Content structure for AI retrieval: headings, chunks and paragraphs, 2026

Direct answer

Content structure for AI retrieval comes down to three habits. Give each heading one specific topic, put the answer in the first sentence under it, and write every paragraph so it still makes sense when copied alone. Retrieval systems cut pages into chunks, and your structure decides where the cuts fall and whether a chunk carries enough context to be matched. Google's guide for generative AI search, last updated on 2026-07-10, adds a limit: you do not need to break content into tiny pieces.

Before you start

How retrieval systems cut a page

Most retrieval-augmented generation (RAG) systems split a document into chunks, turn each chunk into a vector, and search those vectors. Anthropic's contextual retrieval write-up, published September 19, 2024, says chunks are "usually no more than a few hundred tokens." It also says chunk size, boundaries and overlap all affect results. If you want the full mechanics, our explainer on retrieval-augmented generation covers them.

You don't control the chunker. A search engine, an assistant or a customer's own RAG tool decides how to cut your page. What you do control is the material it cuts: headings, paragraph breaks and the words inside each paragraph.

What gets lost when a chunk travels alone

Anthropic gives a plain example. A chunk reads "revenue grew 3% over the previous quarter." It never names the company or the period, so a query about one company's Q2 2023 revenue may not match it.

Anthropic's fix is to prepend 50 to 100 tokens of generated context to each chunk before indexing. Measured as failed retrievals in the top 20 chunks, that cut failures by 35% (5.7% to 3.7%). Adding contextual keyword matching brought the cut to 49%, and adding reranking brought it to 67% (to 1.9%).

That fix belongs to whoever builds the pipeline, and as a publisher you can't add it. You can write the context into the text yourself, so the paragraph names the company, the period and the unit before any system touches it.

What you control as a publisher

Four things are yours: heading wording, paragraph boundaries, the subject each paragraph names, and clean HTML. You do not need special files. Google states that you do not need new machine-readable files, AI text files, markup or Markdown to appear in Google Search, and that it does not use llms.txt. Other assistants may read it, but treat it as optional.

Step-by-step: structure a page for retrieval

Step 1. Give each heading one specific topic

A heading is the label a chunk inherits. Semrush's content chunking guide (January 21, 2026) advises specific headings of roughly four to eight words and warns that "Overview" or "Next Steps" give a system little to work with. The LSEO guide makes the same point: use literal labels, signal intent, and borrow the audience's own vocabulary.

Split any heading that covers two intents. One heading, one question.

Step 2. Keep one idea per paragraph

One main idea per paragraph keeps a chunk from mixing topics. Semrush suggests one to three sentences where possible and clarifies that this does not mean dropping transitions. It means removing dependencies, so a paragraph does not need its neighbor to be understood.

Step 3. Open each section with the answer

Put the answer to the heading in the first sentence. Then add scope and limits, then the mechanics, then an example. If the answer sits at the bottom, a chunk cut earlier may never contain it. This is the same discipline behind answer capsules, applied to every section instead of only the top of the page.

Step 4. Make every paragraph self-contained

Read each paragraph as if it were the only thing on the page. Replace "it," "this" and "the tool" with the name. State the date, the unit and the subject inside the paragraph. Define a key term in one flat sentence the first time it matters, and repeat the definition briefly if a later section depends on it.

Here is a before and after, written as an illustration.

Before: "It also supports caching, which makes this much cheaper after the first call."

After: "Prompt caching in the Claude API lets a repeated prefix be reused across requests, which lowers the cost of every call after the first."

The second version survives being lifted into an answer box. The first doesn't.

Step 5. Size sections with the paste test

Paste the heading and its text into an empty box. If it still reads as a complete answer, the section is the right size. One practitioner guide suggests 100 to 250 words for a standard explanatory block, with up to about 300 words when a topic needs nuance. Treat that as one author's rule of thumb, not a measured threshold. Google says there is no ideal page length.

Step 6. Check the rendered HTML

Use real heading tags in order, with h3 nested under h2. Google says perfectly semantic HTML is not required, but recommends it where practical because it helps screen readers and other tools parse a page. Also confirm that your content is in the HTML the server returns, not added later by script, since a crawler that does not run JavaScript sees an empty page.

Common mistakes

Splitting content into tiny pieces for bots. Google's guide says: "There's no requirement to break your content into tiny pieces for AI to better understand it." Its systems can handle several topics on one page and surface the relevant part. Keep one page per topic and make the sections separable.

Clever headings and mixed intents. A witty heading hides the subject from a retrieval system. So does a heading that covers two questions.

Forcing Q&A everywhere. Chris Green's informal test, reported by Semrush, rebuilt the same articles as dense prose, structured prose and Q&A. Q&A retrieved best, and structured long-form did not perform badly. Green warned against converting content to Q&A when it does not suit the reader. That was one informal test, not controlled research.

Treating structure as a substitute for substance. Semrush puts it directly: "Chunking only works if the substance is there." Structure helps a system find a good answer. It can't create one. For the wider picture, see our guide to getting cited by AI assistants.

Trusting vendor claims as proof. Most chunking advice online is practitioner opinion. The measured figures in this article come from Anthropic's retrieval experiments, and they test a pipeline technique, not a publishing rule. Use the advice here because it also makes pages easier for people to read, which is the baseline Google recommends.

A short audit checklist

Run this on your five most important pages:

• Every h2 and h3 names a specific topic in four to eight words.
• The first sentence under each heading answers it.
• Each paragraph names its subject, with no dangling "it" or "this."
• Each heading plus its text passes the paste test.
• Headings are real HTML tags and present in the server response.

If you want a second pair of eyes on a content library, our SEO and AEO service covers this kind of audit, and you can reach us through the contact page.

Frequently asked questions

How long should a section be for AI retrieval?

No study sets a number. One practitioner guide suggests 100 to 250 words per explanatory block, while Google says there is no ideal page length. Size each section by whether the heading and text still make sense when pasted alone.

Do I need to split a page into many small pages?

No. Google says there is no requirement to break content into tiny pieces for AI. Keep one page per topic and make its sections easy to separate.

Should every article be written as Q&A?

No. In one informal test Q&A retrieved best, but the tester warned against forcing it where it does not suit the reader. Use question-style headings where your readers actually ask that question.

Does structure replace good content?

No. Structure only helps retrieval systems find substance that is already on the page. A well-labeled section with a weak answer still retrieves a weak answer.