Structuring Content for Answer Extraction

Three different things get called answer extraction

Answer extraction is what a search or AI system does to a page when it pulls a span of text out of it and presents that span as the answer, with or without a link back. Featured snippets, People Also Ask boxes and the sentence-level quoting inside a Google AI Overview are all answer extraction. Structuring content for it means writing so that some passage of your page answers a question completely on its own.

Two other things are bundled under the same heading and have entirely different fates. FAQPage structured data is schema.org markup declaring that a page holds a list of questions and answers, each question a Question with an acceptedAnswer. Question headings are an editorial pattern with no markup in them at all: writing an h2 or h3 as the question a reader would actually type, then answering it in the first sentence beneath.

The distinction is the whole subject. As of August 2026 the markup is dead as a Google search feature, the editorial pattern is not, and the two were never the same thing. Almost every piece of advice in circulation on this topic treats them as one, which is how agencies came to sell schema deployment as an answer-extraction deliverable.

The markup died on 7 May 2026, and nobody announced it

Google narrowed FAQ rich results on 8 August 2023, restricting them to "well-known, authoritative government and health websites" and citing a cleaner, more consistent search experience. Most practitioners stopped there, and most 2024 and 2025 advice still describes that carve-out as current. It is not.

FAQ rich results stopped appearing in Google Search entirely on 7 May 2026. A deprecation notice appeared on Google's FAQ documentation the following day; the documentation itself was removed on 15 June 2026, along with the FAQ search-appearance filter and Rich Results Test support; support in the Search Console API ended in August 2026. There was no blog post. As Search Engine Journal reported on 10 May 2026, "Google didn't publish a blog post or explain the reasons behind the removal" — the only announcement was a documentation notice, and that notice was itself deleted five weeks later.

Two practical conclusions, in opposite directions. Deploying new FAQPage markup in 2026 as a Google Search deliverable cannot work, because there is no surface left for it to produce. But existing markup does not need ripping out: FAQPage remains a valid, non-deprecated schema.org type at version 30.0, adopted on somewhere between one and ten million domains, and Google's standing position is that unused structured data "does not cause problems for Search." Removal is optional cleanup with real regression risk and no benefit.

Passage ranking is not chunked retrieval

Two mechanisms sit behind "AI reads your content in pieces," and they are routinely conflated by people giving advice derived from one and applied to the other.

The first is passage ranking, which Google announced in October 2020 and described as improving 7% of search queries across all languages. Its entire public description is one sentence in the ranking systems guide: "Passage ranking is an AI system we use to identify individual sections or 'passages' of a web page to better understand how relevant a page is to a search." Read the last clause carefully. The passage informs the judgment; the unit that ranks is still the page.

The second is chunked retrieval in a retrieval-augmented pipeline, where a document is cut into fragments, each fragment is embedded and stored independently, and retrieval returns fragments with the document attached only as a pointer. Here the unit genuinely is the chunk. Enterprise systems work this way. Google's AI Overviews and AI Mode are grounded on the ordinary web index, which is passage-ranking territory, with a generative layer above it. Google has never described chunking in any AI features documentation, and neither has OpenAI, Anthropic, Perplexity or Microsoft. Assistants that fetch a live URL on request are doing a third thing again: converting a page to text and reading the whole of it under a token budget.

The argument that survives all of it

Whatever the segmentation, a retrieved fragment is read without the rest of the page. That single fact carries almost all of the usable advice on this subject, and it does not require you to know anyone's chunk size, because it holds under every architecture described above.

A paragraph that opens "as we saw above, this depends on several factors" is uninterpretable alone. A paragraph that opens "renewing a passport by standard service takes about three weeks" is not. Two corollaries have real teeth. Antecedents kill passages: it, this, they and the above resolve against text that may not travel with the fragment. And the answer belongs near the top of its own section, because a fragment truncated at a token boundary keeps its front.

This is the strongest practical advice available here and it still deserves an honest label: observed, not documented. The mechanism is described in the retrieval literature and the pattern is consistent practitioner experience through 2025 and 2026, but no operator has confirmed it and no published experiment isolates it on a live search product. Saying so costs nothing and is the difference between a page a reader can check and a page they have to trust.

