AI visibility audit: what a real one contains

182 sites have gone through the same fifteen readiness checks. That is enough to show a pattern: basic access is common, while clear commercial information is much less so. An AI visibility audit needs to measure both.

A real audit has two gates

The first gate is mechanical and editorial. It checks whether crawlers can reach the content, whether the page says what the business does, and whether important facts appear in a form an AI system can use. This is where an AI readiness check helps. It finds sites where robots.txt blocks AI crawlers, where no price is visible, or where the plain statement of audience and cost is missing.

The second gate is external. It asks what appears in answers to buyer questions, which pages are cited and which page types receive attention. That measures an outcome rather than a condition.

A readable site can still be absent from an answer.

A useful audit reports two result sets, not one blended score. The first says what you can fix on your own site. The second shows where assistants currently get their evidence.

Fifteen checks expose readable gaps

The fifteen checks cover access, structure, clarity and useful buying information. In the sample of 182 sites, the easiest check was whether the homepage answered a plain request: 178 sites passed. AI crawlers were allowed in robots.txt on 175 sites, and 153 sites had a published sitemap. Those figures describe the route to the content, not whether the content gives an assistant enough evidence to recommend the product.

The main weakness was commercial clarity.

Only 33 sites passed the check for a plain statement of what the product is, who it is for, what it costs and what it is measured against. The check recorded 49 failures and 96 warnings. Comparison or alternatives pages passed on 91 sites, failed on 41 and warned on 50. Questions answered on the page passed on 99 sites and warned on 79.

The site can be reachable while the offer remains unclear.

Here is a practical checklist that groups the fifteen checks into seven working items:

• Check that the homepage states what the product is, who it serves, its price and the relevant alternative or comparison point.
• Check that one clear H1 names the page topic, rather than leaving the main subject to a slogan or image.
• Check that title and description text say what the page is about.
• Check that important content is present without JavaScript being required to reveal it.
• Check that robots.txt allows relevant AI crawlers and that a sitemap is published.
• Check that product, organisation and question information is marked up where the page supports it.
• Check that buyers can reach pricing, comparisons and direct answers without reconstructing the offer from scattered pages.

The readiness results show why these checks need to stay separate. Organisation schema passed on 113 of 182 sites, while Product or SoftwareApplication schema passed on 102. A clear H1 passed on 151 sites. Price was visible or linked from a pricing page on 145. These are different signals, and a site can pass one while leaving another unresolved.

The readiness check is not a ranking forecast. It reports whether the site supplies usable material and access. That makes it useful, provided you do not treat it as a visibility result.

Take the plain business statement check. Start with the full sample: 182 sites. Next, separate the outcomes: 33 passed, 49 failed and 96 warned. Those figures account for the sample, so the result is more useful than a general comment that the messaging could be sharper.

The result gives you an action split.

Pass, fail and warning describe where each site landed on this check. The largest group was the warning group, so a team should inspect the page rather than applying one automatic rewrite to every site.

Access was rarely the gap.

Then compare it with the access checks. 178 sites passed the homepage request check and 175 passed the AI crawler check. The useful finding is the difference between easy access and clear commercial explanation. More crawler work will not fix a product that never states who it is for.

The scan tests citation reality

The second half starts with buyer questions, not a technical score. CiteGraph has recorded 64,667 citations across 9,791 AI answers. Those answers came from 77 scans of 44 sites in 48 categories, using 507 distinct buyer questions. The scan data covers 9,192 distinct domains and 21,465 distinct pages, which gives the audit something a readiness report cannot: observed source behaviour.

The scan has limits. These figures come from the recorded scans and do not describe every engine, category, question or future answer.

The page mix is sharply uneven. Listicles and roundups account for 68.3% of recorded citations, or 44,164 citations. Competitor product pages account for 24.1%, or 15,575 citations. Together those two kinds of page take 92.4% of citations in the sample.

Other source types occupy much smaller shares.

The scanned product's own site accounted for 1.7% of citations, or 1,107 citations. Community threads accounted for 3%, or 1,930. Documentation accounted for 0.8%, with 509 citations. The scan also records social posts at 0.8%, review platforms at 0.8%, YouTube at 0.4%, news at 0.1% and Wikipedia at 0%, recorded as 25 citations. The smaller categories are included for completeness; the practical contrast is between external comparison pages and the product's own site.

This changes the work plan. A readable pricing page is still useful. It may help an assistant verify a fact once the page is found. It does not, by itself, create the listicle, roundup or competitor page that the scan shows taking most citations. Your audit needs to identify both the pages you control and the source types you need to earn or influence.

A scan makes the question more precise. Instead of asking whether a brand has AI visibility, ask which buyer questions produced citations, which domains appeared, what page type they used and whether your own site appeared. The measurement method guide explains that approach in more detail.

The source pattern should shape the next task.

The audit should end in actions

A useful report turns the two halves into separate work queues. The readiness queue contains fixes such as clearer product statements, comparison pages, question answers, schema and access. The citation queue contains source gaps, buyer questions with weak coverage and page types that appear often in the sample. The queues can inform each other, but they should not be merged into a single score that hides the next action.

Use this order when the team is small. First, fix anything that prevents reading: blocked crawlers, missing content without JavaScript, unclear page titles and inaccessible pricing. Second, make the offer legible: state the audience, cost, use case and comparison point in plain language. Third, publish or improve the pages that answer buyer questions. Fourth, scan those questions and inspect the sources that appeared instead.

Start with the barrier that stops a page being read.

Then look at the gap between your controlled pages and the pages the scan cites. If listicles and roundups dominate the sample, inspect the lists that already answer your category questions. If competitor product pages appear often, make sure your own comparison and alternatives pages give buyers a reason to include you. If your own pages rarely appear, do not assume the problem is one missing schema field. The page may be clear and still lack the external evidence that an answer uses.

A report should show the original question, the answer, the cited page, the source domain and the relevant site finding. That lets you move from “we have low AI visibility” to a task someone can own. A founder can assign the page. A marketer can brief the comparison. An SEO can check access and rendering. Nobody needs another mysterious score in a spreadsheet.

The next question to ask is: which buyer questions are producing citations for other sites, and where does the answer live? The answer lives in a scan of those questions, beside the fifteen readiness checks, not in the readiness score alone.