How we measure

The engines

5 surfaces, chosen because they are what your customers actually use. 3 are assistants we ask directly; 2 are read off live Google results.

System	Mode	Why it's in
ChatGPT	with web search	The assistant most of your customers actually use
Gemini	with Google Search	The default assistant on Android and across Google's apps
Claude	from memory, no web access	Heavily used by business buyers
Google AI Overviews	live search results	The AI answer your customers see without asking for it
Google AI Mode	live search results	Google's dedicated AI answer tab

What we don't query yet

We would rather list this than let you assume it. These are built and disclosed but not running, so they are counted in none of the totals on this page and no report claims them:

System	Status
Gemini (from memory, no web access)

Until it runs, treat every conclusion in a report as a statement about the surfaces in the first table. They lean towards Google's models, and answers there do not automatically transfer to an assistant we have not measured.

We also run every question a second time without web access. The gap between the two is the useful part: if a model names you only when it can search, the internet knows you but the model doesn't. If it names you both ways, you're in the model's baseline knowledge — a much stronger position, and a much slower one to earn.

The 30 questions

We generate the questions from your site, then show them to you for correction before we run anything.

• 25 unbranded. Your name appears nowhere. These are the ones that matter: they measure whether AI reaches for you on its own. They're spread across 4 rungs of buying intent, from someone who doesn't yet know your category exists to someone ready to book:
  • L1 — category discovery, 3 questions. Brand recall at best. Expect no clicks.
  • L2 — solution comparison, 8 questions. Gets you onto the shortlist.
  • L3 — vendor selection, 9 questions. This is where a mention becomes revenue.
  • L4 — high intent, 5 questions. The closest thing to a warm lead.
  Your report breaks your visibility down this way rather than giving you one blended score, because a blended score is what lets a tool report thousands of mentions to someone who is getting one click a day. Most of those mentions were on questions where nobody was choosing anyone.
• 5 branded. Your name is in the question. These test whether AI has its facts right about you, not whether it likes you — one each on identity, facts, pricing, competitor, reviews. The facts one exists to be graded: it pushes the engines to commit to specifics (services, hours, how long you've been going), and every claim that comes back is checked against your own site. A vague answer to a vague question hides a wrong belief. Grading the answer is how we force that belief into the open.

Google has described how its AI features issue multiple related sub-queries behind a single search — what it calls query fan-out. One question is therefore not one test. Asking 30 across the full intent ladder is how we cover the spread rather than sampling one point of it.

Do you query the API or the real chat interface?

Both, and the split matters enough to state precisely. Of our 5 surfaces, 2 are read from live Google results — AI Overviews and AI Mode are captured as they are actually served. The other 3 are queried through official model APIs.

Some vendors market interface scraping as strictly better, and on one point they're right: a chat interface carries personalisation, session history and A/B tests that an API call doesn't see. If what you want to know is "what did one specific logged-in person see on Tuesday", an API cannot tell you and we won't pretend otherwise.

There is a number attached to that, published by the vendor with the strongest commercial reason to publish it. Surfer ran 1,000 prompts through ChatGPT's real interface and through the OpenAI API side by side and found that only 24% of the brands named overlapped between the two, and only 4% of the cited sources. On Perplexity they put source overlap at 8%. Taken at face value that says an API answer and a screen answer are close to different populations, and we are not going to bury a figure because it is aimed at us. (Surfer, 3 Feb 2026.)

Two things we would want you to know before you weight it. It is a study of ChatGPT against the OpenAI API; our model surfaces are ChatGPT, Gemini and Claude, and nobody has published the equivalent for any of them, so the 24% is suggestive for us rather than measured on us. And Surfer did not publish how they decided two brand names or two URLs were the same thing, which is most of the work in any overlap figure. We have not reproduced it.

We query the API deliberately, for the thing it is better at: reproducibility. Same prompt, same parameters, no retrieval noise from someone else's session — which is the only way a second run three months later means anything. A second measurement is how you verify that a fix worked, and it only compares cleanly if both runs were asked the same way. Scraped interface results cannot be reproduced even in principle, which is also why no vendor doing it publishes a confidence interval.

So read our model-surface numbers as the model's tendency under controlled conditions, not as a transcript of any real user's screen. We have never claimed the second thing and this page is where we say so. If you want the "what a user sees" reading, the AI Overviews and AI Mode sections are the ones to trust — those are the real thing.

What the argument does not touch: the blocking findings. Whether your pages are marked noindex or nosnippet, whether an AI crawler is disallowed in your robots.txt, whether your address and phone number agree across the web — none of that changes depending on how a model was queried, and those are the findings that most often turn out to be the reason a business is invisible. When your report opens with blocked, no scraping-versus-API dispute affects it.

