Three Content Interventions Tested in the GEO Benchmark
The Princeton-led GEO study tested several content interventions after starting with the top five Google sources and selecting one source for optimization. It reported changes in that source's relative visibility/impression, not retrieval or citation-inclusion probability. This post explains how statistics, quotations, and source citations can improve the inspectability of a page without promising a current-engine citation.
The research behind the three
In 2023–2024, a research team led from Princeton did what the industry had only speculated about: they systematically modified web content, ran it through generative engines, and measured changes in a selected source's relative visibility/impression within generated responses. The result was "GEO: Generative Engine Optimization" (Aggarwal et al., KDD 2024) — the paper that gave the field its name. The benchmark began with the top five Google sources, randomly selected one source for optimization, and measured that source's relative visibility/impression in generated responses; it did not measure retrieval or citation-inclusion probability, and reported effects varied by domain.
In a benchmark beginning with the top five Google sources, researchers randomly selected one source for optimization. Quotations, statistics, and source citations produced reported relative visibility/impression gains of up to roughly 40% in some settings. The experiment did not measure retrieval or citation-inclusion probability, and effects varied by domain.—Aggarwal et al., "GEO: Generative Engine Optimization," KDD 2024
Three tested interventions involved adding statistics, quotations, and citations to sources. The reported result applies to the benchmark's selected-source visibility/impression measure and varied by domain. It does not establish a current-engine citation mechanism. Editorially, all three can add attributable specificity when they genuinely support the text.
Signal 1: Statistics — give the engine a number to carry
A specific, sourced figure gives a reader a claim that can be checked. It does not require an engine to retrieve, repeat, or credit the page. AI Overview coverage varies by source, cohort, date, locale, device, and definition; this page does not assert a universal rate.
Installation: On each priority page, replace your three vaguest claims with numbers. Your own operational data is best — jobs completed, years measured, results tracked — because it exists nowhere else. Reputable third-party statistics work too, if attributed cleanly: source name, ideally a link, current year. An unattributed number is barely better than an adjective; attribution is what converts a digit into evidence.
Signal 2: Quotations — give the engine a voice to borrow
Engines quote people, not brands in the abstract. A crisp sentence from a named human with stated credentials is retrievable, attributable, and — critically — not the engine's liability. Pages written entirely in anonymous corporate third-person contain no one to quote, so they get quoted exactly that often.
Installation: Add one genuinely quotable statement from a named person to each priority page — the owner, the lead technician, the practitioner. Format it as an actual quotation with attribution, not paraphrase. The test for "quotable" is simple: would the sentence survive being lifted out of the page alone? "Every AI citation is a loan of the engine's credibility" travels; "we take a customer-first approach" does not.
Signal 3: Source citations — show the engine you did the reading
The third signal is the one that feels most counterintuitive to marketers: linking out. Content that references authoritative external sources — research, standards bodies, official documentation — reads to an engine as evidence-backed rather than asserted. It is the difference between a witness and an opinion. According to the KDD 2024 findings, cited content was consistently more visible in generated answers; that benchmark result does not establish that citing sources changes retrieval or citation-inclusion probability on current engines. The benchmark began with the top five Google sources, randomly selected one source for optimization, and measured that source's relative visibility/impression in generated responses; it did not measure retrieval or citation-inclusion probability, and reported effects varied by domain.
Installation: For each priority page, find your two or three load-bearing claims and reference the authority behind them. "According to..." is the load-bearing phrase. If a claim has no findable authority behind it, that is worth knowing too — either it is your original expertise (say so explicitly) or it is filler (cut it).
A worked example, before and after
Abstract advice sticks better with a concrete rewrite. Here is a typical service-page sentence: "We're one of the most experienced roofing companies in the area, known for quality work." Zero signals — nothing to attribute, nothing to check, nothing to carry into an answer. The three-signal version: "We've replaced over 1,400 roofs in this county since 2009, and our repair-versus-replace assessments follow the inspection standards published by the roofing manufacturers' association. As our founder puts it: 'If a repair will honestly buy you five good years, we'll tell you — a roof you don't need is the most expensive one we sell.'" Same business, same facts the owner already knew — but now there is a statistic an engine can cite, a source it can verify, and a sentence it can quote. That rewrite took four sentences, not a content strategy. Multiply it across your five most important pages and you have done more evidence work than most of your industry.
Why the combination matters more than any one signal
No market-wide AI-answer prevalence is established here. Measure the business on a fixed, dated prompt set and record the engines, locale, device or account state where relevant, and whether each result is a mention, linked citation, or recommendation. That sample can identify an observed gap; it does not establish that a category is empty, a citation position is unclaimed, or an early entrant will become a default answer.
The reported effect came from the study's bounded engine and query benchmark and varied by domain. ClickRadius does not claim that the direction holds across every current engine, that the interventions cost almost nothing in every organization, or that the competitive field is empty. Use the content forms where they improve a real claim, then measure the outcome separately.
Frequently asked questions
Where do the three signals come from?
From the study "GEO: Generative Engine Optimization" by Aggarwal et al., published at KDD 2024 — research led from Princeton that experimentally tested content interventions against real generative engines. The benchmark began with the top five Google sources, randomly selected one source for optimization, and reported changes of up to roughly 40% in that source's relative visibility/impression after statistics, quotations, and source-citation interventions. It did not measure retrieval or citation-inclusion probability, and effects varied by domain.
Do the signals work if I don't have original data?
Yes. The signals are about attributable specificity, not exclusivity. Citing reputable third-party statistics with clear attribution, quoting named experts (including your own staff, with their credentials), and referencing authoritative sources all qualify. Original data is stronger where you have it, but properly attributed borrowed evidence beats unattributed assertion every time.
Can I overdo the three signals?
Yes — stuffing pages with irrelevant numbers and decorative quotes recreates keyword stuffing with new material, and engines are good at spotting content that performs evidence rather than containing it. The signals work when they genuinely support claims a buyer cares about. Density should follow substance, not the reverse.
Want to know your evidence density right now? Get your free AI Readiness Score — it measures all three signals across your site — or see plans and pricing.