**TL;DR** — Across 31 client sites through May 2026 we audited a structural choice that lives in the shape of the answer sentence: whether the passage that answers the query is written as a **short, single-claim sentence** that states one fact and stops ("Compression cuts page weight by about 40%.") or as a **long, multi-clause sentence** that bundles the fact with its caveats, mechanism, and context ("Compression, which works by removing redundant data before transfer, can cut page weight by about 40%, though the exact figure depends on your assets and existing optimisation."), and whether sentence length changes how often the AI Overview lifts that sentence into the card. Across 7,810 cited-passage events we joined each cited sentence to its word count and clause count. The headline is that short answer sentences are a real and large citation lever, but the lever is really an extractability-and-self-containment lever wearing a word-count costume. A short single-claim sentence was cited 2.6× more often than a matched long multi-clause sentence carrying the same core fact. The strongest predictor was clause count — a sentence with one main clause was lifted far more than one that subordinated the fact under two or three dependent clauses. The second was front-loading — when a long sentence had to exist, putting the answer in the first clause recovered most of the citation it would otherwise lose. The third, and the warning, was over-fragmentation — chopping every sentence to under eight words stripped the context that made the claim usable, and on 5% of pages produced answer sentences so bare the composer paired them with a competitor's sentence for the missing qualifier. One change — splitting long multi-clause answer sentences into a short claim sentence followed by a separate context sentence — lifted cited-passage rate by 24% on the affected sites over a 30-day follow-up.
Why we ran this audit
The AI Overview composer lifts a sentence and drops it into a card as the answer to a query. The sentence has to survive being read alone, and a long sentence that bundles the fact with three subordinate clauses asks the composer to extract a claim from inside a structure built for a reader who has the whole paragraph. A human reads the long sentence top to bottom and assembles the meaning; the composer has to decide, fast, whether this one sentence is a clean answer it can lift verbatim, and a sentence whose main claim is buried in clause three is a worse candidate than one whose claim is the whole sentence. We suspected the composer was quietly preferring short, single-claim sentences because they extract cleanly, and we wanted to know whether the preference was about length as such or about how many claims the sentence asked the composer to untangle.
The second motivation was a writing habit we kept seeing on expert pages. Skilled writers pack — they fold the fact, its mechanism, its exception, and its scope into one dense, well-built sentence, because a single sentence that carries everything reads as authoritative and avoids choppiness. A human editor scores the packed sentence as excellent prose; the composer, hunting for one liftable claim, finds the claim entangled with three other things and passes for a competitor whose equivalent sentence says one thing. We needed to know whether the density cost the citation, because if it did, the fix is nearly free and improves readability too — split the packed sentence into a short claim and a separate sentence for the context, instead of bundling.
How we ran the measurement
31 client sites — 12 SaaS, 6 publisher, 8 B2B services, 5 DTC — each with a fixed 200-query basket of its real in-market queries, weighted toward factual and explanatory queries where the answer is a single extractable claim. Twice daily through May 2026 we captured every AI Overview card, and for cards citing a client page we identified the specific lifted sentence and recorded its length in words and its number of clauses: short single-claim (one main clause, roughly under 20 words), medium, or long multi-clause (two or more subordinate clauses). For each cited sentence we built a matched control: a comparable sentence on a similar query whose length differed but whose core fact was the same, so the comparison was short-vs-long rather than good-page-vs-bad-page. The cited cohort was 7,810 events.
Two normalisation moves matter. We scored length on the sentence as it would be lifted — alone, with no surrounding context — because that is the unit the composer extracts, and a clause that reads as helpful elaboration in the paragraph reads as noise around the claim in the card. And we matched on sentence citability before comparing length — 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 length is not just the short-sentence pages being better written overall. The 2.6× and 2.3× figures are from those matched comparisons, not raw averages.
The shape of the length pattern
The flat headline first. Short answer sentences are cited more. A short single-claim sentence was lifted 2.6× more often than a matched long multi-clause sentence carrying the same core fact. The effect held through the quality match and the citability control: among sentences our rubric scored as equally liftable, the short ones were lifted far more than the long ones. The composer behaves as though it prefers a sentence it can lift whole over one it would have to trim to a usable claim.
The most decision-relevant cut was that this is about clause count, not raw word count. We tested whether the win was driven by length in words or by the number of claims the sentence bundled, and it was the latter: a long sentence that stayed one main clause ("Compression removes redundant data before transfer to cut page weight by about 40%") was cited nearly as well as a short one, while a short sentence that still crammed two claims into subordinate structure was cited as poorly as a long one. Shortening is a reliable way to drop to one claim, but it is the single-claim shape the composer rewards. Cut length because it is the cheapest way to leave one extractable claim, not because the model is counting words.
Driver one: leave one claim per answer sentence
The single strongest predictor was how many claims the answer sentence asked the composer to separate. Holding the core fact constant, a one-main-clause version was lifted at 2.6× the rate of a version that subordinated the fact under two or three dependent clauses. The composer extracts a sentence and reads it as the answer; a single-claim sentence is the answer, a multi-claim sentence is the answer plus material the composer must strip before the claim is usable, and a composer choosing between candidates prefers the one that needs no surgery. A human reader does the untangling for free and never notices; the composer treats the entanglement as a cost.
We ran a structural test on 27 answer sentences across 14 clients, each a dense multi-clause sentence that bundled the fact with mechanism, exception, and scope. We split each into a short claim sentence followed by a separate context sentence, changing no facts — only unbundling. Over the 45 days that followed, 19 of the 27 claim sentences began being lifted on at least one target query where the bundled version had been skipped. The lever was not new content; it was leaving the answer as one clean claim the composer could lift whole, and moving the mechanism and caveats to the sentence after it.
