Extractable Passages

Every answer surface — featured snippets, People Also Ask, AI Overviews, AI Mode, and citations in ChatGPT, Perplexity or Claude — works the same way. A system finds a passage that answers a question, lifts it out of the page, and presents it on its own.

So the unit of optimisation isn't the page. It's the passage. The question is whether any 40 to 60 word block on your page can be removed from its context and still answer the question completely.

Almost no writing does this naturally, because good prose builds. Extractable prose front-loads.

The core pattern

## [The question, phrased the way a person asks it]

[Direct answer, 40-60 words, complete on its own, no pronouns
pointing at earlier text, no "as mentioned above", includes the
subject of the question by name.]

[Then the expansion: why, how, exceptions, evidence, examples.]

That's the whole technique. What makes it hard is discipline, because the instinct is to set up before answering.

Before

Understanding email authentication can be complex. There are several protocols involved, and they interact in ways that trip people up. Once you've grasped how they relate to each other, the picture becomes clearer. SPF, for its part, is one of these.

After

What is SPF?

SPF (Sender Policy Framework) is a DNS record listing which mail servers are allowed to send email for your domain. Receiving servers check it against the sending server's IP address. If the IP isn't listed, the message fails SPF and is more likely to be rejected or filtered to spam.

The second version can be lifted whole. The first can't be lifted at all — it doesn't say anything.

Match the format to the question type

Extraction systems recognise structures. Using the wrong one costs the placement even when the answer is right.

Check the live SERP before writing. Whatever format currently holds the snippet is the format Google has decided the question wants. If it's a table and you write a paragraph, you're competing for a slot that doesn't exist.

The rules that make a passage liftable

1. The heading is the question. Phrased in natural language. "How much does it cost?" beats "Pricing Considerations". Question-shaped headings match how people ask and how prompts are phrased.
2. The answer is directly under the heading. Not after a transition, an image, or a paragraph of context.
3. No context dependency. No "this", "as we saw", "the above", "it" referring to something in a previous section. Repeat the noun.
4. Name the subject explicitly. "SPF is..." not "It is...". Extracted passages lose their heading sometimes.
5. 40 to 60 words for paragraphs. Under 40 gets padded; over 60 gets truncated mid-sentence.
6. Numbers and dates as digits, stated plainly. "Plans start at 29 euro a month" is extractable. A pricing table rendered as an image is not.
7. One idea per passage. A passage answering two questions gets extracted for neither.
8. In the HTML. Content that appears after a click or after hydration doesn't exist for most retrieval agents. Verify with curl.

Structure a whole page for it

A page should carry six to twelve extractable passages, one per question a reader would actually have. That's what query fan-out rewards: generative systems fire related queries and pull citations from those results, so a page answering the main question plus its follow-ups can be cited for several of them.

Build the question list from People Also Ask, related searches, the subheadings competitors use, and the questions that come up in sales and support. Then give each one its own H2 or H3 with a self-contained answer underneath.

Two additional placements worth having:

• A summary block near the top for long pages, written as substance rather than a teaser. It's frequently what gets extracted for the head query.
• A table wherever anything is comparable. Tables get parsed, extracted and quoted more reliably than the equivalent prose, and they're easier to read anyway.

Verify extractability

# What a non-rendering retrieval agent actually sees
curl -s https://example.com/page \
  | sed -e 's/<script[^>]*>.*<\/script>//g' -e 's/<[^>]*>/ /g' \
  | tr -s '[:space:]' ' ' \
  | head -c 2500

Read it. If the answers aren't in there, in order, near their headings, nothing else in this skill applies yet.

Then the human test: take any single H2 and its first paragraph, paste it somewhere on its own, and ask whether it answers the question with nothing else present. If not, rewrite it.

The honest part about clicks

Winning a featured snippet or an AI Overview citation often reduces clicks, because the answer is now on the results page. This is real and it isn't going away.

What to do about it, rather than pretending otherwise:

• Answer the question completely, then give a reason to visit. The tool, the calculator, the full dataset, the template, the specific case. Extraction gives away the answer; it can't give away the thing.
• Target questions whose answers need a visit. "What is X" gets extracted and you get a brand impression. "How do I do X for my specific situation" needs your page.
• Count the citation as the outcome where the click isn't available. Being the named source in an answer a million people read has value even at low CTR, and it's the only version of visibility available on some surfaces now.
• Don't withhold the answer to force a click. It doesn't work, it costs the citation, and readers resent it.

Don't do these

• Don't add an FAQ block of invented questions. Google reduced FAQ rich results to almost nothing, and a page padded with questions nobody asks reads badly to humans, which still counts.
• Don't write "the answer is" or "in short" as a marker. Extraction doesn't need signposting and it reads like SEO.
• Don't repeat the question verbatim as the first sentence of the answer. Wastes the extractable words.
• Don't optimise passages on a page that doesn't rank. Snippets come from results that already rank, usually in the top ten. Ranking first, extraction second.
• Don't chunk content into tiny pages for AI. Not required, and Google has said so. Multi-topic pages are handled fine.
• Don't rewrite in a special style for LLMs. Clear writing for people is the same thing.

Where the data comes from

The question list is the input that decides whether this works, and the free sources — People Also Ask, related searches, competitor subheadings — cover search but not the prompts people put to assistants, which are phrased differently and often more specific.

Keupera runs prompt research for the questions asked of LLMs, and AI Brand Radar to see which passages actually get cited and by which engine, which is the feedback loop this work otherwise lacks. Over MCP:

claude mcp add --transport http keupera https://mcp.keupera.com/mcp -s user

Related skills

ai-visibility-audit, search-console-analysis, content-refresh, glossary-definition-hub, comparison-alternatives-pages, schema-markup