Structured, self-contained single sentences — statistic lines, definitions, and table rows — not flowing prose. In MaxAEO's 2026 analysis of 3,200 passages quoted by eight AI engines, statistic lines were quoted the most (about 22% of all passages) and carried about 3.4x the "pull" of plain narrative, which trailed every other format. Roughly 45% of quoted passages were reproduced word-for-word. This reference lays out the full by-format breakdown and what it means for how you write for extractability.
Last reviewed: 29 August 2026. All figures below come from MaxAEO's "What Content AI Quotes Most" analysis (2026): 3,200 cited passages hand-classified from a fixed set of 600 informational queries in B2B SaaS, marketing, and tech, sampled over roughly 60 days across ChatGPT, Gemini, Perplexity, Claude, Copilot, Grok, Google AI Mode, and AI Overviews. It is single-vendor and observational — cite "MaxAEO, 2026" with the date, treat the ordering as firm and any single multiple as directional, and note the query set skews technical.
What kinds of content do AI engines quote most?
Short, structured, self-contained sentences — and they win on two separate measures at once. "Share" is how often a format showed up among quoted passages; "pull lift" is how much more likely a passage of that format was to be quoted than plain narrative. Statistic lines led both:
| Passage format | Share of quoted passages | Verbatim rate | Pull lift vs. prose |
|---|---|---|---|
| Statistic lines | ~22% | ~61% | ~3.4x |
| Definition sentences | ~18% | ~44% | ~3.1x |
| List items | ~17% | ~35% | ~2.2x |
| Direct Q&A answers | ~14% | ~40% | ~2.5x |
| Table rows / cells | ~11% | ~58% | ~2.7x |
| Ordered steps | ~9% | ~33% | ~1.9x |
| Attributed quotes | ~5% | ~47% | ~1.6x |
| Plain narrative | ~4% | ~12% | 1.0x (baseline) |
Source: MaxAEO, 2026. The pattern is consistent with how passage-level retrieval works: an engine scores candidate passages and keeps the ones that stand alone cleanly, so a sentence carrying one fact beats the same fact wrapped in three clauses of context. Table rows are the telling case — rare on most pages, so their share is modest, but per passage they punch far above prose, because a row is already a self-contained unit.
How often do AI engines quote you word-for-word?
Close to half the time, and more often the more structured the passage is. Across the 3,200 passages, about 45% were reproduced verbatim rather than paraphrased. The verbatim rate tracked structure directly: statistic lines (~61%) and table rows (~58%) were lifted unchanged far more often than plain narrative (~12%).
- Median quoted passage: 28 words. Long enough to carry one fact, short enough to lift whole.
- 70% of verbatim pulls fell in the 15-to-45-word range. A sentence much longer than that is more likely to be paraphrased or skipped.
- The takeaway: write the sentence you would want quoted. If the fact you care about lives in a 60-word compound sentence, an engine is more likely to reword it — and a reword is where nuance and attribution get lost.
Where in a section does the quoted sentence sit?
Almost always at the front. MaxAEO found 64% of quoted passages were the first or second sentence of their section. That is the same answer-first discipline the structure-a-page-into-extractable-chunks how-to argues for, now visible in the data: leading each section with the direct answer is not a style preference, it is where the quoting actually happens. The companion question — how page length and position interact, and why a fact buried deep in a long page rarely gets pulled — is covered in where on a page AI engines quote from.
Does this contradict "coverage, not length" or the Princeton edit findings?
No — the three line up. This study measures which passage formats get pulled once a page is retrieved; the finding that content length is not the lever, coverage is is about whether the page gets pulled at all. And the 2024 Princeton GEO experiment on what content changes lift AI citations — where adding statistics raised a source's position-adjusted word count by about 41% and keyword-stuffing lowered it — points the same way from a causal angle: statistic-dense, evidence-first passages travel best. It is also distinct from which content formats survive over months: that measures durability of a page's citation; this measures which sentence an engine lifts today.
AI engines quote sentences, not pages. Write each fact as a short, self-contained line — a statistic, a definition, a table row — and you hand the engine something it can lift unchanged.
What should you do to get quoted more?
Reformat the facts you already have into the units engines lift, then confirm it moved your numbers.
- Lead each section with a statistic line or a one-sentence definition. Put the number and the claim in the first sentence, where 64% of quoted passages come from.
- Turn spec-and-comparison prose into tables. Row-and-column facts are lifted verbatim far more often than the same data in a paragraph, part of the discipline in prioritising your structured data.
- Keep quotable sentences to roughly 15-45 words. One fact per sentence, so an engine lifts it whole instead of rewording it.
- Attribute and date your statistics. A specific, sourced number is both more quotable and safer to be quoted — the corroboration signal engines reward.
- Measure per engine. Formatting is a means, not the goal; track your citation coverage after you restructure, engine by engine, because none reports passage-level selection directly.
The honest read for late 2026: getting retrieved is about authority and corroboration, but getting quoted is about extractability — and extractability is a writing choice you control on every page. Watching which of your passages engines actually lift, and whether a rewrite changed that, is exactly what Buffy Intel measures. Questions: [email protected].