Scan free

The method · version 1

Every AI check we run, with its evidence and its source.

We measure what is on a site: what an automated visitor can fetch, read and reach. We never measure what AI products answer. Each check says how we know it matters: a platform’s own document, an independent study, or neither, in which case it is recorded as a fact and never offered as advice.

What we measure

Four questions, asked of the site itself.

When an AI system reads your site for someone, can it get in, does it take away the right facts, and can it reach the page where they act? Each layer is one part of that question, answered by the checks listed under it.

  1. Read

    Can it read your pages?

    Can an AI crawler fetch your pages and understand what is on them?

    7 checks
  2. Represent

    Are your facts the same everywhere?

    Are your name, phone number, address and hours the same on every page an AI crawler reads, and in the business details your site declares?

    4 checks
  3. Transact

    Can it reach the page where people act?

    Can an automated visitor reach contact, booking or checkout, and use the form there?

    9 checks
  4. Trust

    What would it weigh before acting?

    What can an automated visitor see that a careful person would check before acting?

    11 checks

Inputs on the site, never outputs in AI products

We do not track what AI answers say about a site. Those answers are sampled from systems we do not run, and they do not repeat. SparkToro’s study () found that two of the AI tools it tested, asked the same question 100 times, gave the same list of brands in any two answers less than 1 time in 100.

A reading that changes every time it is taken is one nobody else can check. What is on the site can be: the same site, read the same way, gives the same reading, so an owner, an agency or anyone else can repeat it.

Evidence grades

Each check carries one of three evidence grades.

The grade says how we know an AI system reads what the check looks at. It describes the evidence, never the site.

  • Documented

    A platform’s own published document or code says this is read: a standard the crawlers follow, or the documentation or source code of the company that runs the system.

    12 checks at present.

  • Measured

    An independent study observed it. No platform’s own document says so.

    1 check at present.

  • Fact only

    No platform document and no independent study says it matters. We record it because it is true of the site, so it is never offered as advice about AI.

    18 checks at present.

How each source was read

Source grade A
Read at the source: the document, the code or the study itself. 14 checks.
Source grade A†
The source is reported consistently, but we have not yet opened it ourselves. No check at present.
Source grade B
A secondary report, or a remark made on stage. No check at present.
Source grade C
Nothing published shows that AI systems use it. 17 checks.

Every source carries the day we last re-checked it and is due again within 30 days; our own test suite goes red the day one is overdue. A grade we cannot back at its source drops to fact only.

Every check

31 checks, each with its evidence and its source.

Each check has its own page: the question it answers, what we observe, what we cannot see, and a fix where the evidence supports one.

Transact Can it reach the page where people act?

Can an automated visitor reach contact, booking or checkout, and use the form there?

  1. Form fields carry labelsDocumentedLevel 3checked Google Lighthouse’s agent-accessibility-tree audit, read in its source.
  2. Buttons and links carry namesDocumentedLevel 2checked Google Lighthouse’s agent-accessibility-tree audit, read in its source.
  3. A way to contact youFact onlyLevel 2checked No AI vendor documents which pages an automated visitor looks for; recorded as a fact.
  4. A way to bookFact onlyLevel 2checked No AI vendor documents which pages an automated visitor looks for; recorded as a fact.
  5. A way to buyFact onlyLevel 2checked No AI vendor documents which pages an automated visitor looks for; recorded as a fact.
  6. A phone number to callFact onlychecked No AI vendor documents which pages an automated visitor looks for; recorded as a fact.
  7. No challenge on your contact, booking or checkout pagesDocumentedLevel 2checked Cloudflare’s verified-bots documentation: its edge can challenge or turn away an automated visitor on any page.
  8. A page to act on, reached by clicking from your home pageFact onlyLevel 3checked No AI vendor documents which pages an automated visitor looks for; recorded as a fact.
  9. Fields on your contact, booking and checkout pages carry namesDocumentedLevel 3checked Google Lighthouse’s agent-accessibility-tree audit, read in its source.

Trust What would it weigh before acting?

What can an automated visitor see that a careful person would check before acting?

  1. Your secure connection’s certificate, protocol and ciphersFact onlyLevel 1checked No AI vendor documents what it weighs before acting; recorded as a fact.
  2. Every visit kept on HTTPSFact onlychecked No AI vendor documents what it weighs before acting; recorded as a fact.
  3. Safety settings your server sendsFact onlychecked No AI vendor documents what it weighs before acting; recorded as a fact.
  4. Protection against email sent in your nameFact onlychecked No AI vendor documents what it weighs before acting; recorded as a fact.
  5. A privacy policy that matches what the site doesFact onlychecked No AI vendor documents what it weighs before acting; recorded as a fact.
  6. An accessibility statementFact onlychecked No AI vendor documents what it weighs before acting; recorded as a fact.
  7. Marketing claims the page backs upFact onlychecked No AI vendor documents what it weighs before acting; recorded as a fact.
  8. Terms of serviceFact onlychecked No AI vendor documents what it weighs before acting; recorded as a fact.
  9. A refund or returns policyFact onlychecked No AI vendor documents what it weighs before acting; recorded as a fact.
  10. Your ownership of this site, confirmedFact onlychecked No AI vendor documents what it weighs before acting; recorded as a fact.
  11. Code from companies that make AI chat assistantsFact onlychecked Each vendor’s own page describes an AI chat or agent product; a script’s address does not show which of its products runs.

The yardstick

Measured against Google Lighthouse’s Agentic Browsing checks.

Lighthouse is Google’s own page test, and its Agentic Browsing checks are what Chrome’s team says an AI agent needs from a page.

