How Do AI Search Engines Work?

An AI search engine is a search tool that replies to a question with a written answer, built from information it found on the web, and usually credits the pages it used.

That last part is what separates AI search from a plain chatbot. A chatbot with no search step answers only from its training data, which is frozen at a cutoff date. An AI search engine reads current pages first, so it can give today’s facts and show where it got them.

The ones most people use today are Google AI Overviews and AI Mode, ChatGPT search, Perplexity, Microsoft Copilot and Gemini. We look at the biggest ones in their own sections further down.

We’ll use one running, made-up example: a small, family-run Thai cooking school in Chiang Mai. A traveler asks an AI tool: “What’s the best Thai cooking class in Chiang Mai for vegetarians that includes a market tour and hotel pickup?” We’ll follow that question through the pipeline.

If you’re new to the idea of optimizing for these tools, our guide to answer engine optimization is the best place to start. This article goes one level deeper, into the machinery itself.

Traditional search finds and ranks pages; AI search finds passages and writes an answer from them.

Google describes classic search in three stages. Crawling, where Google “downloads text, images, and videos from pages it found on the internet.” Indexing, where it analyzes the content and stores it in “a large database.” And serving, where it returns “the results we believe are the highest quality and most relevant” to the query (Google Search Central).

AI search keeps the first two stages, because the model can’t read the whole web on demand. What changes is the third. Instead of showing you ten links and letting you do the reading, the system reads for you and hands back a summary.

Two rows matter most. The first is the unit that’s ranked. In classic search, a long page that covers a topic broadly can rank well. In AI search, the engine is looking for the one paragraph that answers one part of the question. A page can be strong overall and still have no single passage worth quoting.

The second is the output. Ten links means ten winners. An AI answer usually names a few brands, and there’s no page two. If the cooking school isn’t one of the three it recommends, that traveler may never see its name.

We cover the strategy side of this shift, and how it changes the job, in our comparison of AEO, SEO and GEO.

Retrieval-augmented generation, or RAG, is the technique of fetching relevant information first and then asking a language model to write an answer based on it.

The idea was laid out in a May 2020 paper by Patrick Lewis and eleven co-authors, titled “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks” (Lewis et al., arXiv). Their starting point was a known weakness. Large language models store a lot of facts in their parameters, but their ability “to access and precisely manipulate knowledge is still limited.” They can’t easily show where a fact came from, and they can’t update what they know without retraining.

The fix was to give the model two kinds of memory. Parametric memory is what the model learned in training. Non-parametric memory is an external store of documents it can look things up in. In the original paper, that store was a “dense vector index of Wikipedia,” searched by a neural retriever. The model retrieved the most relevant passages for a question and generated its answer conditioned on them. The authors reported that RAG models produced “more specific, diverse and factual language” than a model working from its parameters alone.

Today’s AI search engines are RAG at web scale. Swap “Wikipedia” for “a search index of hundreds of billions of pages,” add a step that writes good search queries, and you have the basic design of ChatGPT search, Perplexity and Google’s AI features.

You’ll also hear the word grounding. It means the same thing from the other side: the answer is grounded in (tied to) specific retrieved sources. When Google’s API returns a grounded response, it includes metadata listing the search queries it ran, the web sources it used, and which parts of the answer each source supports (Google AI for Developers). That mapping is where citations come from.

Why should a business owner care? Because it tells you what the model does with your content. It isn’t “remembering” your website. It’s reading a few passages from it, right now, and deciding whether they help. That’s far more controllable than most people assume.

The first job of an AI search engine is to work out what you actually want, which usually means pulling several separate needs out of one sentence.

AI search engines get much richer input than classic search. People type complete questions, add conditions (“for vegetarians,” “with hotel pickup”), and ask follow-ups that only make sense in context. The model reads the whole conversation, spots the constraints, and resolves references: if the traveler then asks “which of those does pickup from the Old City?”, it knows “those” means the schools it just listed.

Deciding whether to search at all

Next, the system decides whether a web search is needed. OpenAI says “ChatGPT will choose to search the web based on what you ask,” and users can also force a search with the search button (OpenAI). Google’s Gemini documentation describes the same decision as the model determining “if a Google Search can improve the answer.”

