Statistics, quotes and citations: the content formats that win AI answers

Most GEO advice is folklore: someone noticed Perplexity citing their listicle, wrote a thread about it, and now it's "best practice." There is exactly one widely cited piece of peer-reviewed research that tested content tactics against real generative engines at scale — "GEO: Generative Engine Optimization" by Aggarwal and colleagues from Princeton, Georgia Tech, IIT Delhi and the Allen Institute for AI, published at KDD 2024.¹

It deserves a closer read than the headline number usually gets, because the stakes keep rising: Gartner projected traditional search volume falling 25% by 2026 as buyers shift to assistants,³ and Pew found users click a source inside an AI summary on only about 1% of visits² — the answer itself is increasingly the whole impression your brand makes. The details of the study — what was measured, which tactics flopped, and where the effects varied — matter more for your content plan than the 40% figure everyone quotes.

What the study actually did

The authors built GEO-bench, a benchmark of roughly 10,000 queries drawn from real search and question-answering datasets, and ran them through generative engines that retrieve web sources and synthesize answers with citations. Then they took the source pages, applied one optimization tactic at a time, and measured how the modified source's visibility in the generated answers changed — using position- and word-count-weighted metrics, not just "did it appear."

That design is the crucial part. This wasn't correlation-mining across the web; it was a controlled before/after test of the same content with one variable changed. Nine tactics were tested:

What won — and by how much

The three evidence tactics dominated. Adding citations, quotations and statistics each produced visibility lifts in the neighborhood of 30–40% on the study's combined metrics, and the gains held across most query categories. Fluency and readability rewrites produced smaller but real improvements. The bottom of the table is where it gets interesting: keyword stuffing — the tactic an SEO team would reach for first — was essentially useless, and in some configurations counterproductive.

Why do the evidence tactics win? The mechanics are intuitive once you picture the engine's job. It must produce an answer that sounds grounded, and it must decide which retrieved passages to lean on. A passage with a statistic gives it a concrete fact to state; a quotation gives it pre-packaged, attributable language; a citation signals the claim was vetted by someone. All three reduce the model's risk of saying something unsupported — so passages carrying them get used, and the pages behind them get cited.

Applying the three winners, concretely

The tactics only work when the evidence is real and specific. Here is what each looks like as an edit to an actual page.

1. Statistics addition: replace adjectives with numbers

Before: "Our onboarding is fast and most teams are up and running quickly."

After: "Median time-to-first-report is 2.4 days, measured across the 312 workspaces created in Q4 2025."

The rewrite does three jobs at once: it gives the model a fact it can repeat, it names the method (so the fact survives scrutiny), and it's distinctive enough that when the number shows up in an answer, it's traceable to you. Audit your money pages for "fast," "leading," "trusted," "significant" — each one is a statistic that hasn't been computed yet.

2. Quotation addition: give the model attributable language

Before: "Customers love the reporting features."

After: "'We replaced a four-hour weekly spreadsheet ritual with a report that's just there on Monday morning,' says Dana Kim, RevOps lead at a 200-person logistics firm."

Quotes work because generative answers routinely attribute — "according to," "one user reports." A named, specific quotation is pre-formatted for that move. Real quotes from customers, engineers or published experts all qualify; invented-sounding praise ("This product changed everything!") does not, and models are surprisingly good at ignoring it.

3. Cite sources: let your claims borrow authority

Before: "Air quality has a major impact on productivity."

After: "Indoor CO₂ levels above 1,000 ppm are associated with measurable declines in cognitive performance (Harvard T.H. Chan School of Public Health) — which is why the C2 alerts at 900 ppm."

Citing external research inside your own content converts your page from "vendor asserting things" into "source that synthesizes evidence" — the same category as the journalism and reference content engines already prefer. One rule: cite only sources you've actually read, with claims they actually make. A wrong citation on your page can end up repeated in an answer with your name attached.

Where the evidence goes matters as much as whether it exists

The study modified whole sources, but the unit engines consume is the passage — so placement is leverage. Put a statistic in the first two sentences of a section, not the seventh paragraph, because chunking can split long sections and orphan your evidence from its claim. Keep the claim and its support in the same sentence or adjacent sentences: "Median onboarding is 2.4 days" and the methodology note four paragraphs later may land in different chunks, leaving the number unattributed. And repeat your two or three flagship statistics on every page where they're relevant — product page, comparison page, FAQ — because you don't control which page gets retrieved for which prompt.

The tactics also stack. A quotation that contains a statistic and names its source ("'We cut reporting time 78%,' says…, per the 2025 customer survey") is all three winners in one passage. That's not gaming — it's simply what a well-evidenced paragraph looks like, which is rather the point of the study.

What flopped, and why it matters

Keyword stuffing's failure is the study's most strategically useful finding, because it falsifies the default assumption that GEO is "SEO but more so." A retrieval system may still use keywords to find candidate pages — Google's own guidance for AI features is essentially "standard search indexing applies"⁴ — but the synthesis step — deciding what to actually say and cite — is driven by whether passages carry usable, verifiable content. Repeating "best project management software" eleven times gives the synthesizer nothing.

This is good news for smaller teams. The tactics that work are exactly the ones that reward genuine substance over budget: real numbers from your own data, real quotes from real customers, real citations to real research. None of that requires a link-building program.

Caveats before you copy-paste the playbook

Treat the study the way a good scientist would — as strong evidence with known limits:

1. Benchmark queries are not your buyer prompts. GEO-bench spans broad informational queries. Your revenue depends on a narrower set — "best [category] for [use case]" — where source pools are thinner and effects can be larger or smaller than the benchmark average.
2. Effects varied by category. The paper itself reports that different query types favored different tactics — one more reason a single global "lift" number shouldn't drive your roadmap.
3. Engines evolve. The experiments ran against 2023-era systems. Today's engines retrieve differently, re-rank differently, and are increasingly tuned against manipulation. The mechanism (evidence reduces the synthesizer's risk) looks durable; the exact percentages do not.
4. Visibility isn't recommendation. The study measured presence and prominence in answers. Whether the answer says something good about you is a separate metric you have to track yourself.

Verify on your own prompt suite

Everything above gives you priors, and priors are where you start — not where you stop. The gap between "worked on a 10,000-query benchmark in 2023" and "works on the forty prompts that drive my pipeline in 2026" is exactly the gap that separates teams that compound from teams that cargo-cult. The honest workflow is priors-then-measurement:

• Pick your 3–5 highest-stakes pages and apply the three evidence tactics — real statistics, real quotations, real citations.
• Before shipping, baseline your mention rate on a suite of 20–40 buyer prompts, sampled repeatedly across engines.
• Ship the rewrites, wait for re-crawl and re-indexing (days to weeks on the retrieval loop).
• Re-run the same suite and compare rates with confidence intervals — one run per prompt is noise, not signal.
• Keep what moved your numbers; drop what didn't, regardless of what the benchmark said.

The study hands you an ordering — evidence first, style second, keywords never. Your prompt suite tells you the truth about your category. Run both.

Frequently asked questions