How Recometric measures your AI visibility.
No black box. This page is the whole instrument: the questions we ask, what we read out of each answer, how the three components combine into one number, and what the system is doing while a scan runs.
What we ask the four assistants
We ask what a buyer would ask, in your category and your area, and we never put your business name in the question.
Local discovery
The head term. Somebody wants the best provider in your category and area, and asks for real, named businesses ranked best to worst.
Service intent
The long tail. Somebody names one specific service you offer and says they are ready to hire, which is a different retrieval problem for the assistant.
Problem intent
No ranking is requested at all. Somebody describes their situation and asks who to trust, which is where reputation and public evidence do the work.
Three intents, not three rewordings of one
An earlier version of the measurement asked the same head term three ways. That looked like broad coverage and was not: a local operator was measured only against national players and directory listicles, three times over, and one unlucky query moved the whole number. The three questions above deliberately reach the assistant through three different retrieval paths.
The questions are personalized from what you save in Settings: your category, your location, your service areas, your lead service, who you serve and how you deliver. Every assistant receives the identical strings, which is what makes comparing them fair.
What we read out of each answer
Three things, and only things the answer actually contains. Nothing is inferred about reviews or directories that the answer did not say.
Mention presence
Did this answer name your business at all? Matching runs on the real forms of your name, not on an exact string, so a shortened or reordered form still counts.
Recommendation position
When the answer returns an ordered list, where do you sit in it? Being named third in a list of ten is a different result from being named first.
Answer sentiment
How does the answer describe you? A mention wrapped in a caveat is worth less than a mention the assistant is confident about.
The same three readings are taken for the competitors on your list, up to 5 of them, from the identical answers. That is why the comparison is like for like: nobody gets a friendlier question.
The three components behind one number
Your score is built from three families of evidence. Each one carries a share, each share is shown next to the number, and a family that was not measured contributes nothing rather than a zero.
AI answers
What the 4 assistants actually said about you. This is the thing the product measures, and it carries the largest share of the number.
Four assistant bars, one per model
Website evidence
The verifiable technical and structured-data findings from your site: metadata, schema completeness, crawl files, business details an assistant could read.
One request, to your homepage
Google Business Profile
Public evidence from your listing, when we can resolve it. Strong corroboration for a local business, and deliberately capped so it can never stand in for AI visibility itself.
Never more than 30% of the score
Where the shares come from
Not by judgement. Each assistant's share of the number is derived in code from how much evidence it actually produced: an assistant that answered 3 of our questions with usable evidence carries proportionally more than one that answered one, and an assistant that produced none carries zero and moves the score by exactly nothing. The Google component is capped outright at 30%, and what is left over is what the assistants divide.
A summarizing agent does read the captured answers, and it does score them and explain them in the report you get. What it does not do is decide what its own opinion is worth. That was tried, and it was replaced: the weights a model proposes for itself are kept for audit and are not what the published number is built from.
Every input is stored on your scan. The published score can be rebuilt from those rows with no AI call and no access to our code, and that is on purpose.
What happens while a scan runs
A scan takes a few minutes. This is the order it works in.
- Stage 1
Compile the questions
Your saved business profile is turned into the question set. Nothing is asked yet.
- Stage 2
Ask every assistant
The identical questions go to all 4 assistants. Each one answers on its own; a slow or unavailable model does not hold up the rest.
- Stage 3
Read the answers
Each answer is parsed for your business and for your competitors: named or not, where in the order, and in what tone.
- Stage 4
Audit the website
One request to your homepage collects the technical and structured-data evidence an assistant could read about you.
- Stage 5
Resolve the Google listing
Where your public Google Business Profile can be matched, its evidence is added. Where it cannot, nothing is invented.
- Stage 6
Publish and compare
Shares are assigned, the number is computed, and it is compared against your own earlier measurements on the same basis.
What the measurement refuses to do
Most of the design decisions here are about what NOT to report.
A missing answer is not a zero
If an assistant does not answer, it carries no weight at all. Scoring silence as failure would punish you for our supplier's bad morning.
A partial scan says so, in the number
When fewer than all 4 assistants answered, the result is published with a likely range around it instead of a single confident point.
Nothing is invented to fill a gap
If your Google listing cannot be resolved, or a site check cannot be verified, the component is absent rather than guessed. Absent and zero are different states, and they render differently.
One flip is not a trend
The published number is the average of up to your last 5 comparable measurements, none older than 60 days. A genuine collapse still shows up immediately; a single answer flipping does not.
Free scan and full scan, side by side
Same questions, same assistants, same reading. The difference is how much evidence the number rests on.
| Free scan | Full scan | |
|---|---|---|
| Questions asked | 1 buyer question | 3 buyer questions |
| Assistants asked | 4 (ChatGPT, Gemini, Perplexity, Claude) | 4 (ChatGPT, Gemini, Perplexity, Claude) |
| Minimum evidence to publish | 2 assistants must answer | Published with a likely range when coverage is partial |
| What you get back | A visibility grade and a first fix | A 0-100 score with the share each component contributed |
| Website findings | In your results email | In your dashboard, check by check |
| Repeats | One-time scan | On your plan's schedule |
How often it repeats
AI answers change. A measurement taken once is a snapshot; a measurement repeated on a schedule is a trend.
See the method run on your own business.
One buyer question, four live AI answers, a visibility grade and a focused first fix. No card required.