In our experience, questions about stable facts (“what is pad krapow?”) are often answered from training, while questions about prices, local details, reviews or recommendations usually trigger a search. Recommendation questions are exactly the ones businesses care about.

Google’s AI Overviews add one more filter: they don’t appear on every search. Google says they only show when its systems judge that they add something beyond classic results (Google Search Central).

Query fan-out is the step where an AI search engine turns your one question into several smaller searches, runs them, and pools the results.

Google confirms it uses the technique. Its documentation says AI Overviews and AI Mode may use “query fan-out,” issuing “multiple related searches across subtopics and data sources” to develop a response, which also lets them show a wider and more diverse set of links than classic search (Google Search Central). When Google announced AI Mode’s US rollout at I/O on 20 May 2025, it described the process as “breaking down your question into subtopics and issuing a multitude of queries simultaneously on your behalf.” Its Deep Search mode takes this further and “can issue hundreds of searches” before writing a fully cited report (Google).

Robby Stein, Google’s VP of Product for Search, gave a plain example in a July 2025 interview. For a question about things to do in Nashville with a group, the system “may think of a bunch of questions like great restaurants, great bars, things to do if you have kids,” searching topics the user never mentioned (Search Engine Journal).

For our cooking school, the traveler’s question might fan out into searches like these:

• vegetarian Thai cooking class Chiang Mai
• Chiang Mai cooking school market tour
• Chiang Mai cooking class hotel pickup Old City
• best cooking class Chiang Mai reviews 2026

What the data says about fan-out

Because OpenAI and Perplexity don’t publish their fan-out queries, the best numbers come from tracking companies that log them. Peec AI analyzed about 5 million fan-out queries collected from 1 to 21 April 2026 and found an average of 2.1 searches per prompt in ChatGPT, 1.4 in Perplexity and 6.8 in Grok (Peec AI). In the ChatGPT data, the word “best” was the most common word the model added on its own, appearing in about 15% of cases, followed by “what,” “reviews” and the current year (PPC Land).

An earlier Peec AI analysis of more than 20 million ChatGPT fan-outs, collected from October 2025 to January 2026 across Germany, the UK, Singapore, Thailand and the US, found that the average fan-out query roughly doubled in length, from about 6 words to about 12, while the number of fan-outs per prompt stayed between 2.3 and 2.8. The trend was “virtually identical across all countries and languages,” Thai included (Peec AI).

A word of caution: these are vendor studies. Peec sells AI visibility tracking and hasn’t published a full method for how prompts were chosen, and the numbers will shift as models change. The direction, though, matches what Google says about its own systems.

The lesson: you’re not competing for one query, you’re competing for a cluster of them. A school with one thin “Classes” page might match one of those four searches. A school with clear passages on its vegetarian menu, market tour, pickup zones and reviews has four chances. This is why building topical authority matters more in AI search than it ever did for single keywords, and why researching the prompts people actually use is now part of keyword research.

Once the best passages are selected, they’re handed to the language model along with your question, and the model writes a fresh answer in its own words.

This is the “generation” in RAG. The model doesn’t copy one source. It reads passages from several pages and writes a summary that combines them: this school has an organic farm, that one offers evening classes, a third has strong reviews for vegetarian dishes.

The models doing this writing are custom versions of each company’s flagship model. Google said in May 2025 that it was bringing “a custom version of Gemini 2.5” into Search for both AI Mode and AI Overviews in the US (Google). OpenAI said at launch that ChatGPT’s search model was “a fine-tuned version of GPT-4o,” post-trained with outputs distilled from its o1-preview reasoning model (OpenAI). Both companies have upgraded their models since, so treat those as launch details rather than what’s running today.

Why answers mix retrieved facts with remembered ones

Remember the two kinds of memory from the RAG paper. The model has the passages it just retrieved, but it also has everything it learned in training. When it writes, it uses both.

That’s why an AI answer can name a brand without linking to it. If the model read about a school many times in training data (travel blogs, forums, review sites), it may recommend it from memory and cite some other page for the details. Being named and being cited come from different places. We explain how to earn each one in our guide to AI mentions and citations.

AI search engines cite the pages whose passages actually support the statements in the answer, which is why the cited pages are often different from the top-ranked ones.

