How we measure AI visibility

Why this page exists

An AI visibility number is only worth anything if you can see how it was produced. Most tools in this market publish a score and not the prompts behind it. We publish the whole procedure: the prompt set, the engines, the scoring arithmetic, and the places where the method is weaker than it looks. If you want to run this by hand instead of buying a report, everything you need is on this page and in the step-by-step audit guide.

The prompt set

Every report uses the same ten prompts, with your category and brand substituted in. Seven are category prompts: they describe a buying situation and never name your brand, so a mention is unprompted and therefore meaningful. Three are brand prompts: they name your brand directly, which measures what the engine knows and how it frames you rather than whether it recommends you.

#	Type	Prompt	What it measures
1	Category	What are the best [category] right now? List the top options.	Default recommendation set
2	Category	Which [category] would you recommend for a small business, and why?	SMB segment fit
3	Category	Compare the leading [category] options — who stands out?	Comparative framing
4	Category	What is the best free or affordable [category]?	Price-led discovery
5	Category	What [category] do you recommend for a larger company / enterprise?	Enterprise segment fit
6	Category	What are good alternatives to [competitor]?	Displacement demand
7	Category	I'm choosing a [category]. Give me a shortlist of 5 with one-line pros and cons.	Shortlist inclusion
8	Brand	Is [brand] a good [category]? Give pros and cons.	Sentiment and objections
9	Brand	What do you know about [brand]? Who is it for?	Recall and positioning accuracy
10	Brand	How does [brand] compare with its main competitors in [category]?	Competitive framing

Prompt 6 uses the first competitor you supply. If you supply none, it falls back to asking what buyers in the category are switching to. Supplying competitors makes that prompt sharper, which is why the order form asks for up to five.

The prompts above are the English set. The engines are asked in the language you order in: order from the Dutch, German or Spanish edition of this site and the same ten questions are put to all five engines in that language, phrased the way a buyer in that market would actually type them — not translated back from an English run. That matters because the engines answer a Dutch question from what they have read in Dutch, which is a different measurement, and it is the one a Dutch brand is actually buying. Each edition publishes its own prompt set on this page.

Engines and models

Five engines are queried: ChatGPT, Claude, Gemini, Perplexity and Grok. Each is reached through its vendor current fast production model. That choice is deliberate and worth stating plainly: the fast tiers are what most people are served on free and default plans, so they are the closest available proxy for what an ordinary buyer sees. A report run against each vendor flagship reasoning model would be more flattering to describe and less representative of reality.

Only Perplexity performs live web retrieval by default. The other four answer largely from model knowledge. That distinction matters when you read the results: a gap on Perplexity usually points at a content and citation problem you can fix this quarter, while a gap on the other four points at a slower training-data and brand-presence problem. The generative engine optimization guide covers what actually shifts each one.

How brand mentions are extracted

Answers are recorded as raw text. Brand detection runs in two passes. First, a deterministic matcher looks for your brand and any competitors you named, tolerant of hyphenation and spacing variants so that "Monday.com", "Monday com" and "monday-com" all resolve to the same brand. Second, a model pass reads each answer and extracts every other product or company named as a recommendation, which is what populates the competitor leaderboard with brands you did not think to track.

Two safeguards apply. Answers that come back empty are flagged as failed and skipped entirely, so nothing is ever extracted from a blank response. And a brand-knowledge check reads the three brand prompts for explicit statements of unfamiliarity, so a report can distinguish "the engine knows you and did not recommend you" from "the engine has never heard of you". Those are different problems with different fixes.

The scoring rules

• Mention rate is the share of category answers that name your brand at least once. Brand prompts are excluded from this figure, because a prompt that names you will almost always mention you and would inflate the number.
• Average position is the mean rank of your brand within the list of brands in each answer that mentions you. Position 1 means you are named before every competitor in that answer. Answers that do not mention you do not contribute.
• Share of voice is your total mentions divided by all brand mentions across the report, including competitors surfaced by the extraction pass.
• Sentiment is classified per mention as positive, neutral or negative from the surrounding text, then aggregated. Most mentions in recommendation answers are neutral-to-positive; a negative share above roughly one in five is the signal worth acting on.
• Failed answers are excluded from every denominator. If an engine returns nothing, its rate renders as a dash, never as zero percent. Counting a technical failure as invisibility is the single most common way an AI visibility tool lies to you.

Where the citation data comes from

Paid reports include a section listing the domains that ChatGPT actually cites and consults when answering buyer questions in your category, along with an estimate of monthly AI query volume for that category. This comes from a third-party AI-mentions dataset rather than from the 50 answers, and it is labeled as such in the report. It is category-level data, not brand-level, and it is there to give you a concrete list of publications and pages worth targeting.

Known limitations

Stated plainly, because a methodology page that only lists strengths is marketing.

• Point in time. One report is one moment. It shows you where you stand today and does not track movement. If you need a trend line and alerting, a subscription tracker is the right purchase — the honest breakdown is in the tools comparison.
• Run-to-run variance. Model answers vary. Fifty answers reduce the noise substantially but do not eliminate it, and small differences between two reports of the same brand should not be read as movement.
• Ten prompts is a sample. They are chosen to cover the main buying situations, not to exhaust your category. A prompt set tailored to your specific funnel would find things this one misses.
• One language per report, general-market. A report is measured in the language you order it in — the ten prompts are sent to the engines in English, Dutch, German or Spanish, matching the edition of this site you bought from, and a Dutch report is a Dutch measurement rather than a translated English one. It measures that one language only. There is still no per-country targeting, no persona modeling and no per-country results.
• Extraction is imperfect. The model pass occasionally misses an unusually written brand name. Every report includes the raw answers in an appendix so you can check the scoring against the source text yourself.

Reproducing this yourself

Nothing here is proprietary. Take the ten prompts above, paste them into each of the five engines, record the answers, and count. That is exactly what the report automates, and it takes most people about two hours by hand versus about two minutes paid. The free checker runs one prompt on two engines if you want to see the shape of the output first, and the sample report is a complete real result — Notion in project management software, measured on 20 August 2026, with an overall mention rate of 54 percent that ranged from 43 percent on Gemini, Perplexity and Grok to 86 percent on Claude.

That spread is the most useful thing on this page. One brand, one category, one day, and the engines disagreed by 43 points. Any measurement built on a single engine is measuring that engine, not your visibility. The metric definitions are broken down further in what AI visibility means.