**TL;DR** — Across 30 client sites through July 2026 we audited a structural choice that lives in whether the answer sentence can survive being torn out of its paragraph: whether the passage that answers a query is written as an **anaphora-free answer sentence** that names its own subject outright ("Prerendering cuts load time by 40%.") or as a **pronoun-dependent sentence** that leans on a reference pointing back to an earlier sentence ("It cuts load time by 40%."), and whether resolving the pronoun to the noun it stands for changes how often the AI Overview lifts that sentence into the card. Across 7,720 cited-passage events we joined each cited sentence to whether its subject was a named noun or a reference whose antecedent sat outside the sentence. The headline is that anaphora-freedom is a real citation lever, and it is really a survives-extraction lever wearing a grammar costume. A sentence that named its subject was cited 1.9× more often than a matched pronoun-dependent sentence making the same claim on the same query. The strongest predictor was referent-in-clause — a sentence whose subject was resolvable without reading the sentence before it was lifted far more than one whose subject was an "it" or a "they". The second was demonstrative scope — a "this" or "that" pointing at a whole preceding clause ("This makes pages faster.") was passed over far harder than a pronoun with a single nearby noun antecedent, because a clause-scoped demonstrative loses more when the paragraph goes. The third, and the warning, was over-nouning — repeating the full entity name in every clause ("Prerendering cuts load time because prerendering fetches the prerendering payload early") was cited no more than a naturally-worded sentence, and on 6% of pages the composer took a competitor's cleaner sentence instead. One change — resolving the subject pronoun in the lead answer sentence of each section to the noun it stands for — lifted cited-passage rate by 19% on the affected sites over a 30-day follow-up.
Why we ran this audit
The AI Overview composer lifts a single sentence and drops it into a card as the answer to a query, and the lift is a tearing — the sentence leaves the paragraph that raised it and arrives in front of a reader who never saw the sentences above. Most prose is not built to survive that. Written well, a paragraph names its subject once and then refers back to it: "Prerendering fetches the next page before the click. It cuts load time by 40%. That matters most on mobile." Every sentence after the first is load-bearing only in company; lifted alone, "It cuts load time by 40%" is a claim with a hole where its subject should be — a 40% cut to what? We had spent weeks on the shape of the answer sentence — its polarity, its condition, its rank, whether it dated itself — and anaphora is the natural next structural variable, because it is the one dimension where the sentence does not merely answer a smaller question but stops being a complete statement at all.
The second motivation was that the drafting habit here is not sloppiness — it is good writing. A writer who names "prerendering" in all four sentences of a paragraph produces prose that reads like a form letter, and every style guide, ours included, tells them to pronominalise after the first mention. The habit is correct for a human reading top-to-bottom, who carries the subject forward from sentence to sentence for free. The composer carries nothing forward; it takes the sentence and leaves the antecedent on the page. So the question was whether the cost of a slightly repetitive-reading sentence bought the citation, because if it did, the fix is nearly free and narrowly scoped — resolve the pronoun in the one sentence per section that is actually a candidate for lifting, and leave the rest of the paragraph to read like English.
How we ran the measurement
30 client sites — 11 SaaS, 6 publisher, 8 B2B services, 5 DTC — each with a fixed 200-query basket of its real in-market queries. Twice daily through July 2026 we captured every AI Overview card, and for cards citing a client page we identified the specific lifted sentence and classified its reference structure: anaphora-free (every referring expression in the sentence resolves inside the sentence), pronoun-dependent (the subject or a key object is a personal pronoun — "it", "they", "these" — whose antecedent sits in an earlier sentence), or demonstrative-dependent (a "this" or "that" whose antecedent is a whole preceding clause rather than a noun). For each cited sentence we built a matched control: a comparable sentence on a similar query whose reference structure differed but whose underlying claim was the same, so the comparison was resolved-vs-unresolved rather than good-page-vs-bad-page. The cited cohort was 7,720 events.
Two normalisation moves matter. We scored structure on the sentence as it would be lifted — alone, with no surrounding context — because that is the unit the composer extracts, and a pronoun that is completely unambiguous in the paragraph is a dangling reference in the card. And we matched on sentence citability before comparing structure — we paired each cited sentence with a control our existing cited-paragraph rubric scored as equally liftable (concrete, on the query, factually complete), so the effect we attribute to anaphora is not just the resolved pages being better edited overall. The 1.9× and 2.6× figures are from those matched comparisons, not raw averages.