Google’s API makes the mechanism visible. A grounded response comes with three pieces of metadata: the search queries it ran, a list of “groundingChunks” (the web sources, with title and URL), and “groundingSupports,” which connect specific pieces of the answer text to specific sources (Google AI for Developers). A citation, in other words, is a link between one sentence and one passage. If no passage on your page supported a sentence in the answer, there’s nothing to link.

Ranking on Google doesn’t guarantee a citation

Ahrefs ran 15,000 long-tail prompts through Google, Bing and the major AI assistants in July 2025. Only about 12% of URLs cited by the assistants also ranked in Google’s top 10 for the same prompt. For ChatGPT it was 8%. Perplexity, which runs its own index, was highest at 28.6% (Ahrefs). Overlap with Bing’s top 10 was even lower, at about 10%.

Query fan-out explains much of this gap. The AI searches rephrased versions of your question, so a page that ranks first for “best cooking class Chiang Mai” may never come up for “Chiang Mai cooking class hotel pickup Old City.”

Google’s own AI features sit closer to its rankings: the same Ahrefs article points to its earlier finding that 76% of AI Overview citations came from top-10 pages.

What makes a passage citable

No engine publishes a citation formula, so be wary of anyone who claims to know it. But the research we do have points the same way. In the GEO study by Pranjal Aggarwal and colleagues, presented at KDD 2024, the researchers tested changes to content on a benchmark of 10,000 queries. Adding quotations, statistics and cited sources were among the most effective changes, with relative visibility gains of 30 to 40%, while keyword stuffing performed about 10% worse than doing nothing when tested on Perplexity (Aggarwal et al.). The tests ran on the researchers’ own benchmark, not live ChatGPT, so treat the exact numbers as directional.

Our reading, from the documentation and from checking a lot of AI answers: passages get cited when they are relevant to one sub-question, state a specific fact plainly, make sense on their own, and come from a source that looks trustworthy and agrees with other sources. If you want the full playbook, read what generative engine optimization is.

ChatGPT search works by deciding when a question needs fresh information, running searches through OpenAI’s own crawl and third-party providers, and writing an answer with links to the pages it used.

OpenAI launched ChatGPT search on 31 October 2024, opened it to all logged-in users on 16 December 2024, and made it available without an account in regions where ChatGPT works on 5 February 2025 (OpenAI). Search happens automatically when ChatGPT thinks your question needs it, or manually when you click the search icon. Answers show inline links, and a Sources button opens a sidebar with every reference.

Which crawler ChatGPT uses

This is where most sites get tripped up. OpenAI runs three different bots, and they do different jobs (OpenAI):

OAI-SearchBot is for search. It surfaces websites in ChatGPT’s search features. OpenAI recommends allowing it if you want to appear in ChatGPT search results. Opted-out sites can still appear as plain navigational links, but they won’t be used as sources for the answer.

ChatGPT-User fetches a page when a user’s action needs it, for example when someone pastes your URL into a chat. OpenAI says it is “not used for crawling the web in an automatic fashion,” and that because these visits are triggered by users, robots.txt rules may not apply.

GPTBot collects content that may be used to train OpenAI’s models. Blocking it is a separate decision. OpenAI says each setting is independent, so a site can allow OAI-SearchBot for search while disallowing GPTBot for training.

OpenAI also notes that after you change robots.txt, it can take about 24 hours for its search systems to adjust.

Because ChatGPT uses its own crawl and providers rather than Google’s index, a page that ranks well on Google can still be missing from ChatGPT. If that’s happening to you, our guide on why a site doesn’t show up in ChatGPT walks through the usual causes, and our ChatGPT citation guide covers what to do next.

Perplexity works as an AI-first search engine with its own crawler, its own index of more than 200 billion URLs, and a multi-stage ranking system that feeds passages to the model.

We covered its retrieval pipeline above: hybrid retrieval, self-contained spans, filtering, fast scoring, then cross-encoder reranking (Perplexity). Perplexity adds that a machine-learning model decides which URLs to index and when to recrawl them, based on importance and how often they change.

Perplexity runs two crawlers you should know about (Perplexity):

PerplexityBot “is designed to surface and link websites in search results on Perplexity.” Perplexity says it is not used to crawl content for AI foundation models, and it recommends allowing PerplexityBot in robots.txt if you want to appear in its results.

