When the answer replaces the link

For twenty-five years a search engine answered a query with a list of links, and its job was to send you somewhere else. That contract is changing. Increasingly the engine answers with a written answer assembled on the spot, from an AI reading multiple sources, shown above the links.

Google calls its version AI Overviews, and has a fuller conversational AI Mode; standalone assistants like ChatGPT and Perplexity answer directly by default. The common thread: the response is generated, not just retrieved.

The consequence is measured and large. Analyses of search behaviour, including work by SparkToro, find that a majority of Google searches now end without any click to a website, a figure that has risen sharply as AI answers spread, and Google reported at its 2026 I/O event that AI Mode had passed a billion monthly users.

When the answer is on the results page, the user's need is often met there. The trip to a website, the entire basis of the old bargain, no longer happens. That single shift reorganises what it means to be found.

The name for this is the zero-click search: the query is answered, the user is satisfied, and no website is visited. It is not new, a weather box or a currency conversion always answered in place, but AI answers extend it to the informational queries that were the open web's core traffic.

The uncomfortable arithmetic for publishers: an AI answer is often built from their content and then makes visiting them unnecessary. The site did the work of being the source and receives no visit for it. A business whose model was rank, get clicks, show ads or sell finds the middle step thinning.

This breaks the twenty-five-year deal underpinning the web. That deal was implicit but real: let us crawl your content, and we send you traffic in return. Publishers tolerated being indexed because indexing paid them in visits. When the engine answers directly, it still consumes the content but returns far fewer of the visits, and the exchange that funded a great deal of the web quietly stops balancing.

That tension, not any single feature, is the real story of AI search.

The mechanism is less mysterious than it looks, and knowing it tells you exactly how to be included. Most AI answers use retrieval-augmented generation:

1. The system takes your question and retrieves relevant passages from a search index, the same kind of index from the earlier lessons.
2. It feeds those passages to a language model with an instruction like answer using these sources.
3. The model writes an answer grounded in them, usually with citations back to the pages it drew on.

Two things follow directly. First, classic SEO still matters, because the model can only cite what retrieval surfaced, and retrieval is ordinary search. If you are not in the index and reasonably ranked, you are not a candidate to be cited. Discovery, crawling, indexing, relevance and authority are all still upstream.

Second, being retrieved is necessary but not sufficient. Among the pages pulled in, the model favours those it can extract a clear answer from: passages that state something directly, with facts it can lift and attribute. Ranking well gets you into the room; being quotable gets you into the answer.

The practice of getting cited in AI answers has a name, generative engine optimization, or GEO, and it is best understood as classic SEO plus a shift in what you optimise the content for.

The reported patterns, from early studies and practitioner reports, are consistent and intuitive given how the answer is built:

• Answer directly and early. A passage that states the answer in a sentence is easy to lift; one that buries it under a story is not.
• Structure for extraction. Clear headings, short definitions, step lists, comparison tables and FAQs give a model clean, self-contained units to quote.
• Be dense with attributable facts. Specific, sourced, dated numbers get cited; vague claims do not. Named studies travel further than unattributed assertions.
• Match how questions are phrased. People ask assistants full questions, so content framed around real questions retrieves better than content built for two-word keywords.

Notice how much of this overlaps with simply writing clearly and honestly. GEO is less a new dark art than a reward for the plain, well-sourced, well-structured writing that classic quality guidelines always claimed to want, now enforced by a model that literally has to quote you.

It is easy to over- or under-react. A clean split:

Genuinely new:

• The answer, not the link, is the product on many queries, so meeting a need no longer requires sending a visit.
• Visibility without traffic. You can influence millions of answers and receive far fewer clicks than before, so being cited is now partly its own goal.
• A second surface to rank in. Alongside ten blue links there is now the answer box, retrieved and written differently.

Reassuringly unchanged:

• Retrieval is still search. The index, crawling, relevance and authority all still gate what can be cited.
• Quality and trust still decide inclusion. Models favour sources that are clear, factual and credible, which is what the quality guidelines always rewarded.
• Being the best answer still wins. The surface changed; the goal did not.

The honest summary: AI search did not replace the mechanics from the previous lessons. It added a layer on top of them and changed the payoff, from a click to a citation. The site that is genuinely the clearest, best-sourced answer was already winning, and is now winning in two places at once.

The old scoreboard was traffic. When the goal shifts from clicks to citations, measurement has to change with it, because a large part of your impact now happens off your own site, inside an answer you never see in your analytics.

The emerging practice is to track a different set of things:

• Do you appear in AI answers for the questions you care about? This means periodically asking the assistants your target questions and checking whether, and how accurately, you are cited.
• Is the citation accurate? A model can represent you correctly, incompletely, or wrongly, and a confident wrong summary is a real risk that no traffic number reveals.
• Share of the conversation. Across a set of questions in your area, how often does your source get used versus competitors?

The deeper adjustment is conceptual. Analytics counted visits to you. The new question is your presence in answers about your topic, wherever they are generated, which your own logs cannot see. This mirrors exactly the shift the advertising path describes: from a clean click-based ledger to triangulated, partly-inferred estimates. Search measurement is undergoing its own version of the same move from counting to estimating.

Search has changed surface many times, ten links, then boxes and snippets, now AI answers, and one strategy has survived every change, because it is the one the engine is ultimately trying to reward.

Be the genuinely best, clearest, most trustworthy answer to a real question, and make it trivial to crawl, understand, and quote.

Every tactic in this path is a facet of that sentence. Crawlable, so you can be discovered. Well-structured, so you can be indexed and extracted. Relevant to real intent, authoritative, and honestly sourced, so you are trusted and cited. The list of specific techniques will keep changing, faster now than ever, but the target underneath does not, because the engine's own incentive, to satisfy the person asking, does not.

The closing reframe: stop optimising for the algorithm, which you cannot see and which changes weekly. Optimise for the job the algorithm is trying to do, which is stable, public, and pointed at exactly the thing you would want to be anyway, the best available answer.

An AI answer still starts with ordinary retrieval from the index, then a model writes from the passages it pulled, favouring quotable, well-sourced ones. Classic SEO gets you retrieved; extractable content gets you cited.