How AI search works

From ten links to one answer

For twenty years, search meant a ranked list of links and you chose where to click. AI search flips that. The tool reads the pages for you and hands back one written answer, often with a few cited sources beside it.

That changes what visibility means. Position three on a list still got clicks. A business left out of a written answer may never be seen by that searcher at all.

Under the surface, the main products work in a similar order. Google's guide to optimising for its generative AI features names the method: retrieval-augmented generation, which Google also calls grounding, relies on its core Search ranking systems to retrieve relevant, up-to-date web pages for the AI response. Anthropic's help page on Claude's web search describes the same pattern, with Claude invoking a search tool to ground its answer in content from the live web and citing the sources. None of the providers publish their full selection rules. What they do publish is enough to explain the stages in this guide.

Search engines used AI long before they wrote answers. In October 2019 Google announced BERT, a language model it used to understand queries better, especially longer, conversational ones where small words such as 'for' and 'to' change the meaning. That kind of AI interprets the question and still returns a list of links. Generative AI search adds a second job: writing the answer itself. When people say AI search today they usually mean the second kind, and that's what this guide covers. These terms come up in every explanation of it:

• Large language model (LLM): the AI that writes the answer. It learned language patterns from training data collected up to a cut-off date.
• Training cut-off: the date a model's training data stops. Anthropic's help centre lists a cut-off for each Claude model and says the models may not know about anything after it.
• Retrieval: fetching current pages from a search index or the live web at the moment someone asks.
• Grounding: tying the written answer to the retrieved pages so it reflects sources, not memory alone. Microsoft's Bing team describes grounding as the system that connects AI to current, authoritative information beyond the model's training.
• Query fan-out: running several related searches behind one question. Google documents it for AI Overviews and AI Mode.
• Citation and mention: a citation is a linked source shown with the answer; a mention is your business named in the text, with or without a link. AI recommendations versus mentions explains why the difference matters.

Step one: what the model remembers and what it looks up

An AI answer can come from two places, and knowing which one you're looking at explains a lot of odd results. The first is the model's training data, the text it learned from before its cut-off date. The second is retrieval, the pages fetched when the question is asked. Questions about current, local or changing details, such as opening hours or which businesses serve a suburb, are the ones that most need retrieval. OpenAI's help page on searching the web says ChatGPT may search automatically when a question would benefit from current information, and Anthropic says Claude searches when a topic benefits from current information.

When a tool answers from memory, it can describe a business as it looked months or years ago. Say a hypothetical cleaning company in Maroochydore changed its phone number last year. A model whose training data predates the change may repeat the old number if it answers without searching, while the same tool with search turned on can read the new number from the current website and listings. That's why fixing your own pages matters, and also why a fix won't show in every answer straight away.

The providers split these jobs at the crawler level too. Each names one bot that collects pages for search answers and a separate control for model training, so a business can opt out of training without dropping out of search answers. The table sets out what each provider's own documentation says.

OpenAI, Anthropic and Perplexity also run fetchers that visit a page only when a user's question calls for it: ChatGPT-User, Claude-User and Perplexity-User. OpenAI says robots.txt rules may not apply to ChatGPT-User because a user initiated the visit, and Perplexity says Perplexity-User generally ignores robots.txt for the same reason. Anthropic says disabling Claude-User stops its system retrieving your content in response to a user query.

Two practical checks follow. First, open your robots.txt and look for any of these user agents under a Disallow rule; a line added years ago by a developer or a security plugin can quietly keep a site out of one product's answers. Second, check firewall and CDN bot settings, because Google lists crawling being allowed by any CDN or hosting infrastructure among its basics for AI features, and Perplexity's crawler page gives firewall instructions for the same reason. A crawler visit in your logs proves a visit and nothing more; why a crawl doesn't prove an AI citation explains the gap.

