What an AI Visibility Audit Actually Checks in 2026

An AI visibility audit checks four things: what ChatGPT, Claude, Perplexity, and Google’s AI Overviews currently say when a buyer asks for a recommendation in your category; which sources those answers are built from; whether your own pages and business records are structured so the engines can read and trust them; and which reputation signals are missing. Anything that skips the first step is not an audit — it is a checklist applied to a site without evidence.

The reason this became a distinct piece of work is that ranking and being recommended have come apart. SparkToro’s 2026 analysis found that fewer than a third of Google searches still send a click to any website, which means a growing share of buyers form a shortlist without ever landing on a page you control. A site can rank respectably and still be absent from the sentence that actually decides who gets called.

What does an AI visibility audit actually measure?

It measures your presence inside generated answers, not your position in a ranked list. The core deliverable is a sample: a set of real buyer questions run against each assistant, with the answer text, the businesses named, and the sources cited recorded for every run. Everything else in the audit exists to explain that sample.

That distinction matters because the two measurements disagree more often than most buyers expect. Position tells you where a blue link sits; an answer sample tells you whether you appear at all in the text above it. In our own weekly sampling we have watched our Salt Lake City page move from position 73 to position 35 in Google over four weeks — a genuine, measurable gain — during a period when the same business was named in roughly a third of the ChatGPT samples we ran. Both numbers are real. Neither predicts the other, which is exactly why an audit has to collect both rather than inferring one from the other.

A competent audit therefore reports two separate scoreboards. The search scoreboard covers impressions, position, and the queries you already surface for. The answer scoreboard covers how often you are named, which competitors are named instead, and what the engine cited to reach that conclusion. Read together they tell you whether your problem is reach, structure, or reputation — three problems with three different fixes.

Which sources do the engines actually cite?

The cited-source mix rotates, and that rotation is the single most useful thing an audit can document. Over nine weeks of running the same Utah buyer question against ChatGPT, we have logged four distinct patterns: weeks where it quoted agencies’ own service pages, weeks where it leaned on directories, weeks where it built the answer from third-party “best agencies” roundups, and one week where it returned a Google Business Profile panel ranked by review count.

That variation has a practical consequence. If your audit samples once and reports the result as your standing, it has measured a single frame of a moving picture. The correct output is a series with a date on every entry, so you can tell a genuine trend from a phrasing artifact. When we sampled this week, the answer opened with a local-business panel in which every named agency carried between 22 and 113 Google reviews — a format in which no amount of on-page work would have qualified us, because the panel was assembled from business listings rather than from web pages.

An audit should name the specific sources, not summarize them. “Third-party mentions matter” is not actionable. “This week the engine cited Expertise.com’s Provo roundup, two competitors’ own city pages, and a review-ranked business panel” tells you precisely which three doors exist and which one you are closest to opening.

What on-page factors does the audit look at?

It checks whether your pages answer questions in extractable, self-contained blocks, and whether the claims on them are sourced. Generative systems retrieve passages, not whole documents, so a page that only makes sense read top-to-bottom tends to lose to one built from paragraphs that stand alone.

The research here is unusually specific. Aggarwal et al. (Princeton, 2023), the paper that named generative engine optimization, found that adding citations, quotations, and statistics to a page lifted its visibility inside generated answers by up to 40%. Those are structural changes rather than keyword changes, and they are straightforward to check: does each section open with a direct answer, is there at least one verifiable statistic per section, and is every external claim linked to a named publisher.

The audit should also flag the opposite problem, which is more common than under-optimization. Unsourced performance claims — the “40% more citations” and “5x ROI” numbers that decorate most agency sites — are a liability in an environment where the assistants themselves now coach buyers to ask for evidence. We removed four such claims from our own service pages earlier this year for exactly that reason, and we would flag yours.

What the audit checks	Where it lives	How fast it moves
Answer-first structure, sourced claims	Your own pages	Weeks
Entity records, schema, canonical facts	Your site + listings	Weeks
Business listing completeness, categories	Google Business Profile	Weeks
Reviews and rating volume	Third-party platforms	Months
Roundups, directories, press mentions	Sites you do not control	Months

Why do reputation signals show up in a visibility audit?

Because in several answer formats they are the only qualifying criterion. Research from Trustpilot, reported by TechRadar Pro (2026), found businesses with 80 or more reviews appeared in over 75% of AI answers, while those with none appeared in roughly 1%.

We are a live example of the gap that creates, and it is worth stating plainly. Our own Google Business Profile has zero reviews. When this week’s sample returned a review-ranked local panel, that single fact excluded us from the list regardless of how well the rest of the work had been done. An audit that reported only our content and schema — both of which are in good shape — would have told us nothing about the actual constraint.

This is also where an honest audit becomes uncomfortable. Business Insider’s August 2026 reporting documented small businesses losing traffic to AI Overviews with no clear recourse, and some of what an audit surfaces genuinely sits outside your direct control. Saying so is more useful than a report that finds only problems the auditor happens to sell a fix for.

What should an AI visibility audit cost?

Ours is free, and we think a first audit generally should be, because the first pass is a diagnostic rather than a deliverable. You can run our free AI visibility audit on your own site and get the scan back in minutes; the parts that require judgment — reading the answer samples, identifying which competitor is being named instead of you and why — are what we walk through on a call.

We are not going to quote market rates for other firms’ audits, because prices vary enough by scope that any range we published would be a guess dressed as a fact. What we will say is how to judge one. Ask whether the deliverable includes dated answer samples with the cited sources listed, or only a site crawl. Ask which findings the firm cannot fix, and treat a report with no such section as incomplete. Ask what the audit would look like re-run in eight weeks, because a single snapshot cannot show a trend. Our own published rates are on the pricing page if you want to see what the ongoing work costs after the audit.

What we’d actually do

If you are buying one audit, buy the one that samples the assistants and shows its working. The scan portion — schema, headings, structure, listing completeness — is largely commoditized and several tools do it adequately. The part that is hard, and the part that changes what you do next, is the evidence: which businesses the engines name in your category this month, what they cited to get there, and which of those doors is actually reachable from where you stand.

Then re-run it. The most valuable thing in our own file is not any single sample but the nine-week series, because the series is what revealed that the answer format rotates and that our binding constraint was reviews rather than content. One audit tells you where you are. A series tells you what to do. If you want to see what that looks like against your own category, book a 30-minute call — no charge, no obligation, and you keep the findings either way.