Perplexity-User visits a page when a user asks a question that needs it, and may link that page in the answer. Because the fetch is triggered by a person, Perplexity says this fetcher “generally ignores robots.txt rules.”

For site owners, Perplexity’s numbered inline citations make it easy to see who got credit for each claim. It also had the highest overlap with Google’s top 10 in the Ahrefs study (28.6%). Our guide on getting cited by Perplexity goes further.

Each AI search engine relies on a specific crawler to find pages, and that crawler is different from the one used to collect training data.

This is the most practical point in this guide. If the search crawler can’t reach your page, retrieval never sees it. Here’s the map, from each company’s documentation:

Note that Google doesn’t use a separate AI crawler for AI Overviews and AI Mode. Google’s documentation says the Googlebot robots.txt rules remain the main access control, and that Google-Extended is for limiting use of content in training and grounding in some of Google’s other AI systems (Google Search Central).

How much do AI crawlers send back?

Cloudflare, which sits in front of a large share of the web, publishes data on this. In August 2025 it reported that, over the previous 12 months, about 80% of AI crawling was for training, 18% for search and 2% for user actions. It also measured “crawl-to-refer” ratios, the number of pages crawled for every visit sent back. In July 2025, Google’s ratio was about 5 to 1, Perplexity’s about 195 to 1, OpenAI’s about 1,091 to 1, and Anthropic’s about 38,066 to 1 (Cloudflare).

That’s why many sites block AI bots wholesale. Our advice: decide on training bots by your own content policy, and keep search crawlers open.

You can test all of this in about a minute with our free AI Crawler Checker, which checks whether Googlebot, Bingbot, OAI-SearchBot, PerplexityBot and the training bots can reach your pages. If you need to fix the file itself, the robots.txt generator builds a clean one.

AI search engines make mistakes when retrieval finds weak sources, when the model misreads a passage or the question, or when there simply isn’t enough good content on a topic.

The most-cited test of this comes from Columbia University’s Tow Center for Digital Journalism. In March 2025, researchers Klaudia Jaźwińska and Aisvarya Chandrasekar gave eight AI search tools excerpts from news articles and asked them to identify the original article, publisher and URL. Across 1,600 queries, the tools gave incorrect answers more than 60% of the time. Perplexity did best at 37% wrong; Grok 3 did worst at 94% (Nieman Lab). The test was narrow (finding the source of a quoted excerpt, not answering everyday questions), but it shows that citations aren’t always reliable.

Google has been open about its own failure modes too. After AI Overviews launched in May 2024, some answers went viral for bad advice, including one that suggested using glue to get cheese to stick to pizza, drawn from a forum post. Google explained that errors usually come from misreading a query or a nuance of language, or from a “data void,” a topic with little high-quality content available, and it said AI Overviews “generally don’t ‘hallucinate'” in the way other LLM products can (Google).

Put those together and you get four main ways AI answers go wrong:

Bad retrieval. A satirical, outdated or weak page was the best match for a sub-search.

Misreading. The model takes a passage out of context or merges facts from two businesses.

Data voids. Niche or local topics may have only one or two pages to draw from.

Randomness. Generation varies from run to run, so a correct answer today can be a slightly wrong one tomorrow.

For businesses, data voids are the interesting one. If nobody has written a clear page about vegetarian cooking classes with Old City pickup, the AI makes do with whatever it finds and may describe your school wrongly. Publishing that page yourself fills the void with correct facts.

Once you know how AI search engines work, the to-do list writes itself, because every step in the pipeline maps to one job you control.

Decide and retrieve: be reachable. Let Googlebot, Bingbot, OAI-SearchBot and PerplexityBot fetch your pages, get key pages indexed in Google and Bing, and keep main content in the HTML, not only in JavaScript, images or PDFs.

Fan-out: cover the cluster. List the sub-questions behind your most valuable queries (for the cooking school: vegetarian options, market tours, pickup zones, price, reviews) and give each a clear passage, in Thai and English if your customers use both.

Rerank: write passages that stand alone. A direct answer in the first sentence under each heading, headings that match how people ask, and plain specifics: numbers, times, places, who it’s for.

Generate: give the model facts it can trust. Cite sources, show who wrote the page, keep dates and prices current.

