What Is Passage Ranking and Chunking?
Passage ranking and chunking mean that search systems judge parts of a page, not only the whole page. Google has a passage ranking system that identifies and judges individual sections 1. Perplexity says it splits documents into self-contained spans that are retrieved and ranked one at a time 3, and Microsoft describes Copilot doing something similar 7. To be quoted, each section must make sense on its own.
- Google ranks individual passages as well as whole pages, and its passage ranking system is still listed as current 1.
- Perplexity and Microsoft both describe breaking pages into small, self-contained pieces that are ranked one by one 37.
- Lost context hurts retrieval. In Anthropic's own tests, adding context to chunks cut top-20 retrieval failures by 35%, and by 67% with BM25 and reranking 5.
- Google says you do not need to chop content into tiny pieces for AI. Clear headings and one idea per section are enough 6.
What is passage ranking and chunking?
Chunking is splitting a document into smaller pieces so a search system can store, find and score each piece on its own. Passage ranking is scoring those pieces against a question, so the best section of a page can win even when the rest of the page is a weaker match.
Think of a long report in a filing cabinet. A person asks for the paragraph on late-payment fees. A good clerk does not hand over the whole report. They pull the one paragraph that answers the question. Search systems now work like that clerk.
Google documents passage ranking, but it does not publish how its AI features cut pages into pieces. The chunking details here come from Perplexity, Microsoft and builders' guides.
For a site owner this changes the unit of competition. Your page is not the only thing being judged. Each section, with its heading and the sentences under it, is judged too.
How do AI search engines split web pages into chunks?
They use a mix of splitting rules, and the major engines do not publish their exact settings. What is public tells us the pattern: pages are cut into small pieces, each piece is turned into a number list called an embedding, and the closest pieces to the question are fetched and re-ranked 38.
Perplexity says it breaks documents into self-contained spans that can be retrieved and ranked one at a time. It searches with both keyword matching and meaning-based matching, merges the two candidate lists, and then uses a stronger reranker model to pick the final set 3. Microsoft says assistants such as Copilot parse content into smaller structured pieces, and those pieces are what get ranked and assembled into answers 7.
The table lists the splitting methods that builders of search tools commonly describe. Which one a given engine uses is not public.
| Method | How it cuts | What it may mean for your page (our reading) |
|---|---|---|
| Fixed size | Every N tokens, often near the embedding model's limit | A cut can land mid-thought, so short, tight paragraphs survive better |
| Sentence or paragraph | Breaks at sentence ends, line breaks or blank lines | Clear paragraph breaks give clean edges |
| Document structure | Follows headings, HTML tags, Markdown or PDF sections | If the splitter is structure-based, your headings become boundaries |
| Semantic | Starts a new chunk where the topic shifts, judged by embeddings | One idea per section helps the shift show up |
| Contextual | An AI model adds a short note saying where the chunk sits in the document | Naming the subject in each section helps if the note is missing |
Why does passage-level retrieval matter now?
Google started judging passages in 2020. It said the change would improve 7 percent of search queries across all languages 2. Its ranking systems guide, last updated 10 December 2025, still lists passage ranking as an AI system that identifies individual sections of a page 1.
AI answers raise the stakes. An assistant does not show ten links. It builds one answer from a handful of pieces and links the sources. Microsoft says assistants can often lift question-and-answer pairs word for word into their responses 7.
Lost context is the main failure. Anthropic gives the example of a chunk that says revenue grew by 3% without naming the company or the year. That chunk is hard to retrieve. In Anthropic's tests, adding a short context note to each chunk cut relevant chunks missing from the top 20 results from 5.7% to 3.7%, and combining it with BM25 and reranking took it to 1.9% 5. Those are results for Anthropic's own test set, not for your site, but the lesson carries over.
What did an AI agent search for when we asked this question?
On 20 October 2026 we gave the question 'how do AI search engines split web pages into chunks' to a Claude agent with a web search tool. It ran 7 searches. The record sits in our Query Fan-Out Explorer 9. This is one run by one tool on one day, so a repeat would differ.
The agent never searched our exact words. It split the question into definition, Google, retrieval-system, citation, size and heading angles. It also added names we never typed: Perplexity, ChatGPT and RAG. It listed 8 sites, which were SEO tools, agencies and cloud tutorial sites. None was Google, Microsoft or Perplexity documentation.
| Search it ran | What it seems to be after |
|---|---|
| how AI search engines chunk web pages | The question restated as a plain definition |
| passage ranking Google how it works | Google's own passage system |
| RAG chunking strategies for web content | The builder's view of splitting |
| optimize content for chunk retrieval LLM citations | Advice for publishers |
| ideal chunk size tokens semantic chunking | A number to aim for |
| how to structure headings for AI passage retrieval | Heading rules |
| Perplexity ChatGPT search retrieval passages from pages | How named assistants behave |
What makes a chunk quotable?
