Understand, measure and improve a brand’s presence in AI answers.
of the sources ChatGPT cites are third-party; only 6.5% come from official sites. A brand optimized only for classic SEO can become invisible before the first click.
When the interface becomes an answer, visibility must be measured in the answer, then improved at the source, continuously.
Engines rank pages.
Engines select and cite.
Intent sharpens over the turns.
AI recommends, then acts.
Surfaces and sources keep changing.
Three complementary studies — 990 CEA LIST queries, 602 prompts and 21,143 citations from China, 4 AI engines and Google in Toronto.
Discoverable, Identifiable, Selected, Absorbed, Recommended, Verified — where does the chain break for your brand.
ChatGPT, Gemini, Perplexity, Claude: each engine selects and reuses its sources differently.
Tick 1 point for each yes answer to find out whether your base is fragile, partial, or ready for a structured test.
From the 15-day sprint to continuous governance of your visibility in AI answers.
PDF · 18 pages · English and French editions
990 queries · 5 configurations · 79 attributes · citation and rank
Head-to-head comparisons won by the most consistent domain
Ranking well depends on the consistency of the whole site, not only on the page (59–66% on the main collection).
602 prompts · 21,143 citations · 18,151 pages · selection and absorption
Yet ChatGPT’s retained pages show the highest average influence: 0.2713, against 0.0646 for Perplexity. Counting citations does not measure influence.
4 AI engines and Google · verticals and languages · paraphrases · official, third-party and social sources
Third-party sources, ChatGPT vs Gemini
Each engine combines third-party, official and social sources differently: engines don’t consult the same Internet.
The foundation is shared. Priorities change with how each engine selects and reuses its sources.
Control review platforms, media and the consistency of brand information. Measure browsing and non-browsing modes separately.
Describe services, uses, places, sensations and concrete differences. Avoid generic commercial wording.
Create precise evidence blocks, comparisons and sources that are easy to verify. Track the diversity of cited domains.
Strengthen recognized sources, canonical identity and local coverage. Re-run tests by language.
Tick 1 point for each yes answer. The result predicts neither citation nor rank. It is meant to spot the missing foundations.
The results describe observed behaviors. They do not give access to the models’ internal mechanisms.
Studies [2] and [3] are preprints.
Results are observational, except explicitly described intervention tests.
Engines, models and interfaces evolve. Every measurement must be dated.
No metric guarantees a future citation, rank or recommendation.