An answer engine does not rank your page. It splits your page into chunks, embeds them, retrieves the handful that look relevant to the question in front of it, and then decides which of those it is willing to attribute. The page is the container. The chunk is the unit that competes.

That distinction sounds academic until you watch it happen to a page you wrote. We audited a client’s pricing methodology page last spring — 2,400 words, ranked third for its head term, cited by nobody. The reason was visible the moment we chunked it the way a retriever would: every specific number in the piece lived three paragraphs away from the sentence that explained it. Individually, no chunk said anything a model could stand behind. The page was excellent and unquotable.

The retrieval unit is a paragraph, not a page

Most production retrieval pipelines chunk on something between 200 and 500 tokens, with a small overlap, and they respect structural boundaries when the HTML gives them any. That is roughly one to three paragraphs of normal prose, or one section under an h3.

Two consequences follow, and both of them are unintuitive if your instincts were formed by classical SEO:

  • Context does not carry. A chunk arrives at the model with no memory of your introduction. If paragraph fourteen says “this approach cuts crawl waste by about a third”, the retriever has no idea what this approach refers to and the model has nothing safe to cite.
  • Coverage beats length. A 3,000-word page that states six claims well produces six retrievable units. A 3,000-word page that states one claim and then circles it produces one — and it is competing against every other page that states the same thing more plainly.

The practical test is brutal and takes ten seconds: take any paragraph out of your page, paste it into a blank document, and ask whether a stranger could quote it without needing anything else on screen. If they could not, no retriever will either.

The anatomy of an evidence block

We structure the load-bearing paragraphs of every page we build as an evidence block: a claim, the data behind it, and the conditions under which it holds. Three parts, in that order, physically adjacent.

The third part is the one everybody skips, and it is the one that does the work. A claim with a stated boundary is safer to quote than a claim without one, because the model can reproduce the hedge along with the assertion. An unqualified superlative gives it nothing to hide behind, so it either drops your sentence or paraphrases it into mush and cites somebody else.

The boundary condition is also the sentence your competitors will not write. It is the single cheapest way to sound like a practitioner rather than a content brief.

Five habits that make a passage unretrievable

These are, in order, the things we find most often in audits.

  1. Pronoun openers. “It also means…”, “This is why…”, “That said…”. Every one of these severs the chunk from its antecedent. Open with the noun.
  2. Numbers orphaned from their claim. A statistic in one paragraph and its interpretation in the next produces two half-chunks, neither citable.
  3. Headings that are teasers. “The hard part” tells a retriever nothing. “Why hreflang clusters break when one URL 404s” tells it everything. Headings are frequently prepended to their chunk as metadata; a clever heading is a wasted signal.
  4. Definitions in a glossary at the end. If a term needs defining to make a passage stand alone, define it in the passage. Duplication across chunks is not a penalty; it is how retrieval works.
  5. The narrative build. The classic structure — set up tension for 400 words, pay it off in the conclusion — produces a page whose only quotable paragraph is the last one. Front-load the assertion in each section, then support it.

Retrofitting a page you already have

You do not need to rewrite the library. Pick the ten pages that already earn non-branded impressions and do this to each:

  • Chunk it yourself. Split at every h2 and h3, then at every 300 words inside a long section. Read each fragment cold.
  • For every fragment that fails the stranger test, do the minimum: replace the opening pronoun, pull the relevant number up, and add one sentence naming the condition under which the claim holds.
  • Give every claim a heading it can inherit. If a section makes two distinct claims, it is two sections.
  • Delete the throat-clearing paragraph at the top of each section. It is always there, it is always the first chunk retrieved, and it always says nothing.

Budget an hour a page. This is not a content project; it is an editing pass, and it is the highest-return editing pass available on an established site.

What to measure, and what not to

The honest answer is that citation measurement is immature. Answer engines do not publish impression data, sampling is expensive, and the same prompt returns different sources on different days. Anyone selling you a clean citation-share dashboard is selling you a sample of forty prompts with a confidence interval they have not shown you.

What we track instead, in rough order of usefulness:

  • Prompt-set citation frequency. A fixed set of 60–120 buyer-intent prompts, run monthly against the major engines, scored for whether the client appears and in what role. It is a sample, we say so, and we watch the trend rather than the level.
  • Passage-level entrances. Classical analytics, filtered to landings on URLs with a fragment. A rise means people are arriving at a section, which is the same structural property retrieval rewards.
  • Referral traffic from assistant surfaces. Small, noisy, and directionally honest, which is more than most GEO metrics manage.

What we do not track is a “GEO score”. There is no such number, and inventing one would be exactly the kind of unfalsifiable claim this whole structure exists to avoid.

The uncomfortable part

Writing this way is less enjoyable. You give away the conclusion in the first sentence of every section, you repeat definitions you have already written once, and you spend your best sentences on caveats. It reads flatter than the essay you wanted to write.

It is also the only structure that survives being cut into pieces by a machine that will never see your introduction — which, increasingly, is the only way your writing gets read at all.