1. Five engines, in native search
We query five engines, each configured to use its native web search (not just its training memory) so it reflects what a user would actually get today:
- ChatGPT — OpenAI · native web search
- Claude — Anthropic · native web search
- Gemini — Google · native Search grounding
- Perplexity — sourced answer by default
- Grok — xAI · real-time search
2. Multiple runs per engine (K runs)
AI answers vary from one call to the next. So as not to judge on a lucky shot, we ask the same set of questions several times (K runs) to each engine, then aggregate. An engine counts as PRESENT if it describes the brand by name, ABSENT if it doesn't know it, NEGATIVE if the tone is unfavourable.
The source stability shown per engine is a confidence indicator: the more the same citations recur across runs, the more reliable the signal. It is not a grade.
3. Entity purity (homonyms)
A brand can share its name with another entity (a homonym). Before keeping an answer, we check that it really refers to your organisation — via a fingerprint built from your site (domain, sector, official signals). Citations are classified: on-brand, off-topic, or uncertain. Purity is the share of citations genuinely tied to you.
4. Accuracy: the "≥ 2 engines" rule
A claim is flagged as a confirmed error only if at least two engines state it in a concordant way. An isolated claim (a single engine) is filed under "to verify" and is not counted as an error. The goal: avoid turning a one-off hallucination into a verdict.
5. Score & dimensions
The overall score aggregates three dimensions — presence (are you described?), image & accuracy (sentiment and errors), sources (quality and purity of citations). When too few engines describe you, we don't invent a bad score: we display "insufficient signal" — that's an absence of presence, not a condemnation.
A mirror, not an endorsement
The report is an automatic rendering of public, un-endorsed AI answers, as of the analysis date. The engines' answers vary over time and constitute neither an opinion nor a verification by Levier IA. We faithfully reproduce what the engines say (including, for Gemini, the search-suggestions widget shown as-is, in line with Google's terms) — without editorialising: an absence is not a flaw.
Is something inaccurate?
If an engine spreads incorrect information about your brand, see the corrections page: it explains how to request a correction on the Levier IA side and how to act directly with the engines' publishers.