How does AI search change SEO?
By putting an answer where a list of links used to be — which changes what traffic reaches a site, what content is worth producing, and how success should be measured.
What actually changed. Generative answers synthesise a response from multiple sources and cite some of them. For a query the answer fully satisfies, the user has no reason to click anything. This does not remove all traffic; it removes a particular kind — the quick factual lookup that was never commercially valuable anyway — while leaving queries where people want to compare, buy, verify or go deeper.
What follows practically:
Thin informational content loses most. Pages that restate what everyone else says are exactly what a model can synthesise without you. Content that survives contains something not already in the training data: original research, first-hand experience, proprietary data, genuine expertise, and specific recent facts.
Being cited becomes a goal in itself, even without a click. A citation carries brand exposure and authority, and cited pages appear to skew towards clear structure, direct answers near the top, and verifiable claims.
Brand and direct demand matter more. If people search for you by name, or arrive without a search at all, the intermediary matters less. This is an argument for the brand-building work performance marketers have historically underweighted.
Measurement gets harder. Impressions in an answer are not reported the way link positions were, so flat traffic with rising visibility is a genuinely plausible state — and attributing it requires looking at branded search volume, direct traffic and assisted conversions rather than sessions alone.
What has not changed: the underlying need for crawlable, fast, well-structured pages; matching search intent; internal linking; and topical depth. AI systems are reading the same web, and a page that cannot be crawled or parsed cannot be cited.
What is genuinely uncertain: how citations are selected, how much traffic is lost in aggregate, whether publisher licensing deals become the norm, and whether conventions like llms.txt gain real adoption. Treat confident claims about ranking in AI answers with scepticism — the field is young and evidence is thin.