AI search products differ in their tools, sources, citation formats, and use of context. Choose which products to measure from evidence about your audience. Record the product and settings used in every test.
The platforms
ChatGPT
ChatGPT combines a conversational interface with optional search and other tools. OpenAI describes search as using third-party providers and partner content. A response with web access should be measured separately from one generated without it. Usage figures in the cited study describe July 2025.The NBER working paper "How People Use ChatGPT", built on OpenAI's internal usage data.OpenAI's "Introducing ChatGPT search" announcement. OpenAI names the partner publishers but not the search providers.
Google AI Overviews and AI Mode
Google’s AI Overviews and AI Mode can issue related searches to prepare an answer. Google says its existing SEO requirements apply to both. They can return different responses and supporting links, so keep their measurements separate.Pew Research Center's measurement of real browsing behavior, including which searches triggered summaries.Google's AI Mode announcement from I/O 2025, which describes the fan-out technique and the custom Gemini 2.5 model behind it.
Gemini
Gemini is Google’s assistant product. When evaluating a Gemini answer, record whether it used Google Search or another source. Sharing a provider with Google Search does not make the user experience or generated response identical.
Perplexity
Perplexity provides search-based answers and operates a crawler. Record the mode and sources used in the tested response. Citations identify attributed sources; they do not necessarily expose every result considered during retrieval.Perplexity's crawler documentation, covering PerplexityBot and the user-triggered Perplexity-User agent.
Copilot
Microsoft offers Copilot experiences in several products. For a search test, identify the particular Copilot product and whether web grounding is enabled. Do not combine results from a public search experience with an enterprise assistant using private work documents.
Claude
Claude can use web search when that tool is available. Identify the tested configuration and retain its citations or tool records. A model name alone does not establish which sources were available to an answer.
Other assistants
Other assistants appear in social networks, browsers, and devices. Include them when customer evidence supports doing so. Verify current source access and product behavior before treating one assistant as equivalent to another.
Search providers and other sources
Search providers affect which pages can be retrieved. The model, product instructions, and retrieved content also affect the answer. Maintain Google and Bing search access where relevant, and use provider documentation or logs to identify other sources. Avoid assuming an undocumented provider relationship.
Record the type of task
Distinguish answers generated without retrieval, answers supplied with retrieved information, and tasks that involve browsing or actions. A single product may support all three. The relevant configuration is the one your test actually used.
Choosing what to measure
Measure both answer visibility and business outcomes. Referral analytics can record visits that include attribution data. They cannot count every exposure inside an answer. Attribution and Business Impact explains the limits.Cloudflare Radar's crawl-to-refer data; in the June 19-26, 2025 sample, ratios ranged from Anthropic's 70,900:1 to Mistral's roughly 0.1:1.
Begin with the products your customers report using, the referrals you can observe, and the questions you need to test. A small, documented set is easier to compare than a broad set of inconsistent tests. Expand coverage when it informs a business decision.