The AI Visibility Audit: 10 Checks You Can Run Yourself
A ten-point audit of whether AI assistants can find, read and accurately describe your AI product. Each check has a pass condition and a fix.
Published 2026-10-05 · 5 min read
Whether an AI assistant can describe your product accurately depends on a short list of technical and editorial conditions. All ten are checkable in an afternoon with no tools beyond a browser and an assistant account.
Each check below has a pass condition and a fix. Run them in order — the early ones gate the later ones, so a failure at check two makes checks five to ten moot.
Check 1: can a crawler reach the page at all
Pass: robots.txt permits the retrieval crawlers, and the product page returns a 200 to a request with no cookies and no JavaScript.
- How to test: fetch your robots.txt and read it. Then load the product page in a private window with scripts disabled, or fetch it with a plain command-line client.
- Common failure: a default-deny robots file, a cookie banner that blocks the first request, or content rendered only client-side.
- Fix: explicitly permit the retrieval agents; server-render the substance; move consent walls off the public pages.
Check 2: is the page indexed
Pass: a site-restricted search for your domain returns your product page.
- How to test: search for your exact domain restricted to that site, and separately for your exact product name in quotes.
- Common failure: a noindex tag inherited from a staging configuration, or nothing linking to the domain so it was never discovered.
- Fix: remove the tag, submit a sitemap, and get ten inbound links from directory listings to force discovery.
Check 3: does the first sentence define the product
Pass: reading only the first sentence of your product page, a stranger could say what the product does and who it is for.
- How to test: copy the first sentence into a blank document and read it with no context. Better: give it to someone who has not seen the site.
- Common failure: the page opens with a value proposition, a slogan, or a rhetorical question. All three require context that a quoted sentence will not carry.
- Fix: rewrite as "[Product] does [specific thing] for [specific person]." Move the value proposition to the second sentence.
Check 4: are facts in real structure
Pass: pricing, specifications and comparisons are in actual table elements with headers, not in styled divs or prose.
- How to test: view source and search for table elements around your pricing and spec content.
- Common failure: a component library that renders tables as nested divs. Visually identical, far less reliably extracted.
- Fix: use semantic table markup, or add the same facts as structured data alongside.
Check 5: is the entity consistent
Pass: product name, canonical URL and description match across your site, every directory listing and every social profile.
- How to test: open three listings and your homepage side by side. Note every difference in the name spelling, the URL they point to, and the first sentence of the description.
- Common failure: name variations, www versus non-www, an old domain still resolving, descriptions written months apart.
- Fix: pick one canonical form, redirect every variant, and standardise the description everywhere. This is the highest-value fix in the whole audit.
Check 6: does the markup declare it
Pass: the product page carries SoftwareApplication markup with name, url, applicationCategory and sameAs; the site carries Organization markup once.
- How to test: view source and look for the JSON-LD block, or paste the URL into a structured-data validator.
- Common failure: no markup at all, or sameAs missing. sameAs is the property that connects your site to your profiles, and it is the one most often omitted.
- Fix: add JSON-LD generated from the same data that renders the page, so the two cannot drift. Omit any property you cannot fill truthfully.
Check 7: is there a machine-readable summary
Pass: an llms.txt exists at the root, describes the site in one paragraph and links the important pages.
- How to test: fetch /llms.txt.
- Common failure: absent. Adoption is still low, so this is a check most products fail without penalty — but passing it is cheap.
- Fix: generate it at build time from the same content that renders your pages. Hand-maintained versions go stale.
Check 8: what does an assistant actually say
Pass: asking an assistant about your product by name produces an accurate description, and asking about your category mentions you where appropriate.
- How to test: two sets of questions. By name: "what is [product]" and "what does [product] do". By category: "best tool for [specific job]" and "[category] alternatives". Ask each assistant you care about and record the answers verbatim.
- Common failure: the product is unknown, or it is described using stale or wrong information — usually sourced from an old listing rather than your site.
- Fix: trace the wrong fact to its source. It is almost always an outdated listing or an old article. Correct the source, not the assistant.
Check 9: is the traffic measurable
Pass: your analytics has a saved segment for AI referrers, with a recorded baseline.
- How to test: look for a segment matching the assistant domains, and check whether it has any history.
- Common failure: no segment, so AI traffic lands in direct and referral and cannot be distinguished. UTM parameters do not survive assistant links, so referrer is the only method.
- Fix: build the segment now and record the baseline. Without a before, you cannot tell whether any later change worked.
Check 10: is there a date, and is it honest
Pass: important pages carry a visible publication or review date that reflects when the content actually changed.
- How to test: look at your product page, comparison pages and any guides. Is there a date? Does it correspond to a real change?
- Common failure: no date at all, or an auto-updated "last modified" that changes on every deploy and therefore means nothing.
- Fix: show a date that changes only when the substance changes. Inflated freshness is detectable and destroys the trust the date exists to create.
Scoring it
| Passes | Reading |
|---|---|
| 0-3 | Retrieval is likely blocked entirely. Checks 1 and 2 gate everything else |
| 4-6 | Reachable but not quotable. The writing and structure checks are where the work is |
| 7-8 | Competitive. Entity consistency and markup are probably the remaining gaps |
| 9-10 | Unusual. Maintain it and re-run quarterly |
Nine of these ten checks are things a model can only guess at if you have not done them. That is the whole opportunity: the work is unglamorous, mostly unclaimed, and mostly free.
More in SEO & GEO
- Entity Consistency: Why AI Models Fail to Attribute Your ProductWhen a model cannot tell your mentions describe one product, it attributes none of them. The signals that resolve an entity, and how to audit yours.
- How to Measure AI Referral Traffic When UTMs Are StrippedAI assistants strip UTM parameters, so referrer segmentation is the only reliable method. The setup, the domains to watch, and how to read the numbers.
- The 10-Page SEO Starter Kit for a New AI ProductThe ten pages a new AI product should have before writing anything else, what each one is for, and the order to build them in.