TL;DR
- LLMs extract spans, not pages. If your answer isn't isolated as a discrete block, the model picks someone else's.
- Lead every section with a one-sentence answer, then expand. Bury the lede and you lose the citation.
- Definition blocks, tight Q&A, and comparison tables are the three highest-yield formats for extraction.
- Freshness compounds with structure: Profound's analysis and other 2025 studies show AI engines favor recently updated pages, so reformat old posts instead of writing new ones.
- Markup matters less than visual hierarchy: a short paragraph under a clear H2 beats elaborate schema on a wall of text.
When an AI engine answers a query, it doesn't read your page the way a human does. It chunks the document, embeds the chunks, retrieves the top-matching ones, and stitches a response from the cleanest extractable spans. "Answer-block formatting" is the practice of writing those spans deliberately — so the chunk that retrieves for your target query is yours, not a competitor's. This post covers the patterns that consistently get lifted into ChatGPT, Perplexity, Gemini, and Google's AI Overviews, with before/after markup.
Why extractability beats prose
Long-form prose was rewarded by classic SEO because Google's ranking systems valued depth and dwell time. Generative engines invert the incentive. They need a quotable unit: typically 40–80 words that directly answer the prompt, can stand alone without surrounding context, and contain the entities the query is about.
If your answer to "what is llms.txt" is buried in paragraph four after a personal anecdote, the retriever will still find your page — but the chunk it surfaces will be the anecdote, not the definition. Worse, the model may merge your chunk with a cleaner one from another domain and cite the other domain.
Two practical implications:
- Lead with the answer. First sentence of every section should be the literal answer to the implicit question in the H2.
- Make the unit self-contained. Don't write "As mentioned above…" — the chunk that gets retrieved won't have "above."
Freshness amplifies this. Recent crawls of AI Overview citations indicate the overwhelming majority of cited URLs were published or updated within the last two years, with a significant share from the current year. That's a tailwind for anyone willing to reformat existing posts into extractable blocks.
The five answer-block patterns that get lifted
1. Lead-with-answer paragraph
Before:
There's been a lot of debate in the GEO community about whether llms.txt actually does anything. Some practitioners swear by it, others think it's snake oil. After running tests across 12 client sites over six months, my view is…
After:
llms.txt is a plain-text file at your domain root that tells AI crawlers which URLs to prioritize for training and retrieval. It is not an official standard, but Anthropic, Perplexity, and several smaller engines have indicated they read it. Below: when it helps, when it doesn't.
The "after" version is extractable as a single chunk. The "before" version forces the retriever to skip three sentences of throat-clearing.
2. Definition blocks
Use a bolded term followed by an em-dash and a one-sentence definition. Engines love this format because it mirrors dictionary structure, which is heavily represented in their training data.
**Answer block** — a self-contained passage of 40–80 words
that directly answers a specific query, formatted so a
retriever can extract it without surrounding context.
3. Tight Q&A
The FAQ pattern works only if questions are phrased the way users actually ask them and answers stay under three sentences. Long FAQ answers underperform short ones in retrieval tests because the model truncates.
4. Comparison tables
Tables are disproportionately cited when the query contains "vs", "best", "compare", or "difference between." A two-column table with 4–6 rows outperforms a 200-word paragraph covering the same ground.
| Format | Best for | Typical chunk size |
|---|---|---|
| Lead paragraph | Definitional queries | 40–80 words |
| Table | Comparison queries | Whole table |
| Tight Q&A | Long-tail questions | 1–3 sentences |
| Numbered list | Process / how-to | 1 sentence per step |
5. Numbered process lists
For "how to" queries, a numbered list with one verb-led sentence per step extracts more reliably than a flowing prose walkthrough. Keep each step independently meaningful.
Before/after: a real section rewrite
Before (prose, buries answer):
When clients ask me about schema markup, I usually start by explaining the history. JSON-LD became the recommended format around 2015, and since then there's been an explosion of types. For AI search specifically, the question of which schema to use is complicated by the fact that different engines weight different signals…
After (extractable):
For AI search, prioritize three schema types:
Articlewithauthor,FAQPage, andOrganizationwithsameAs. These three cover authorship signals, direct Q&A extraction, and entity disambiguation — the building blocks engines use to decide whether to cite you. Other types (HowTo, Product) help in narrow cases. See Google's structured data docs for full reference.
Same information. The second version puts the answer in the first 30 words, names the entities the query is about (Article, FAQPage, Organization), and links a canonical reference. It is also visually scannable, which matters because research from Princeton and Georgia Tech on Generative Engine Optimization found that citation-worthy content tends to be statistic-dense, source-linked, and structurally clear.
A 20-minute audit you can run today
- Open your five highest-priority pages.
- For each H2, read only the first sentence. Does it answer the question implied by the heading? If not, rewrite it.
- Find any paragraph longer than 80 words. Break it into a lead sentence + supporting sentences, or convert to a list/table.
- Add one definition block per page using the bolded-term-em-dash pattern.
- Check that your FAQ answers are under three sentences. Trim ruthlessly.
- Update the publish date if you've made substantive edits — freshness is a real ranking signal in AI retrieval, per Ahrefs' 2025 citation study.
This usually moves the needle within one to three crawl cycles. You don't need new content; you need extractable content.
FAQ
Does schema markup replace good formatting?
No. Schema helps engines parse entities and relationships, but the retrieval step still operates on visible text chunks. A clean H2 with a one-sentence answer beats elaborate JSON-LD wrapped around a wall of prose.
How long should an answer block be?
Aim for 40–80 words for definitional answers and one to three sentences for FAQ-style answers. Longer blocks get truncated mid-thought; shorter blocks lack the context the model needs to trust the source.
Should I add a TL;DR to every page?
Yes, if the page is over ~600 words. A TL;DR at the top gives engines a pre-summarized version of your argument, which is frequently the chunk that gets retrieved for broad queries. Keep it to three to five bullets.
