Methodology

How Recall measures AI visibility

Everything below is reproducible from the description alone. Where a figure appears, it comes from running this method against our own brand between 19 September 2026 and 21 September 2026 — including the result that does not flatter us.

Sampling

A fixed set of 22 unbranded buyer questions — "best AI visibility tools", not "is Recall any good" — is sent to ChatGPT, Perplexity, Gemini with web search enabled, 8 times per question per engine, every 24 hours. Branded questions are excluded: they measure how an engine describes a company already named in the prompt, which is a different quantity from whether it volunteers that company unprompted, and pooling the two inflates every score.

Repetition is not redundancy. The same question asked twice returns different answers, so a single run is an anecdote. 8 runs is the floor at which a brand's share of answers can be estimated closely enough for a change to be detectable at all.

What counts as an answer

Every response is classified before it is counted: answered, hedged, refused, not grounded, no answer, or error. Only measurable outcomes enter a denominator. A refusal is not evidence that a brand is invisible, so counting it against every brand in the set would understate all of them equally and hide the refusal itself. In the run below, 66% of probes produced a usable answer.

Statistics

Each share of answers is reported as a Wilson score interval, not a bare percentage. A lift is only called when the intervals for two matched samples separate; when they overlap, the verdict is "inconclusive" and says so. This is the difference between measurement and reassurance, and it is why a Recall report will sometimes decline to tell a customer their work succeeded.

Four causes of absence

A brand missing from an answer is not one problem. It is four, and they have four different fixes — which is why a single visibility score cannot tell anyone what to do next.

Retrieval absence

The engine cannot fetch the pages. Blocked crawlers, JavaScript-only rendering, or a site too thin to answer from.

Entity ambiguity

The engine cannot tell which company is meant. No structured data, or nothing linking the site to the same entity elsewhere.

Corpus absence

The engine reads sources that never mention the brand. Nothing on the brand's own site fixes this, which is why diagnosing it matters.

Framing loss

The brand is described in terms its buyers do not use, so it is not retrieved for the question they asked.

Our own result

We ran this against dorecall.com before running it for anyone else. 1,608 probes produced 9,800 citations across 620 distinct domains, at a measured cost of $1.83. Share of answers naming each tool:

Recall was named in none of them, and our domain appeared in none of the 620 sources the engines cited. Our own site is not the problem: crawlers are allowed, pages render without JavaScript, the organisation is declared and linked. This is corpus absence— the engines read sources that do not mention us. We are publishing the number because a measurement tool that only reports flattering results is not a measurement tool.

Limits

• These figures are a recorded run, not a live view. Re-measuring per visitor would cost money per page load and produce a number nobody could cite twice.
• Answers vary by locale. This run was global English; a brand selling in one country should be measured there, because cited sources barely overlap across languages.
• Share of answers is not market share. It measures what engines say, which is downstream of what has been written about a company, not of how good the company is.

Citing this

Every heading here has a stable fragment ID, and the figures are dated. Quote them freely with a link back to this page. If you want the underlying runs for a claim, ask and we will send them.