Skip to content
Learn

How AI search works

In this lesson

Section titled “In this lesson”
  • Describe how AI search changes clicks and visits, with real numbers.
  • Compare the main AI search products and the index that each one uses.
  • Explain the RAG pipeline: retrieve passages first, then write an answer from them.
  • Explain embeddings, dense retrieval and chunking in plain words.
  • Explain query fan-out and multi-step reasoning, and what they mean for your content.

On day 01 you met GEO, grounding and vector search. This lesson opens the machine. All AI search products follow the same basic pattern. When you know the pattern, you can see why the base of SEO still matters, and where AI search is different.

Section titled “How AI is changing search”

People now ask questions in a chat box, not only in a search bar. The roadmap.sh item says that content “now needs to work for two audiences”: search crawlers and AI systems that summarize information.

The change is large. In August 2026, Google said that AI Overviews “has over 2.5 billion monthly active users” and that AI Mode “has surpassed one billion monthly users”. At I/O in May 2026, Google said that total Search queries reached an all-time high.

But more searches do not mean more clicks. In July 2025, Pew Research Center studied the browsing of 900 US adults in March 2025:

  • About 18% of Google searches showed an AI summary.
  • With an AI summary, users clicked a normal result in 8% of visits. Without one, they clicked in 15% of visits.
  • Users clicked a link inside the AI summary in only 1% of visits.

This is one study of US users, from one month. Google says that clicks from pages with AI Overviews are “higher quality”, because people spend more time on the site. Both can be true: fewer clicks, but better ones. Day 08 shows how to measure this for your own site.

What does Google tell site owners? Its AI optimization guide (July 2026) says that optimizing for generative AI search “is optimizing for the search experience, and thus still SEO”. The rest of this lesson shows why: the AI products sit on top of a search index.

The AI search products

Section titled “The AI search products”

Chat-based search means that a person asks an AI assistant in natural language, and the assistant replies in text, not with a page of ten links. The assistant can answer from its training data, from a web search, or from both. These are the main products:

Product Who Where the web results come from How it shows sources
AI Overviews Google Google’s Search index A summary at the top of the results page, with links
AI Mode Google Google’s Search index, with query fan-out A full chat answer in Search, with links
ChatGPT search OpenAI Third-party search providers and partner content; OpenAI’s own crawler, OAI-SearchBot Inline citations and a sources list
Perplexity Perplexity Its own web search Numbered citations in every answer
Copilot Microsoft Bing Inline links; “Copilot Search cites its sources prominently”
Gemini Google Grounding with Google Search, when the model decides that a search helps Inline citations

Some details for each:

  • AI Overviews show only “when our systems determine that it is additive to classic Search”, so they “often don’t trigger”.
  • AI Mode is for “further exploration, reasoning, or complex comparisons”. Since May 2026, its default model is Gemini 3.5 Flash. Google says that the two features “may use different models and techniques”, so their links can differ.
  • ChatGPT search launched on 31 October 2024. OpenAI said that it “leverages third-party search providers, as well as content provided directly by our partners”. OpenAI’s crawler docs say that sites that block OAI-SearchBot “will not be shown in ChatGPT search answers”.
  • Perplexity calls itself an AI search product. Its help center (August 2026) says that it searches the web “in real-time” and lets paid users pick the model, for example from OpenAI or Anthropic.
  • Copilot is Microsoft’s assistant. Copilot Search in Bing launched in April 2025 and “blends the best of traditional and generative search”.
  • Gemini is Google’s assistant, a separate product from Search. The Gemini API docs describe its grounding: the model “determines if a Google Search can improve the answer”, runs “one or multiple search queries”, and returns an answer with citations.

What matters most here: each product depends on an index. For Google’s features, a page “must be indexed and eligible to be shown in Google Search with a snippet”. For ChatGPT, the page must not block OAI-SearchBot. For Copilot, the page must be in Bing. Day 07c covers these crawlers and the controls for them.

RAG: the shared pattern

Section titled “RAG: the shared pattern”

Retrieval-augmented generation (RAG) is a method where a system first retrieves documents, then a language model writes an answer from them. The name comes from a 2020 paper by Lewis and others. It joined a language model with “a dense vector index of Wikipedia”. The authors found that RAG models wrote “more specific, diverse and factual language” than a model without retrieval.

Google uses the same idea. Its AI optimization guide defines RAG as a technique “also known as grounding”. It relies on “our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index”. So day 01’s grounding and RAG are the same pattern. The two names come from different fields.

