DEV Community

PageLoom Studio
PageLoom Studio

Posted on

How we QA 2,000 programmatic pages before delivery

Last week we generated 2,010 pages in one batch: 402 five-page sets, one set per agency, each set written in that agency's own brand and language. Before any of it went out, every page passed the same gate. This post is the gate, in the order it runs, with the numbers we actually got.

Programmatic SEO fails quietly. A template bug repeats 400 times, a translation leaks English into an Arabic page, two cities produce the same paragraph, and nobody notices until a client does. So the rule is simple: a page is not "done" because it renders. It is done when it survives the checks below.

1. Duplicate content is a threshold, not a feeling

We compare every pair of pages inside a batch using five-word shingles and Jaccard similarity, with a hard ceiling of 0.8. Anything above it fails the build.

The 2,010-page batch came in under 1% maximum body overlap. With real data that is not easy: same service, same city, same "why choose us" paragraph. The fix is structural — each page is built from its own facts (prices, areas, hours, nearby places), not a shared skeleton with a swapped city name.

We also enforce a word floor of 350. The shortest page in the batch had 369 words; medians across sample sets landed between 420 and 575. Thin pages are not cheaper pages, they are future rewrites.

2. Every page is a real page

Rendering is the easy part. The checks that catch real problems are boring:

  • one <h1>, a unique <title>, a meta description inside length limits
  • a self-referencing canonical and correct hreflang pairs
  • every internal link resolves, no orphan pages, sitemap entries match the files on disk
  • structured data that matches what is visible on the page
  • no placeholder text, no lorem ipsum, no "TODO" in the copy

A broken internal link on page 37 of 300 is invisible in a spot check, so the link check runs over the whole set.

3. Language and direction are tested, not assumed

Our batches cover Arabic, Turkish, Spanish, Portuguese, Indonesian and English. Each language has its own failure modes: Arabic needs dir="rtl" and logical CSS properties, Turkish needs correct casing rules, and every locale needs its UI labels translated — a stray English "View all" on an Arabic page is a bug, not a detail.

We check direction at the page level, mirror the navigation, and verify that numbers and dates read naturally in each locale. For RTL the test is visual too: a layout that flips because the browser guessed breaks the first time a component is reused.

4. A browser gate, not just an HTML gate

HTML that validates can still look broken. Every batch runs through a real browser at 360, 390 and 1280 pixels wide. We fail on horizontal overflow, console errors, and touch targets under 44 pixels. Screenshots are kept per breakpoint so a regression is a diff, not an argument.

This is also where template changes get caught: a CSS tweak that fixes one card can push a table off-screen on mobile, and you only see it at 300 pages across three widths.

5. Delivery is part of QA

The clock starts when payment and the intake form are complete, not when the first email arrives. Delivery is a zip with a manifest, a seven-day rework window, and data deletion after 30 days. If a client's data changes mid-batch, we regenerate the affected pages and rerun the same gate.

What still needs a human

Programmatic does not mean automatic publishing. A person approves the data source, the offer and the brand voice before generation. The machine checks consistency; a human decides whether the page deserves to exist. When those two disagree, the human wins.

The batch, in numbers

  • 2,010 pages generated across 402 agency-branded sets
  • 0 errors, 0 warnings at delivery
  • shortest page: 369 words; medians 420–575
  • maximum body similarity inside the batch: under 1%
  • browser checks at 360 / 390 / 1280 px, no overflow, no console errors, no touch target under 44 px

If you want to see what the output looks like before talking to us, the gallery is public: pageloom-samples.surge.sh. Every set there was produced and checked by the same pipeline.

If you run an agency and want a small paid test, the pilot is 10 pages in 72 hours for $150, credited in full toward a larger sprint. The details and payment options are here: pageloom-samples.surge.sh/pay.html.

We also build a free five-page sample in your brand within 48 hours — same gate, same numbers, no obligation.

Top comments (0)