Step two: a page must be eligible before it can be retrieved

Retrieval only works on pages the system can reach and is allowed to use. For Google the rule is written down: to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet. Google's AI features page adds that there are no additional requirements and no special optimisations needed. So the base checks are the same ones that decide whether a page can appear in ordinary search at all.

Eligibility is a minimum, never a promise. Google's guide to how Search works says it doesn't guarantee it will crawl, index or serve a page, even one that follows Search Essentials. Other providers publish less, but OpenAI's publisher FAQ says any public website can appear in ChatGPT search and asks site owners to make sure they aren't blocking OAI-SearchBot. Work through this list for each page you want cited:

• The page loads normally and isn't blocked in robots.txt for the search crawlers in the table above.
• It isn't marked noindex, and its canonical tag points to itself, so the URL you want cited is the one that gets indexed. Noindex, canonical and redirects explained covers the differences.
• Search Console's URL Inspection tool reports the page as indexed on Google.
• No nosnippet rule or very low max-snippet value limits what Google can show. Google lists nosnippet, data-nosnippet, max-snippet and noindex as controls that apply to its AI features.
• The important content is in the page text, not only inside images, PDFs or scripts. Google lists making sure important content is available in textual form among its basics for AI features.
• The page is linked from other pages on your site. Google lists findability through internal links in the same set.
• Your Google Business Profile and any Merchant Center data are current, which Google also names in that list.

Step three: query fan-out turns one question into several searches

AI search rarely runs your exact words as a single search. Google says AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources to develop a response, and that while the response is being generated its models identify more supporting web pages. Google doesn't publish how many searches it runs, so treat any figure you see quoted elsewhere as someone else's estimate.

OpenAI describes something similar for ChatGPT. Its help page says that when ChatGPT search uses partner search providers, it typically rewrites the question into one or more targeted queries, may send further, more specific queries after reviewing the first results, and can add a general location based on the user's IP address. Its own example turns a request for good restaurants near me into a search for top restaurants in the user's city.

Fan-out explains a pattern that puzzles many owners. A page gets cited for a detail it never targeted as a keyword, while the page built for the main phrase is left out. Each sub-search finds its own match, so a site where each page answers real questions plainly gives the tool more chances to pull it in.

Worked example, hypothetical. Say a plumbing business in Caloundra, which we'll call Coastline Plumbing, wants to appear when someone asks an assistant for an emergency plumber in Caloundra who can come tonight. The sub-searches below are illustrative only; no provider publishes its real ones, and the actual set will vary between products and from one search to the next.

1. Emergency plumber Caloundra: the tool looks for businesses that say they do emergency work in that suburb. Coastline's emergency page names Caloundra and the nearby suburbs it attends, so it's a candidate.
2. Plumber Caloundra open now: the tool looks for hours. If the website says 24 hours but the Business Profile shows weekday business hours, it has two conflicting answers and may name a business whose details agree instead.
3. Licensed plumber Queensland: the tool looks for licensing. A page that states the licence number in plain text gives it something specific to cite.
4. Caloundra plumber reviews: the tool reads the Business Profile and review sites. Recent reviews that mention burst pipes or blocked drains in Caloundra match the question closely.
5. What an emergency call-out involves: an explainer question. A clear FAQ on Coastline's site could be cited here even though that page never ranked for the main phrase.

Step four: the answer lifts passages and checks them against other sources

Retrieved pages are raw material; the answer is built from small parts of them, checked against other sources. That's why Coastline's fix list comes straight from the fan-out breakdown: make the hours agree everywhere, state the licence plainly, keep the suburb list accurate and answer the call-out question on the page. No one can promise a citation from those changes. Each one removes a reason for the tool to pass the business over.

Once the sub-searches return their pages, the model writes from the parts that answer each sub-question. In practice that's usually a sentence or a short paragraph, rarely the whole page. Claude's help page, for example, says its responses include direct citations, source links and relevant quotes where appropriate. The pages that hold up are ones where the answer to a question sits in a sentence that makes sense on its own.