Google says not to shred the page

The genre of advice that tells you to break content into micro-sections so machines can parse it has a direct refutation, from the operator, in a document written for exactly this question. In its generative AI optimization guide, under a section Google titled "Mythbusting generative AI search: what you don't need to do," Google writes: "There's no requirement to break your content into tiny pieces for AI to better understand it." It adds that its systems "are able to understand the nuance of multiple topics on a page." The guide is at Google's generative AI optimization guide.

The only structural guidance Google does give is framed around people rather than machines: "People generally appreciate it when web pages are organized by paragraphs and sections, along with headings." That framing is worth noticing. In a document about optimizing for machines, Google declines to make a machine-facing argument for structure.

The research points the same way. Qu, Tu and Bao tested semantic chunking — splitting on detected topic boundaries — against plain fixed-size chunking across document retrieval, evidence retrieval and answer generation in October 2024, and concluded that "the computational costs associated with semantic chunking are not justified by consistent performance gains." The study is at arXiv 2410.13070. The entire structure-your-content-for-chunking genre assumes better-bounded chunks retrieve better. In the one controlled comparison available, they did not reliably do so.

Where the numbers in circulation came from

Three families of figure dominate the advice sold under generative engine optimization (GEO), the practice of writing and marking up pages in the hope of being quoted by AI answer systems. None of the three measures what it is quoted as measuring.

The 112% FAQ schema lift. It comes from a TechCognate study of roughly 1,000 queries on Gemini, a different surface from AI Overviews, whose authors write in the study itself: "This is correlational, not causal. We can observe patterns, but we cannot definitively prove Gemini uses any specific signal." It is re-quoted constantly as a causal finding about AI Overviews. Two errors in one statistic, and the authors pre-empted the second one.

Recommended chunk sizes and answer lengths. Keep sections to 200 to 300 words, answer in 40 to 60 words, target 500-token chunks — none of these traces to a search operator. They trace to default parameters in retrieval libraries such as LangChain and LlamaIndex, and to agency posts citing each other. A library default is a configuration value in third-party tooling. It measures nothing about any search product.

Schema as an extraction lever generally. Ahrefs added JSON-LD to 1,885 pages against roughly 4,000 matched controls in May 2026 and measured −4.6% on AI Overviews, +2.4% on AI Mode and +2.2% on ChatGPT. No uplift anywhere. On featured snippets, the longest-running form of answer extraction, Google's answer to "how do I mark up a passage for this" has always been two words: "You can't."

How to actually write the page

What remains is editorial work, and it is unglamorous enough that it does not sell well.

• Use the question the reader would type as the heading, then answer it in the first sentence or two beneath. Not because markup rewards it — there is no markup — but because it forces the answer to the front of a self-contained span.
• Open every section with a noun, not a pronoun. This is the single highest-yield edit on most pages, and it takes minutes.
• Give each substantive question its own section rather than distributing the answer across four. A page can cover a topic thoroughly and still contain no single span that answers a specific question in it. Thoroughness at page level does not guarantee retrievability at passage level.
• Do not bolt a question block onto the bottom of the page to add FAQ content. Questions nobody asked, answered in terms already used above, add retrievable text that is redundant with the body. That dilutes the page rather than sharpening it.
• Leave the existing FAQPage markup alone unless you are already editing the template. It is harmless, and Google says so.

Notice what is absent from that list: word counts, chunk sizes, markup and any promise about citations. The advice is narrower than the genre it sits in, and that is the point.

What would change this verdict

The page carries an Observed verdict because its central recommendation rests on a mechanism that is well described in the retrieval literature and undocumented by every company that operates a live system. Two things would move it.

The first is a controlled test isolating question headings and direct answers, holding the visible content constant across matched page sets and measuring citation and extraction change across surfaces. That is the design Ahrefs used for structured data, and it returned a null result there, which is exactly why the same design applied here would be worth something. The second is any operator statement about how a fetched or indexed document is segmented for retrieval. As of August 2026 no such statement exists from Google, OpenAI, Anthropic, Perplexity or Microsoft — a silence that is uniform across the industry and is itself the most important fact on this subject.

Until one of those arrives, treat the writing pattern as a sound editorial default with a mechanism behind it, and treat anyone quoting a specific percentage lift for it as quoting something nobody measured.

Frequently asked questions