Saved answers. Counted carefully.

The first study is a small pilot. It is a plan, not a result. Its scope, including the country and the categories, is still subject to owner approval and has not been fixed publicly.

5 categories x 5 questions x 3 assistants x 2 repetitions = 150 planned answers. Planned is not the same as successful: any failed request is reported separately, never silently dropped.

• Answers come from search-enabled provider APIs. They are not replicas of the consumer chat apps, which may answer differently.
• Each study records the requested pinned model ID. Provider-returned model metadata, when supplied, is preserved in the downloadable observations.
• Gemini answers come from the Google Gemini API. That is not Google AI Overviews in search results, and we do not treat it as such.
• One fixed country and location for the whole study. Prompt text is written once and left unchanged.
• Every request is independent, with no shared conversation history, and is timestamped.
• Two rounds, so we can see whether the same assistant repeats itself.
• Source links the answer actually cites are kept apart from sources the search tool retrieved but the answer did not cite.
• Only valid HTTP(S) citations with identifiable domains are counted. For Google grounding redirects, the provider's source-domain label is used; destinations are not independently verified.

• A person reviews each successful answer and annotates which businesses or brands are named, and which are actually recommended.
• Each annotation is backed by verbatim evidence from the answer. That evidence is saved privately for audit, not published.
• We do not publish full answer text, and nothing on this page is an endorsement from any assistant.

• Brands and domains are counted once per answer. A business named five times in one answer counts as one answer.
• Percentages use successful answers as the denominator, and the denominator is always shown next to the figure.
• Values that cannot be computed are shown as not available, never as zero.

• Categories are self-selected, not a random sample of the economy.
• One location only. Answers elsewhere may differ.
• Two repetitions show some variation, not its full range.
• Results are not representative market share and support no causal claims about why a business is or is not recommended.
• Nothing here guarantees a ranking, traffic or revenue change for anyone.

Research data is collected separately from the product. Customer scans, email addresses and report IDs are never used in, or published as part of, any study.