How we measure AI visibility
We record how a defined panel of AI models answers buyer questions about a business, with and without live web search. The report keeps the observations, context, and limits visible.
Three questions, kept apart
A business can be named, described inaccurately, and never recommended, all in the same run. Each of these is a different observation with a different fix, so the report never adds them into one number.
Does AI name the business?
Presence
For each buyer question, whether the business is named in the answer at all. The report gives the count of answers where it appeared and the number of answers asked for, per mode. When that count is too small to say anything about, the count is the finding and no score is made from it.
How does AI describe it?
Perception
When a model does name the business, the sentence it wrote: services, area, price band, tone. Each description is compared with facts you can verify, and a wrong fact is recorded with the source the model gave, if it gave one. This is an observation of what was said, not a share of answers.
Would AI recommend it?
Preference
The buyer asks for a recommendation or a comparison and the model chooses. The report records who it put forward, in what position, and the reason it stated. Being named is not the same as being recommended, and neither is a measure of leads.
Two test configurations
Each question is asked of six models with search disabled and seven with web search enabled. Perplexity is measured with web search only. The two configurations are reported side by side and never combined.
What this does not tell us: exactly what a model has retained, or that a provider actually consulted a given page because search was on. A business missing from both modes is not, on its own, proof of a page problem, and a change between two runs is not, on its own, proof that a correction caused it.
The protocol: expected, obtained, usable
8 buyer questions per angle and per target × (6 models without search + 7 with search) = 104 expected answers per angle and target. Three angles = 312 expected answers per target.
Panel: ChatGPT, Claude, Gemini, Grok, Perplexity, Mistral, Muse Spark. The versions in a report are the ones recorded during that run.
312 per target is the number asked for, not the number obtained. A call can fail or return nothing usable; the report says so rather than counting it as an answer. The buyer questions, also called prompts, are eight questions per angle repeated across the eligible models, not 312 different questions.
How much do answers move on their own?
Before calling anything a change, I measured chance. On 8 October 2026, I asked the same questions to the same models twice on the same day in Montréal. With web search, a business moved by up to 6 answers out of 56. Below 5 mentions, names come and go from one run to the next. So in the 30-day proof loop, a change counts only beyond that margin, and only for a business mentioned at least 5 times in one of the two runs.
Limit: one city (Montréal), one pair of runs per mode. This rule will be measured again in other cities and in memory mode, then updated with its date.
Web control measured on 8 October 2026 in Montréal. Source audit IDs: cmuznhmvf008njmzhazw5p5a7, cmuznjcnm00nbjmzhk8ixzigc, cmuzq2ibj011xjmzh6gxguz24, cmuzq2ryz013vjmzhidg3ybeh
How each finding is read
Every observation carries the question, the model and version available, the date, the mode, the answer or a faithful excerpt, the source the model provided when it did, and a separate interpretation. The raw text is kept so a line a client disputes can be checked against what the model wrote rather than against a summary of it.
A priority carries the action, the evidence it rests on, who does it, and how it will be checked. Where the wording of a next step has been improved after the report was produced, the page says so; it is not passed off as a quote from the PDF.
Limits, and what a retest means
Models are not stable: the same question asked twice can move the wording and sometimes the recommendation. A retest keeps the questions, the panel and the context the same and records which model versions changed in between. It is the only way a before and after means anything, and it still shows movement, not cause.
We do not measure leads, traffic or how many people saw an answer; no model publishes that. A mention, a citation, a presence or a preference is what the models said when asked, and none of it is a direct measure of a prospect or a sale.
The measurement date and the date the report was generated are recorded separately.
Company facts
- Legal name
- VectorGap
- Brand
- VectorGap
- Founded
- May 2026
- Based in
- Brussels, Belgium
- Contact
- gap@vectorgap.ai
Also on the About page.
