Trial, limits set by vendor
Profound	$99/mo billed yearly (Starter)	$399/mo billed yearly (Growth)	"Try for free", no published limits
Semrush AI Visibility	$99/mo per domain, billed annually	Enterprise, price not published	None stated
Ahrefs Brand Radar	$50/mo (custom-prompt packages)	$699/mo (all AI platforms)	Free public preview
Scrunch	$250/mo (Core)	Enterprise, no published ceiling	7-day trial
Athena HQ	$295/mo, or $245/mo billed annually	Enterprise, custom	Free Essential plan with $25 credit

Three things worth noticing in that table.

The advertised number is not the number you pay. Profound's $99 and Semrush's $99 are both annual-billing prices, so the real commitment is roughly $1,200 up front. Several vendors show a lower "annual-equivalent" monthly figure that only applies if you buy a year.

The useful tier is higher than the entry tier. Entry plans typically cover one brand, a small prompt allowance and one or two models. Brands that want competitor tracking across several engines land in the $199–$495 range.

Free means trial, not free. Athena HQ is the exception with a genuinely free plan carrying a $25 credit. Ahrefs gives a public preview of Brand Radar data. Everyone else wants a card eventually.

The question the dashboards can't answer

Here's the honest limit of the entire category, including the free scan on this site.

A visibility tool measures the prompts you gave it. If the prompt list is wrong, the dashboard is confidently wrong. It will show a clean 0% and you will have learned nothing, because your customers never asked those questions in the first place.

Nobody outside your company knows what your buyers actually type. Not us, not an agency, not a $495/mo tracker. That knowledge sits with the person who reads your support inbox and your product reviews.

So the sequence that works is:

Get the questions right first. They come from your own customer conversations, not from a keyword tool.
Measure once against those questions, to get a baseline.
Publish an answer to the highest-value unanswered question.
Re-measure the same questions later so the change means something.

A daily dashboard is worth paying for at step 4, when you already have a publishing cadence to attribute movement to. Bought at step 1, it becomes a $3,000-a-year graph of a number you can't move.

Before you buy any tracker

[ ] Write down 20 questions your customers actually asked you
[ ] Check whether your pages answer them in text a crawler can read
[ ] Publish one answer
[ ] Only then start paying to watch a trend

Where Bulti sits, plainly

We are not a tracker, and we'd rather say that than sell you a worse version of one.

The free report reads one of your public pages, tests it against 30 buyer questions for your category, and shows you the gaps and real US search volume behind them. No sign-up, no card.
The paid product is a $1,500 one-time 60-day operating program: the full report, 60 pre-written lessons, and the same automated diagnosis re-run at Day 30 and Day 60 so you can compare like with like. No subscription, no renewal.
There are no consulting hours in it. It's software plus a fixed curriculum, and it's priced against a year of a mid-tier subscription rather than against an agency retainer.

If what you need is a daily dashboard across nine models, buy one of the tools above — several are good. If what you need is to know which page to write this week and to have proof it changed anything, that's the gap we built for.

Not sure what a GEO audit tool measures before you spend anything? It walks through the four checks, including a 20-minute version you can run by hand. To measure actual model mentions, use the AI search visibility tracking routine.

FAQ

Q. Which AI visibility tool is best?
For continuous multi-model monitoring with competitor benchmarking, Profound, Peec AI and Scrunch are the most commonly cited in that order of maturity, and Athena HQ has the only meaningful free tier. "Best" depends on whether you already publish regularly — if you don't, no tracker will help yet.

Q. Is there a free AI visibility checker?
Athena HQ has a free plan with a $25 credit, Ahrefs offers a public Brand Radar preview, and our own scan is free with no sign-up. Everything else in the category is a time-limited trial.

Q. What's a realistic budget for AI visibility tracking?
If you want more than one brand, more than one engine and competitor data, plan for $200–$500 a month. Entry tiers at $29–$99 are for a single brand with a small prompt set, and several of them require annual billing.

Q. Do I need a tracker to appear in ChatGPT?
No. Trackers measure; they don't change what AI reads. Appearing depends on whether your pages answer buyer questions in crawlable text, and on what third-party sources say about you.

Originally published at en.bultibulti.com. We run a free GEO audit there — it checks one of your pages against 30 buyer questions, no signup.

Answer engine optimization: the same work, a clearer name

Bulti — Sun, 16 Aug 2026 14:29:10 +0000

⏰ The 30-second version

AEO (answer engine optimization) means making your content easy to lift into a direct answer — in an AI assistant, an AI Overview, or a featured snippet.
GEO (generative engine optimization) is the same intent, aimed specifically at generative systems. In practice the two labels describe overlapping work.
SEO is still the foundation. Answer engines mostly read pages that regular crawlers can already reach.
The acronym doesn't decide anything. The work is: answer a real question, in extractable text, on a page that can be found and corroborated.
If you have to pick one thing to do first, make sure your buying-decision detail exists as text rather than inside an image.

