Skip to content

How we measure

How a VectorGap figure is made.

Buyer questions, seven AIs, two modes and two runs. Every figure comes with the answers behind it.

  • 24buyer questions per target
  • 7AIs with web search, 5 of them also without it
  • 384answers per target at every measurement

The protocol

24 questions, 12 series, 384 answers per target.

8 questions for each angle. Each question goes to 7 AIs with web search: ChatGPT, Claude, Gemini, Perplexity, Grok, Muse Spark from Meta, Google AI Mode. 5 of them also answer without web search. That makes 288 answers. The Presence questions are asked a second time: 96 more answers. Perplexity and Google AI Mode are measured with web search only.

With web search · 7 AIsWithout web search · 5 AIs
ChaClaGemPerGroMusGooChaClaGemGroMus
Presence8 questions
Perception8 questions
Preference8 questions

384answers per target: 288, plus 96 for the second Presence run

Two modes

With web search and without it, read side by side.

Without web search

What the AI learned

The AI answers with what it learned during its training. It changes when the model changes.

With web search

What the AI finds today

The AI searches the web before it answers. It can change as soon as your pages change.

The two modes are shown side by side. They are never added into one number.

Two runs

Asked twice, to measure the noise.

Two identical measurements on the same day do not give identical answers. So at every measurement, VectorGap asks the Presence questions a second time. The figure covers both runs. The difference between the runs gives the margin of your own market.

First run
Second run

17answers out of 56 changed between the two runs.

Simulation for a brand named in one answer out of two. The rate of change, 17.5%, is measured.

17.5%of the answers change on the same day (84 of 480)
±7.7points of margin with 7 AIs in one run (56 answers)
±5.5points of margin with 7 AIs in two runs (112 answers)

The answers that differ between the two passes give each measurement its own margin. In the 30-day proof loop, a change counts only beyond that margin. It counts only for a business mentioned at least 5 times per pass, on average, on one of the two dates.

Chance test measured on 14 November 2026 on 7 companies (480 answers compared). A measurement run in a single pass keeps the previous rule: up to 6 answers out of 56, measured on 8 October 2026 in Montréal.

Read the full test in the article “How AIs interpret a brand”

Seven AIs

One AI says little about the others.

Which AI you ask explains 27% of the citation gaps. That is as much as the brand itself. The wording of the question explains 10%. A brand that measures one AI sees nothing of the six others.

Yago

Named by 5 of 7 AIs in French.

FR
NL
EN

Measured on 29 October 2026, with web search and the panel of that date. One dot is one AI.

9%of the companies named are shared by two different AIs for one question (20,629 comparisons)
±30points of margin when ChatGPT alone answers 8 questions
±5.5points of margin when 7 AIs answer in two runs

The proof

Every figure has its answers.

Each finding carries the question, the AI and its version, the date, the mode and the answer. You can check any line against what the AI wrote.

Each figure with its count of answers
Each figure with its count of answers

Before and after

What counts as a change.

When I compare two dates, a difference counts only when it is larger than the margin measured on your market. The questions, the AIs and the mode stay identical.

Start

Start with one measurement.

$250, once. Two markets, measured now and measured again in 30 days. If the report teaches you nothing, I refund it.

By ordering, you accept the Snapshot terms.

See a real report
Johnny Lagneau, founder of VectorGapI run the measurement and I present the report to you myself.