A quotable chunk answers one question in its first sentence, names its subject, and adds the detail that proves it. Headings help here: a section under a heading that carries its subject survives being cut out of the page. The invented example below shows a heading hierarchy where each H3 carries its own subject. For a wording rewrite, see atomic content.
- BEFORE
- ## Pricing
- ### Basic
- ### Plus
- ### Extras
- AFTER
- ## How much does a website audit cost?
- ### What does the basic website audit cost?
- ### What does the plus website audit add?
- ### Which extras can be added to a website audit?
What is the ideal chunk size, and should you write to it?
No engine publishes an ideal chunk size for web pages, so do not write to a token count. Pinecone, a vector database company, suggests testing sizes from 128 or 256 tokens up to 512 or 1,024 for its own search systems 4.
Anthropic describes its chunks as no more than a few hundred tokens 5. Those are builders' settings, not rules for your site.
Google is direct on the other side. Its guide says there is no requirement to break content into tiny pieces for AI to understand it, and that you should write for your audience 6. The same guide recommends headings that give a clear structure.
Our decision rule: write sections a human would find complete, usually 80 to 250 words. If a section runs past 300 words, check whether it holds two ideas and split it under two headings. If a section is under 40 words, check whether it needs the subject and a source to stand alone. These are our editing thresholds, not measured results.
How can you apply this to your pages this week?
The full editing checklist and a before-and-after example are in our guide to atomic content. For a quick check, paste your opening paragraph into the free Answer Block Checker.
What an AI assistant searches when it answers this
We gave the question below to AI search tools and recorded the web searches they ran behind the scenes. This is called query fan-out. Each phrase is a page you could publish to be found and cited.
“how do AI search engines split web pages into chunks”
The searches it ran- how AI search engines chunk web pages
- passage ranking Google how it works
- RAG chunking strategies for web content
- optimize content for chunk retrieval LLM citations
- ideal chunk size tokens semantic chunking
- how to structure headings for AI passage retrieval
- Perplexity ChatGPT search retrieval passages from pages
Frequently asked questions
Is passage ranking the same as chunking?
Do I need to split my pages into small chunks for AI?
No. Google's generative AI guide says there is no requirement to break content into tiny pieces for AI 6. Write complete sections under clear headings. The engine does the cutting, and clear structure makes its cuts cleaner.
What is the best chunk size in tokens?
No engine publishes one for web pages. Pinecone suggests testing from 128 or 256 tokens up to 512 or 1,024 in your own search system 4. For publishers, a complete section of about 80 to 250 words is a practical target, not a rule.
How do headings help AI find the right passage?
Does Perplexity really retrieve passages?
Its engineering team says it breaks documents into self-contained spans that are retrieved and ranked individually 3. The post describes its Search API and does not state span length, so treat exact sizes as unknown.
Sources
- A guide to Google Search ranking systems · Google Search CentralOpened October 20, 2026. Supports: Passage ranking is listed as an AI system that identifies individual sections or passages of a web page. Page footer: last updated 2025-12-10.
- How AI is powering a more helpful Google · Google (The Keyword blog)Opened October 20, 2026. Supports: On 15 October 2020 Google announced it can judge passages in addition to the overall page, and said this would improve 7 percent of queries across all languages.
- Architecting and evaluating an AI-first Search API · Perplexity ResearchOpened October 20, 2026. Supports: Perplexity breaks documents into self-contained spans retrieved and ranked individually, merges lexical and semantic retrieval into a hybrid candidate set, and uses cross-encoder rerankers.
- Chunking Strategies for LLM Applications · PineconeOpened October 20, 2026. Supports: Fixed-size, content-aware, structure-based, semantic and contextual chunking; suggests testing sizes from 128 or 256 tokens to 512 or 1024 tokens. Published 28 June 2025.
- Introducing Contextual Retrieval · AnthropicOpened October 20, 2026. Supports: Chunks can lose context; adding a 50 to 100 token note cut top-20 retrieval failures from 5.7% to 3.7%, 2.9% and 1.9% for the three methods tested; chunks are usually no more than a few hundred tokens. Published 19 September 2024.
- Optimizing your website for generative AI features on Google Search · Google Search CentralOpened October 20, 2026. Supports: No requirement to break content into tiny pieces for AI; recommends headings that give clear structure; warns against separate content for every search variation. Page footer: last updated 2026-07-10.
- Optimizing Your Content for Inclusion in AI Search Answers · Microsoft Advertising Blog (Microsoft Bing)Opened October 20, 2026. Supports: Published 8 October 2025: assistants like Copilot parse content into smaller structured pieces that are ranked and assembled into answers, headings mark where one idea ends, and assistants can lift question-and-answer pairs word for word.
- Vector embeddings · OpenAIOpened October 20, 2026. Supports: An embedding is a vector whose distance to another vector measures relatedness; small distances suggest high relatedness.
- Query Fan-Out Explorer (observed examples) · Chalk LabsOpened October 20, 2026. Supports: The 7 searches a Claude agent with a web search tool ran on 20 October 2026 for the prompt 'how do AI search engines split web pages into chunks', and the 8 sites it listed.