Facts and citations for AI: evidence, not tone

There is a reliable, measured way to make a page more likely to be quoted by AI answer engines: write it like a well-sourced reference. Add concrete facts and statistics, cite authoritative sources for the claims you make, quote credible voices directly, and say what you know plainly.

The caveat the hype guides skip is that this is about adding evidence, not chasing a density score. How much it helps also depends on your topic and on where you already rank. The promise and its limits come out of the same published experiments, and this guide keeps them together.

Why evidence gets you cited

Answer engines like ChatGPT, Perplexity, Gemini and Copilot are built on retrieval-augmented generation: in principle a system retrieves candidate pages, selects passages and writes an answer citing a few. The vendors do not publish their pipelines, and the research runs on stand-in engines: the GEO study answered with GPT-3.5 over the top five Google results and also tested Perplexity.ai; AutoGEO built engines on Gemini, GPT and Claude models. What such a system can use is a passage that is on topic, checkable and easy to lift whole. That is a mechanism, not a published ranking rule. A paragraph built around a specific, verifiable fact with a source attached is strong on all three counts, which is why the measured results below are unsurprising. Grounding can be sentence-level. Anthropic's Citations API cites the exact sentences of documents a developer supplies: "Claude can now provide detailed references to the exact sentences and passages it uses to generate responses" (Anthropic announcement). That is a developer feature, not a description of web search. It nevertheless shows how precisely a supplied document can be cited.

This is also the best-measured territory in generative engine optimization. The GEO study (Princeton, 2024) tested nine content changes across 10,000 queries and reported that "we demonstrate that GEO can boost visibility by up to 40% in generative engine responses." (GEO study) The winners were the evidence methods: "our top-performing methods, Cite Sources, Quotation Addition, and Statistics Addition, achieved a relative improvement of 30-40% on the Position-Adjusted Word Count metric and 15-30% on the Subjective Impression metric."

The same experiments set the limits. Rewriting pages in a more persuasive, authoritative tone produced no significant gain, and keyword stuffing did not help either. Evidence moved the needle; swagger did not. That leaves four levers worth pulling: concrete statistics, authoritative sources at the page level, a followable citation beside each checkable claim, and plain statements of what you actually know.

Add concrete facts and statistics

The first lever is replacing vague claims with specific numbers, dates, named entities and measured quantities. "Statistics Addition" was one of the top methods in the GEO study, part of that same 30-40% range. A sentence like "prices rose a lot" gives an answer engine nothing to quote; "prices rose 12% between 2023 and 2025" is a fact it can lift, attribute and verify. The pattern is always the same: take the claim you were going to make anyway and make it specific, sourced and checkable.

One distinction matters here, because most guides get it wrong. What the research measured is presence, not density. The intervention was adding concrete facts, citations and quotes to a page, not hitting a words-per-fact ratio or an "entity density" percentage. Neither study defines an ideal number of statistics per hundred words, and chasing one produces pages that read like almanacs. A page can be packed with numbers and still say nothing, and it can be dense with confident adjectives while containing no checkable claim at all. The test for each sentence is simpler: is the claim specific enough that a reader could verify it, and is the source of the number close by?

Concrete detail shows up in the machine-extracted rules too. AutoGEO, a 2025 study that mined the preferences generative engines reward, extracted two rules for its Gemini generation engine that say it plainly: "Substantiate claims with specific, verifiable data, statistics, or named examples." and "Use specific, concrete details and examples instead of abstract generalizations." Those rule sets were mined per engine, so this is Gemini's measured preference rather than a law of AI search.

Cite authoritative sources

The second lever is linking the claims on your page to authoritative external sources: primary data, standards bodies, government and academic publications, DOIs, recognized references. This is the method the GEO study calls "Cite Sources", and it is the one rule that shows up across engines. AutoGEO's cross-engine finding reads: "Attribute all factual claims to credible, authoritative sources with clear citations." (AutoGEO) In the GEO experiments, adding credible sources raised a page's visibility. The papers measure that outcome; they do not show that an outbound link is itself a ranking signal.

How much this helps depends on what you write about. The GEO study found the effect strongest where a source can verify a fact: "the addition of citations through Cite Sources is particularly beneficial for factual questions, likely because citations provide a source of verification for the facts presented". Citations and statistics did the most for factual, statement-style and law-and-government topics. For pure opinion or narrative writing there is far less to verify, so there is far less for a citation to do.

Three practical nuances. A citation marked rel="nofollow" still counts: the attribute shapes how link equity flows for search engines, but the citation still tells a reader, and a model, where the claim comes from. Second, put the citation in the main content, next to the claim it supports. A link in the footer, nav or sidebar sits away from any claim, so it does no evidential work. Third, this is a lever for long-form editorial content. A landing page or product page that cites little is normal and fine.