Why three names exist

Search changed shape faster than the vocabulary settled.

Once assistants started answering instead of listing, practitioners needed a word for "getting picked as the answer". Several appeared at once: answer engine optimization, generative engine optimization, LLM SEO, AI search optimization. Vendors then attached themselves to whichever label they'd bet on.

US search volume, measured with Google's keyword planner in August 2026, tells you roughly where the market's language is:

Term	US monthly searches
generative engine optimization	5,400
ai search optimization	3,600
answer engine optimization	2,400
geo optimization	1,300
llm seo	880
llm optimization	720

Nobody has won the naming yet. Pick the one your team understands and move on — the underlying checklist doesn't change.

What AEO actually asks of a page

An answer engine is doing something narrower than a search engine. It's looking for a passage it can lift, attribute and stand behind. That imposes real constraints.

One question per passage. A section that answers exactly one question is quotable. A section that covers five is not.

Conclusion before reasoning. Snippet extraction takes the top of a section. If your answer arrives in the fourth paragraph, the first three are what gets read.

Self-contained sentences. "As mentioned above, it's usually the second option" is unusable out of context. Answers get lifted away from their neighbours.

Text, not pictures. This remains the most common blocker we see in scans of small ecommerce sites: the ingredient list, the sizing guidance, the "who this is for" — all of it inside a JPEG. A model reading the page sees a headline and a buy button.

Something to corroborate it. Assistants weigh what other sources say about you. A claim only you make is a claim, not evidence.

The AEO pass on one page

[ ] Each H2 is a question a customer would type
[ ] First sentence under it answers the question outright
[ ] No sentence depends on the paragraph above to make sense
[ ] Key buying detail is real text not an image
[ ] At least one external source backs the numbers

Where AEO and SEO stop being the same thing

They overlap more than the marketing suggests. Both want crawlable pages, clear structure and credible sources. Two differences matter.

Ranking versus being named. SEO competes for a position in a list. AEO competes to be the sentence. There's no second page in an answer, and often no third brand. The distribution is winner-takes-most.

Clicks versus mentions. A user can read an answer that names you without visiting. That breaks the measurement habits SEO built — you'll see fewer clicks per unit of visibility, and your analytics won't show the impression at all. Track mentions on a fixed question set, not just sessions.

That second point is why measurement here has to be deliberate. If you don't write down the questions before you start, you have no baseline to compare against later.

The order of work

Collect 20 real customer questions. From support tickets, reviews and sales conversations. Not from a keyword tool — the phrasing customers use in an assistant is conversational and specific.
Check which ones your site already answers in text. Most brands are surprised here. Our free scan does this against 30 category questions in about 30 seconds, and the audit checklist covers the manual version.
Fix extractability before writing anything new. Move the buying detail out of images, split the multi-topic sections, put conclusions first.
Write the highest-value missing answer. One page, one question, honestly answered — including the cases where you're not the right choice.
Re-ask the same questions after 30 days. Same wording, same models. That's the only way to tell whether the work moved anything.

Notice that none of those steps requires choosing between AEO, GEO and SEO. That's the point.

FAQ

Q. What is answer engine optimization?
It's structuring content so an answer engine — ChatGPT, Perplexity, Google's AI Overviews — can lift a direct answer from your page and attribute it to you. In practice: one question per section, conclusion first, self-contained sentences, real text instead of images.

Q. Is AEO different from GEO?
Barely. GEO usually refers to generative systems specifically, AEO to any direct-answer surface including featured snippets. The work overlaps almost completely, and most vendors use them interchangeably.

Q. Does AEO replace SEO?
No. Answer engines mostly draw from pages that conventional crawlers can already access. Broken indexing, blocked crawlers and thin pages hurt both. AEO is a layer on top, not a replacement.

Q. How do I measure AEO results?
Fix a question set, record how often you're named per model before you change anything, publish, then re-ask the identical questions. Mention rate on a fixed question set is the honest metric — traffic alone will understate the effect.

Originally published at en.bultibulti.com. We run a free GEO audit there — it checks one of your pages against 30 buyer questions, no signup.

llms.txt: the evidence says nobody is reading it

Bulti — Sun, 16 Aug 2026 14:28:44 +0000

⏰ The 30-second version

llms.txt is a proposed Markdown file that lists your important pages for AI systems. It is a proposal, not a standard anyone has adopted.
Ahrefs analysed 137,000 sites and found 97% of llms.txt files received zero traffic in May 2026. The bots are not fetching them.
SE Ranking looked at 300,000 domains and found no link between having llms.txt and being cited in AI answers.
Google has said repeatedly that it does not use the file, and OpenAI, Anthropic and Perplexity document robots.txt and their crawler user-agents instead.
Writing one takes an hour and costs nothing. Believing it will make AI recommend you costs a quarter.

