An AI visibility audit tests whether ChatGPT, Claude and Perplexity can reach, read and trust your site, across four groups: access, rendering, structured data, content and authority. A trustworthy one also separates that from the recognition probe, real buyer questions asked across all three engines, and hands back the observed count per engine, not a single score.
An AI visibility audit is the practical, repeatable version of AI brand intelligence: the same inputs, recognition, description, share of voice, sources and accuracy, turned into a fixed test you can run on demand and compare over time, instead of a one-off impression.
It checks five separate layers, each answering a different question about whether AI can use your site at all. Access asks whether AI crawlers can even reach your pages; rendering asks whether they can read what loads; structured data asks whether they can understand who you are; content asks whether they can quote you directly; authority asks whether the rest of the web corroborates what you claim about yourself.
| Layer | Question it answers | What a failure looks like |
|---|---|---|
| Access | Can AI reach your pages at all? | robots.txt blocks GPTBot, ClaudeBot or PerplexityBot |
| Rendering | Can AI read what loads? | Key content only appears after JavaScript runs |
| Structured data | Can AI understand who you are? | Missing or inconsistent Organization schema |
| Content | Can AI quote you directly? | No answer-first paragraph for the buyer's actual question |
| Authority | Does the web corroborate you? | No independent mentions beyond your own site |
Bottom line: a page can look perfect to a human and still fail an audit at layer one or two, before content quality ever enters the picture.
Access and rendering come first because they gate everything else: if AI can't get into a page or can't read what's on it, the remaining three layers never get a fair test. Reflexa's own audit is built this way on purpose, the first two layers gate the score, so a blocked or unreadable site never gets a flattering composite regardless of how good its structured data or copy is (Reflexa, How it works, 2026).
Do this: fix access and rendering findings first, always, even if a content or authority finding looks more urgent on the page.
The recognition probe is the part you run yourself, by asking the questions a buyer would actually type, not your brand name, and counts how often each engine names you in the answer. A buyer question is the unit this whole exercise runs on: a real, purchase-shaped question like "what's the best [category] for [customer]," asked more than once, across more than one engine, so a single lucky or unlucky answer doesn't get mistaken for a trend.
Proof: in the Reflexa AI Visibility Index's September 2026 project management edition, 100 fixed buyer questions were asked three times each across ChatGPT, Claude and Perplexity, and Jira appeared in the top three recommendations in only 59 of its 882 eligible answers, an observed count, not an estimate.
Why it matters: a recognition count built on real buyer questions tells you what AI actually says today, not what you assume it says based on your Google ranking.
Be careful of any audit that blends a hard failure with a soft one into a single number, because that number can't tell you what's actually broken. A blocked crawler and a thin paragraph are not the same kind of problem, and averaging them into one composite score hides which layer to fix first. A trustworthy audit keeps them separate: a finding that reads "blocked" stays "blocked" in plain language instead of turning into a softer-looking grade, and low recognition counts are shown as counts, not as a manufactured percentage out of 100 (Reflexa, What to Fix, 2026).
Rule of thumb: if a report can't show you the raw evidence behind its headline number, treat the number itself with suspicion.
Read the recognition probe first to see whether you're named at all, then read the layer findings behind any low count before touching anything. An access or rendering finding is worth fixing before a content or authority one, because a structured-data improvement can't help a page AI can't reach in the first place. A full section-by-section walkthrough of a report is covered in how to read a Reflexa AI-visibility report, if you want the order past this first pass.
Shortcut: on a first read, you only need two sections: the recognition probe to see where you stand, and the top finding under Site checks to see what to change first.
None of this requires buying anything, and each check takes a few minutes:
Bottom line: you can test three of three of the four groups yourself with free tools in under ten minutes, before running anything that touches your whole site.
It is a test of whether ChatGPT, Claude and Perplexity can reach, read, understand, quote and trust your website, run across four groups (crawler access, page content, brand identity, external sources) plus the buyer questions asked across the engines.
Crawler access (can the engines reach the pages), page content (can they read and quote them), brand identity (do they know who you are), external sources (do others vouch for you). Access and content gate the other two, so a blocked or unreadable site never gets a flattering result.
The part you run by hand or with a tracker: ask real buyer questions, such as 'what's the best X for Y,' across ChatGPT, Claude and Perplexity and count how often each engine names you. A site audit cannot produce this number; it tells you whether the engines could read you when they answered.
Be cautious of any audit that blends a hard failure, like a blocked crawler, with a soft one, like thin content, into a single number. A trustworthy audit reports each layer's result plainly and shows observed answer counts rather than a synthetic score.
Minutes for the site half. Reflexa's free page check fetches a site the way AI crawlers do and runs the 22 checks in one pass, with the exact finding behind each; the recognition half — asking the engines — you run yourself or with a tracker.
The free check reads your pages the way ChatGPT, Claude and Perplexity read them — 22 checks, evidence attached, no card needed.