What content AI quotes most? Across 3,200 passages we watched eight AI engines pull into their answers, the formats quoted most are statistic lines, definition sentences, and list items — not the polished narrative paragraphs most brands spend the most time writing.
That gap is the point of this study. Most advice on getting cited tells you to "write AI-ready content." We did the reverse: we looked at what engines actually lifted, classified each quoted passage by its shape, and reported the distribution. No prescriptions — just what got pulled, in what form, by which engine.
Below is the full breakdown: which passage formats win, which get quoted word-for-word versus reworded, which engine prefers which shape, and how long a quoted passage really is.
What content AI quotes most, in one table
The short answer: AI quotes structured, self-contained single sentences most — statistic lines, definitions, list items, and table rows — and quotes flowing narrative prose least.
The table below ranks every format by its share of all quoted passages, with how often each was reproduced word-for-word (verbatim rate) and how often a passage of that shape was pulled versus a plain prose sentence of similar length (pull lift).
| Passage format | Share of quoted passages | Verbatim rate | Pull lift vs. plain prose |
|---|---|---|---|
| Statistic line | 22% | 61% | 3.4× |
| Definition sentence ("X is…") | 18% | 44% | 3.1× |
| List item (bullet) | 17% | 35% | 2.2× |
| Direct Q&A answer | 14% | 40% | 2.5× |
| Table row or cell | 11% | 58% | 2.7× |
| Ordered step (how-to) | 9% | 33% | 1.9× |
| Attributed quote / expert claim | 5% | 47% | 1.6× |
| Plain narrative paragraph | 4% | 12% | 1.0× (baseline) |
Two patterns run through the table. Statistic lines lead on both share and efficiency — they are common and they get pulled. Table rows are the mirror image: rare on most pages, so their share is modest, but per passage they punch well above prose. And the format brands produce most — flowing narrative paragraphs — sits dead last on every column.
How we ran the study
We tracked 3,200 cited passages drawn from AI answers across eight engines: ChatGPT, Gemini, Perplexity, Claude, Copilot, Grok, Google AI Mode, and Google AI Overviews. The passages came from a fixed set of 600 informational queries in B2B SaaS, marketing, and tech, sampled daily over roughly 60 days in early 2026.
For each answer, our team logged every passage that mapped to an identifiable source, then hand-classified it into one of eight micro-formats and tagged it verbatim or paraphrased by comparing the engine's wording against the source HTML. A "quote" here means any passage we could trace to a source — both linked citations and unlinked lifts where the text match was unambiguous.
Limits worth stating. This is observational, not causal: we measured what got pulled, not what caused a pull. The query set skews B2B and technical, so consumer or YMYL topics may behave differently. And it is a snapshot — engine behavior shifts week to week, which is why a one-time check ages fast and ongoing AI visibility monitoring beats a single report. Treat the numbers as directional, not laws.
The formats AI quotes most, ranked
Statistic lines: the most-quoted format
A statistic line is a single sentence built around a figure — a percentage, a count, a dollar amount, a ratio. It was the single most-quoted shape in our sample at 22% of all pulls, and 61% of those were verbatim, the highest of any prose format.
The reason is mechanical. A number is the unit of an answer, and an engine can't safely reword "37%" the way it can reword a sentence of opinion. So it copies. The Princeton-led GEO study (Aggarwal et al.) found that adding statistics boosted a source's visibility in generative engines by up to 40%; our passage-level read is sharper because we measured individual sentences rather than whole pages. A high-pull example: "Brand mentions in ChatGPT rose from 4% to 14% of relevant answers within 45 days."
Definition sentences ("X is…")
A definition sentence states what something is in one self-contained line, usually in "X is…" or "X refers to…" form. It took 18% of pulls at a 3.1× lift — engines lift a clean opening definition to anchor an explanation before they build around it.
These work because they are portable: a tight definition carries its own context and needs nothing above or below it to make sense. This is also where strong entity signals pay off — when your brand facts are structured with schema for AI search, your definition of a term is the one engines borrow. Example: "Answer engine optimization is the practice of structuring content so AI systems quote it directly in their answers."
List items
Each bullet is a discrete, bounded unit — exactly how retrieval systems chunk a page. List items took 17% of pulls, the third-largest share, but only a 2.2× lift, the widest gap between share and efficiency in the study.
