How we measure what AI tells your buyers.

We ask AI assistants the questions a buyer asks before contacting a supplier, then record which companies each answer names, which sources it cites and what it says about the company we audit. Nothing in a report comes from a model’s opinion of the company: every number is a count over recorded answers.

• Three buyer questions per company, phrased around what the company actually makes and where it sells, for example “Recommend the best industrial filtration manufacturers in Ontario.”
• One direct question about the company: what it makes, who runs it and where it is based.
• The question set is frozen and versioned. The current set is QUALIFY_PANEL_V2, in use since 2026-09-24. A changed question gets a new version, so reports built on different sets are never compared as if they were the same measurement.

• Starter audits: ChatGPT with web search on, an OpenAI model with web search off, and Perplexity with web search on, through its Agent API. Perplexity supported its Sonar API until 2026-09-27; audits before that date used Sonar.
• Full audits add Google AI Overviews, Google AI Mode, Gemini and, on request, Microsoft Copilot.
• Each report names the assistants it used, the date of every answer and the model each one reported. Where a provider does not report the model that answered, the report says so.

• A company counts as named when an answer mentions it as a company or cites its website.
• Names that are also everyday words, places or trade phrases count only when the answer cites the website or clearly presents the name as a company.
• “Named in 4 of 9 answers” means exactly that: four recorded answers out of nine. We report counts rather than a single ranking, because answers change between runs.
• A statement counts as wrong only when the recorded answer contradicts the company’s own details, for example a different city or a website address that is not its own. Starter audits check two details, location and website; they do not fact-check anything else the assistants say.

• Every answer used in a report is stored in full, with the assistant, the question, the time and a sha256 fingerprint.
• Each report ships with that evidence ledger, and its machine-readable description (schema.org) points to this page.
• Automatic counts are checked by a person against the recorded answers before a report is sent.

1. Open ChatGPT with web search on, or Perplexity.
2. Paste a question from the report, word for word.
3. Compare which companies it names with what the report recorded. Your answer may differ in detail; the pattern across several runs is what the report measures.

• An audit is a snapshot of answers on the dates shown, not a guarantee of future answers or rankings.
• Answers can differ by account, location and time.
• A starter audit does not measure Google search positions.
• We cannot make an assistant name a company. We can show what it reads, and whether that changes, by asking the same questions again after a fix.

• Hide text or instructions for AI readers in a report or on a client’s site.
• Publish numbers we did not measure.
• Promise a ranking or a revenue figure.

Write to hi@arbling.com with the report date and the company name, and we will show you the recorded answer behind any number in it.