**TL;DR** — Across 28 client sites through May 2026 we audited a grammatical choice that lives in the verb of the answer sentence: whether the passage that answers the query is written in **active voice**, with a named subject doing the action ("Compression reduces page weight by 40%", "Schema markup tells search engines what your content means"), or in **passive voice**, with the actor demoted or dropped ("Page weight is reduced by 40%", "Content meaning is conveyed to search engines"), and whether voice changes how often the AI Overview lifts that sentence into the card. Across 7,420 cited-passage events we joined each cited sentence to whether its main clause was active or passive. The headline is that active voice is a real and underused citation lever, but it is really an agency-and-directness lever wearing a grammar costume. A sentence in active voice was cited 2.3× more often than a matched passive sentence stating the same fact. The strongest predictor was actor presence — a sentence that named who or what performs the action was lifted far more than one whose actor was demoted into a "by" phrase or dropped entirely. The second was verb directness — a strong active verb close to the subject beat a passive construction that buried the verb behind an auxiliary chain. The third, and the warning, was forced-active awkwardness — mechanically converting every passive sentence to active produced stilted prose that was cited no more, and on 4% of pages slightly less, when the active version invented a vague actor ("you", "the system") that added nothing. One change — rewriting passive answer sentences into active voice with a named actor, on the sentences that answered the query — lifted cited-passage rate by 22% 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 for who does what. A passive sentence hides the doer — "bounce rate was reduced", "the index is updated nightly", "results were improved" — and a human reader fills the actor back in from context without noticing. The composer, extracting that sentence alone into a card next to three competitors, has no context to fill from, and a sentence missing its actor is a less complete answer than one that states it. We suspected the composer was quietly preferring active answer sentences because they arrive in the card whole — subject, verb, object — while passive sentences arrive with a hole where the doer should be, and we wanted to know whether that preference was real or whether the model parses passive constructions well enough that voice does not matter.
The second motivation was a failure mode we kept seeing on technical and academic-leaning pages. Writers in those registers reach for passive voice as a marker of objectivity — "it has been shown that", "the metric is calculated as", "performance is degraded when" — and the result is a page full of relevant, accurate answer sentences whose actor is missing. A human grader scores them as correct; the composer, hunting for one liftable sentence that states who does what to whom, finds the answer sentences agentless and passes for a competitor whose equivalent sentence is active. We needed to know whether the cost was the passive construction, because if it is, the fix is almost free — name the actor and let it do the verb, instead of demoting it behind the action.
How we ran the measurement
28 client sites — 11 SaaS, 6 publisher, 7 B2B services, 4 DTC — each with a fixed 200-query basket of its real in-market queries, weighted toward the how and why queries ("how does X work", "what does X do", "why does X happen") where the answer sentence has an actor that could be named or demoted. 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 classified its main clause: active (a named subject performing the verb), agentless passive (the actor dropped entirely), or by-phrase passive (the actor demoted into a trailing "by" clause). For each cited sentence we built a matched control: a comparable sentence on a similar query whose voice differed, so the comparison was active-vs-passive rather than good-page-vs-bad-page. The cited cohort was 7,420 events.
Two normalisation moves matter. We scored voice on the sentence as it would be lifted — alone, with no surrounding context — because that is the unit the composer extracts, and a passive sentence whose actor is obvious from the paragraph is actorless in the card. And we matched on sentence citability before comparing voice — 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 active voice is not just the active-voice pages being better written overall. The 2.3× and 2.1× figures are from those matched comparisons, not raw averages.
The shape of the active-voice pattern
The flat headline first. Active sentences are cited more. A sentence whose main clause was active was lifted 2.3× more often than a matched passive sentence stating the same fact. The effect held through the quality match and the citability control: among sentences our rubric scored as equally liftable, the active ones were lifted far more than the passive ones. The composer behaves as though it prefers a sentence that arrives in the card with its actor intact over one that arrives with the doer demoted or missing.
The most decision-relevant cut was that this is about agency, not grammar tags. We tested whether the win was specifically about active voice or more broadly about the actor being present and doing the verb, and it was the latter: an agentless passive ("bounce rate was reduced") was cited far worse than a by-phrase passive ("bounce rate was reduced by lazy-loading"), which was cited nearly as well as the fully active version ("lazy-loading reduced bounce rate"). Active voice is a reliable way to put the actor in front of the verb, but it is the actor-and-action completeness the composer rewards. Write active because it is the cheapest way to keep the doer in the sentence, not because the model is scoring voice as a category.
Driver one: keep the actor in the sentence
The single strongest predictor was whether the answer sentence named the actor performing the verb. Holding the sentence constant, a version with the actor present and doing the action was lifted at 2.3× the rate of an agentless passive version where the doer was dropped. The composer extracts a sentence and reads it for a complete proposition — actor, action, object; an active sentence hands it all three, an agentless passive hands it two and a gap. A human reader never experiences the gap, because the actor was established earlier in the paragraph — but the composer always reads the sentence cold, with nothing earlier attached, so the dropped actor is simply missing.
We ran a structural test on 23 answer sentences across 12 clients, each written as an agentless passive whose actor sat in an earlier sentence. We rewrote each into active voice with the actor named, changing no claims — only promoting the doer back into the subject slot. Over the 45 days that followed, 16 of the 23 sentences began being lifted on at least one target query where they had previously been skipped. The lever was not new content; it was making the answer sentence carry its own actor, so that when the composer pulled it out of the page the proposition was complete.