Read that carefully: lists get quoted a lot mostly because lists are everywhere, not because any single bullet is unusually magnetic. On a per-passage basis, a stat line or table row beats a bullet. Lists still earn their place — just don't assume volume of bullets equals volume of citations. The bullets that pulled best were themselves mini-stat-lines or mini-definitions, not vague phrases.
Direct Q&A answers
A direct Q&A answer pairs a question-shaped heading with a 40–60 word reply. It took 14% of pulls at a 2.5× lift. The structure mirrors how people actually query AI, so the match between question and answer is tight.
The pattern that pulled was strict: a heading phrased the way a user would type it, immediately followed by a complete answer in the first sentence — no throat-clearing. When the answer was buried two sentences down, the pull rate dropped sharply — a pattern we return to in the section on page position below.
Table rows and cells
A table row pairs a label with values — a price, a spec, a yes/no. Tables took 11% of pulls, lower than lists because tables are rarer on the average page, but at a 2.7× lift and a 58% verbatim rate, the second-highest. When a table exists, its rows punch hard.
The mechanism mirrors stat lines: labels and numbers travel safely without rewording, so engines copy them into comparisons. A row like "MaxAEO — 8 engines tracked — daily refresh" can drop straight into an AI answer intact. Comparison and "X vs. Y" pages are quoted disproportionately for this reason.
Ordered steps
Ordered steps are numbered how-to instructions. They took 9% of pulls at a 1.9× lift — useful, but lower than other structured formats. The reason is dependency: a single step often doesn't stand alone because it assumes the steps before it, so engines either pull the whole sequence or none of it.
Steps that pulled well were self-describing — each one named its own action and outcome rather than relying on "then do this." When a how-to query surfaced, ordered lists were strongly preferred over the same instructions written as a paragraph.
Attributed quotes and expert claims
An attributed claim carries a named source: "According to X…" These took just 5% of pulls at a 1.6× lift. Lower share, but a real role — engines use them to add authority and a citation hook, and they reproduce them at a 47% verbatim rate because the attribution is part of the value.
This is the format earned media feeds. When your claims are quoted by sources AI already trusts, those become the lines engines repeat — the mechanism behind digital PR that gets you cited by the sources AI trusts.
Plain narrative paragraphs: the least quoted
Here is the uncomfortable finding. Plain narrative paragraphs — flowing prose with no number, definition, list, or table structure — took just 4% of pulls and set the 1.0× baseline. Only 12% were reproduced verbatim; the rest were reworded beyond recognition.
This is the format brands write most and the format AI quotes least. Narrative still does work — it carries voice, builds an argument, and gives the structured passages something to sit in. But on its own it is the least likely shape to be lifted into an answer, and when it is used, the engine almost always rewrites it.
Numbers travel verbatim, prose gets reworded
Across the full sample, about 45% of quoted passages were reproduced word-for-word or with only trivial edits; the rest were paraphrased or fused with other sources. But that average hides a sharp split by format.
Numeric formats get copied. Prose gets reworded. Statistic lines (61% verbatim) and table rows (58%) topped the verbatim ranking, while plain paragraphs sat at the bottom (12%). The logic is consistent: an engine can rephrase an idea freely, but it cannot reword "8 engines" or "$49/month" without risking accuracy — so it copies the exact characters.
The practical read is about control. If you want specific wording preserved — a brand name, a positioning line, an exact claim — attach it to a number or a tight definition rather than leaving it in a sentence of prose. That is the format that comes out the other side intact.
Which AI engine quotes which format
Engines don't quote the same way. Below is the top format each engine pulled in our sample, plus its tendency to copy versus reword. The spread is wide enough that "AI" is the wrong unit of analysis — track each engine separately.
| Engine | Top format pulled | Second | Verbatim tendency |
|---|---|---|---|
| ChatGPT | Definition sentences | Direct Q&A | Medium |
| Perplexity | Statistic lines | List items | High |
| Google AI Overviews | List items | Table rows | Medium-high |
| Google AI Mode | Definition sentences | Table rows | Medium |
| Gemini | Definition sentences | Statistic lines | Medium |
| Claude | Narrative synthesis | Definitions | Low |
| Copilot | Statistic lines | Direct Q&A | High |
| Grok | List items | Statistic lines | Medium |
Perplexity and Copilot quoted most verbatim — both surface inline citations, so faithful copying is part of their interface. Claude rewrote the most, leaning toward synthesis over extraction, which means exact wording survives least often there. Google's surfaces favored scannable structure — lists, rows, and steps — consistent with their roots in classic search snippets.