Driver two: front-load the claim when the sentence must stay long
Holding clause count roughly constant, the second driver was where the claim sat. Some sentences genuinely cannot be split — a conditional fact whose condition is inseparable from the claim — and for those the position of the main clause mattered. A long sentence that opened with its claim and trailed the qualifiers ("Page weight drops about 40% with compression, depending on your asset mix") was cited far more than the same content with the claim last ("Depending on your asset mix and existing optimisation, page weight drops about 40% with compression"). The reading consistent with the data is that the composer weights the opening of the sentence most heavily, so a claim in the first clause survives extraction even when the sentence as a whole is long.
We ran a structural test on 18 answer sentences across 11 clients that could not be cleanly split and led with their qualifiers. We rewrote each to open with the claim and trail the conditions, changing no facts. Over the 60 days after the change, 13 of the 18 sentences improved their cited-passage rate. The two drivers compound: a long sentence that both buries the claim in clause three and leads with qualifiers is the worst case, and a short single-claim sentence is the best — when length is unavoidable, front-loading recovers most of what the length would otherwise cost.
Driver three: over-fragmentation, and the sentence too bare to stand alone
The third driver was the warning. Shortness is a tool, not a rule, and chopping every answer sentence to the bone backfires when the cut strips the context that made the claim usable. A sentence like "Compression cuts page weight by about 40% on text-heavy pages" became, under a blind shortening pass, "Compression cuts page weight 40%" — and the bare version lost the scope that made the figure trustworthy. These over-fragmented sentences were cited no more often than the well-formed short version, and on 5% of audited pages the composer paired the bare sentence with a competitor's sentence that supplied the missing qualifier, so the citation was shared rather than won outright. The reading consistent with the data is that the composer rewards a sentence that is one complete claim, not one stripped of the scope it needs to be true; a single-claim sentence still has to carry the qualifier that makes the claim hold, and cutting that qualifier to hit a word count manufactures a sentence that is short but incomplete.
We confirmed this on 15 sentences across 9 clients where an earlier optimisation pass had shortened everything indiscriminately. We restored the scope-defining qualifier to the claim sentences that had lost it, keeping the unbundling but not the over-cut. Over the following 45 days the restored sentences regained their solo citation while reading completely, and none drew a shared-citation pairing. The actionable rule is blunt: split the bundled sentence into a short claim, but the claim sentence must still be a complete, true statement — keep the one qualifier that defines its scope rather than cutting to the shortest possible string.
What changed in our content checklist
Three changes. We added a clause-count pass: before publishing, we read each section's lead answer sentence and count its claims, and a sentence that subordinates the fact under two or three clauses gets split into a short claim sentence and a separate context sentence — because the composer lifts a sentence whole and a single-claim sentence needs no trimming. We added a front-load check to the same pass: when a sentence genuinely cannot be split, the claim moves to the first clause and the qualifiers trail it, because the composer weights the opening most. And we added a completeness guard: a short claim sentence must still be true on its own, so the qualifier that defines its scope stays even when it costs a few words.
We dropped one habit. For years our most expert writers had packed — folding fact, mechanism, exception, and scope into one dense, well-built sentence, because density reads as command of the material and avoids choppiness. The audit removes that default for answer sentences: the one sentence the composer would lift is the one that should carry a single claim, and a packed sentence spends the citation to buy prose elegance the composer cannot read. So packed answer sentences left our playbook — we now write the answer as one short claim and let the sentence after it carry the mechanism and the caveats, and accept that the cited sentence reads plainer than our writers would choose because it is built to be lifted alone.
- 01Leave one claim per answer sentence. A short single-claim sentence was cited 2.6× more than a matched long multi-clause one carrying the same fact — the composer lifts a sentence whole and prefers one that needs no trimming.
- 02Front-load the claim when the sentence must stay long. The composer weights the opening most; a claim in the first clause survives extraction, while the same content with the claim last was cited far less.
- 03Do not over-fragment. A claim sentence still has to be true alone; blindly cutting the scope-defining qualifier left bare sentences cited no more, and on 5% of pages the composer shared the citation with a competitor that supplied the missing context.
- 04Split, do not cram. 19 of 27 bundled sentences were lifted after the fact was left as one clean claim and the mechanism and caveats moved to the sentence after it.
Where this argument breaks
For conditional facts whose condition is inseparable from the claim — an effect that only holds under a stated qualifier, a figure that is false without its scope — the sentence cannot be split and the lever becomes front-loading rather than shortening, so length is unavoidable and position is the move. For navigational and brand queries there is no answer sentence whose length matters. For narrative and persuasive passages — case studies, opinion, story-driven content — sentence length is a rhythm and craft choice, not a citation lever, and the clause-count pass is for the answer sentences only. For some languages the effect may differ — in our parallel Chinese-language audit (文心一言, 元宝, 通义) the single-claim win was present but the word-count threshold was different, since Chinese packs more meaning per character and a clause that would be long in English stays compact, so we counted clauses rather than words there. The 5% shared-citation figure is small and noisy; we are confident an over-cut sentence does not help and mildly confident it splits the citation, 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 31 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 a short single-claim sentence far more readily than a long multi-clause one, the unit is the individual answer sentence rather than the page, and the cheapest citation win on a factual query is to leave the answer as one clean claim the composer can lift whole — and move the mechanism and the caveats to the sentence after it.