An AI Overview does not quote your page. It quotes one passage from it. So the unit of work is not the article, it is the paragraph: a question-shaped heading, a direct answer in the first 40 to 60 words, and at least one fact a model can lift and attribute without guessing.
This guide stays at that level deliberately. The site-wide work — crawler access, business consistency, schema, off-site mentions — is a separate job, and we cover it in how to optimize your website for AI search. What follows is the craft: how a paragraph has to be built to survive extraction, shown through a rewrite rather than a rule.
What makes a passage quotable to an AI Overview?
A quotable passage answers one specific question in its first 40 to 60 words, carries at least one checkable fact, and still makes sense with everything around it deleted. That last condition is the strict one. Models extract passages, not pages, so your paragraph has to hold up alone.
The original GEO research — GEO: Generative Engine Optimization by Aggarwal et al., published on arXiv in November 2023 and presented at KDD 2024 — found that adding statistics, citations and quotations to a source raised its visibility in generated answers by up to 40%, while keyword stuffing produced no benefit at all. Caveat worth stating: that study ran against a Bing-Chat-style system in 2023, not today’s AI Overviews.
More recent data points the same way. An Ahrefs study published December 2025, covering 174,048 pages cited across 560,346 AI Overviews, found the correlation between word count and citation was 0.04 on a Spearman scale — statistically nothing — and that 53.4% of cited pages were under 1,000 words. Length is not the lever. Density of answer is. Google’s AI optimization guide describes the same mechanism from the other side: its systems “understand the nuance of multiple topics on a page and show the relevant piece to users.” The relevant piece. Your job is to make sure there is one.
Why does specificity matter more in 2026 than it did in 2023?
Because the supply of competent, generic writing has collapsed in value. Graphite’s Common Crawl analysis — 55,400 English-language URLs, checked against three AI detectors and extended through March 2026 — found AI-generated articles jumped to 35.9% of sampled articles within twelve months of ChatGPT’s November 2022 launch, then leveled off close to an even split with human-written work: 49.6% in Q1 2025, a brief 50.9% majority in Q4 2025, back to 49.9% by Q1 2026. Fluent prose is now free. Only first-hand specifics are scarce.
The share has plateaued since early 2025, and Graphite’s reading is that practitioners found AI-generated articles do not perform well in search — these articles, they note, “largely do not appear in Google and ChatGPT.” Publishing volume and visibility turned out to be different measures. Google’s guide says the same thing plainly: don’t “just recycle what others on the internet have already said, or could easily be produced by a generative AI model,” and offer instead “unique expert or experienced takes that go beyond common knowledge.” Read that as an instruction, not a values statement. If a model could have written your sentence without reading your site, that sentence gives it no reason to cite your site.
The scarce things are yours: your prices, your turnaround times, your failure rates, your unglamorous edge cases, the question you get asked four times a week. Nothing else on the page is hard to reproduce.
What does a rewrite actually look like?
Here is the pattern we rewrite most often — a pricing section on a service page. The “before” is the shape of the copy, not any one client’s words. Note that it is not badly written. It is grammatical, warm, and completely unquotable.
Before:
Our Pricing
We believe every business is different, which is why we offer flexible, tailored cleaning packages designed around your exact needs. Our experienced team works to the highest standards and we pride ourselves on competitive rates alongside a genuine customer-first approach. Get in touch today for a free, no-obligation quote.
After:
How much does office cleaning cost per month?
Office cleaning in Manchester costs most small offices between £180 and £450 per month. Frequency is what moves the number: a 1,500 sq ft office cleaned twice a week sits near the bottom of that range, five times a week near the top. Prices are fixed for 12 months and there is no minimum contract.
Six changes are doing the work:
| Before | After | Why it matters |
|---|---|---|
| “Our Pricing” | “How much does office cleaning cost per month?” | The heading now matches a query someone types or says |
| Answer arrives never | Answer arrives in sentence one | The extractable passage starts at word one |
| “competitive rates” | “£180 to £450 per month” | A range is quotable; an adjective is not |
| No variables named | “frequency”, with two worked points | Explains the range instead of hiding behind it |
| No units, no place | “1,500 sq ft”, “twice a week”, “in Manchester” | Lets a reader locate themselves; anchors local queries |
| 49 words, 0 facts | 55 words, 7 checkable facts | Same length. That was never the variable |
The “after” survives the extraction test: cut it out of the page, show it to someone who has never seen the site, and it still answers the question it claims to answer. The “before”, extracted, says nothing at all.
One honest note. Publishing a price range means committing to it, and some businesses genuinely cannot. If that is you, publish the structure instead of the number — what the price depends on, what a typical project involves, what the minimum engagement is. Structure is still specific. “It depends” is not.
Which formats get lifted most reliably?
Four formats travel better than prose, because their structure tells a model what each part is with no interpretation required. You do not need to convert your whole page. Use them where the content is already shaped that way and currently buried inside a paragraph.
| Format | Where it belongs | Why it extracts cleanly |
|---|---|---|
| Direct definition (“X is…”) | first sentence of a section introducing a concept | Self-contained by construction; nothing to resolve from context |
| Comparison table | packages, pricing tiers, option A vs option B | States relationships explicitly instead of leaving them to be rebuilt from prose |
| Numbered step list | any process or instruction | Reproducible in full, in order, with attribution |
| Q&A block | end of a page, mirrored in FAQPage schema | Literally the shape of the interaction the user is having with the model |
The reason a table wins is narrow but real: “our standard tier includes X while premium adds Y and Z” makes a model reconstruct which feature belongs to which tier. A table has already done that.
