Content Structure for Passage Retrieval

Passage retrieval is the step where a search system finds the short piece of text that answers a question, instead of judging the page as one block. A passage can be a paragraph, a list or a few paragraphs under one heading. AI search engines such as Google AI Mode and ChatGPT search read passages from several pages, then write one answer and cite the pages they used. A citation is the link shown as the source.

This matters more because of query fan-out, the way AI search splits one buyer question into several smaller searches. Each smaller search looks for its own passage. Our service article, query fan-out explained, shows how to list those sub-questions. This guide shows how to lay out a page so each sub-question has one clear passage to find.

Retrieval systems cut pages into pieces before they compare them with a question. OpenAI's Retrieval guide says its file search splits every file into 800-token chunks with 400 tokens of overlap by default. A token is a short piece of text, often part of a word. Web search engines do not publish their exact cut points, but the idea is the same: the system ranks pieces, not whole pages.

Google has ranked passages in classic search for years. In its Search On post of October 2020, Google said it could "better understand the relevancy of specific passages" and expected this to improve 7 percent of search queries. On the answer side, OpenAI's web search guide ties each citation to a span of the answer text. If the right fact sits in a passage that makes sense alone, it is easier to find and easier to cite.

Start from the buyer prompt the page serves, then list the sub-questions behind it: what it is, what it costs, who it suits, how it compares, what proof exists, where and how fast it is delivered. The free Query Fan-Out tool expands one prompt into likely sub-questions. Its list is simulated by Leadee's rules, so treat it as a checklist, not as a record of real searches.

Give every sub-question one owner page, as the service article's mapping step explains. Then, for each page, keep only the questions it owns. A page that tries to answer every question in the cluster ends up with long, mixed sections that answer nothing clearly. If you track AI visibility already, use the same buyer questions as your prompt set.

Turn each question into a heading (an H2, or an H3 under it), written the way a buyer would ask it. "How long does a prototype take?" tells a reader and a machine exactly what the section answers. "Our process" does not. Use real HTML heading tags, not bold text styled to look like a heading, so the page outline is in the code.

Keep the heading levels in order. The W3C accessibility tutorial on headings says headings "communicate the organization of the content on the page" and that skipping ranks, such as an H2 followed directly by an H4, can be confusing. Give each heading a short, stable id so you can link straight to the section. The same outline helps screen-reader users, search engines and AI systems.

Open every section with the answer, then add the detail. A first sentence that states the prototype lead time and the part sizes it applies to works as a passage. "Speed has always been one of our core values" does not, even if the real answer comes three sentences later. If a system lifts only the first lines of the section, it should still get the answer.

Keep the opening sentence specific and checkable: a range, a standard, a place, a named product. How to write that sentence, with a number, who says so and a source, is covered in our guide on how to write pages ChatGPT can quote. This guide is about where that sentence sits and what surrounds it.

A retrieved passage arrives without the rest of the page. Words such as "it", "this service" or "as mentioned above" lose their meaning when the section is read alone. Repeat the subject by name at the start of each section: "CNC machining at the Ohio plant" instead of "it". Name your company or product where a reader might not know who "we" are.

Explain terms inside the section that uses them, not only in the first section of the page. If a section relies on a table or an earlier definition, add one short sentence that restates the key point. A reader who lands on the section from an AI answer link should be able to act on it without scrolling up.

Each section should answer one question. When a section starts to answer two, split it and give the second question its own heading. Keep paragraphs short (our own rule is 120 words at most), and lists for things that really are lists, such as steps, specifications or certificates. Long sections are more likely to be cut in the middle, with the answer in one piece and its context in another.

Do not write for one chunk size. The 800-token default in OpenAI's file search is one system's setting, developers can change it, and web search engines do not publish theirs. Short, self-contained sections work across chunk sizes: a small chunk still holds a full answer, and a large chunk holds a few complete ones instead of fragments.

Google's AI features guide lists "making sure that important content is available in textual form" among its good practices, and says no special markup or AI text file is needed. Prices, tolerances, service areas, lead times and certificates that live only in an image, a slider or a PDF download are hard to retrieve as a passage. Put them in the page text, and keep the image or PDF as a backup.

Use HTML tables for comparisons and specifications, with a clear header row. Make sure the text is in the HTML the server sends, not added later by JavaScript; our guide to JavaScript rendering and AI crawlers shows how to check with view-source. Structured data, the schema.org code that labels facts for machines, should repeat what the visible text says, not replace it.

Navigation, cookie notices, footers and repeated sales blocks can end up in the same pieces as your answers. In Leadee's test of 2026-11-09, a rule-based extractor split 8 Netbase JSC pages, including the CMSmart homepage (cmsmart.net, part of Netbase JSC), into 766 passages. Menu blocks were counted as passages too, because they used the same paragraph and list tags as the content. Our ChatGPT citation test has the details.

Use semantic HTML so machines can tell the parts apart: one main element for the content, nav for menus, footer for the footer, and aside for side notes. Google's blog post of 2025-05-21 on AI search lists, among page experience factors, "whether visitors can easily distinguish main content from other content". Do not repeat the same promo paragraph after every section; it adds noise to every chunk.

Check each page before and after the change. Open view-source and confirm the headings and answers are in the HTML. Read each section on its own and ask: does it answer its heading without the rest of the page? The free ChatGPT Citation Score splits a page into passages and flags the ones with no direct answer, so you can rewrite those first.

Then measure the buyer prompts, not the sections: buyers only see the final AI answer. Run each prompt several times per engine and compare periods, as the query fan-out service article explains. The Visibility Tracker runs those prompts for you through the engines' official APIs. Every other ChatGPT guide is on our ChatGPT SEO hub.

Example. A contract manufacturer has one "Capabilities" page with a single long block of text. A buyer asks an AI assistant for "CNC machining suppliers for aluminum prototypes in Ohio". The fan-out sub-questions include materials, tolerances, lead time and certifications. Before the change, all four answers sit in one long paragraph that opens with the company history. After the change, the page has four sections:

Leadee is an AI-native platform that stands on the shoulders of giants: it reads what Google, OpenAI and other engines publish, and it asks the engines through their official APIs rather than scraping consumer chat apps. The free Query Fan-Out tool gives you the sub-questions to build headings from, and the ChatGPT Citation Score shows which passages stand on their own.

Your team decides which questions each page owns and writes or approves the changes; Leadee never edits your website on its own. If you want experts to map the sub-questions and plan the page changes, start with the AI Visibility Audit. For B2B buyers, our guide on how AI chooses which suppliers to recommend explains what else the engines look at.