  1. AppliedButtons, links and form fields carry names an automated visitor can read agent-accessibility-treeApplied to the home page, read at phone size: Lighthouse’s own rule (the check is not met when any of its axe-core rules finds a problem) over our axe-core run, configured the way Lighthouse configures it.
  2. Not appliedForms offered to automated visitors through WebMCP webmcp-form-coverageWebMCP needs a browser feature we do not run.
  3. Not appliedTools offered to automated visitors through WebMCP webmcp-registered-toolsWebMCP needs a browser feature we do not run.
  4. Not appliedWebMCP tool descriptions are valid webmcp-schema-validityWebMCP needs a browser feature we do not run.
  5. Not appliedThe layout stays still while the page loads cumulative-layout-shiftLighthouse measures this in a lab test on a simulated phone, which we do not run. The real-visitor figure we read is a different measure, so we do not use it here.
  6. AppliedAn llms.txt file follows the suggested format, if the site has one llms-txtApplied when the site has the file: we request /llms.txt once and apply Lighthouse’s three format checks. With no file, Lighthouse counts the check as not applying, and so do we.
  7. AppliedAn agent resource list (ARD) is valid, if the site has one ard-schemaApplied when the site has an agent resource list: we request /.well-known/ard.json once and run Lighthouse’s validator over it. With no list, the check does not apply.

Calibration

Our reading of Lighthouse’s accessibility-tree check (agent-accessibility-tree) matched Lighthouse 13.4.0 on 8 of 8 hand-built test pages, on the verdict and on the failing rules alike (). The calibration against real sites is pending. Until it is done we print no number for this area, only which of Lighthouse’s checks are met.

Where we go further: We look at what an automated visitor receives on every page we read, and we keep the record week to week. Where we do less: Lighthouse says these checks are still being developed: its newest version has seven, its current release six. From what we read we can apply up to three of them: the accessibility check on your home page at phone size, and the two that read your llms.txt and agent resource list when you have them.

The 33 rules Lighthouse’s accessibility-tree check reads

Lighthouse’s check is not met when axe-core finds a problem under any of these rules. We run axe-core 4.12.1 with Lighthouse’s own configuration and apply the same list, read from Lighthouse’s source at commit 21501c1 (the category itself: its configuration).

button-nameinput-button-nameinput-image-altlabellink-nameselect-namedocument-titlearia-allowed-attraria-allowed-rolearia-command-namearia-conditional-attraria-dialog-namearia-hidden-bodyaria-hidden-focusaria-input-field-namearia-prohibited-attraria-required-attraria-required-childrenaria-required-parentaria-rolesaria-textaria-toggle-field-namearia-tooltip-namearia-treeitem-namearia-valid-attraria-valid-attr-valueduplicate-id-ariadefinition-listtable-duplicate-nametabindexautocomplete-validpresentation-role-conflictsvg-img-alt

Levels

A level is a set of checks, never a mark a site earns.

Each level asks one question and is answered by its own set of checks. A report prints, for each level, how many of its checks raised a finding, were checked with nothing found, were not seen or were not measured. It never says a site is at a level.

  1. Level 1

    Can an automated visitor that says who it is get your pages and read them?

  2. Level 2

    Is what it reads about your business the same everywhere, and can it find where to act?

  3. Level 3

    Can it get all the way to the submit button?

    It runs only once the owner confirms the site is theirs, because we only click through a site with permission. It stops at the submit button: we never send a form, so it never shows whether a booking or a message goes through.

The other 13 checks sit in no level. New formats such as llms.txt and the content-use lines stay facts until they are an agreed standard, because requiring them would fail nearly every site on something nobody has agreed on; most of the rest are Trust checks, which belong to their own areas.

What we never do

What this method will not tell you.

  • No score. A site gets no number and no rating for AI; Lighthouse’s own category reports how many of its checks are met, and so do we.
  • No verdict. We never say a site is ready, safe or at a level; a level is a set of checks with counts.
  • No citation tracking. We do not track whether AI answers mention, cite or recommend a site: those answers are sampled from systems we do not run, and they do not repeat.
  • No detection of AI-written content. We do not judge who or what wrote a page.
  • No named AI product as able or unable. We fetch under our own name, so we never say what a particular AI product can or cannot read on a site.
  • No advice without evidence. A check with no documented or measured effect is recorded as a fact and never offered as something to do.

What this is, and isn’t. Invara reports observable technical facts about a website with a confidence level for each finding. It does not determine legal compliance, does not assert violations of any law, and is not legal advice. Items we can’t observe from the outside (such as whether you have a contract with a vendor) are flagged to verify, never asserted. Remediation offers are technical changes, not legal outcomes.

Version

A version you can cite.

Method version
1
Part of
The standard we measure every site against, version 0.5
Evidence last re-checked
Next re-check due by
Fingerprint
7b9fc43ace94 — it changes whenever any check, grade, source, date or level on this page changes.

Cite as: Invara, “How AI sees your site”, method version 1, evidence re-checked November 6, 2026 (fingerprint 7b9fc43ace94). https://clearsitedigital.com/how-ai-sees-sites

When a platform changes what it documents, the grade changes when we re-check that source, and the version and fingerprint change with it. Each check’s own page has its own address.

See your home page the way an AI crawler reads it.

Free and read-only: your home page’s text with and without JavaScript, and in a free account, what we checked in each of the four layers.

Free and read-only: we only look, and there is nothing to install. By pressing the button you confirm this is your site or you have permission to check it, and you agree we may keep the results, with your site’s name removed, for research on websites in general (how we handle it). See a sample report.