Methodology

Last updated 2026-09-07

What we check, how, and what we don’t know. Every result page shows its own questions and timestamp; this page is the standing version.

Which questions

Twelve per product, derived from your site: the questions a buyer in your category would type into an AI assistant. Buyer phrasing, no marketing words, at most three that name a rival (“X alternatives”), never your own name. The free check asks the first six of them, once. Customers get all twelve, twice a week. They are listed on your result, verbatim. You can correct them; corrections re-run the check.

Which engines, exactly

• ChatGPT — OpenAI Responses API with the hosted web_search tool, default settings.
• Claude — Anthropic Messages API with the hosted web_search tool (up to five searches per answer).
• Gemini — Gemini API with Grounding with Google Search.
• Perplexity — Perplexity Sonar via the chat completions API, default search settings.

Two tiers, and the result says which one ran. The free check uses the cheapest capable model on each engine: gpt-5.4-mini, Gemini Flash, Claude Haiku 4.5, Perplexity Sonar. Customers’ weekly checks use the flagship tier a consumer sees by default: gpt-5.5, Gemini Flash, Claude Sonnet 5, Perplexity Sonar Pro. The exact model id and settings used for each answer are stored with that answer and shown under “the questions we asked” on your result.

What we count

• Named — your product’s name (or a known alias or your domain) appears in the answer text.
• Cited — a source the engine attached to its answer points at your domain.
• Sources — every URL an engine cited, and whether that answer named you or a rival. The “why you’re absent” list is the sources cited in answers that named a rival and not you.

Cadence and sampling

The free check asks six questions once per engine, on the lean tier. Weekly checks for customers ask all twelve twice, on the flagship tier, and report the majority, because engines do not answer the same way every time. A single free check is a snapshot, not a trend; the weekly re-check is the trend.

Known limits

• These are the engines’ search-grounded API endpoints, not the consumer apps. Vendors do not state that the two match, and independent studies find they often don’t. We say “named in ChatGPT” as shorthand for “named by OpenAI’s search-grounded model”.
• Google AI Overviews and AI Mode are not checked. They cannot be queried faithfully, so we leave them out rather than approximate.
• Name matching is literal. A product with a common-word name can be over-counted; a product engines misspell can be under-counted. Aliases are derived from your site and can be corrected.
• We check twelve questions, verifiably, rather than hundreds unverifiably.

Corrections

Wrong about something? Tell me. Corrections ship within 24 hours and are logged.