A RAG answer has five steps:

  1. Question. The system reads the question and decides if a search can help.
  2. Fan-out. The model writes one or more search queries.
  3. Retrieve. A search returns pages or passages for each query.
  4. Generate. The model writes the answer from the retrieved text.
  5. Cite. The answer links to the sources that support it.

Example data. Press Next step to move through the pipeline. Then turn off web search, or remove your page from the index. Notice that your page can get a link only if it was retrieved in step 3.

The pipeline shows the most important rule of this lesson. You cannot be cited if you are not retrieved. Retrieval comes before generation, as it comes before ranking in normal search. This is why crawling, indexing and relevance still decide most of the result.

Embeddings and dense retrieval

Section titled “Embeddings and dense retrieval”

An embedding is a list of numbers that represents the meaning of a piece of text. Google’s machine learning course says that embeddings “place similar items closer together”, so the distance between two embeddings shows how similar they are. Two passages about the same idea get close embeddings, even when they share no words.

Dense retrieval finds passages with embeddings. The system turns the query into an embedding, then finds the stored passages whose embeddings are nearest. “Dense” means that every number in the list has a value. “Sparse” methods, such as BM25, use mostly empty lists of word counts. Day 01 called these two methods vector search and lexical search.

The key paper is Dense Passage Retrieval (Karpukhin and others, 2020). It used two encoders: one for questions and one for passages. Similarity was the dot product of the two vectors. Its dense retriever beat a strong BM25 system “by 9%-19% absolute” in top-20 passage accuracy on question-answering tests.

Dense retrieval does not replace words. As day 01 showed, many systems use hybrid search, because exact terms such as product names still need lexical matching. No AI search company publishes the full mix that it uses.

Chunking

Section titled “Chunking”

Chunking means splitting a long document into smaller pieces, called chunks, before the system embeds and stores them. The system then retrieves chunks, not whole documents.

Why split? Pinecone, a vector database company, gives two reasons. An embedding model can take only a limited amount of text. Also, a chunk must hold the information that a search needs. Large chunks “may introduce noise or dilute the significance” of single sentences. Pinecone’s rule of thumb: “if the chunk of text makes sense without the surrounding context to a human, it will make sense to the language model as well.”

How large is a chunk? There is no standard. The DPR paper split Wikipedia into “disjoint text blocks of 100 words”. Other systems use a fixed number of tokens or split at headings.

Google’s normal search has a similar idea. Its passage ranking system identifies “individual sections or ‘passages’ of a web page” to judge relevance.

So should you cut your pages into small pieces? Google says no. Its AI optimization guide lists “chunking” content under things that you “can ignore”: “Google systems are able to understand the nuance of multiple topics on a page and show the relevant piece to users.” Many third-party GEO guides say the opposite. For Google, follow Google. For other products, no company publishes its chunking method, so the claims are not confirmed. A safe middle path is the one that helps human readers too: clear sections, each with a heading and one main point. Day 07b covers how to structure content for AI extraction.

Query fan-out

Section titled “Query fan-out”

Query fan-out means that the AI system turns one question into several related searches, runs them, and joins the results. Google’s I/O 2025 post says that AI Mode breaks “your question into subtopics” and issues “a multitude of queries simultaneously on your behalf”. Google’s guide says that both AI Overviews and AI Mode may use it.

Google gives an example. For “how to fix a lawn that’s full of weeds”, fan-out queries “might include ‘best herbicides for lawns’, ‘remove weeds without chemicals’, and ‘how to prevent weeds in lawn’”.

How many searches? Ahrefs (March 2026) reports that AI Mode “typically makes 5 to 11” for a shopping task, and that ChatGPT Deep Research made 420. It also cites research from Seer Interactive and Nectiv with “an average of 9-11 fan-out queries per prompt”. These are third-party observations. The companies do not publish their numbers.

Fan-out changes how you think about one page. A page can be retrieved for one sub-question, even if it never ranks for the main question. A page that answers several sub-questions clearly has more chances.

Sections on your page
How the system splits your page

Toy model with example data. Tick the sections that your page has, then change how the system splits the page. Notice that each new section can match one more fan-out query, and that one big chunk with many topics loses matches.

Google adds a warning. Do not create a separate page for every fan-out query. Doing that “primarily to manipulate rankings or generative AI responses” breaks the scaled content abuse policy from day 02. Cover the real sub-questions of your readers on good pages instead.