The shape of the anaphora pattern
The flat headline first. Sentences that named their subject were cited more. An anaphora-free sentence was lifted 1.9× more often than a matched pronoun-dependent sentence making the same claim on the same query, and the gap widened to 2.6× where the antecedent sat two or more sentences back. The effect held through the quality match and the citability control: among sentences our rubric scored as equally liftable, the resolved ones were lifted far more than the unresolved ones. The composer behaves as though it prefers a sentence that can stand alone, which is unsurprising once you notice that standing alone is the entire job of the sentence it is choosing.
The most decision-relevant cut was that this is about the sentence being self-contained, not about pronouns being bad. We tested whether the win came from pronouns being absent or from every reference resolving inside the sentence, and the second was the whole story: a sentence with a pronoun whose antecedent sat inside the same sentence ("Prerendering works because it fetches the page early") was cited as well as a fully nouned one, while a sentence whose pronoun pointed outward was passed over. The composer is not allergic to "it". It is unwilling to lift a sentence with a hole in it. Resolve the references that point off the end of the sentence, and leave the ones that point inward alone.
Driver one: name the subject the claim is about
The single strongest predictor was whether the subject of the claim was named in the sentence. Holding the claim constant, a sentence whose subject was a noun was lifted at 1.9× the rate of one whose subject was a pronoun. The composer extracts a sentence and puts it in front of a user who did not read the page; "It cuts load time by 40%" is not a worse answer than "Prerendering cuts load time by 40%" so much as it is not an answer, because the thing the sentence is about was left behind in the previous line. A human reading in order supplies the subject without noticing they are doing it; the reader in front of the card has nothing to supply it from, and the composer appears to model that reader rather than the one who read the paragraph.
We ran a structural test on 31 answer sentences across 17 clients, each a section's lead answer sentence whose subject was a pronoun pointing back at a heading or an earlier line. We rewrote each to name its subject, changing no claims and touching no other sentence in the paragraph — only replacing the pronoun with the noun it already stood for. Over the 45 days that followed, 23 of the 31 sentences began being lifted on at least one query where the pronoun version had been skipped. The lever was not new content; it was one word per section, and the content that won was content that had always been there and had simply been unliftable.
Driver two: the demonstrative that points at a paragraph, not a noun
Holding subject-naming constant, the second driver was what the reference pointed at. A demonstrative whose antecedent was an entire preceding clause — "This makes pages faster", "That is why the check fails" — was passed over 2.6× harder than a personal pronoun with a single noun antecedent one sentence back. The reading consistent with the data is that the two failures are different sizes. A dangling "it" is missing a noun, and a composer working from the query and the page title can sometimes guess it; a dangling "this" is missing an idea that took a full clause to state, and no amount of surrounding signal reconstructs it. These sentences are also the ones writers produce most freely, because a clause-scoped "this" is the cheapest way to chain an argument forward, and it is exactly the chain the composer cannot lift a link out of.
We ran a structural test on 19 answer sentences across 12 clients, each opening with a clause-scoped demonstrative on a query our logs showed it was being passed over on. We rewrote each to restate the antecedent as a noun phrase inside the sentence — "This makes pages faster" became "Fetching the next page before the click makes pages faster" — accepting a longer sentence that reads slightly redundant to anyone who just read the line above. Over the 60 days after the change, 14 of the 19 improved their cited-passage rate. The two drivers compound: naming the subject is one half and restating what the demonstrative pointed at is the other — the sentences that won were the ones that owed the paragraph nothing.
Driver three: over-nouning, and the sentence that reads like a robot
The third driver was the warning. Resolution helps up to the point the sentence still reads like a person wrote it, and pushing it further backfires. A sentence that repeated the full entity name in every clause — "Prerendering cuts load time because prerendering fetches the prerendering payload before the click" — was cited no more often than a naturally-worded resolved sentence ("Prerendering cuts load time because it fetches the page before the click"), and on 6% of audited pages the composer passed over the over-nouned sentence for a competitor's cleaner one making the same claim. The reading consistent with the data is that the composer is choosing a sentence to show a human, and a sentence that reads like keyword stuffing is a bad thing to show a human — the goal was never to delete pronouns, it was to close the holes, and once a sentence has no holes, further nouning only costs it.
