CrawlCheck

Findings · 2026-10-14 · By · 0 views

Release notes 14 Oct 2026: doors, signed weekly figures, trial

What changed on crawlcheck.io on 14 October 2026: the site reorganised around who is asking, every live figure became a signed weekly edition with a feed, Watch gained a free 30-day trial, and the defects found on the way were fixed the same day.

On 14 October 2026 crawlcheck.io was reorganised around four doors (For agents, For sites, Developers and Data) plus Learn. Every page that shows a live figure now publishes a signed weekly edition, an Atom feed and a JSON twin, listed at /api/v1/instruments. Pricing shows two ladders, one site and many sites, and Watch starts with a free 30-day trial, one per domain and email, no card. Every scan report carries the same ladder. Prices did not change. One scoring rule changed: the instruction check now flags command wording only (revision 32).

Release notes for one day. Each section says what changed, where to see it, and how to check it yourself. Earlier notes: 3 October and 29 September.

ChangeWhereHow to check it
Four audience doors and Learn/for-agents, /for-sites, /developers, /data, /learnEvery page in the sitemap is three clicks or fewer from the homepage
Signed weekly editions and feeds/api/v1/instrumentsEach edition verifies offline with the published key
Two-ladder pricing and the 30-day Watch trial/pricingPrices unchanged; one trial per domain and per email
The ladder inside every reportany scan reportSame words and prices as /pricing
Datasets catalogue/datasetsOpen CC BY 4.0 snapshots listed apart from the paid archive
Customer sign-in moved/signin/login is now the owners' door only
A date per page in the sitemap/sitemap-pages.xmlStanding pages carry ten different dates, not one
Instruction check: command wording only/rulesScoring revision 32

The site is organised by who is asking #

Until today the navigation grouped pages by what we had built. It now opens on the question each visitor brings. An agent, or the person building one, starts at For agents: what to establish before acting on a site, the signed answer that does it in one request, and how to cite that answer. A site owner starts at For sites: the free scan, the fix and the monitoring that keeps it fixed. A developer starts at Developers: the API by lane, the specifications, the SDKs, keys and quotas. Anyone checking our numbers starts at Data. Learn holds the guides, findings and glossary.

The homepage asks one question, whether an agent can trust a site, with the scan box and the four doors under it. Two older landing pages became sections of For agents and redirect there permanently, under anchors that keep their old meaning (#agentic-web and #verify-a-website). The verifier kept its own address, because signed answers, the SDKs and earlier posts link to it.

The navigation, llms.txt and agents.md are now generated from one site tree, so a page added to the navigation appears in all three at once. We checked the result rather than the code: every page in the sitemap is listed in llms-full.txt, and a breadth-first crawl from the homepage reaches every sitemap page in three clicks or fewer. The four doors are also served in German, French, Spanish and Portuguese, with no untranslated text left on any of them.

Every live figure has a signed weekly edition #

The pages that show live numbers (status, metrics, the agent safety benchmark, the Agent Traffic Index, assurance and the north-star measure) now share one frame: what the figure measures, the chart, the dataset it is computed from, and what it does not prove. Each has a machine-readable twin and an Atom feed. Once a week each figure is sealed as an edition: a signed JSON document named by its ISO week and served unchanged from then on. The first edition, 2026-W42 of the north-star measure, verifies with the same offline verifier as a Resolve answer.

The reason is citation. A live number is easy to read and hard to cite, because it moves after you quote it. An edition is the number as it stood that week, signed, so a reference in a paper, a report or an AI answer points at something that will not change underneath it. The agent conduct record, the proof log, the weekly census and share of preflight carry the same twin-and-editions box.

Pricing: two ladders and a free trial #

Pricing opens on two ladders. One site: the free scan, then the fix list implemented ($749 once), then Watch ($29 a month). Many sites: the Developer, Growth and Agency plans. Add-ons sit between the two. No price changed. Two questions were added to the page because buyers asked them: whether the scoring method is for sale (it is not; the rulebook is public) and how refunds work (in full within seven days of a first subscription charge, as the refund policy says).

