**TL;DR** — Across 29 client sites through May 2026 we audited a certainty choice that lives in the modal words of the answer sentence: whether the passage that answers the query is written as a **direct claim** that states the fact plainly ("Compression reduces page weight by roughly 40%", "Schema markup helps search engines read your content") or as a **hedged claim** that wraps the fact in qualifiers ("Compression may reduce page weight by up to 40% in some cases", "Schema markup can sometimes help search engines read your content"), and whether hedging changes how often the AI Overview lifts that sentence into the card. Across 7,540 cited-passage events we joined each cited sentence to a hedge score counting its modal qualifiers ("may", "can", "might", "often", "typically", "in some cases", "up to"). The headline is that direct answer sentences are a real and underused citation lever, but the lever is really a commitment-and-precision lever wearing a grammar costume. A direct sentence was cited 2.4× more often than a matched hedged sentence stating the same fact. The strongest predictor was claim commitment — a sentence that asserted the fact was lifted far more than one that buried it under stacked qualifiers. The second was hedge stacking — one honest qualifier cost almost nothing, but two or three compounded the penalty sharply. The third, and the warning, was false confidence — stripping every hedge from a genuinely uncertain claim produced overconfident prose that was cited no more, and on 5% of pages drew a visible accuracy correction from the composer, because the hedge had been carrying real epistemic information. One change — rewriting over-hedged answer sentences into direct claims, keeping only the qualifiers that carried real uncertainty — lifted cited-passage rate by 23% on the affected sites over a 30-day follow-up.
Why we ran this audit
The AI Overview composer lifts a sentence and reads it as an answer. A query is a request for a fact, and a sentence that wraps the fact in "may", "can", "in some cases", "depending on a number of factors" answers the request while declining to commit to it. A human reader treats the hedge as ordinary professional caution and reads straight through to the fact underneath. The composer, extracting that sentence alone into a card next to three competitors, reads the hedge literally: a sentence that says the fact "may" be true is a weaker answer than one that says it is true, and when a direct competitor sentence states the same fact plainly, the hedged one is the worse card. We suspected the composer was quietly preferring committed answer sentences because they answer the query outright, while hedged sentences answer it with a shrug, and we wanted to know whether that preference was real or whether the model discounts hedging the way a human reader does.
The second motivation was a failure mode we kept seeing on cautious, well-edited pages. Writers who fear being wrong — legal-adjacent, medical-adjacent, technical pages reviewed by careful editors — hedge reflexively, so "X reduces Y" becomes "X may, in certain configurations, help to reduce Y", and the page fills with relevant, accurate, fully-hedged answer sentences. A human grader scores them as correct and appropriately careful; the composer, hunting for one liftable sentence that commits to an answer, finds the answer sentences non-committal and passes for a competitor whose equivalent sentence states the fact. We needed to know whether the cost was the hedging, because if it is, the fix is almost free — state the fact and reserve the qualifier for the cases that genuinely need it, instead of hedging by reflex.
How we ran the measurement
29 client sites — 11 SaaS, 6 publisher, 7 B2B services, 5 DTC — each with a fixed 200-query basket of its real in-market queries, weighted toward the factual and explanatory queries ("does X do Y", "how much does X change Y", "what causes Y") where the answer sentence states a fact that could be committed to or hedged. 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 scored its hedging: direct (the fact asserted with no qualifier), single-hedge (one modal or qualifier), or multi-hedge (two or more stacked qualifiers). For each cited sentence we built a matched control: a comparable sentence on a similar query whose hedge level differed, so the comparison was direct-vs-hedged rather than good-page-vs-bad-page. The cited cohort was 7,540 events.
Two normalisation moves matter. We scored hedging on the sentence as it would be lifted — alone, with no surrounding context — because that is the unit the composer extracts, and a hedge that reads as reasonable caution in the paragraph reads as non-commitment in the card. And we matched on sentence citability before comparing hedge level — we paired each cited sentence with a control our existing cited-paragraph rubric scored as equally liftable (concrete, right length, directly on the query), so the effect we attribute to directness is not just the direct-claim pages being better written overall. The 2.4× and 2.2× figures are from those matched comparisons, not raw averages.