We confirmed this on 15 sentences across 8 clients where an earlier optimisation pass had stripped every pronoun on principle. We rewrote each back to resolve only the references that pointed outside the sentence, restoring the ones that pointed within it, changing no claims. Over the following 45 days the naturally-worded versions held or improved their citation rate while reading like prose again, and none drew the competitor-substitution we saw in the over-nouned cohort. The actionable rule is blunt: resolve what dangles, keep what resolves inside, and stop — a sentence that repeats its own subject three times reads as machine output to the composer as readily as it does to you.
What changed in our content checklist
Three changes. We added a standalone read to the publishing pass: we take each section's lead answer sentence, read it with the rest of the page covered, and check that nothing in it points at something we cannot see — because that is literally the state the composer will read it in, and a sentence that fails this read is a sentence that cannot be lifted no matter how good the page around it is. We added a demonstrative check to the same pass: any "this" or "that" opening an answer sentence gets its antecedent restated as a noun phrase, because a clause-scoped reference loses more in the tearing than a noun-scoped one. And we added an over-nouning guard: once a sentence resolves, we stop resolving, so a within-sentence pronoun stays a pronoun and the sentence still reads like English.
We dropped one habit, narrowly. Our style guide had a flat rule — name the subject once, pronominalise after — and it is a good rule that makes paragraphs readable, so we kept it everywhere except one sentence per section. The audit carves out the lead answer sentence: that one is written to be torn out, so it repeats a subject the reader three lines up already has, and it reads very slightly redundant in place. That is the trade, and it is worth naming honestly rather than pretending the resolved sentence is better prose. It is not better prose. It is prose that still means something after the paragraph around it is gone, which is a different property, and on the one sentence the composer might lift it is the property that pays.
- 01Name the subject in the lead answer sentence. An anaphora-free sentence was cited 1.9× more than a matched pronoun-dependent one — the composer lifts the sentence and leaves the antecedent on the page.
- 02Restate what a demonstrative points at. A clause-scoped "this" was passed over 2.6× harder than a noun-scoped pronoun — a missing idea is unreconstructable in a way a missing noun is not.
- 03Resolve outward references only. A pronoun whose antecedent sits inside the same sentence cited as well as a fully nouned one — the lever is self-containment, not pronoun deletion.
- 04Do not over-noun. Repeating the entity in every clause was cited no more, and on 6% of pages the composer took a competitor's cleaner sentence making the same claim.
Where this argument breaks
For sentences that were never citation candidates — the connective tissue of an argument, the second and third sentences of a paragraph — the standalone read is wasted effort and the style guide should win, so this is a lever for the lead answer sentence of each section and nothing else. For navigational and brand queries there is no answer sentence whose structure matters. For narrative and persuasive passages — case studies, opinion, story-driven content — anaphoric chaining is how prose builds momentum, and stripping it to make sentences extractable would trade a good essay for a quotable one. For some languages the effect may differ, and here the carve-out is large rather than cosmetic: Chinese drops pronouns freely, so a zero-subject answer sentence is ordinary rather than marked, and in our parallel Chinese-language audit (文心一言, 元宝, 通义) the resolution win was present but the baseline was worse across the board, since the unresolved version is far more common and the composers appear to tolerate it more. The 6% competitor-substitution figure is small and noisy; we are confident over-nouning does not help and mildly confident it loses the citation outright, but it is the weakest finding here and we would not restructure a page on it alone. Our window was 60 days and the cohort was 30 sites; the multipliers are point estimates that will move by vertical and query type. Outside those carve-outs the lesson holds: in 2026 the AI Overview lifts an answer sentence that names what it is about — subject resolved, demonstratives restated, inward pronouns left alone — more readily than one that borrows its subject from the line above, the unit is the individual answer sentence rather than the page, and the cheapest citation win available is to read your one candidate sentence with the paragraph covered and give back whatever the covering took away.