What the file is

Put a Markdown file at yourdomain.com/llms.txt containing a short description of your site and a curated list of links to the pages you consider important. The pitch is that AI crawlers, which struggle with JavaScript-heavy pages and sprawling navigation, could read the curated list instead of guessing.

It's often compared to robots.txt, which causes most of the confusion. robots.txt controls access. llms.txt controls nothing — it's a suggestion. Search Engine Land's framing is the clearest one going: it isn't a crawl-blocker, it's a treasure map.

The idea is reasonable. The question is whether anyone is holding the map.

What the evidence shows

This is the part most articles skip, so here it is with sources.

Server logs, at scale. Ahrefs analysed 137,000 sites and found 97% of llms.txt files received zero traffic in May 2026. Not "low engagement" — zero requests.

Citations, at scale. SE Ranking's analysis of 300,000 domains found no clear relationship between llms.txt adoption and how often a domain was cited in LLM answers.

Semrush's own logs. Semrush checked their server logs from mid-August to late October 2025 and recorded zero visits to their llms.txt from Google-Extended, GPTBot, PerplexityBot or ClaudeBot. Googlebot and Bingbot fetched it a handful of times and treated it as an ordinary file.

Google, on the record. John Mueller: "AFAIK none of the AI services have said they're using LLMs.TXT (and you can tell when you look at your server logs that they don't even check for it). To me, it's comparable to the keywords meta tag." Gary Illyes added that Google has no plans to support it. When someone spotted an llms.txt on a Google developer domain and asked whether that was an endorsement, Mueller answered "no".

The deeper objection. On the Search Off the Record podcast, Mueller made a structural argument: even a perfectly written llms.txt is self-reported. If every site claims to be the best source, the file gives an LLM no way to tell them apart. That's the meta-keywords story again, and it ended the same way.

OpenAI. GPTBot has been observed fetching /llms.txt on some sites, which reads like exploration rather than endorsement. OpenAI's published bot documentation covers OAI-SearchBot, GPTBot and ChatGPT-User, and manages access through robots.txt. llms.txt is not in it.

97% — of llms.txt files got zero bot traffic (Ahrefs, 137,000 sites, May 2026)

So should you make one?

Make one if you want. It takes an hour, it can't hurt, and if the standard is ever adopted you're already done. Anthropic, Cloudflare and Vercel publish one, mostly for documentation sites where an agent already landed and needs a map.

Just be precise about what you're buying: an option on a future, not a visibility lever today. Two failure modes to avoid.

Don't let it substitute for the work. An hour on llms.txt feels like progress. It isn't the thing that gets you named when a customer asks ChatGPT which brand to buy.

Don't let an LLM write it for you and call it strategy. Mueller's aside is hard to argue with: if an LLM can generate the file by reading your site, then an LLM can read your site.

What actually moves AI visibility

The signals with evidence behind them are unglamorous:

Crawlable text. If your answer only exists inside an image, a PDF or a JavaScript tab, it does not exist. This is still the single most common blocker we find in scans.
A direct answer to a real question. Conclusion first, in a paragraph that stands on its own without the three paragraphs above it.
Third-party corroboration. What review sites, forums and retailers say about you often carries more weight in an answer than what your own homepage claims.
Structure that survives extraction. Question-shaped subheads, FAQ blocks, comparison tables — formats that quote cleanly.

Google's own AI-optimization guidance says the same thing in fewer words: you do not need special AI files to appear in generative search.

The hour spent better

[ ] Find one buyer question your site answers only in an image
[ ] Rewrite it as text with the conclusion first
[ ] Add three FAQ entries in the words customers use
[ ] Re-check the same question in 30 days Want to see which questions your site currently fails? The free scan tests one of your pages against 30 real buyer questions and shows the gaps, no sign-up needed. If you'd rather check by hand first, the GEO audit checklist takes about 20 minutes.

FAQ

Q. Does llms.txt help with AI citations?
There's no evidence that it does. Two large studies (Ahrefs, 137,000 sites; SE Ranking, 300,000 domains) found zero traffic to the files and no citation effect, and no major AI provider documents it as a signal.

Q. Is llms.txt a Google ranking factor?
No. Google has stated it does not use the file, and that you do not need any special AI text file to appear in AI Overviews or AI Mode.

Q. Why do tools like Yoast and Lighthouse mention llms.txt then?
Because it's cheap to support and popular to ask about. Chrome's developer docs describe it as an "emerging convention" that may help an agent already on your site. That's a hedge, not a finding.

Q. If I already have one, should I remove it?
No need. It costs nothing to keep. Some people noindex it so it doesn't compete with real pages in search. Just don't count it as work done on visibility.

Originally published at en.bultibulti.com. We run a free GEO audit there — it checks one of your pages against 30 buyer questions, no signup.