Mentions: build your reputation where the model reads. Reviews, travel forums, Pantip threads, “best of” lists and your Google Business Profile shape whether your brand gets named. Keep brand facts consistent everywhere; entity SEO explains why.

Cite: measure across engines, over time. Every engine uses a different index and answers vary between runs, so track a fixed set of prompts in each engine every month.

How to see what AI search engines say about you

The only reliable way to see how AI search engines treat your brand is to ask them your customers’ questions, several times each, and record what comes back.

Here’s the process our team uses. A spreadsheet is enough to start.

1. Write down 20 to 50 questions that lead to a purchase in your category, in the words customers use (and in both languages if your market is bilingual).
2. Run each question in ChatGPT with search, Google AI Mode, Gemini, Perplexity and Copilot.
3. Repeat each run at least three times, because generation and fan-out vary.
4. Record whether your brand is mentioned, whether your pages are cited, and which competitors and third-party pages appear.
5. Repeat monthly and watch the trend, not individual answers.

Two things usually stand out. The engines disagree: a brand can be recommended by Perplexity and absent from ChatGPT, which is what you’d expect when they search different indexes. And many cited pages are third-party review sites, forum threads and comparison lists. Those are your outreach targets.

Doing this by hand gets slow once you have more than a few prompts. That’s why we built GrixTrack, which tracks Google rankings by country, city and language alongside mentions and citations in ChatGPT, Gemini and Google AI Mode, and runs each prompt three times to smooth out the randomness. The free plan covers 10 keywords and 5 prompts. For the full method, including which metrics to report, see how to measure AI visibility.

If you’d rather have someone else look first, our AEO team can run the check for you and show you which sources AI cites for your most important questions.

Google Search Central · YouTube

Google Search Central · Google crawlers behind the scenes · watch on YouTube

1. Google Search Central: AI features and your website (updated 10 Dec 2025)
2. Google Search Central: In-depth guide to how Google Search works (updated 18 Dec 2025)
3. Google: AI Mode in Google Search, updates from Google I/O 2025 (20 May 2025)
4. Google: AI Overviews, about last week, by Liz Reid (30 May 2024)
5. Google Thailand: AI Mode in Google Search is now available in Thailand (22 Aug 2025)
6. Google Thailand: AI Mode now available in Thai, plus Google AI Plus (8 Oct 2025)
7. Search Engine Journal: Query fan-out technique in AI Mode, new details from Google (30 Jul 2025)
8. Google AI for Developers: Grounding with Google Search (Gemini API)
9. OpenAI: Introducing ChatGPT search (31 Oct 2024, updated Feb 2025)
10. OpenAI: Overview of OpenAI crawlers
11. Perplexity: Architecting and evaluating an AI-first search API (25 Sep 2025)
12. Perplexity: Perplexity crawlers
13. Microsoft: Bringing the best of AI search to Copilot (7 Nov 2025)
14. Microsoft Learn: How to use Grounding with Bing Search in Foundry Agent Service
15. Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (arXiv, May 2020)
16. Aggarwal et al.: GEO: Generative Engine Optimization (KDD 2024)
17. Ahrefs: Only 12% of AI-cited URLs rank in Google's top 10 for the original prompt (Aug 2025)
18. Peec AI: ChatGPT fan-outs have doubled in length in 4 months, 20 million fan-outs analyzed (12 Feb 2026)
19. Peec AI research: Fan-outs per prompt by engine (May 2026)
20. PPC Land: What ChatGPT actually searches for, 5 million fan-out queries analyzed (Jun 2026)
21. Nieman Lab: AI search engines fail to produce accurate citations in over 60% of tests, Tow Center study (Mar 2025)
22. Cloudflare: The crawl-to-click gap, Cloudflare data on AI bots, training and referrals (29 Aug 2025)
23. Search Engine Land: Google zero-click searches reach 68% in early 2026 (SparkToro / Similarweb data)
24. Pew Research Center via Search Engine Roundtable: Google users click less when AI summaries appear (Jul 2025)

Keep reading

• AEO basicsWhat Is AEO? Answer Engine Optimization Explained
• AEO basicsWhat Is GEO? Generative Engine Optimization Explained
• AI platformsHow to Rank in Google AI Overviews
• AI platformsWhy Is My Website Not Showing in ChatGPT?