Name the layer you are investigating
An AI answer can show a link without revealing the infrastructure that supplied it. Investigating search therefore starts by separating four activities: fetching a page, storing information for later lookup, selecting material for a request, and composing an answer. Evidence about one activity does not automatically establish the others.
Use precise working definitions. Crawling is an observed or documented fetch; indexing concerns retained searchable information; retrieval concerns material selected for a particular request. Answer generation concerns the final response. A search product may combine multiple systems, so avoid assuming that every answer follows one fixed pipeline.
Turn a broad claim into a testable question
Instead of asking whether an assistant owns an index, ask whether a newly published public page can be found under specified conditions. That narrower question has an observable outcome. It still cannot establish who operates the storage layer or whether an intermediary supplied the result.
Write down competing explanations before running the test. A cited page might come from live retrieval, a cached copy, a partner service, or a link supplied in the conversation. Listing alternatives prevents the first plausible story from becoming the conclusion merely because it fits one answer.
Collect observable evidence with provenance
Keep the prompt, answer, linked destinations, collection time, language, and search settings together. Inspect whether a citation actually supports the adjacent claim. A displayed domain may be useful evidence of a destination while saying little about how that destination entered the system.
For a site you control, server logs can add another observation. Verify crawler identity against the provider’s documented method rather than trusting a user-agent string alone. A request establishes access to a resource at that time; it does not prove successful indexing, retrieval, or inclusion in a later answer.
Design a bounded page probe
An illustrative probe could use a genuinely useful public explanation containing a distinctive, truthful phrase. Avoid hidden bait or fabricated facts. Record the page’s publication and revision times locally, then ask natural questions that could reasonably lead to it. Do not paste the destination into the discovery prompt.
- Confirm that the page is publicly accessible and that its canonical URL is stable.
- Save the question and collection conditions before checking the answer.
- Record both successful and unsuccessful attempts without changing the question mid-series.
- Separate page discovery, phrase recognition, and citation support in the observation sheet.
Compare explanations without overstating them
Finding a page through several assistants does not prove a shared index. Similar results can arise from common public sources or similar query interpretation. Likewise, different results do not prove independent infrastructure. Repeated observations can narrow hypotheses, but architectural ownership usually needs direct documentation or stronger independent evidence.
Use operational tools for their actual scope
Kept AI’s native Agent can help analyze crawler activity, crawlability, robots rules, and paths on the connected site. Use that evidence to investigate access problems. It does not expose a provider’s internal index, and changing your own robots overrides cannot force a provider to retrieve or cite a page.
Publish the narrow conclusion you earned
Report the tested conditions and what was observed, then state what remains unknown. A missing result can reflect query interpretation, access, ranking, freshness, or sampling. Keep the evidence packet so another reviewer can challenge the interpretation. Useful research reduces uncertainty without turning a limited observation into an infrastructure announcement.