Conversational search

Conversational search is search in which each query is interpreted as a turn in an ongoing exchange: the system resolves pronouns, ellipsis and implied subjects against what was already said, instead of treating the input as a self-contained string.

The test is simple. "How does that affect coastal cities?" is not a query. It has no referent, no retrievable meaning, and no keyword worth matching. It becomes answerable only when the previous turn is available to resolve "that". Everything else about conversational search follows from the cost of keeping that previous turn available.

The mechanism: history is re-sent, not remembered

There is no memory in the ordinary sense. The earlier turns are re-supplied as input on every subsequent turn.

Anthropic's documentation states it directly: "As the conversation advances through turns, each user message and assistant response accumulates within the context window, and previous turns are preserved completely." Each turn's input phase "contains all previous conversation history plus the current user message", and the output "becomes part of the input for the next turn". The window itself — up to 1M tokens depending on the model — "holds the conversation history plus the new output Claude generates".

Two consequences fall out of that arrangement, and they are the ones that matter to a publisher.

First, the effective query at turn five is not the five words the person typed. It is a resolved question the system constructed from the accumulated exchange, and it is that resolved question, not the literal input, that retrieval runs against.

Second, history competes for space with everything else in the request — the system prompt, tool definitions, and any documents retrieved for this turn. Anthropic notes that "as token count grows, accuracy and recall degrade, a phenomenon known as context rot", which is why long sessions get compacted or truncated rather than simply carrying on. A retrieved passage of yours arrives into a window that already has a conversation in it.

One turn is not one search

Google documents that both AI Overviews and AI Mode "may use a 'query fan-out' technique — issuing multiple related searches across subtopics and data sources". A conversational turn can therefore expand into several underlying searches whose results are combined into a single answer, and it describes AI Mode as "particularly helpful for queries where further exploration, reasoning, or complex comparisons are needed", where "people can ask nuanced questions that might have previously taken multiple searches".

So the mapping from what the reader asked to what was searched is neither one-to-one nor visible to you. A page can be retrieved for a sub-question the reader never typed. This is the same passage-level selection that governs retrieval-augmented generation and AI search generally.

There is no markup that opts you in. Google states there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", and that you "don't need to create new machine readable files, AI text files, or markup to appear in these features".

Failure modes

• Writing for the literal follow-up. Nobody searches "how does that affect coastal cities". The resolved sub-question is what gets searched, and it looks like an ordinary explicit question.
• Passages that depend on their neighbours. A section retrieved for a fan-out sub-question arrives alone. A caveat two paragraphs earlier does not travel with the claim it qualifies.
• FAQ pages built as keyword bait. Comprehensive question lists help only where each answer is independently correct and specific; a list of thin answers is a list of thin passages.
• Assuming a mention in turn one persists. Each turn retrieves again. Being cited early in an exchange does not reserve a place later.
• Treating conversation length as free. Under context window pressure, earlier material is compacted or dropped, including earlier citations.

What to do about it

Write sections that survive being lifted out. One subject per section, a heading phrased as the question it answers, the qualifier in the same sentence as the claim, and identifiers — versions, product names, numbers — written literally rather than implied by context.

Then check the outcome by sampling actual multi-turn exchanges rather than single questions. A brand that appears when asked about directly, and disappears two turns later when the conversation narrows to a specific comparison, has a passage-level problem that a single-question test will never surface.