The snapshot quotes a sentence, not a page
Structure and markup refreshed for current answer engines; the original analysis is preserved.
At I/O this month Google previewed a search that, instead of listing ten links, builds one answer by pulling pieces from several sources and knitting them into a narrative. The quiet consequence for your writing: a single sentence of yours can be lifted out of its page and set beside strangers. When that happens, it has to stand on its own and make clear whose it is — or be misread, miscredited, or dropped.
the short answer
Generative search synthesizes one answer from several sources, so the unit it uses shrinks from the page to the sentence. A sentence that leans on its neighbours — "it cut costs 40 per cent" — breaks when extracted: dropped, or credited to the wrong subject. A self-contained, attributable claim — "Acme’s platform cut send costs ~40 per cent for mid-market teams" — survives intact. So write your load-bearing claims to name their subject and hold their meaning alone, inside good prose. It is not schema and not just structure; it is the sentence itself, written to travel.
key takeaways
- At I/O this month Google previewed a search that synthesizes one answer from several sources — knitting extracted pieces into a narrative — instead of listing ten links.
- The unit of use shrinks from the page to the sentence: your sentence can be lifted out of its page, placed beside others, and shown without the paragraph that explained it.
- A claim survives extraction when it is self-contained (carries its own subject and scope) and attributable (obvious whose it is) — so it cannot be misread or credited to someone else.
- The breakers are vague pronouns, implied subjects, numbers with no referent, and claims that only make sense after three prior paragraphs. Concrete subject, concrete scope, your entity named.
- This is not about choppy fragments; it is about making your load-bearing claims stand alone inside good prose — which usually makes the writing clearer for people too.
the same fact, two sentences
Both sentences state the same result. Pulled into a synthesized answer, the first loses its subject and may be attached to whatever sits nearby; the second carries its subject, scope, and owner with it. The difference is not the fact — it is whether the sentence can survive on its own.
Why synthesis changes the writing, not just the ranking
Most advice about being found treats the page as the thing that wins or loses: rank it, structure it, mark it up. That still matters, but synthesis adds a layer underneath it that older advice does not reach. When an answer is assembled from fragments of several sources, the contest is no longer only between whole pages; it is also between individual sentences competing to be the clear, usable, creditable expression of a point. A page can win the right to contribute and still contribute badly, because the specific sentences it offers fall apart once separated from their surroundings. The new failure mode is not "my page did not rank" but "my page was used, and the sentence it gave was so dependent on its context that the system either discarded it or quietly handed my point to someone else".
It is worth being measured about how far this goes today. The feature Google previewed is experimental, limited at first to people who opt in, and it deliberately declines to generate an answer when it is not confident — so this is an early signal, not a settled regime, and most of your traffic this month still arrives the old way. The reason to act on an early signal is that the underlying shift is not really about one product: any system that answers by synthesizing sources rewards sentences that survive extraction, and punishes ones that do not. Writing your key claims to stand alone costs little, helps human readers regardless, and quietly prepares you for a mode of search that is clearly arriving even if it has not fully arrived.
The shift, in three parts
The unit of use is shrinking from the page to the sentence; self-contained, attributable claims survive while context-dependent ones break; so write your load-bearing claims to travel without wrecking the prose. Open each part for where it changes the work.
01 The unit of use is shrinking from the page to the sentence
For twenty-five years the deal was simple: search found your page, the reader arrived, and your sentences were read where you put them, in order and in context. The generative search Google previewed at I/O this month changes the deal. Instead of returning a ranked list of pages, it builds one answer by drawing relevant pieces from several sources and knitting them, in Google’s own description, into a single narrative — reformulating the elements it extracts rather than handing you a page to read. The practical effect is that the unit being used is no longer the page; it is the sentence, or the claim, pulled out and recombined with material from other sites. Your page can still be excellent, and you still need it, but its sentences now have to be ready to perform on their own, away from home, surrounded by other people’s words. That is a different writing problem from "make a good page", and this month is the moment it stopped being hypothetical: from here on, some of your readers meet your sentences one at a time, chosen by a machine.
02 Self-contained and attributable claims survive; context-dependent ones break
When a sentence is lifted out of its paragraph, everything it was quietly borrowing from its neighbours disappears, and you find out fast whether it could stand on its own. "It cut costs by forty per cent" is meaningless alone — what is "it", whose costs, in what setting? Recombined into a synthesized answer, a sentence like that is either dropped as unusable or, worse, attached to whatever subject happens to sit near it, which may be a competitor. Contrast it with "Acme’s platform cut send costs by about forty per cent for mid-market teams": that sentence carries its own subject, its own scope, and its own owner, so it survives extraction intact and credits the right company. The two properties that matter are being self-contained — the claim names what it is about without leaning on the previous sentence — and being attributable — it is obvious whose claim it is. The breakers are predictable: vague pronouns, implied subjects, bare numbers, and points that only resolve after several earlier paragraphs. None of these are bad writing on a page read in order; they are simply not built to survive being read out of order.
