Passages that survive extraction

How should I structure a page so an assistant can lift an answer out of it?

Write in blocks that answer one question completely and survive being removed from the page. An assistant does not read your article; it retrieves a chunk of it, drops it into a prompt alongside chunks from four other sites, and generates prose. Anything in your chunk that depends on the rest of the page - a pronoun, a "as described above", a table with its explanation two scrolls away - arrives broken and loses to a competitor's chunk that arrived whole.

That is the whole principle. The rest of this chapter is what it implies.

What the experiment actually found

Structural advice in this field is mostly assertion. There is one large experiment, so start there. The GEO study applied nine content modifications to real source pages and measured the change in visibility inside generated answers. Three won, and they are strikingly unglamorous:

Two rows deserve emphasis because they contradict standard practice.

Tone does nothing. Writing more confidently, more authoritatively, more persuasively - the default instinct of every marketing rewrite - produced no significant improvement, and the authors conclude that generative engines are "already somewhat robust to such changes."[^geo-authoritative-null] The model is not persuaded by confidence. It is not the audience for your voice.

Substance does the work. The paper describes its winning methods as requiring minimal changes while significantly improving visibility, by "enhancing both the credibility and richness of the content."[^geo-minimal-changes] A number, a quotation, a source link. That is the intervention.

There is also a result for choosing which of the three to lean on: the study reports that the best-performing method varies by topic, with Statistics Addition strongest in law, government, debate and opinion topics and Quotation Addition strongest in people, society, explanation and history topics.[^geo-domain-variation] For a B2B software category the honest read is that both apply and neither has been measured specifically for you.

The four shapes that extract cleanly

Each of these exists because it produces a chunk that stands alone.

1. The definition block

Open every page and every major section with a direct answer in one or two sentences, before any context. Bold the term, define it, then elaborate.

Crawl-to-refer ratio is the number of pages an AI platform's crawlers fetch from a site for every visitor that platform sends back to it.

That block can be quoted verbatim with nothing lost. A version that begins "To understand this, we first need to consider..." cannot be quoted at all.

The heading above it should be the question a person would type. Not "Our approach to measurement" but "How do you measure AI visibility?" This is not a keyword tactic - the study says keywords do not help - it is that a heading phrased as a question matches a sub-query directly, and the text under it is scoped to answer exactly that.

2. The comparison table

Tables extract better than prose comparisons for a mechanical reason: each row is already a complete statement with its subject attached. "Perplexity-User ignores robots.txt for user-initiated fetches" survives extraction. "It ignores them, unlike the previous one" does not.

Rules that make a table citable: every cell says what it is about, no cell relies on the row above, and no cell contains only a checkmark. A grid of green ticks carries no text for the model to use, so it contributes nothing to an answer even when your page is the source.

3. The question-and-answer block

An FAQ is the extraction-friendly shape in its purest form: an explicit question, followed by a complete answer, repeated. The value is not in the schema markup around it - it is that the shape matches what retrieval is looking for.

Two conditions. The questions must be real ones people ask you, taken from support tickets, sales calls and search suggestions rather than invented to fill a section. And each answer has to be complete: 40 to 80 words, self-contained, no "see above".

4. The sourced statistic

The single highest-return edit available, and the one with the most obvious failure mode. Every number needs a source URL and a date, and the number has to actually appear in that source. Fabricated statistics are the fastest way to be useless to a model that can check, and they compound: once your page is cited for a number that is wrong, you have published misinformation under your own byline.

If you cannot source it, write the sentence qualitatively. "Most teams have never checked their crawler logs" is honest. The same sentence with a confident, unsourced percentage in front of it is a liability wearing a lab coat - and it is the exact failure this guide's own content pipeline blocks in continuous integration, because a percentage in a paragraph with no matching sourced entry fails the build rather than the review.

What breaks extraction

The failure modes are the mirror image of the shapes above, and every one of them is common in otherwise well-written content.

• Cataphoric openings. "There are three reasons for this." Which this? The

chunk does not include your previous section.

• Narrative build-up. The classic feature-article opening - an anecdote,

then the point four paragraphs later - buries the answer below the retrieval window and offers the model nothing quotable up front.

• Answers split across a heading boundary. Half the definition under one

heading, the qualifier under the next. Each chunk is now wrong on its own.

• Content that only exists after JavaScript runs. Google's own guidance for AI

features asks site owners to make sure important content is available in textual form.[^google-textual-form] An answer inside a client-rendered accordion is an answer no fetcher sees.

• Images carrying the argument. A diagram or a screenshot of a table is

invisible to text extraction. Put the same content in prose or a real table and keep the image as reinforcement.

• Gated or paywalled passages. A model cannot cite what it cannot read.
• Undated pages. Nothing on the page says when it was true, which makes it a

weak source for anything time-sensitive - most of this category.

A revision pass you can actually run

On any page you want cited, in this order:

1. Read only the first 60 words under each heading. If they do not answer the

heading, rewrite them until they do.

1. Search the page for "this", "that", "these" and "it" as sentence openers.

Replace each with the noun. This one edit does more for extraction than any schema change.

1. Find every qualitative claim that could be a number, and either source a

number or leave the claim qualitative. Never invent the middle option.

1. Add the source link inline where a claim came from somewhere. The study's

Cite Sources method is exactly this and it was a top performer.

1. Convert the longest comparison paragraph into a table, with subjects in

every cell.

1. Add a visible last-updated date and make it true.

Six passes, no rewrite. The study's framing is worth repeating here: these are minimal changes that significantly improve visibility.[^geo-minimal-changes] The temptation is always to rebuild the page. The evidence says to add substance to the one you have.

The chunk boundary is a real constraint

Retrieval systems split documents before they embed them, and you do not control where the split lands. You do control how much damage it does.

A page written as a sequence of complete units degrades gracefully: whichever way it is cut, most chunks still say something true and whole. A page written as one continuous argument degrades badly: cut anywhere, and both halves are partial. This is the practical reason for short sections with descriptive headings, and it is also why front-loading matters. If the answer is in the first sentence under a heading, it survives almost any boundary. If it is in the last sentence of a long section, it survives only the cut that happens to include the whole section.

The same logic argues against the two structural habits marketing sites lean on hardest: the long scrolling landing page where every claim depends on the visual context around it, and the "ultimate guide" that answers forty questions in one undifferentiated flow. Both read fine to a human scrolling and both extract badly, which is a reasonable summary of why plenty of well-made sites are invisible in AI answers while a plain documentation page outranks them.

Where the position weighting bites

One subtlety about how visibility is scored. The study's primary metric weights a source by how much of the answer it contributed and by how early that contribution appears, decaying exponentially with position.[^geo-position-weighting] Being the fourth source in a long answer is worth much less than being the first.

Practically, that argues for depth on the narrow question rather than breadth on the broad one. If you are the most complete source on one specific sub-question, you are likely to be quoted early in the answer to it. If you are the seventh adequate source on a broad question, you are a footnote. Pick the questions where you can be the best answer that exists, and be visibly, checkably that.

What to take from this chapter

Write blocks that survive being cut out. Lead with the answer. Put a real number in it, with a link to where the number came from. Say when you checked. Everything else in this chapter is a special case of those four sentences - and the one large experiment that has been run on the question says the same thing, in a different vocabulary, with 10,000 queries behind it.