Quotations, Statistics, Citations: The Three GEO Signals

Published May 14, 2026 · ClickRadius Institute

Most advice about AI-search visibility is folklore — plausible, untested, recycled. There is one important exception: a peer-reviewed, large-scale study that actually measured which content changes make generative engines cite a source more often. Its answer was strikingly specific. Three interventions — adding quotations, adding statistics, and adding source citations — consistently outperformed everything else tested. This article examines that research in depth, explains why these three signals work at the level of how language models compose answers, and lays out an implementation playbook for each.

The study that started the field

"GEO: Generative Engine Optimization," led by researchers from Princeton University with collaborators from Georgia Tech, IIT Delhi, and the Allen Institute for AI, was presented at KDD 2024 — one of the top academic venues for data-mining research. The team coined the term GEO and did something the SEO industry had never had: a controlled benchmark.

Their method, briefly: they assembled a benchmark of roughly 10,000 queries spanning multiple domains, took source content that generative engines drew on, applied one of nine defined content modifications to it, and measured how the modified source's visibility changed in the engines' generated answers — using position- and word-count-weighted metrics rather than a simple mentioned/not-mentioned flag. In other words: same content, one variable changed, visibility measured. It remains the closest thing GEO has to laboratory conditions.

We demonstrate that GEO methods can boost visibility by up to 40% in generative engine responses.—Aggarwal et al., "GEO: Generative Engine Optimization," KDD 2024

Three of the nine interventions consistently led the results: quotation addition, statistics addition, and citing sources. Just as instructive, the study found that keyword stuffing — the core reflex of legacy SEO — delivered little to no benefit, and in some configurations performed worse than doing nothing. The optimization target had genuinely changed.

Why these three, mechanically

The three winning signals are not arbitrary — they map directly onto what a language model needs when it writes an attributed answer:

• An answer needs assertable facts. Statistics are maximally assertable: "costs rose 12% year over year" is a complete, useful sentence. Vague prose gives the model nothing crisp to carry into the answer.
• An answer needs safe attribution. A quotation arrives pre-attributed — the model can restate it with the credit built in. Attribution risk is why engines shy away from repeating anonymous, unsupported claims.
• An answer needs verifiability. Content that cites its own sources signals that its claims can be checked. For a system whose primary failure mode is being caught confidently wrong, verifiable inputs are systematically preferred.

Put simply: generative engines are attribution machines, and the three signals are the three forms content-level evidence can take. ClickRadius's scoring kernel weights exactly these signals — quotations, statistics, and source citations — within its 6-category, 0–100 AI Readiness Score, because they are the on-page factors with published evidence behind them.

Signal one: statistics

The highest-leverage and most abused signal. Done right:

1. Use real numbers only. Attribute every figure — "according to [source]," "industry surveys suggest," "in our own project records." A fabricated statistic repeated by an AI engine to thousands of users is a reputational liability that no visibility gain covers.
2. Place numbers in citable passages. A statistic buried in a footer helps nothing. Put it in the first two sentences under the heading that matches the question it answers.
3. Prefer decision-relevant quantities. Costs, durations, ranges, thresholds, failure rates, adoption rates — numbers a buyer would act on. These are the sentences engines are trying to write.
4. Date your data. "As of 2026" or "in the twelve months ending March 2026" makes a number safer to cite for time-sensitive queries and easier to maintain.
5. Mine your own operations. Your job records, response times, and project outcomes are statistics no competitor can copy — first-party data is the most defensible statistical moat a business has, provided it is stated honestly and framed as your own experience.

That last point deserves expansion, because most businesses sit on unpublished statistical assets without recognizing them. A service company's scheduling system knows its true average response time and job duration. An e-commerce operation knows return rates by category and the questions that precede purchases. An agency knows how long onboarding actually takes across its last fifty clients. Each of these, honestly computed and framed ("across our last 200 installations..."), is a statistic that exists nowhere else on the web — which means any AI answer wanting that kind of concrete texture has exactly one source to cite for it. Fifteen minutes with your own records typically yields five publishable numbers; that is a better statistical arsenal than most competitors' entire sites carry.

Signal two: quotations

Quotations transfer authority from a person to a page. Implementation that works:

• Quote genuine expertise. Your licensed practitioner, your founder, your engineers — real people saying substantive things. "The most common mistake we see in inspections is X, and it accounts for maybe a third of our emergency calls" is a citable sentence; marketing copy in quotation marks is not.
• Quote external authorities with attribution. Published statements from researchers, official guidance, standards bodies. This overlaps with signal three and compounds with it.
• Use real quote markup. <blockquote> with a <cite> element gives the chunker an unambiguous quote-plus-attribution unit.
• One strong quote per key passage. The research added relevant quotations in measured doses; a page that is wall-to-wall quotes dilutes its own voice and its coherence for retrieval.