How long is a quoted passage?
The median quoted passage in our sample was 28 words, and 70% of verbatim pulls fell in a 15–45 word band. That is shorter than the "75–150 word" figure repeated across most GEO advice — and the gap is a definition problem worth fixing.
There are two different lengths in play. The answer capsule — the surrounding block that earns the pull — typically ran 60–150 words and gave the engine context to decide the passage was relevant. But the sentence actually lifted out of that block was much shorter. So the advice "write 75–150 word chunks" describes the container, not the quote. Build a tight capsule, but make sure the one sentence you most want pulled stands alone inside it at sentence length.
Where on the page AI pulls from
Position matters as much as format. 64% of quoted passages were the first or second sentence of their section or paragraph. Answers buried in the middle of a block were pulled far less often, even when they were the better answer.
This is front-loading, and it compounds with format. A stat line or definition that opens its section is doing two things at once — using a high-pull shape in a high-pull position. The corollary is blunt: a great answer in sentence four of a paragraph is, for quoting purposes, a hidden answer. Lead each section with the conclusion, then explain.
What the data suggests for your content
This study is descriptive, not a checklist — but a few implications follow directly from the numbers:
- Lead sections with a stat line or a definition. They top both share and verbatim rate, and front position adds a second multiplier.
- Put exact wording you care about next to a number or in a table. That is what survives word-for-word; prose gets rewritten.
- Don't mistake list volume for citation volume. Bullets earn share through ubiquity, not per-item magnetism — make each bullet a mini-stat or mini-definition.
- Build comparison tables where the topic allows. Rare on most pages, high-lift when present.
- Treat narrative as the connective tissue, not the quotable unit. It carries the argument; the structured sentences inside it get pulled.
These line up with the broader playbook in how to optimize for AI search — the difference is that here you can see which moves the engines actually rewarded.
How to see what AI quotes from your site
Knowing which formats get quoted in general is one thing; knowing what gets quoted from your pages — and what gets quoted from competitors instead — is the part that defends a budget. That requires monitoring the answers themselves, daily, across engines.
This is what MaxAEO does: it tracks how ChatGPT, Gemini, Perplexity, Claude, Copilot, Grok, AI Mode, and AI Overviews mention, rank, and describe a brand, and surfaces which passages get pulled and which competitor lines get quoted in your place. Pair that with a measure of AI share of voice and you can watch a format change show up as a citation change rather than guessing.
Frequently asked questions
Does AI quote my content word-for-word or paraphrase it?
Both, and it depends on format. About 45% of pulls in our sample were near-verbatim. Numeric formats are copied most — statistic lines (61%) and table rows (58%) — because numbers can't be safely reworded. Plain prose is reworded most, with only 12% reproduced word-for-word.
What content format is most likely to get quoted by ChatGPT?
In our data, ChatGPT pulled definition sentences most often, followed by direct Q&A answers. A clean "X is…" sentence that opens a section is its highest-probability target. Perplexity and Copilot, by contrast, leaned hardest on statistic lines.
Do lists really get cited more than paragraphs?
Yes — list items took 17% of pulls versus 4% for plain paragraphs. But the lift per item is only 2.2×, lower than stat lines or tables. Lists win on volume because they're everywhere, not because any single bullet is unusually quotable. Make each bullet a mini-stat or definition to raise its odds.
How many statistics should a section have to get quoted?
Our study didn't test a threshold, so we won't invent one. What we can say: statistic lines were the most-quoted and most-verbatim format, and the academic GEO study found adding statistics lifted source visibility by up to 40%. Lead with a relevant figure rather than counting them.
Does what AI quotes most differ across engines?
Substantially. Google's surfaces favored lists, rows, and steps; ChatGPT and Gemini favored definitions; Perplexity and Copilot favored stat lines and quoted most verbatim; Claude rewrote the most. Treating all engines as one behavior is the most common mistake — track them separately.