Driver two: put a strong verb close to the subject
Holding actor presence constant, the second driver was verb directness. A sentence with a strong active verb sitting right after its subject — "Caching cuts response time" — was lifted more than the same fact wrapped in an auxiliary chain — "Response time can be seen to be cut through caching" — even when both named the actor somewhere, because the composer reading for the proposition found it cleanly in the first and had to dig past "can be seen to be" in the second. The reading consistent with the data is that the composer prefers a sentence whose verb is doing visible work next to its subject, because it can extract the claim without unwinding a stack of auxiliaries and nominalations first.
We ran a structural test on 18 answer sentences across 10 clients that buried their verb behind auxiliaries or turned it into a noun ("the reduction of", "the calculation is performed"). We rewrote each to lead with subject and a direct verb, keeping the meaning identical. Over the 60 days after the change, 13 of the 18 sentences improved their cited-passage rate. The two drivers compound: a sentence that names its actor but buries the verb is half-built, and one with a direct verb but a dropped actor is the other half — the sentences that won named the actor and put a strong verb right behind it.
Driver three: forced-active awkwardness, and the invented actor that backfires
The third driver was the warning. Active voice is a tool, not a rule, and mechanically converting every passive sentence backfires when the conversion invents an actor that was not really there. A sentence like "INP is measured in milliseconds" became, under a blind active-voice pass, "Browsers measure INP in milliseconds" or worse "You measure INP in milliseconds" — and the invented actor ("browsers", "you", "the system") added a vague doer the sentence did not need. These forced conversions were cited no more often than the original passive, and on 4% of audited pages slightly less, because the vague actor read as filler and the sentence lost the clean focus on the fact. The reading consistent with the data is that the composer rewards a real actor doing a real action, not the mere grammatical shape of active voice; when the true sentence has no meaningful agent — a definition, a measurement convention, a state of being — passive is often the honest construction and forcing an actor in degrades it.
We confirmed this on 15 sentences across 9 clients where an earlier optimisation pass had blindly de-passivised everything. We reverted the ones whose active version had invented a hollow actor, keeping active voice only where a genuine doer existed. Over the following 45 days the reverted sentences held or improved their citation while reading more naturally, and none lost the chip they had. The actionable rule is blunt: write active when there is a real actor doing a real action, which is most of the time on how-and-why queries — but when the honest sentence has no agent, leave it passive rather than bolt on a fake one.
What changed in our content checklist
Three changes. We added an actor pass: before publishing, we read each section's lead answer sentence and ask who or what performs the verb, and if the actor is dropped into nowhere we rewrite the sentence to put it back in the subject slot — because the composer reads the sentence cold and a dropped actor is simply gone. We added a verb-directness check to the same pass: the answer sentence should carry a strong verb close to its subject, not a nominalisation or an auxiliary stack, so the proposition extracts cleanly. And we added a real-actor guard: we only convert to active where a genuine doer exists, and we leave definitional and measurement sentences passive rather than invent a hollow "you" or "the system" that the composer reads as filler.
We dropped one habit. For years our technical and B2B writers had reached for passive voice as a register marker — it read as measured, objective, institutional, so "performance is improved", "the metric is calculated", "results were observed" became the house default for anything that sounded like analysis. The audit removes that default for answer sentences: passive in the one sentence the composer would lift spends the actor for a tone the composer does not read. So reflexive passive voice left our playbook for answer sentences — we now name the actor in the sentence built to be cited, and reserve passive for the genuinely agentless cases and for the sentences around the answer where no citation is at stake.
- 01Write the answer sentence in active voice. An active sentence was cited 2.3× more than a matched passive one stating the same fact — the composer lifts a sentence whose actor and action arrive intact.
- 02Keep the actor present. The win is agency, not grammar; an agentless passive that drops the doer was lifted far worse than a by-phrase passive that keeps it — 16 of 23 sentences were cited after the actor was promoted back into the subject.
- 03Put a strong verb next to the subject. A direct verb beat an auxiliary chain or a nominalisation; 13 of 18 sentences improved after the verb was pulled close to its actor — the composer extracts the proposition without unwinding a stack first.
- 04Do not force a hollow actor. Mechanically de-passivising definitions and measurement conventions invented vague doers and was cited no more, and hurt on 4% of pages — leave the genuinely agentless sentence passive rather than bolt on a fake "you" or "the system".
Where this argument breaks
For definitional and measurement sentences with no real actor — "INP is measured in milliseconds", "the term is defined as", "Schema is a vocabulary" — passive or copular constructions are honest and forcing an agent degrades them, so the lever is for how-and-why queries where someone or something genuinely does the action. For navigational and brand queries there is no answer sentence whose voice matters. For narrative and persuasive passages — case studies, opinion, story-driven content — voice is a craft choice serving the prose, not a citation lever, and the actor pass is for the answer sentences only. For languages with different voice norms the effect may differ — in our parallel Chinese-language audit (文心一言, 元宝, 通义) the active-voice effect was present but weaker, since Chinese marks passivity less heavily and a topic-prominent sentence already foregrounds the actor; promoting the doer still helped on the hardest extractions. The 4% forced-active penalty is small and noisy; we are confident a hollow invented actor does not help and mildly confident it hurts, 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 28 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 names its actor and lets it do a direct verb far more readily than one whose doer is demoted or dropped, the unit is the individual answer sentence rather than the page, and the cheapest citation win on a how-or-why query is to write the one sentence you want cited in active voice with a real actor — and to leave the genuinely agentless sentence alone.