Attribute every checkable claim

The third lever is finer-grained than the second. It is not enough for a page to cite something somewhere; each substantive, checkable claim (a statistic, a finding, a quote) should carry a source that a reader, and a model, can follow. The unit of trust is the linkage between one claim and one source.

Two failure modes break that linkage. The first is the uncited claim: a concrete statistic or finding sitting in the text with no source anywhere near it. The number may be true, but nothing on the page lets anyone confirm it. The second is attribution without a source: writing according to a 2024 study and never linking or naming the study. This one is worse than it looks, because it imitates the shape of sourced writing while giving the model nothing to verify. Vague attribution reads as evidence to a human skimmer and as an empty pointer to anything that tries to follow it. Academic writing has held the same line for far longer: Harvard's guide to using sources asks that a source be integrated so a reader can see "not only which ideas come from that source, but also what the source is adding to your own thinking".

When you have a strong source, quoting it directly is the strongest single move measured. "Quotation Addition" was the top performer among the GEO study's methods, at the high end of that 30-40% range: a short, credited, verbatim quote from a credible source is about as extractable as text gets. It arrives pre-packaged as a claim, an authority and an exact wording, which is precisely what an answer engine wants to reuse.

One honest limit. Matching every claim to a followable source does not prove the source actually supports the claim; that judgement, called entailment, is a separate and harder problem. The practical aim is followable, well-matched citations: each checkable claim near a source that plausibly backs it, with the quote or number faithful to the original.

State solid claims plainly, hedge where it counts

A plain, definitive sentence ("the pump delivers 30,000 Pa") is, in principle, easier for a model to extract and reuse than a hedged one ("it might deliver around 30,000 Pa, depending"). A hedged sentence leaves the system with a choice: carry your qualifiers through, or reach for a competitor's flat statement instead. This is a plausible extraction advantage, but the cited experiments did not test it.

It is also not a tone trick, and the research is blunt about this. The GEO study tested exactly that: rewriting content in a more persuasive, authoritative voice, as its "Authoritative" method. The result: "we find no significant improvement, demonstrating that Generative Engines are already somewhat robust to such changes." Sounding confident is not a substitute for having the evidence. The lever is stating a solid, sourced claim in plain declarative form, not dialing up the rhetoric on a weak one.

And there is a place where hedging is simply correct: medical, legal, scientific and financial writing. "This may interact with your medication" must stay hedged, because the uncertainty is real and the reader's safety depends on it being stated. Flattening genuine uncertainty into false certainty is bad E-E-A-T, the expertise-and-trust standard Google's quality raters apply most strictly to exactly these topics, and it can be dangerous to the people acting on the advice. The working rule: state solid claims plainly, and keep every qualifier that reflects real uncertainty. Confidence should track the evidence, in both directions.

Where this helps most, and where it can backfire

The gains are not evenly distributed, and the GEO study is unusually honest about it. The lift is biggest for pages that do not already rank well: "the Cite Sources method led to a substantial 115.1% increase in visibility for websites ranked fifth in SERP, while on average, the visibility of the top-ranked website decreased by 30.3%." Evidence is an underdog's lever. If your page already sits at the top, piling on citations is not a guaranteed win, and the measured average for the top slot went the wrong way.

Match the lever to the topic as well: citations for factual and law-and-government material, statistics for data-driven and policy subjects, quotations where people and their words are the story.

The contrast case is keyword stuffing. In the same experiments it produced "little to no improvement" and landed below the un-optimized baseline on the objective metric. Repeating your target phrase persuades no reranker of anything. Evidence beats keywords, measurably.

Frequently asked questions

Does adding statistics and citations really get me cited by AI?

It is one of the better-supported tactics available: the GEO study measured the evidence methods (statistics, citations, quotations) as its biggest visibility gains. The caveats: the effect depends on your topic and your current rank.

How many statistics should a page have?

No. Make each claim specific and checkable, put its source nearby, and stop there. A page can be statistically dense and still vague where it counts.

Should I remove all hedging and sound totally confident?

No. Keep genuine qualifiers in medical, legal, scientific and financial writing. The uncertainty there is real information. Overclaiming where caution is warranted signals lower trustworthiness, not higher, and the GEO study found authoritative tone alone produced no significant gain anyway.

Does a nofollow link still count as a citation?

Yes. A citation carrying rel="nofollow" is still a citation for this purpose: it tells a reader where the claim comes from, which is the point. Whether a model treats it differently from any other link is not documented either way.

Will adding citations help a page that already ranks first?

Not necessarily. Evidence helps most when you are the underdog. If you already lead, keeping the page accurate and current is the better investment.