03 So write your load-bearing claims to travel — without wrecking the prose
The wrong lesson is to chop everything into blunt, disconnected fragments; that would damage the writing for the humans who still read it and still matter most, and it is a real temptation to resist. The right lesson is narrower and kinder to your prose: identify the handful of load-bearing claims on a page — what you do, for whom, with what result, what makes you different — and make sure each of those can stand on its own, naming its subject and holding its meaning even if quoted in isolation, while the writing around them stays natural and human. Most good editing already does this for topic sentences; you are just extending the habit, on purpose, to the claims you most want carried correctly. A sentence that needs no run-up is usually a clearer sentence anyway, so this tends to improve the page for people at the same time. That double benefit — better for the reader, sturdier for the synthesizer — is exactly the kind of work the AC Group has favoured for 27 years: nothing gimmicky, just claims written so they cannot be mistaken about who they belong to.
Two paragraphs, one gets miscredited
Consider a page that builds its case carefully: a paragraph sets up a problem, names the company in the topic sentence, and then, two sentences later, lands the result with "and that brought onboarding time down from three weeks to four days". To a human reading top to bottom, the subject is obvious — the company introduced earlier owns that number. To a synthesizer that lifts the result sentence on its own, the subject is gone: "that brought onboarding time down" refers to nothing, and the impressive figure floats free, ready to be attached to whatever company the surrounding fragments happen to mention. The page did its job for the human and failed the extraction, and the company’s best proof point risks being quoted in support of a rival who merely sat nearby in the answer.
Now rewrite that one load-bearing sentence so it carries itself: "Acme cut customer onboarding from three weeks to four days." Nothing about the paragraph for human readers needs to change much; the surrounding prose can stay warm and connected. But that single sentence now names its subject and its result in one breath, so when it is lifted into a synthesized answer it arrives whole and credited correctly. The fix was not more content, a new schema type, or a structural overhaul — it was writing the claim that mattered most so it could not be separated from its owner. That is the entire move, and it is small, repeatable, and almost free: find the sentences you most want carried, and make each one able to travel alone.
What to do with this
Start by finding your load-bearing claims, not by rewriting everything. On each important page, mark the few sentences that state what you do, who you do it for, what result you produce, and what sets you apart — the sentences you would most regret seeing misquoted. Read each one as if it were the only sentence that survived the page, with no paragraph around it. If it still says clearly what it means and whose it is, leave it; if it leans on a pronoun, an implied subject, or a number with no owner, rewrite it to carry its own subject and scope.
Then hold the line on the prose. The goal is not to turn your writing into a list of blunt declarations; it is to make the handful of claims that carry your meaning sturdy enough to travel, while everything around them stays readable and human. Where naming your entity in the sentence reads naturally, do it, so attribution is unmistakable; where it would be clumsy, make sure the subject is at least unambiguous. Pair this with the entity and schema work that tells machines who you are, and you cover both halves: the markup that identifies you and the sentences that keep identifying you even after they are pulled apart. That combination — clear entity, claims that travel — is the unglamorous, durable craft the AC Group has practised for ' + years + ' years, now aimed at a search that quotes a sentence rather than a page, and that will quote it whether or not you wrote it ready to be quoted.
A two-minute test for any claim
You do not need tooling to find your fragile sentences; you need a cover sheet, real or imagined. Take a claim you care about and hide everything around it — the heading, the sentence before, the paragraph it sits in — so only those words remain. Read them cold, as a stranger would, and ask three things: do I know who this is about, do I know exactly what is being claimed, and could this be confused for a statement about some other company? If the answer to any of those is no, the sentence is leaning on its surroundings, and the surroundings will not travel with it into a synthesized answer. The fix is almost always to put back the one missing piece the sentence was borrowing — usually the subject — so it carries itself.
Run that test only on the claims that matter, not on every line; the goal is targeted, not total. A useful shortcut is to imagine each candidate sentence appearing, by itself, inside an answer that also quotes two of your competitors — because that is the actual situation it may face. If, standing among rivals, your sentence would clearly be read as yours and as true, it passes. If it could just as easily be read as describing the competitor quoted next to it, it fails, and a failing sentence is one you would rather catch now than discover later as a competitor’s borrowed talking point. The whole check takes a couple of minutes per page and turns a vague worry into a short, fixable list.