The shape of the hedging pattern
The flat headline first. Direct sentences are cited more. A sentence that stated its fact plainly was lifted 2.4× more often than a matched sentence that wrapped the same fact in qualifiers. The effect held through the quality match and the citability control: among sentences our rubric scored as equally liftable, the direct ones were lifted far more than the hedged ones. The composer behaves as though it prefers a sentence that answers the query outright over one that answers it while declining to commit.
The most decision-relevant cut was that this is about commitment, not modal words as such. We tested whether the win was specifically about removing hedges or more broadly about the sentence committing to its claim, and it was the latter: a sentence that kept one qualifier but still committed to a definite fact ("Compression typically reduces page weight by about 40%") was cited nearly as well as the fully direct version, while a sentence with no modal word that still refused to commit ("the effect depends on many factors") was cited as poorly as a multi-hedge. Removing hedges is a reliable way to make a sentence commit, but it is the commitment the composer rewards. Cut qualifiers because they are the cheapest way to make the sentence answer the query, not because the model is blacklisting the word "may".
Driver one: commit to the claim
The single strongest predictor was whether the answer sentence committed to its fact. Holding the sentence constant, a version that asserted the claim was lifted at 2.4× the rate of a version that hedged it into a maybe. The composer extracts a sentence and reads it as the answer to a question; a committed sentence is a usable answer, a hedged one is a usable answer with a disclaimer that, read literally and alone, says the answer might not hold. A human reader never reads the hedge literally, because the surrounding page makes clear how confident the claim really is — but the composer always reads the sentence cold, so the literal "may" is all the certainty information it has.
We ran a structural test on 25 answer sentences across 13 clients, each written as a confident fact dressed in reflexive hedging — "may", "can", "in many cases" — where the underlying claim was actually firm. We rewrote each into a direct claim, changing no facts — only removing the qualifiers the writer had added out of caution rather than necessity. Over the 45 days that followed, 18 of the 25 sentences began being lifted on at least one target query where they had previously been skipped. The lever was not new content; it was letting the answer sentence commit to the fact it already stated, so that when the composer pulled it out of the page it read as an answer rather than a hedge.
Driver two: do not stack qualifiers
Holding commitment roughly constant, the second driver was hedge stacking. One qualifier cost almost nothing — "typically reduces" was cited nearly as often as "reduces" — but the penalty was sharply non-linear: "may typically, in some cases, help to reduce" stacked four hedges and was cited far worse than any single-hedge version, because the composer reading the sentence for a claim found the claim buried under a pile of escape hatches. The reading consistent with the data is that each added qualifier moves the sentence further from a usable answer, and a sentence with three or four of them reads less as a careful claim and more as a refusal to make one.
We ran a structural test on 19 answer sentences across 11 clients that carried three or more stacked qualifiers. We rewrote each to keep at most one genuinely needed hedge and cut the rest, keeping the meaning the same. Over the 60 days after the change, 14 of the 19 sentences improved their cited-passage rate. The two drivers compound: a sentence that commits but stacks four qualifiers is half-built, and one with a single clean hedge but no real commitment is the other half — the sentences that won committed to the fact and carried at most one qualifier that earned its place.
Driver three: false confidence, and the hedge that was carrying real information
The third driver was the warning. Directness is a tool, not a rule, and mechanically stripping every hedge backfires when the qualifier was carrying real uncertainty. A sentence like "Results may vary depending on your configuration" became, under a blind de-hedging pass, "Results vary by configuration" or worse "This works for every configuration" — and the second invented a certainty the fact did not have. These over-confident conversions were cited no more often than the honest hedged version, and on 5% of audited pages the composer appended its own correction or caveat to the card, because the claim was now stated more strongly than the page could support and the composer flagged the gap. The reading consistent with the data is that the composer rewards a committed claim that is actually true, not the mere grammatical shape of confidence; when the fact is genuinely conditional — an effect that really does depend on context, a result that really is uncertain — the hedge is honest information and removing it manufactures a false claim the composer can catch.
