AI search is a form of web search in which a language model reads pages pulled from a search index and writes a single answer, usually with links to the pages it used. The person asking gets a paragraph or a short list instead of ten blue links. ChatGPT search, Perplexity, Google AI Overviews, Google AI Mode, Gemini, Claude with web search, Grok and Microsoft Copilot all work this way.
The model does not know the answer in advance. It knows language, and it has seen a lot of text during training, but for a question about a product, a price, a local business or anything recent it depends on what the retrieval step hands it. That is why the pages a search engine can find and read decide what the answer says.
How does AI search work?
Most AI search products follow the same five steps, even when the user interface looks different.
- Understand the question. The model reads the prompt and the conversation before it. "Is it worth it?" only makes sense together with the previous turn.
- Write search queries. The model turns the question into one or more queries for a conventional search index. A broad question is often split into several narrower ones, a technique called query fan-out.
- Retrieve pages. The index returns candidate pages for each query. The system fetches them, strips the navigation and splits the text into passages.
- Select passages. A ranking step scores the passages against the question and keeps the few that answer it best. How AI search chooses which sources to cite covers this step in detail.
- Generate the answer. The model writes a response from the selected passages and attaches citations to the sentences it took from them.
The pattern of retrieving text first and generating from it second is called retrieval-augmented generation. AI search is retrieval-augmented generation running on top of a web index.
How is AI search different from a traditional search engine?
A traditional search engine ranks documents. AI search ranks documents, then ranks passages inside them, then hands the best passages to a model that decides which ones to quote. The visible result is different in four ways.
| Traditional search | AI search | |
|---|---|---|
| Result | A ranked list of pages | One written answer |
| Unit of competition | The page | The passage, and the brand named in it |
| Queries per question | One | Often several, written by the model |
| Position | A stable rank for a query | A mention or citation that varies between runs |
The last row matters most for anyone measuring their presence. Two people asking the same question a minute apart can get different answers, because the model samples its words and the retrieval step may return a slightly different set of pages. A single check tells you little. Repeated checks over time tell you how often a page or a brand appears.
Which products use AI search?
Some products are built around a generated answer, and some add one to a classic results page.
- Answer-first products. Perplexity, ChatGPT search and Google AI Mode return a written answer as the main result, with sources listed beside or below it.
- Answers above a results page. Google AI Overviews and Bing Copilot show a generated summary on top of the familiar list of links.
- Assistants with browsing. Claude, Gemini and Grok answer from training data most of the time and run a web search when the question needs current or specific information. When do AI assistants search the web? explains how they decide.
The difference affects what a website can expect from each. A product that always searches reads the live web on every question. An assistant that searches only sometimes may answer from memory, and the memory reflects what was on the web when the model was trained.
What does AI search mean for a website?
Three things have to go right for a page to influence an AI answer.
The page has to be retrievable. AI crawlers must be allowed to fetch it, and the search index the product uses has to contain it. A page blocked in robots.txt for the relevant bot is invisible no matter how good it is.
The page has to be quotable. The passage-selection step prefers text that answers a question on its own: a clear first sentence, a definition, a short list, a table with real values. A paragraph that only makes sense after reading the three above it rarely gets picked.
The brand has to be named where the model looks. When someone asks for the best tool for a task, the model lists the names it found in the pages it read. Review sites, comparison articles, forums and documentation are the usual sources. A brand that is absent from those pages is absent from the answer. Tracking that presence across platforms and prompts is what AI visibility measures.
