Jul 28,2026 
When Google Gemini, AI Overviews or AI Mode cites a page, it sometimes does something no other answer engine does: it tells you exactly which sentences it used.
If you look closely at citation URLs from Google’s AI surfaces, many of them carry a text fragment (#:~:text=). That’s a deep link that highlights the precise passage the answer was built from. ChatGPT, Perplexity, and Claude just give you a bare URL. Google actually gives us much more to work with.
We’ve been collecting this data at Hoverboard AI, the best AEO tool for agencies and brands, and recently added a new feature to our citation URLs and individual prompts that allow users to look deeper at these passages. Over the past month we resolved 230,000 text-fragment citations back to their source pages, pulled out the exact quoted passage, and classified each one against a set of structural and content signals.
What we ended up with is something close to a ground-truth dataset of what Google’s AI actually lifts from a page. Not which pages get cited. Which sentences get quoted. And the patterns are consistent enough that you can build a content playbook around them.
How we did this
A quick note on methodology first, because it shapes how you should read the numbers. Every passage in this dataset comes from Google’s AI surfaces, meaning Google Gemini, AI Overviews and AI Mode. That wasn’t a sampling choice on our part. Text-fragment citations are only emitted by Google. In every sample we checked over a 30-day window, the fragment-bearing citations traced back to Google’s AI, while other answer engines returned plain URLs with no passage to resolve.
Honestly, that limitation is also what makes the data valuable (especially considering that Google AI is increasing in answer engine market share). Improving visibility on Google AI mode, Google AI overviews and Google Gemini is increasing in importance. Normally you have to guess what an AI system found quotable. Here, Google marked the exact text for us. So read everything below as a study of Google’s AI quoting behavior specifically. Our working assumption is that other LLM-based engines select passages in broadly similar ways, but we can’t prove that from this dataset.
For each resolved passage, we tagged whatever characteristics we could detect: where it sits on the page, whether it lives under a heading, whether it contains a statistic, a credential, pricing, a review, and so on. A single passage can carry several tags, or none at all.
What 230,000 quoted passages have in common
Here’s how often each signal shows up across the resolved passages:
| Signal | Share of quoted passages |
|---|---|
| Intro paragraph (first ~15% of the page) | 23.2% |
| Quantified claim / statistic | 18.2% |
| Sits under an H2–H6 heading | 15.4% |
| FAQ answer | 15.2% |
| Credential (license, certification, award, “since 19xx”) | 11.2% |
| List item | 7.8% |
| Review content | 7.1% |
| Ranking (“top N”, “best of”) | 3.4% |
| Recent year mentioned | 2.7% |
| Pricing | 2.5% |
| In the H1 block | 2.1% |
| Table row | 0.6% |
These signals stack. About 40% of quoted passages carry exactly one of them, 20% carry two, and 9% carry three or more. Another 31% carry none at all, which turns out to be one of the more interesting findings in the dataset. More on that bucket later.
Six things this tells you about writing for AI citations
1. Front-load the answer
Nearly 1 in 4 quoted passages sits in the first 15% of the page. Google’s AI has a strong preference for quoting from openings rather than conclusions.
If you only make one change to an existing page, make it this: add a summary paragraph directly under the H1 that answers the target query in 40 to 60 words. Skip the “in this article we’ll explore” throat-clearing and just state the answer plainly, up top. The rest of the page can elaborate all it wants, as long as that intro can stand on its own when lifted out of context.
2. Put a number in the sentence
18.2% of quoted passages contain a quantified claim of some kind: a statistic, a dollar figure, “X years of experience,” “over N patients.” That makes numbers the second most common signal in the entire dataset, ahead of headings, FAQs, and credentials.
The rewrite here is simple. Vague authority claims like “extensive experience” or “trusted by many clients” almost never get quoted. “38 years in practice” or “$2.1B recovered for clients” does. Go through your key pages and look for any claim that could carry a number and currently doesn’t. Each one is a missed citation opportunity.
3. Use question-shaped headings
15.4% of quoted passages sit under an H2 through H6, and another 15.2% are FAQ answers. Add those together and roughly 30% of all quotes are anchored to heading structure. We see the same thing over and over in the data: when a heading matches the phrasing of the user’s query, the model lifts the block immediately after it.
Generic headings like “Our Services” or “Why Choose Us” waste this mechanic entirely. Convert them to the literal question your audience asks, something like “How much does a concierge doctor cost in San Diego?”, and make the first paragraph underneath a complete, self-contained answer.