Multi-step AI intent

Section titled “Multi-step AI intent”

Some questions need more than one round of search. Multi-step AI intent describes a question where the system searches, reads the results, finds a gap, and searches again before it answers.

Google first described this for AI Overviews in May 2024. It said that a custom Gemini model with “multi-step reasoning” lets people ask a complex question “all in one go”, instead of many separate searches. Google’s example: a yoga studio that is popular with locals, near your commute, and has a discount for new members.

Research modes take this further:

  • Google’s Deep Search in AI Mode “can issue hundreds of searches” and write “a fully-cited report”.
  • Perplexity’s Research mode “performs dozens of searches” and “reads hundreds of sources”.

For content, this means that the system may read your page as one step in a longer task. It may use your page for one fact, such as a price or a specification, and other pages for the rest. Facts that are clear, current and easy to find on the page are easier to use. Day 07c covers agents that act on such tasks.

What matters most

Section titled “What matters most”
Priority Idea Why
1 Be in the index Every AI product retrieves from an index. No index, no citation
2 Answer the real sub-questions Fan-out and multi-step search look for parts of an answer
3 Write clear, self-contained sections Retrieval works on passages, and a clear passage matches better
4 Offer something that others do not Google’s guide asks for “non-commodity content” with a unique point of view

There are also things to ignore. Google says that you do not need llms.txt files, special markup, or rewriting “just for AI systems”. For Google, good SEO is good AI search optimization.

Key terms

Section titled “Key terms”
  • Chat-based search: finding information by asking an AI assistant in natural language.
  • AI Overviews: Google’s AI summaries at the top of some search results pages.
  • AI Mode: Google’s full chat search experience, with fan-out and follow-up questions.
  • Retrieval-augmented generation (RAG): a method where a system retrieves documents first, then writes an answer from them.
  • Embedding: a list of numbers that represents the meaning of a piece of text.
  • Dense retrieval: finding passages by comparing the embedding of a query with the embeddings of stored passages.
  • Chunk: a small piece of a document that a system embeds and retrieves on its own.
  • Chunking: splitting a document into chunks.
  • Query fan-out: turning one question into several related searches, then joining the results.
  • Multi-step AI intent: a question that needs several rounds of search and reasoning before the answer.

Exercise

Section titled “Exercise”

Exercise · about 15 min

Watch fan-out and citations in one AI answer

Use Perplexity (free, no account needed) or Google AI Mode.

  1. Ask: “How do I fix a lawn that is full of weeds?”
  2. Open the list of sources or steps. In Perplexity, look for the steps or sources panel. Write down any search queries that it shows.
  3. Write down the domains of the first five cited sources.
  4. Search the same question in a normal Google search. Write down the domains of the top five organic results.
  5. Compare the two lists. How many domains are in both?
  6. Open one cited page. Find the exact passage that supports the answer. Is it under a clear heading? Does it make sense on its own?
Show a possible solution

Your results will vary by product, place and date. That is part of the lesson.

In step 2, expect queries that are not your exact words. They often match Google’s example: weed killers, methods without chemicals, and prevention.

In step 5, expect a partial overlap. The AI answer uses a search index, so many cited pages also rank well. But fan-out searches for sub-questions, so some cited pages rank for a sub-question, not for your full question.

In step 6, the supporting passage is usually a short section with one main point, for example a list of steps under a heading. That is a chunk that makes sense on its own. If you cannot find the passage, the system may have used the page for a smaller fact, or the page may have changed since the system read it.

Check yourself

Section titled “Check yourself”
A page blocks no crawlers, but it is not in Google's index. Why can AI Overviews not cite it?
A page uses the words 'affordable plane tickets'. A user asks for 'cheap flights'. Which method is most likely to retrieve the page?
An AI Mode answer to 'how to fix a lawn full of weeds' cites a page that does not rank in the top ten for that query. What is the most likely reason?
A third-party GEO guide says: 'Split every page into 50-word blocks so that Google AI can chunk it.' What does Google's own guidance say?
A user asks AI Mode for a studio that is near their commute, popular with locals and has a new-member discount. Why may your page about discounts be used, even though it says nothing about location?

Sources

Section titled “Sources”

Next

Section titled “Next”

Day 07b covers how to write for AI answers: structure for extraction, TL;DRs, data tables, direct answers, citation resources and Google’s preferred sources.