We confirmed this on 16 sentences across 10 clients where an earlier optimisation pass had blindly de-hedged everything. We restored the qualifiers on the claims that were genuinely conditional, keeping directness only where the fact was firm. Over the following 45 days the restored sentences held or improved their citation while reading honestly, and none drew a corrective caveat from the composer. The actionable rule is blunt: state the fact directly when the fact is firm, which is more often than cautious writers assume — but when the claim genuinely depends on conditions, keep the one qualifier that carries that information rather than manufacturing a confidence the page cannot back.
What changed in our content checklist
Three changes. We added a commitment pass: before publishing, we read each section's lead answer sentence and ask whether the underlying claim is firm, and if it is and the sentence still hedges, we cut the reflexive qualifier so the sentence commits — because the composer reads the sentence cold and a literal "may" is read as doubt. We added a hedge-count check to the same pass: an answer sentence should carry at most one qualifier, and a sentence stacking three or four gets trimmed to the one that earns its place. And we added a truth guard: we only de-hedge where the fact is genuinely firm, and we keep the qualifier on conditional claims rather than manufacture a confidence the page cannot support and the composer will flag.
We dropped one habit. For years our most careful writers had hedged reflexively, because a qualifier feels safe — "may", "can", "in some cases" cost nothing in a paragraph a human reads in context, and they insulate the writer from ever being precisely wrong. The audit removes that default for answer sentences: a reflexive hedge in the one sentence the composer would lift spends the citation to buy a caution the claim did not need. So reflexive hedging left our playbook for answer sentences — we now commit to the fact in the sentence built to be cited, and reserve qualifiers for the claims that are genuinely conditional and for the sentences around the answer where no citation is at stake.
- 01Commit to the claim in the answer sentence. A direct sentence was cited 2.4× more than a matched hedged one stating the same fact — the composer lifts a sentence that answers the query outright, not one that answers it with a maybe.
- 02Do not stack qualifiers. One hedge costs almost nothing, but three or four compound sharply; 14 of 19 sentences improved after stacked qualifiers were trimmed to the one that earned its place.
- 03Keep the hedge that carries real uncertainty. The win is commitment to a true claim, not the grammatical shape of confidence; blindly de-hedging conditional facts manufactured false claims and drew a corrective caveat from the composer on 5% of pages.
- 04De-hedge only firm facts. 18 of 25 confident-but-hedged sentences were cited after the reflexive qualifier was cut — but restore the qualifier on the genuinely conditional claim rather than manufacture a confidence the page cannot back.
Where this argument breaks
For genuinely uncertain or conditional facts — a result that really does depend on configuration, a medical or legal claim that really is case-specific, an effect with real variance — the hedge is honest and removing it manufactures a false answer the composer can catch, so the lever is for firm facts stated cautiously rather than for facts that are genuinely soft. For navigational and brand queries there is no answer sentence whose certainty matters. For narrative and persuasive passages — case studies, opinion, story-driven content — hedging is a craft and tone choice, not a citation lever, and the commitment pass is for the answer sentences only. For YMYL pages the trade-off tilts: an over-direct medical or financial claim that the page cannot fully support is worse than a lost citation, so on those pages we keep the qualifier whenever the fact is anything short of settled, and accept the citation cost. The 5% corrective-caveat figure is small and noisy; we are confident a false-confident claim does not help and mildly confident it draws a correction, 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 29 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 sentence that commits to its answer far more readily than one that hedges it, the unit is the individual answer sentence rather than the page, and the cheapest citation win on a factual query is to state the firm fact plainly in the one sentence you want cited — and to keep the qualifier only where the uncertainty is real.