Where the breakage costs you most
Some sentences are more painful to lose than others, so it helps to know where to look first. The costliest are your quantified results — the "cut costs by", "grew by", "reduced time from X to Y" claims — because they are exactly what a synthesizer loves to surface and exactly the ones most often written with an implied subject. A floating, impressive number is worse than a missing one, since it can be handed to a rival and read as theirs. Close behind are comparative claims, where you say you are faster or more accurate or better suited than some alternative, because once the comparison is extracted the question "than what, said by whom" decides whether the line flatters you or someone else.
Your differentiators belong on the same shortlist: the sentences that explain what makes you not interchangeable. Those are the ones you least want recombined into a generic answer that could describe any provider, stripped of the part that made them about you. The pattern across all three — results, comparisons, differentiators — is that they are high-value and subject-hungry: they only mean something attached to a specific company, so they suffer most when the company drops out. If you fix nothing else, fix these, in that order, and you will have protected the claims a buyer is most likely to remember and most likely to act on.
Writing for synthesis: quick answers
What changed about how my content gets used?
The unit of use is shrinking from the page to the sentence. A classic search result sends a reader to your page, where your sentences are read in order, in context, surrounded by everything that gives them meaning. The generative search Google previewed this month does something different: it builds a single answer by pulling relevant pieces from several sources and knitting them into one narrative. In that mode your sentence may be lifted out of its page, set next to sentences from other sites, and presented without the paragraph that explained it. So the question is no longer only "is my page good and findable" but "does this individual sentence still make sense, and still clearly belong to me, once it is standing alone among strangers". A page can be excellent and still contribute sentences that fall apart the moment they are extracted.
What makes a claim survive being extracted?
Two things: it has to be self-contained, and it has to be attributable. Self-contained means the sentence carries its own subject and scope rather than leaning on the sentence before it — "it cut costs by 40%" is helpless on its own, while "Acme’s platform cut send costs by about 40% for mid-market teams" stands up anywhere. Attributable means it is obvious whose claim it is and what it is about, so that when a synthesizer places it beside other sources it credits the right entity rather than blurring it into someone else’s. Vague pronouns, implied subjects, numbers without a referent, and claims that only make sense after three prior paragraphs are the things that break under extraction. Concrete subject, concrete scope, your entity named where it matters — that is a sentence that can travel without losing its meaning or its owner.
Doesn’t this just mean writing in bland, choppy fragments?
No, and that is a real risk worth naming. Writing everything as disconnected, keyword-stuffed sentences would make your content worse for the humans who still read it, and the humans still matter most. The aim is not to chop prose into fragments; it is to make sure the load-bearing claims — the few sentences that state what you do, for whom, with what result — can each stand on their own even inside good, flowing writing. Most well-edited prose already does this for its topic sentences; the discipline is to extend it deliberately to the claims you most want carried correctly. You can write naturally and still ensure that your key sentences name their subject and hold their meaning alone. Done well, it improves the writing for people too, because a sentence that needs no run-up is usually a clearer sentence.
How is this different from schema or general structure advice?
Schema marks up your entity and facts for machines; structure organizes a page so it is easy to parse; both matter and we have written about them. This is a narrower, complementary point about the prose itself: the actual words of a claim, and whether they survive being separated from their neighbours. You can have perfect schema and a clean structure and still write body sentences that collapse into ambiguity when one of them is quoted alone in a synthesized answer. Schema tells a machine "this entity is called X and does Y"; self-contained prose ensures that when the machine paraphrases or lifts your sentence about Y, that sentence already says who and what on its own. Think of it as writing for a world where any one of your sentences might be the only one that travels — so each important one should carry its own passport.
A note on sources and timing
This is written in May 2023, just after Google previewed at I/O a generative search experience that builds an answer by drawing on several sources and knitting the extracted pieces into one narrative, rather than returning a single ranked page. We have described only what was public and experimental as of this writing — it was an opt-in preview, limited at first, and designed to stay quiet when it is not confident. The durable point does not depend on that one product: any search that answers by synthesizing sources uses sentences, not whole pages, so a claim has to stand on its own and say whose it is to survive being extracted and recombined. Writing your load-bearing sentences that way is cheap, helps human readers too, and is exactly the kind of plain, durable craft the AC Group has worked in for 27 years — make sure each claim that matters can travel without losing its meaning or its owner.