Watch now starts with a free 30-day trial that opens every locked section of the report on one domain: one trial per domain and one per email address, no card. Every scan report carries the same ladder as /pricing, in the same words and at the same prices, as a sidebar on a desktop and a bottom sheet on a phone, with one line under it: paying never changes a reading. After a purchase, the welcome page ends on the one thing that shows the purchase worked, and the account page is laid out as a record with the account details beside it.

Defects found on the way, and fixed the same day #

Rebuilding the site meant crawling all of it first: every page in the sitemap and all 4,433 internal links. That sweep and the checks after it found these, each fixed before the day ended:

  • Customer sign-in behind the owners' gate. For about an hour, every Sign in link opened a code form only two people could pass. What happened and what changed.
  • The directory stopped rebuilding for about 20 hours because a scheduled job checked the clock instead of the schedule. The write-up.
  • 103 sitemap entries carried one date. Why engines ignore that, and how each page is now dated.
  • Our own llms.txt failed our own instruction check, and the check was wrong too. Both fixes.
  • 31 glossary links pointed at anchors that no longer exist; they now open the term's own page.
  • Signed answers got faster. Resolve answers now take a median of 80 ms, with 96% under 300 ms, and preflight a median of 57 ms. The status page still reads degraded because its target is measured over 30 days; it clears as older samples roll off.
  • Phone layout. Long headings and wide tables no longer push pages sideways at 390 pixels.

What changed in scoring #

One rule. The instruction check, which reads llms.txt, agents.md and hidden page text for wording aimed at a model, now matches only the command form of the verb: never say and always recommend, not never says or always recommends. A site flagged only for describing something in the third person is no longer flagged. The change is scoring revision 32, so a report scored before it is labelled with the older revision and a finding that disappears for this reason shows as a rule change, not as a fix. Every other rule, weight and grade boundary is as it was on 13 October.

Every figure above came out of CrawlCheck.

Point it at your own domain and see the same measurements, free.

Scan a domain — free

The main product

Found this on your own site? We fix it for $749.

Scan free to see where you stand. The fix is one site, every finding implemented and re-measured, with a sealed before and after.

Questions this post answers

What changed on crawlcheck.io on 14 October 2026?

The site was reorganised around four audience doors (For agents, For sites, Developers, Data) plus Learn; every live figure gained a signed weekly edition, a JSON twin and an Atom feed; pricing became two ladders with a free 30-day Watch trial; the same ladder appears in every scan report; and customer sign-in moved to /signin. Prices did not change.

What is a signed weekly edition of a CrawlCheck figure?

A JSON document holding the figure as it stood in one ISO week, signed with CrawlCheck's Ed25519 key and served unchanged after publication. It can be cited without the number moving later, and it verifies offline with the open verifier script.

How do I cite a CrawlCheck number so it does not change?

Cite the weekly edition rather than the live page. Every instrument lists its editions at /api/v1/instruments, each at a permanent URL named by its week, for example /api/v1/instruments/north-star/2026-W42.json.

Is there a free trial of CrawlCheck Watch?

Yes. Since 14 October 2026 Watch starts with a 30-day trial that opens every locked section of the report on one domain. There is one trial per domain and one per email address, and no card is needed.

Where do CrawlCheck customers sign in?

At crawlcheck.io/signin. The /login address is reserved for the site's owners and is behind an access gate.

Did CrawlCheck change its scoring on 14 October 2026?

One rule changed: the instruction check now matches only the command form of the verb, such as "never say" or "always recommend", and no longer flags descriptions such as "never says" or "always recommends". It is scoring revision 32. Everything else is unchanged.

Related findings

How anything measured in this article was measured15client identitiesone second, one address5machine filesapex and www114named agentsresolved from robots.txt24sections scoredreach, read, quoteHow anything measured here was measured15 client identities5 machine files114 named agents24 sections scoredone second, one addressapex and wwwresolved from robots.txtreach, read, quote
No account, nothing installed, and the same sequence on every domain — which is what makes one scan comparable to another. Run it on your own site.

Comments

Comments are read before they appear. Nothing is published automatically, and no account is needed.

Writing about this? Facts, live figures and marks — every number on that page is dated and traceable to a scan.

All findings · The dataset · How the dataset works