How AI search finds sources
Separate finding source material from generating an answer.
Finding source material and composing an answer are different activities. Keeping them separate helps you diagnose what an observation can tell you. The distinction is a conceptual aid, not a claim that every product executes one fixed sequence. Some responses use web search; others may rely on supplied information or model knowledge without a fresh search.
Discovery is not the final answer
A search-enabled product needs some way to locate candidate information. Access rules and the availability of a page can matter at this stage, but discovering a resource does not establish that it will be selected, accurately represented or cited. A crawler entry in a log is evidence of a request, not evidence of inclusion in a generated response.
Google says AI Overviews and AI Mode may issue related searches across subtopics using a technique called query fan-out. It also says the features can use different models and techniques, producing different responses and links. These are Google-specific descriptions, not a published blueprint for every answer engine.
Retrieval and synthesis answer different questions
Retrieval asks which information is found for a request. Synthesis concerns how information is combined into a response. A relevant source can be found without appearing in the visible answer. A visible reference can also fail to support the precise wording next to it. Inspect the source and the claim together instead of treating the presence of a link as sufficient evidence.
Hypothetical example: a heat-pump question could involve a selection guide, technical specifications and installation notes. The illustration on this site uses these document types to make the distinction tangible. It does not describe an actual product query, a measured retrieval event or an engineering recommendation. The relevant evidence for a real decision would need to be inspected directly.
Check what the cited page supports
When you observe an answer, open the linked page and find the relevant passage. Does it support the whole claim or just one part? Are important conditions missing? Does the source describe the same product, date or location? If the link is too broad to establish support, record that limitation rather than filling the gap with an assumption.
For your own site, write the explanation so that a person can follow it in its original context. State the subject, conditions and evidence together. Google recommends making important information available as text and matching structured data to visible content in its AI-search guidance. Neither recommendation guarantees faithful synthesis by another system.
Diagnose the layer you can observe
If your page cannot be opened, investigate access. If it is readable but leaves the reader’s question unanswered, investigate content. If a generated answer cites it inaccurately, preserve the exact answer and source context before changing the page. A missing citation by itself does not reveal which layer was responsible.
OpenAI separately documents a search crawler and a training crawler in its crawler overview. That separation is another reason to avoid describing training access as proof of live retrieval. Keep the provider, the request type and the observed evidence distinct when deciding what to investigate next.