Ten links to one answer
For twenty-five years, searching meant the same thing: type a query, get a ranked list of links, click one. Answer engines - ChatGPT with browsing, Google's AI Overviews, Perplexity, Gemini, Claude - change the last step. Instead of handing you links to read, they read the pages for you and write a single answer, often with citations.
The natural assumption is that this replaces search. It does not. The crawl, index, rank pipeline that classic search is built on is still doing its job; the answer engine bolts a generation step onto the end of it. Understanding that is the whole point of this article - because it tells you what actually changed for anyone who publishes a page.
Where an AI gets facts
A large language model can answer from two very different places, and the difference decides everything downstream:
- Trained-in knowledge - facts baked into the model during training. Fast and always available, but frozen at the training cutoff and prone to hallucination: the model states something fluent and wrong because it is predicting plausible text, not looking anything up.
- Retrieved knowledge - facts the model fetches at question time from an outside source, then reads before answering. Current, checkable, and attributable to a source.
Ask a plain chatbot "what changed in this framework last week" and trained-in knowledge fails - it has no memory of last week. The fix, and the thing that connects AI back to search, is retrieval.
Grounding: search inside the AI
When an answer engine fetches live information before responding, it is doing retrieval-augmented generation (RAG) - often called grounding. The loop is short:
- Retrieve - run the user's question against a search index and pull the most relevant pages. This is literally the rank step from classic search.
- Augment - paste those pages into the model's context alongside the question.
- Generate - the model writes an answer grounded in that fetched text, and cites the pages it used.
Read that again: the retrieval step is search. The answer engine does not have a magic new way to know things about the web - it queries an index, takes the top results, and reads them. Which means every page that wants to inform an AI answer must first be findable and rank well in exactly the way SEO basics describes. The AI layer sits on top of the pipeline; it does not go around it.
The answer on top of search
Google's AI Overviews make the stacking visible. For many queries, the results page now opens with a generated summary above the familiar blue links. That summary is produced by the RAG loop above - retrieve the ranked results, read them, write the paragraph - and it links out to the sources it drew from.
So the ranked list did not disappear; it got a generated preface. The same is true, in different framing, of a standalone tool like Perplexity: a search runs behind the scenes, and the model narrates the top results with footnotes. Whichever product you use, the shape is the pipeline you already know plus a writing step.
Writing for two readers
Here is what changes if you publish pages. You are now writing for two audiences at once: the human who might click, and the model that might quote you in an answer. The good news is that they want mostly the same things, and this cluster of practices has picked up the name GEO - generative engine optimization - the answer-engine cousin of SEO.
- Be retrievable first. An answer engine can only quote a page its retrieval step found. Everything in robots.txt and sitemaps and SEO basics is the price of admission - unindexed pages are invisible to the AI too.
- Answer the question plainly. Models lift clear, self-contained statements. A page that states its answer directly is easier to quote than one that buries it.
- Be a citable source. Clear authorship, dates, and structured data help a model identify and attribute you - the same signals that earn rich results.
The through-line: there is no separate trick for ranking in AI answers. Being genuinely useful and clearly structured is what works for the crawler, the ranker, the reader, and now the model.
The zero-click tradeoff
The honest downside, named plainly: when the answer appears in full at the top, fewer people click through. This is the zero-click shift - the AI reads your page, uses it, and satisfies the user without sending them to you. You may inform the answer and still get no visit.
That reframes what you measure. A citation in an AI answer is a form of visibility even when it drives no click, and the UTM tracking and analytics from the rest of the series capture only the visits, not the influence. There is no settled playbook here yet - the fair summary is that being the source an answer engine trusts is becoming its own goal, distinct from the click it used to guarantee.
What has not changed
Step back and the picture is calmer than the headlines. Answer engines did not throw out crawling, indexing, or ranking - they consume them. A page still has to be crawlable, indexed, and ranked before any AI can retrieve it, and the work that earns that standing is the same work the whole search series describes.
What is new is one layer on top: a model that reads the top results and writes the answer, and a reader - the model - you now write for as well as the human. Learn the pipeline underneath (how internet search works is the place to start) and the AI layer stops looking like a replacement and starts looking like what it is - the newest thing sitting on a foundation that has not moved.