The Honest Measurement Charter.

The questions are fixed before the test

Your buyers’ questions are written down, dated and fixed before the first measurement. We do not choose the questions after seeing the answers: that would be cheating.

We tell you which engine actually answered

The report names the engine that was actually queried, ChatGPT, Claude or Gemini, and how it was queried. We never present an intermediary tool as though it were the engine itself.

In practice, our readings go through the vendors’ own official tools, the command-line interfaces from OpenAI, Anthropic and Google: official surfaces from the same vendors, with each vendor’s official web search tool, and the exact model is named on every reading. No vendor documents whether these surfaces and the consumer applications share the same search infrastructure. We do not assume it: we measure the gap, and we publish it. We chose this surface because it can be measured honestly: the session is neutral, with no history and no personalisation, and anyone can rerun the test under the same conditions and check our numbers as of the reading date. Engine answers themselves move from day to day: a reading is a dated photograph. An answer from the application depends on the account asking the question; it cannot serve as a baseline. We have not audited our peers’ tools and we publish no comparison of them; what we guarantee is that we name the surface and the model queried instead of simply writing “ChatGPT”. That gap has been measured on the OpenAI surface: at the reading of the night of 14 to 15 August 2026, the ChatGPT application and our instrument, then running gpt-5.5, returned the same presence verdict on all 30 frozen questions, 30 out of 30. Since 16 August 2026 the instrument queries gpt-5.6-sol; the concordance of that pair will be established at the November 2026 reading, and the Claude and Gemini surfaces remain to be checked the same way. The concordance study is published, method, limits and evidence included, and will be rerun quarterly. And to judge on evidence rather than promise: the full example of a delivered diagnostic is public, reading taken from 7 to 8 August 2026, 270 AI answers read one by one, raw answers included.

We repeat and count in front of you

Each question is asked several times because the engines do not give the same answer twice. You read “cited 2 times out of 3” in black and white. When an aggregate score exists, its formula, its weights and every figure that feeds it are published next to it; a score without its formula never leaves a HENUSSE file.

Failed tests remain visible

A test that failed or could not be run remains recorded as such in the report. When a result is too weak to support a conclusion, the report says “inconclusive”, not a flattering figure.

Cited, mentioned, recommended: three different things

Being used as a source, being named in passing and being recommended to a buyer are not equivalent. The report separates all three, so you know exactly where you stand.

We keep everything, and you can repeat the test

The question, the engine’s full answer, the pages it used, the date and the time: everything is retained. If a result surprises you, we can repeat the test under the same conditions, either ourselves or in front of you.

In every audit, you receive the AI answers exactly as they were returned, word for word, with the date, the question and the model queried. Nothing is rewritten and nothing is cut. If an AI system cites no one, we write that down too. You can ask the same question yourself. The answer may differ because AI systems move: that is also what we measure, on known dates.

Every figure has a date

The engines change quickly: a result from March says nothing about June. Every measurement carries its date, and no undated claim leaves a HENUSSE file.

If it is zero, you read zero

If your company appears nowhere, the report says so. A measurement problem is reported as a measurement problem; we never replace it with a more convenient figure.