Signal three: source citations

The counterintuitive one: citing others makes you more cited. It reads as generosity; it functions as verification. In practice:

• Name the source in the sentence. "According to Google's Search Central documentation..." carries the verification signal even when a reader never clicks. Naked links carry less.
• Cite upward. Government data, peer-reviewed research, official documentation, recognized industry bodies. The credibility of what you cite rubs off on the claim you attach it to.
• Cite for your load-bearing claims. Prioritize the passages where you most want citations — your money pages' key facts — rather than sprinkling references decoratively.
• Keep citations current. A page citing 2021 data for a 2026 question loses to the page citing this year's. Source freshness is part of passage freshness.

The other six interventions — and what they teach

The three winners are best understood against the full field. The study's nine tested methods also included fluency optimization (smoother, better-organized prose), easy-to-understand rewriting (simplification), authoritative tone, unique-word and technical-term enrichment, and keyword stuffing. The pattern in the results is instructive:

• Fluency and readability improvements produced real, secondary gains. Cleaner, better-organized writing helped visibility — consistent with how retrieval chunks and models process text — just not as much as adding evidence. Editing quality is a multiplier on the three signals, not a rival to them.
• Tone and vocabulary manipulations were inconsistent. Sounding authoritative is not the same as being attributable; the engines rewarded substance over style, and effects varied by query domain.
• Keyword stuffing sat at the bottom. The single most practiced SEO tactic of the prior two decades transferred worst to the generative setting.

The study also found that effectiveness varied meaningfully by domain — different query categories responded best to different combinations, and combining complementary methods (for example, fluency plus statistics) often outperformed any single one. The transferable lesson: treat the three signals as the core of a stack, applied over clean structure and clear writing, and expect your own category's optimal mix to be an empirical question your monitoring answers — not a constant copied from a benchmark.

What the study did not say

Intellectual honesty about the evidence keeps a GEO program grounded:

• "Up to 40%" is a ceiling, not a promise. Effects varied by domain, query type, and engine; lower-visibility sources gained the most. Expect meaningful but variable lift, not a fixed multiplier.
• It measured content signals only. The study held off-site factors constant. Industry data since suggests the majority of what separates cited brands from invisible ones is off-site — entity authority, directory presence, third-party corroboration. The three signals are the strongest on-page levers, operating within an authority context built elsewhere.
• Engines evolve. The benchmark predates current model generations. The mechanism it exposed — engines prefer attributable, verifiable material — has proven durable, but specific effect sizes should be re-verified against your own citation monitoring, not assumed from a 2024 paper.

The three signals are not a trick that fools engines. They are the properties that make content worth citing — the research merely proved engines can tell the difference.—ClickRadius Institute

A 30-day implementation plan

1. Week 1 — select targets. Choose the five pages tied to your highest-value buyer questions. Baseline their current AI visibility by querying the engines directly.
2. Week 2 — arm the passages. For each target question on each page: one attributed statistic, one genuine quotation, one named source, placed in an answer-first passage under a question-shaped heading.
3. Week 3 — verify plumbing. AI crawler access, clean HTML structure, structured data, visible dates.
4. Week 4 — measure. Re-query the engines with the same questions and log changes in mentions and descriptions. Retrieval-heavy engines move first; treat their movement as the leading indicator.

This loop — score, fix, publish, monitor — is precisely what ClickRadius automates: the AI Readiness Score grades these signals across six categories, auto-fix and content generation apply them, and citation monitoring across five live AI engines (ChatGPT, Gemini, Perplexity, Claude, and Grok) measures whether they convert.

Frequently asked questions

Where do the three GEO signals come from?

From the Princeton-led study "GEO: Generative Engine Optimization," presented at KDD 2024 — the first large-scale academic benchmark of how content changes affect visibility in AI-generated answers. Among nine tested interventions, adding quotations, statistics, and source citations produced the strongest gains, with reported visibility improvements of up to 40%.

Can I just add any statistics and quotes to rank in AI answers?

No — relevance and honesty both matter. The research added material relevant to the content's subject, and engines corroborate claims across sources. Fabricated or off-topic figures do not survive corroboration, and being repeated incorrectly by an AI engine is a reputational risk. Use real data, attribute it, and place it in the passages you most want cited.

Do the three signals work on every AI engine?

Effect sizes vary by engine and query type, but the direction holds broadly because all major engines share the same underlying need: claims they can safely attribute. Retrieval-heavy engines like Perplexity reflect the signals fastest; engines that lean on model memory show slower, less direct effects. That is why measurement across multiple engines matters.

Want to know how your pages score on these exact signals? Get your free AI Readiness Score — quotations, statistics, and citations are graded directly — or see plans and pricing.