The verdict, and the rule that picks it

Your report opens with one of exactly 4 answers to the question you bought it to settle: is this worth your money right now? The answer is chosen mechanically from the numbers, not written by hand, so the same run always produces the same verdict and you can check the working.

Verdict	When we say it
Leave this alone	You were named in 33.0% or more of the answers to questions where a customer is choosing a supplier (tiers L3 and L4). This is checked first, because the most valuable thing this report can tell you is to keep your money.
Something is blocking you	We found one of 6 specific settings that make a page ineligible for AI answers outright — a noindex or nosnippet directive, an AI crawler blocked in robots.txt, or a page that only exists after JavaScript runs. A known mechanical cause outranks any statistical reading of the same run, because you can verify it yourself and fix it once.
Invisible in key segments, but closing the gaps is quick and low-cost	You are not being named, and there is evidence the engines can already reach your material — they read your site and recommended someone else, or you surface occasionally. Better position than it feels like.
Invisible, and we don't think spending fixes it	No mechanical cause, no sign the engines are reaching you, competitors named throughout. We will say so. It is the least comfortable of the 4 verdicts and the one most likely to save you a year of retainer fees.

Each verdict carries a confidence level with the reason attached. We borrow the floor from Trakkr, who publish the only public statistical rule we could find in this field: under 20 mentions, a rate is a direction and not a measurement. Below that line your report says so rather than rounding up to confidence it hasn't earned. If any part of the method could not run for your site, the confidence drops a step automatically, whatever the numbers say.

How a mention is counted

String matching is not enough, and this is where most free tools go wrong. Ask “tell me about Acme Dental” and the answer will contain “Acme Dental” whether or not the model knows anything. That's an echo, not recognition.

A separate model pass classifies every response into one of:

Result	Meaning
Recommended	Named as an option someone should consider
Mentioned only	Named, but not as a recommendation
Cited, not recommended	Your site was used as a source while a competitor got recommended. This one hurts, and it's common.
Echoed	Your name appears only because it was in the question
Absent	Not there

Alongside each: your position in the list when named, and your share of voice against every other business that came up.

Repeat sampling

AI answers change. In one 30-day Semrush tracking study, 36% of the 81 pages it tracked were cited by AI Mode on day one and 26% on day 30 (Semrush, accessed 2026-10-27). Citation sets also shift with every model update. You pick the 3 questions that matter most to you; we ask each of those 3 times and report how consistent the answer was. A finding that held all 3 times and a finding that appeared once are labelled differently in your report.

We do not repeat all 30 questions — that would roughly triple the cost of the report without changing any decision you'd make. We also don't sample across multiple days, because you'd be waiting a week. Both limits are stated in your report.

If you serve a local area

Google describes local ranking as relevance, distance and prominence. We measure the first and third directly. For distance, we run your key queries from several coordinates around your address and record where in the map pack you appear from each — which shows you the practical edge of your reach rather than a theoretical radius.

We also check your Google Business Profile primary category (the single highest-leverage field there is), your review volume, rating, recency and response rate, and whether your name, address and phone number agree across your own site and the directories AI reads.

What Google itself says

There are “no additional requirements … nor other special optimizations necessary” for a page to appear in Google's AI features.
— Google Search Central, AI features documentation (accessed 2026-10-27)

We quote this on purpose, even though it undercuts the premise of the entire GEO industry. Google's position is that a page needs to be indexed and eligible to show a snippet, and that you don't need to create new machine-readable files or special markup. We think that's broadly right.

Which is exactly why the valuable thing is not optimization advice — it's finding out what AI currently says about you. Nobody can tell you that without running the queries, and running them properly costs money and takes 302 API calls. That's the product.

What we cannot measure

Stated here rather than buried, and repeated in an appendix in every report:

• Precision. 30 questions is a sample, not a census. The direction (whether AI knows you, who it prefers instead) is solid. Treat an exact percentage as approximate, and compare it against your own later re-run rather than against someone else's number.
• Personalised results. We query logged-out. A signed-in user with search history may see something different — for some categories, very different.
• Volume. We can tell you whether AI names you. We cannot tell you how many people asked.
• Revenue impact. Nobody can currently attribute a sale to an AI mention with any rigour. Anyone claiming otherwise is guessing.
• The future. A finding is true on the date stamped on it. Models update; some of it will be stale in a month.
• Every engine, everywhere. Google AI Mode isn't available in all countries. Where a check couldn't run for your market, your report says so instead of quietly dropping it.

Version policy

This method is frozen. Question counts, the classification rules, and every metric definition are fixed. If we change any of them, the version changes, and reports run under the old version keep saying so. Two reports carrying the same version stamp are therefore always comparable. (The version identifier itself is printed inside every report, where it belongs.)