Systems that generate answers from web pages typically retrieve passages, not whole pages, and Google confirms that its generative features ground responses in retrieved pages. But Google also says you need not break content into tiny pieces or write in a special way. The sound response is ordinary good writing: clear headings, answers in real text and sections that still make sense when read alone.
A passage, read outside its page
Retrieval-augmented generation, introduced by Lewis and colleagues in 2020, pairs a language model with a retriever that fetches relevant text from an outside source, so the answer can draw on that text and show where it came from.[1] A study published the same year on dense passage retrieval showed the retrieval step working on passages, with learned representations rather than keyword matching.[2] Google describes its generative search features in the same terms: they retrieve pages from the Search index to ground responses, may issue several related searches through query fan-out and show supporting links.[3] Even in classic search, Google’s 2020 passage-understanding work was about the idea that a single sentence answering a specific question can be buried deep in a page.[4]
What Google says about chunking. Its guide to generative AI search says there is no requirement to break content into tiny pieces and no ideal page length. It also says you do not need to write in a specific way just for generative AI search.[5] This chapter is therefore not about engineering chunks. It is about writing sections that deserve to be read alone.
Use headings that identify the question
A heading is the label a section carries when it is read, searched or extracted. WCAG success criterion 2.4.6 asks that headings and labels describe a topic or purpose,[6] and the same rule serves readers who scan, people using screen-reader heading lists and any system that works out what a section is about. State the question or the point.
| Vague | Specific |
|---|---|
| Overview | What a ten-page website costs and what changes the price |
| Our approach | How we run a redesign in four stages |
| Details | Which file formats the booking system accepts |
| More information | What happens after you send the enquiry form |
Keep the hierarchy honest: one h1 for the page, h2 for the main sections, h3 only for real subsections. The heading and the first sentence under it should work as a pair: the heading names the question, the sentence answers it. The answer-first chapter covers that pairing.
Keep the answer available as text
Google’s list of practices for AI features includes keeping important content in text form and supporting it with high-quality images and video where relevant.[3] An answer inside an image, an unlabeled chart or a video with no transcript cannot be read by a system that reads text. WCAG’s criterion on images of text makes the same point for accessibility: use real text to convey information rather than images of text, with narrow exceptions.[7]
| Content | Problem | Fix |
|---|---|---|
| A price list as an image | The numbers are unreadable to text systems and to screen readers. | Use an HTML table with a header row. |
| A chart | The data lives only in pixels. | Add a text summary and the data table beneath it. |
| A video explainer | The answer is spoken, not written. | Publish a transcript or a written summary on the page. |
| Content loaded on click or after JavaScript | A crawler that does not click or render never sees it.[8] | Put it in the initial HTML. |
| A PDF as the only version | Harder to link to a section, and often not updated. | Publish an HTML page and offer the PDF as an extra. |
Keep evidence and exceptions close
An extracted passage should carry what makes it true. Put the number with its unit and date, the condition beside the claim, and the source link in the same section. Avoid references that break when the passage is lifted out, such as “as above,” “this” without a noun and “the previous method.” If an exception changes the answer, write it into the same paragraph.
Position matters too, in at least one respect that has been measured. Liu and colleagues found that language models’ performance can degrade significantly when the position of relevant information in a long input changes, and is often highest when it appears at the beginning or end and lowest in the middle.[9] That study concerns the text given to a model, not how pages rank, so do not read it as a rule for web pages. It does support a habit you already have from the answer-first approach: say the most important thing early, and do not bury the conclusion in the middle of a long section.
Test the public document
Check the page the way a system will meet it: as a public document, without your browser’s help. Copy each section into a blank file and see whether it still makes sense. Fetch the page with curl and read the HTML you get. Use the URL Inspection tool’s rendered view to see what Google sees. Then ask a colleague who has never seen the page a question it should answer, and see whether they can find the answer by reading one section.
Passage audit
- Heading
- Does it name the question or the point?
- First sentence
- Does it answer, and make sense alone?
- Condition
- Is the main exception in the same paragraph?
- Evidence
- Are the number, unit, date and source in the section?
- Text
- Is the answer in real text, present in the initial HTML?
- Fix
- The smallest edit that fixes the weakest answer.
The aim is not a prescribed length or format. A section that passes these checks is longer or shorter as its subject demands. If you find yourself splitting good prose into artificial pieces to satisfy a theory, stop. Google has said that is unnecessary, and readers will notice.
Apply it to your website
- Headings state the question or the point, and the first sentence under each answers it.
- Each important section makes sense when copied out of the page alone.
- Prices, specifications, steps and data are real text, in tables and lists, not images.
- Charts and videos have a text equivalent on the page.
- The answer is present in the initial HTML, without a click, a scroll event or a script.
- Numbers carry units and dates, and conditions and sources sit in the same section as the claim.
- You have not split the page into artificial fragments or padded it to reach a length.
Sources and further reading
Sources were consulted in October 2026 and the numbered notes above point here. Platform documentation changes often, so check the linked page before you act on a platform-specific detail. Worked scenarios are illustrative unless identified otherwise; methods and worksheets are Vantle’s editorial guidance.
- Lewis et al.: Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (NeurIPS 2020)The original retrieval-augmented generation paper.
- Karpukhin et al.: Dense Passage Retrieval for Open-Domain Question Answering (2020)Retrieval of passages with learned dense representations.
- Google Search Central: AI features and your websiteGrounding in retrieved pages, query fan-out, and best practices including text content.
- Google: Search On 2020, passage understandingGoogle’s description of ranking specific passages within a page.
- Google Search Central: Optimizing your website for generative AI featuresNo requirement to chunk, no ideal length, and no need to write specially for AI.
- W3C: Understanding WCAG 2.2, Headings and LabelsSuccess criterion 2.4.6: headings and labels describe topic or purpose.
- W3C: Understanding WCAG 2.2, Images of TextSuccess criterion 1.4.5: use text, not images of text.
- Vercel: The rise of the AI crawler (December 2024)Observed behavior of AI crawlers with JavaScript-rendered content.
- Liu et al.: Lost in the Middle: How Language Models Use Long Contexts (TACL)How the position of relevant information affects model performance.