You don't need to rewrite your site into fragments. Google's guide says you don't need to break content into tiny pieces for AI, or write in a specific way for generative AI search. The useful habit is ordinary clear writing: the direct answer first, then the detail. A hypothetical before and after shows the difference.

Before: 'We pride ourselves on decades of combined experience and a commitment to quality across a wide range of residential and commercial services throughout the region.' That sentence answers no question, so there's nothing to lift. After: 'We attend burst pipes and blocked drains in Caloundra, Golden Beach and Pelican Waters, and we take after-hours emergency calls. Every call-out is quoted before work starts.' Each clause answers something a searcher might ask, and the suburbs, services and hours can be checked against other sources.

That checking is the last step. For current or local questions, assistants read a batch of results and decide which businesses they can describe without guessing. Your website says you serve Caloundra. Does your Google Business Profile agree? Do directories show the same phone number? Do reviews mention the same services? When the sources agree, the answer has less to reconcile and your business is easier to name. Google AI Overviews and the assistants differ in plumbing, but they reward the same underlying signals:

• A plain answer near the top of each page.
• Service and location details that match everywhere online. Citations and business listing accuracy covers the directory side.
• Steady, genuine reviews with replies from the business.
• Mentions on sites other than your own, such as local news, suppliers or industry bodies.
• Structured data that matches the visible page. Google says it gives explicit clues about a page's meaning, and also that no special schema is needed to appear in AI Overviews or AI Mode.

Google AI search and chat assistants compared

The products share the retrieve-then-write pattern but differ in when they search, where they look and how they show sources. The table sets out what each provider's documentation says. Google's own features get a fuller treatment in AI Overviews and AI Mode for local businesses; this table is for telling the products apart, and how AI visibility differs by product goes further on each one.

Accuracy is the question people raise most about these products, and every provider answers it the same way: check the sources. Google's AI Overviews help page says AI responses may include mistakes. OpenAI says search results and citations can be incomplete, outdated or incorrect, and suggests opening a cited source to check it supports the answer. Anthropic says the quality of Claude's responses depends on the sources it references. A classic results page leaves the judging of each source to you. An AI answer does that judging for you, which saves time and adds a step where errors can creep in.

When an answer is wrong about your business, and how to measure

Answers are generated fresh, so they change. The same question can bring a different answer, different sources or no overview at all from one day to the next, and Google says AI Overviews appear only when its systems decide they add to classic Search. One check proves little. A routine tells you far more.

Say a hypothetical electrician in Nambour finds an assistant quoting old opening hours. The error usually traces back to a source the tool read, such as a directory listing nobody updated. The routine below fixes that source, which is the part a business controls. AI source correction and retraction goes further on stubborn errors.

Step one: ask the same questions in each product you care about, with search on, and save the answer, the date and the cited links. Step two: open every cited source and note which one carries the wrong detail. Step three: fix it at the source, starting with your website and Business Profile, then the directory or review site involved. Step four: use the product's feedback tool on the wrong answer; Google puts thumbs up and thumbs down icons at the bottom of each AI Overview, and Anthropic points users to the thumbs down button in Claude. Step five: repeat the same questions monthly so you compare like with like, and expect corrections to take time, because cached pages and training data both lag.

Measurement is partial, and each provider reports differently. The table lists the official reports available now and the limit of each.

No report tells you why a page was chosen. Google says no third-party tool has access to its internal ranking or AI systems, so treat any product claiming to show Google's selection logic with caution. How to measure ChatGPT referrals and how to interpret Google AI impressions walk through the reports in detail.

What a local business controls is the input: pages that can be crawled and indexed, answers stated plainly, and details that agree everywhere. Whether a product cites you is outside anyone's control, and no one can guarantee it. Start with one service page. Make the answer obvious in the first two sentences, check your details match across the web, then repeat for the next page. Our guide on ranking in AI search lists the full set of actions, and AI SEO work covers the same checks for businesses that want help with them.