How to run an AI visibility audit (a repeatable method)
An AI visibility audit has four steps: write down the questions your buyers actually ask; put them to several engines repeatedly and record who is named, recommended and cited; check whether your site can be read and cited by AI crawlers; then rank fixes by which failure you actually have. The traps are sampling once, measuring one engine, and asking about yourself by name.
Updated 22 July 2026 · By Eli
You can buy this as a product — ours is the index — but the method is not a secret, and a rough audit run honestly beats a polished one run wrong. Here is the repeatable version, with the traps marked.
Step 1: write the questions buyers ask
Not questions about you — questions a buyer with money asks before knowing you exist: "where can I buy X online in the UK", "best shop for Y", "who delivers Z by Friday". Ten to twenty of them, phrased the way real customers phrase things. Mine them from your search-console queries, your customer emails, and the forums in your category.
Trap: asking "what do you know about [your brand]?" That measures recognition, not visibility — an assistant can describe you fluently when asked and still never volunteer you to a buyer who didn't.
Step 2: sample, don't ask once
Put each question to each engine several times, ideally across days. Answers vary run to run — the same question can name you Monday and skip you Tuesday — so a single ask is an anecdote. Record, for every answer: which brands were named, which were actively recommended, and which sources were cited.
Trap: one engine. They disagree seven-fold about the same brand, so a ChatGPT-only audit is a ChatGPT audit. Cover at least ChatGPT, Gemini and one citation-heavy engine like Perplexity.
Step 3: count honestly
Three numbers per brand: mention rate (share of answers naming it), recommendation rate (share actively steering to it — far rarer, and the one that moves sales), and citation rate (share citing its site). Then a share of voice — your mentions over everyone's — remembering the denominator must include every brand the answers named, not just the rivals you already watch. Write down who the assistants cited: that list of sources is your actual battleground.
Trap: treating small differences as findings. At modest sample sizes, a few percentage points is noise. Trends across repeated audits mean something; wobble does not.
Step 4: check the readable half
While the sampling runs, audit whether AI can read you at all: crawler policy (including your CDN's bot defaults), JavaScript-dependence, structured data, and whether your key pages answer a buyer's question in a liftable first screenful. This is the fast half to fix, and it gates everything else.
Turning it into decisions
The audit's output is a diagnosis: unreadable site → fix this month; absent from the roundups and threads the engines cite → start the editorial and community work now, it is slow; named but never recommended → study what the cited sources say about the brands that do get the nod. Re-run quarterly with the same questions, same engines, same counting — an audit's value compounds only if successive readings are comparable.
Or start with the readable half in two minutes: run a free scan.