See why AI assistants do not cite your site. Measure what changes when you fix it.

Each stage files its output in your project folder. Baseline, diagnose and fix run in /site-check or /aeo-baseline. The re-run is /aeo-recheck.

Baseline

You give a URL, the engines to test, the market and language, and the competitors to track. Site Check builds a fixed set of buyer questions from real phrasings, each with its source. It runs every question on every engine you name, repeated at least as many times as the minimum you set. Where your Claude Code session can reach an assistant, it runs the question itself; where it cannot, you run the question there and paste the answer, and the record says which. It records whether each answer mentioned you and whether it cited a page of yours as two separate facts, plus every source the answer drew on, whether or not you appeared. The questions, engines, market, language, competitors and results go into a panel file, so the re-run uses the identical frame. Every number comes with its sample size, engine, market and date. An engine the run cannot reach, or too few repeats, reads NOT CHECKED.

Diagnose

For each question where your site was not cited, Site Check records who was cited instead and sorts the gap into a cause, with the evidence. The causes are in the next list.

Fix

Each AI-related fix names the questions it targets, the cause it addresses, and a check that proves the fix was made, such as a curl line or a page that now carries the answer. The list goes to whoever built the site. No fix comes with a promise that citations will change.

Re-run

Run /aeo-recheck (or /geo-recheck) with the earlier report and its panel file. Site Check first checks each AI-related fix and marks it made, not made or NOT CHECKED. It then re-runs the identical panel on the same engines at the same repeat minimum, and gives you a before-and-after table for every question, with mention and citation on their own rows and the change in cited sources. For each fix it shows how the questions it targeted changed, and labels that change “moved with”, never “caused”. It flags an engine or model version change, drift in the panel, and too few repeats. The report suggests a re-run date, 30 days after the fixes by default. That is a product default, not a known interval for any engine to update. Site Check has no server and watches nothing, so you start the re-run.

The causes Site Check looks for.

Reach

The engines cannot get the page: an AI crawler blocked in robots.txt, a noindex, a page that is not indexed, content that appears only after JavaScript, or a missing page. The evidence comes from the search eligibility check.

Answer

No page of yours answers the question in a form an assistant can lift: no page, the answer buried, or no plain standalone answer. The evidence is your closest page, quoted, next to the cited page’s answer, quoted.

Corroboration

The cited answers lean on third-party sources such as reviews, directories, forums, lists or news, and you are absent from them. The evidence is the source URLs and whether you appear on each.

Accuracy

An answer states something wrong about your business. The evidence is the quote and the page of yours that should state the fact.

A cause is a diagnosis from the evidence gathered. It is not proof of how any assistant decides what to cite.

Site Check checks a website or landing page before launch, or before promoting it, inside Claude Code. The same run that asks the AI assistants checks the site itself. When visibility is poor, it records whether the search eligibility check passed before it blames anything else. Each check names its evidence and its rule for pass, fix or stop.

Search eligibility

Can search engines find and read the site? Pages that answer, content present in the raw HTML, robots.txt, noindex directives, canonicals, the sitemap, redirects, titles and descriptions, internal links, retired files that still answer, the policy for AI crawlers, and whether llms.txt is accurate. Speed runs through PageSpeed Insights. Search Console rows are checked when you give it access.

Structured data

Does the markup parse, match what a visitor sees, and point to profiles that exist? Ratings, review counts, prices and regulated fields go to you to confirm before the section can pass. Site Check does not treat markup as a way to get cited by AI assistants. Google states that no special markup is needed for its AI features.

Design

Does the page read as AI-generated? Typefaces, palette, repeated structure and copy, plus the cover-the-logo test: hide the name, read the headline, and see whether it could belong to any company. Font licenses and borrowed assets are flagged.

Conversion structure

One primary action, a hero that names the visitor’s problem or outcome, proof you can attribute, a plain offer and price, answers to the likely objections, the page at phone width, and whether header links give a visitor an exit.

Page claims

Every factual claim on the page: prices, counts, results, comparisons, guarantees, statements about third parties, legal or regulatory statements and testimonials. Each row carries a tag, a source and the date it was read.

Content decay

For an existing site with a Search Console export, it compares equivalent periods, separates a sitewide drop from a page-specific one, and protects pages that convert or hold links. Without the export this section reads NOT CHECKED and says why.

AI readiness, when it applies

Every run asks one question: does the site use AI to make or influence decisions about people, or to generate content shown to customers? If so, the report gains a section that lists what to check and the questions for a licensed professional. It returns only Not verified or Needs expert review, never a pass.

Handed a GEO or AEO vendor’s proposal, Site Check quotes each promise, tags it vendor claim, and grades it supported, unsupported or contradicted by what the run found and by what the engines publish about their own features.

/site-check

The full check on a site or landing page, with the citation gap leading the summary. It asks what you want to know, whether the site is new or existing, and whether it uses AI to make decisions about people or to generate content shown to customers. The report and the panel file are filed in your project.

/aeo-baseline

The AI answers baseline and the citation gap alone, with the panel file saved so you can re-run it later. /geo-check runs the same check under the name many people use for it.

/aeo-recheck

The re-run. Give it the earlier report and its panel file. /geo-recheck runs the same command.

/page-claims

The claim ledger alone for one page, when you only need to know which claims hold.

Lost? Type /bb, or /next for where the project stands.

Markdown files in your project folder, filed by date and site. This is the layout of a report. It contains no findings; those come from your site.

Before you see it, the report goes to a verifier running in a fresh context. Confirmed findings get fixed. Judgment calls are listed for you.

See a sample report. It is a real run of Site Check 0.2.0 on backbrief.ai, Backbrief’s own site, dated 2026-10-13. It shows the citation gap and names the saved panel. The answer engine ran through a signed-in session, and the report says so. The recheck follows once the fixes are made.