How AI search works: from a question to sources and an answer

Five concepts to keep separate

Training shapes a model using earlier data; publishing a page does not immediately retrain it. Retrieval obtains material at answer time, potentially from a search index or another available source. Browsing is a tool action that may open a page. Generation composes the response. A citation is an output reference. A source may be consulted without receiving a visible citation, and a named business may appear without a link to its own website.

Trace a fictional Danish buyer task

Imagine a small manufacturer asking how to choose a service provider for a bilingual website migration. A search-assisted system might investigate language URLs, redirect planning, ecommerce dependencies, measurement continuity, supplier scope and implementation responsibilities. These six subtopics are an illustration, not captured internal queries. The answer could combine multiple sources and then suggest selection criteria. A page offering concrete migration acceptance checks has a distinct purpose within that research task.

What retrieval-augmented generation changes

Grounding can bring more current source material into a response than pretrained knowledge alone. It does not make every answer correct or every cited source authoritative. The source can be outdated, the extraction incomplete or the synthesis mistaken. A careful reviewer checks whether the linked page actually supports the adjacent statement, preserves conditions and refers to the right entity. “The answer contains a link” is a starting point for inspection.

Handle a conflict without inventing certainty

Suppose an old directory says a business sells gifts while the current official service page describes SEO consulting. Record both sources and dates. Review whether they refer to the same entity and whether the old activity has ended. Correct the official facts and pursue legitimate source corrections where appropriate. Do not assume that repeated prompts will erase model knowledge or that a new schema field overrides every external source.

Why two tests can disagree

Wording, conversation history, language, location, account settings, product updates and retrieval availability may differ. A generated response is not a fixed search position. Preserve the complete test conditions you can observe, repeat a balanced panel and distinguish no result from a failed run. If a model version is not visible, record “not disclosed” instead of guessing. Avoid attributing every difference to an undocumented sampling setting.

What website owners can improve

Make important source pages public, accurate, understandable and linked from relevant navigation. State business identity and scope clearly, support consequential claims and remove stale contradictions. Test initial HTML, status codes and appropriate crawler access. These improvements help visitors and source accessibility; they do not buy inclusion. Review platform documentation when it changes, and evaluate actual answers separately from crawl logs.

Continue with the related decision

• AEO learning paths
• Query fan-out: turn a complex question into useful topic coverage
• Discuss the scope of AEO and AI-search work