Method: how X-ray measures
X-ray measures whether an AI assistant names your company when a real buyer asks, who it names instead, and whether what it says is true. This page shows how the measurement is made and what it does not tell you.
1. The questions
The measurement starts from buying situations: the moments when a buyer starts looking for what your industry sells. For each situation, questions are written the way a buyer would ask them, in the market's own language.
By default there are 50 questions, split across the situations (for example 8 situations × 7 questions). Below about 50, a visibility figure swings with the wording of a single question.
2. Five AI engines, each on its own
Every question goes to five AI engines that search the web while they answer:
- ChatGPT (OpenAI, with web search)
- Perplexity (sonar)
- Gemini (grounded in Google Search)
- Claude (Anthropic, with web search)
- Grok (xAI, with web search)
Each engine's result is counted separately. Answers are never pooled into one figure, because that would quietly change every denominator. The report names the engine and the model it measured.
If an engine cannot be measured, it is left unmeasured. It is never replaced by another.
3. What is read from each answer
- Whether your company is named, and in which place.
- Who is named instead of you.
- How your company is described.
- Which sources the answer relies on.
- Whether the claims are true: prices and facts are checked against your own site. A price inside a range your site publishes is not an error.
4. The sources are checked
The pages an engine cites are fetched and read for whether your company appears on them. Every page read is stored with its time and SHA-256.
If a page could not be fetched, its row is marked as a lower bound. The report says, row by row, which one it is.
5. The buying-decision test
A buyer agent is made to pick one supplier in each buying situation and say at which step, and why, every other one dropped out. The reason comes from a closed list, and every reason names a fix.
This is a different measurement from visibility and it is reported in its own unit: buying decisions.
6. The seal
The SHA-256 of every delivered file is written to MANIFEST.sha256. Every answer, every decision and the log of pages read are attached inside the PDF report, with the measurement time and the engine version.
Anyone can recompute the hashes and compare them with the manifest. If a single byte has changed, the hash does not match.
The delivery's MANIFEST.sha256 also gets an RFC 3161 timestamp from a third-party authority (the file MANIFEST.sha256.tsr). It proves when the files existed, and anyone can check it with OpenSSL, without NOX.
7. What the measurement does not tell you
- Each question is asked once per engine per measurement. The figure is directional, not a statistical interval. The same question can get a different answer on a different day.
- No euros are estimated unless the client supplies their own figures (average order and the number of buying situations). Without them the report shows the probabilities and the formula.
- A price, a route or a location is never guessed. Missing information shows as missing.
- AI engines change every week. That is why the measurement repeats every week.
See this on your own company →
Version history
| 1.1 | 25 Oct 2026 | The seal gets an RFC 3161 timestamp. |
| 1.0 | 25 Oct 2026 | First published version. |