What is not worth your time?
Four things get sold as AI-content optimization and do not hold up when someone measures them: chunking your text, adopting a special “AI style”, bolting on schema to win citations, and hitting a word count. Skipping all four is the cheapest win in this guide, because each one costs hours that belong in the specifics instead.
- Chunking your content into fragments for the model. Google’s guide is explicit: “There’s no requirement to break your content into tiny pieces for AI to better understand it.”
- A special “AI writing style.” Same guide: “You don’t need to write in a specific way just for generative AI search. AI systems can understand synonyms and general meanings.”
- Schema as a citation lever. Ahrefs tracked 1,885 pages that added JSON-LD between August 2025 and March 2026 against roughly 4,000 controls (published May 2026). Citations moved −4.6% in AI Overviews, +2.4% in AI Mode, +2.2% in ChatGPT — no meaningful uplift anywhere. Schema still earns its place for rich results and for removing ambiguity about your business. It is just not what gets a passage quoted.
- Word-count targets. See the 0.04 correlation above. Padding to 2,000 words “for the AI” only adds text a model has to read past.
Notice the pattern: every discarded tactic is a formatting trick, and everything that works is content you had to actually know.
The pre-publish test: five questions per section
Before publishing, run each section through five questions. If any answer is no, the section is not finished, however well it reads. This takes about ninety seconds a section and catches most of what would otherwise go out unquotable.
- Is the heading a question a real person asks, in their words? Not your category name. Not your service name. The phrasing from your inbox.
- Do the first 40 to 60 words answer it, standalone? Cover the rest of the section with your hand. Does what remains still answer the heading?
- Is there at least one checkable fact? A number, a price, a duration, a threshold, a named tool, a named place. Adjectives do not count.
- Could a competitor publish this identical paragraph? If yes, you have written the commodity version. Add the part only you know.
- Would you be comfortable seeing this quoted with your name on it, out of context? This catches vagueness and overclaiming at once, and it is the closest you can get to simulating what the model does.
Question four fails most often, and it matters most now that fluent generic prose is free. Question five is the one that keeps you honest.
How do you know whether it worked?
You will not see a same-day change, because AI systems refresh their answers on their own schedule. Measure two ways: check Search Console’s generative AI reporting if you have it, and re-ask the same customer questions you asked before the rewrite, on a fixed cadence, logging what comes back.
Google launched Search generative AI performance reports in Search Console on 3 June 2026, showing impressions, pages, countries, devices and dates for AI Overviews and AI Mode. Two caveats: there is no click data, and it rolled out to a subset of sites starting in the UK, with data beginning around mid-May 2026. For most businesses the manual method is still the primary measurement — our step-by-step version is in how to measure AI visibility.
One structural finding sets expectations. In an Ahrefs study of 863,000 keyword SERPs published March 2026, only 37.9% of pages cited in AI Overviews also ranked in the top 10 for that query, down from 76% in their July 2025 study, and 31.0% of citations came from pages beyond position 100. That is a passage-level system, not a ranking-level one — which is the practical argument for doing this work before you rank, and a reminder that behaviour shifted this much in under a year. Nobody can guarantee a citation. What you control is whether there is a passage worth quoting when the model comes looking.
Where this fits with everything else
Passage craft is one layer, and it will not compensate for a failure in the layers around it. If a crawler cannot reach the page, the best paragraph you have ever written is invisible. If four AI tools describe your business three different ways, a clean answer will not settle the contradiction.
So work in this order: the site-level checklist in how to optimize your website for AI search, then this guide applied to your money pages, then a gap check with the DIY AI SEO audit checklist. If AI tools ignore you entirely today, why ChatGPT doesn’t recommend your business covers the usual causes, and GEO vs AEO vs SEO clears up the terminology in one table.
If you would rather have someone read your pages and tell you which passages are failing and in what order to fix them, that is what our AI visibility audit does: 39 criteria across 6 areas, the same questions run across ChatGPT, Gemini, Perplexity and Claude, and a prioritized fix list with instructions. From PLN 499 net (around $125), with a refund if we don’t find at least five things worth fixing. Prefer to talk it through first? Book a free 20-minute consultation, or get a wider read with our free 3-minute marketing audit.
FAQ
Do I have to rewrite every page on my site this way? No, and you shouldn’t try. Start with the pages that answer buying questions — pricing, service detail, process, FAQ — because those are what customers ask AI tools about. Brand pages, manifestos and case narratives can keep a normal voice. Ten well-built passages on the right pages beat a whole site rewritten thinly.
Won’t answer-first writing make my copy sound robotic? Only if you stop after the answer. The structure constrains the first two sentences under a heading; the rest of the section is yours. Most people find the discipline improves the writing for humans too, because it forces you to know what you’re claiming before you decorate it. Google’s guide is clear that you don’t need a special style for AI.
Should I add an FAQ section to every page? Add one where you genuinely have repeat questions, and answer them with real specifics. An FAQ invented to hit a format is padding, and it competes with the answers that matter. Two real questions answered concretely beat eight generic ones.
Does using AI to draft my content hurt my chances of being quoted? The drafting tool isn’t the issue; the absence of anything original is. Graphite’s data suggests AI-generated articles are published at scale yet largely don’t appear in Google and ChatGPT, and Google asks for takes that couldn’t “easily be produced by a generative AI model.” Use AI to structure and tighten. Supply the prices, numbers and experience yourself — that part can’t be outsourced to a model, because the model doesn’t have it.