4. Write credentials as standalone sentences
11.2% of quoted passages contain a credential: board certifications, licenses, awards, years in operation. Google’s AI leans on these when answering trust-oriented queries like “who are the best…”
The failure mode we see constantly is credentials buried three sentences deep inside a long bio paragraph. What actually gets quoted is a short, declarative sentence: “Dr. Nanos is a board-certified family physician with advanced training in epidemiology and public health.” Write your credentials the way you’d want them quoted, because that’s how they’ll be quoted.
5. Reviews, rankings, and pricing are undersupplied
Reviews appear in 7.1% of quoted passages, rankings in 3.4%, and pricing in just 2.5%. Those are low numbers considering how often users ask commercial questions (“best X,” “how much does Y cost”). Our read is that this reflects low supply rather than low demand. Most sites keep ratings in widgets, prices behind contact forms, and “top N” claims in JavaScript-rendered components, none of which produce a quotable sentence in the HTML.
The pages that do state this information in plain prose (“Rated 4.9 across 300+ patient reviews,” “Membership ranges from $2,000 to $5,000 per year”) get disproportionate pickup, precisely because so few competitors give the model anything to quote.
6. Tables qualify the page, but prose gets the quote
Table rows account for just 0.6% of quoted passages. Google’s AI almost never quotes a table cell directly. At the same time, our earlier research found that pages containing tables get cited more often overall.
Put those two facts together and a clear rule falls out: tables help the page get selected, and the quoted sentence still has to live in prose. Pair every table with a sentence that states its takeaway, along the lines of “Of the five plans compared above, only two include home visits,” so there’s something liftable sitting next to the structured data.
The 31% with no signal at all
Nearly a third of quoted passages carry none of the signals above. They’re plain descriptive sentences with no number, no heading, no credential, no structure. Our interpretation is that these get quoted because nothing better existed for that query, and the model had to pick something.
That makes the zero-signal bucket a useful competitive diagnostic. If a rival is being cited for an unremarkable sentence, that citation is weakly held. Publish the same fact with a number, a date, or a credential attached, put it under a question-shaped heading near the top of the page, and you’ve given the model a strictly better passage to quote. In our data, that’s the profile of content that wins citations.
What this data can and can’t tell you
Two honest caveats before the checklist, because we’d rather flag them ourselves.
First, these are correlations. We can see that quoted passages contain statistics and credentials far more often than chance would suggest. We can’t prove from this dataset alone that adding a statistic to a paragraph causes it to get quoted, since pages that are already authoritative tend to contain these features naturally. Our zero-signal finding is suggestive here (a third of citations are weakly held, and better-structured passages appear to displace them), but the clean test would compare quoted passages against the passages Google skipped on the very same pages. We have that data, and it’s the follow-up study we’re working on now.
Second, passage structure is not step one. Google’s pipeline has to retrieve your page before it ever evaluates your paragraphs, so topical authority, indexing, and page reputation still come first. A perfectly structured answer on a page Google never retrieves gets quoted exactly as often as you’d expect. Think of passage optimization as what converts retrieval into citation, and the rest of your SEO as what earns the retrieval.
What to do with this
The passage-level view changes the unit of work. You’re not really optimizing pages anymore. You’re manufacturing quotable sentences.
The checklist that falls out of this dataset: a 40 to 60 word direct answer under the H1. Numbers in place of adjectives wherever a claim can carry one. Question-shaped H2s with self-contained answers beneath them. Credentials as short standalone sentences. Ratings, rankings, and prices stated in visible prose rather than widgets. A takeaway sentence next to every table.
None of this is exotic. It’s the discipline of writing every important fact as a sentence an AI could lift verbatim and stand behind. Google’s text fragments show us exactly which sentences clear that bar. Hoverboard tracks these cited passages, and the signals they carry, for your domain and your competitors’, so you can see which of your sentences are getting quoted and where a weakly-held, zero-signal citation is sitting there waiting to be taken.
Methodology: 230,000 text-fragment citations (#:~:text=) collected from Google Gemini, AI Overviews and AI Mode over 30 days, resolved to their source pages and classified by Hoverboard. Fragment citations are exclusive to Google’s AI surfaces; other answer engines return bare URLs. Percentages reflect the share of exactly-resolved passages carrying each signal, and passages can carry multiple signals.