Ask ten marketing blogs how to get cited by ChatGPT or Google's AI Overviews and you'll get ten different, mostly unverified guesses—shorter paragraphs, more bullet points, a friendlier tone, more keywords. Almost none of it is tested against anything. Then there's the paper that actually is: "GEO: Generative Engine Optimization," published by researchers from Princeton University, the Allen Institute for AI, Georgia Tech, and IIT Delhi, which built a 10,000-query benchmark and controlled-tested nine specific content tactics against each other to see which ones actually moved the needle.
It sounds like exactly the kind of dense academic paper that never makes it into a real content decision. It shouldn't stay that way—the findings are concrete, specific, and directly usable, and three of the nine tactics the researchers tested outperformed the rest by a wide margin while one of the oldest tricks in traditional SEO barely moved anything at all. This article walks through what the study actually tested, what it found worked, what it found didn't, and the one nuance most summaries of this paper leave out entirely: the tactics didn't help every source equally.
Citing sources, quoting real people, and backing claims with statistics are the tactics a controlled 10,000-query study found actually moved AI citation visibility—while old SEO habits like keyword stuffing barely moved it at all.
The Study Behind the Buzzword
The researchers built GEO-BENCH, a 10,000-query benchmark drawn from nine different sources—including real anonymized search queries, university exam questions, and debate-style prompts—specifically to test how "generative engines" (their term for AI systems like ChatGPT and Google's AI-powered search) decide what to cite. For each query, they simulated a generative engine pulling the top five real Google search results, generating a response with those sources, and then testing what happened when one of those source pages was rewritten using a specific content tactic—while everything else about the query and the competing sources stayed the same.
That controlled design is what makes the paper worth taking seriously over a blog post's anecdote: it isolates the effect of one content change at a time, repeated across thousands of queries, rather than reporting what happened to one website once.
The Citation Triad: The Three Tactics That Actually Moved the Needle
Call it the Citation Triad—the three tactics that measurably outperformed everything else the study tested: citing sources (attributing specific claims to a named origin), adding quotations (including a direct quote from a real, identifiable person or document), and adding statistics (backing a claim with an actual number instead of a vague qualitative statement). The researchers report that these three methods achieved a relative improvement of 30–40% on their primary visibility metric, and 15–30% on a separate metric that scored how relevant, influential, and prominent a citation appeared within an AI-generated answer, compared to unmodified content.
It's worth being precise about what that means and doesn't mean. These are gains observed across a 10,000-query benchmark under controlled conditions—not a guarantee that rewriting any single page will produce a 30–40% lift in citations for your specific business. What the study does establish is a real, ranked hierarchy of which content changes are worth making first: cite something specific, quote someone real, and back claims with numbers, before spending effort on softer stylistic changes.
Why Keyword Stuffing Still Doesn't Work
The study also tested keyword stuffing—loading content with more of the exact terms a query might use, a technique that's been part of traditional SEO folklore for two decades. Against generative engines, the researchers found it had little to no measurable effect on citation visibility, placing it among the weakest of the nine tactics tested. The mechanism that made keyword density matter for older, ranking-based search algorithms doesn't transfer cleanly to a system that's summarizing and selecting sources based on the substance of a claim, not the frequency of a term.
The practical read for a Thai business: if your team is still optimizing pages by cramming in search-term variations, that effort is better redirected toward the Citation Triad. It won't hurt you outright, but the controlled evidence says it won't help either. It's the same substance-over-decoration lesson behind what separates a genuinely useful AI-powered website from AI Theater—the label or the trick doesn't earn trust, the underlying substance does.
The Underdog Advantage
Here's the nuance most summaries of this paper skip: the tactics didn't help every source equally. The researchers also tested what happened when sources across a full range of search rankings were optimized at the same time, and found the effect was strongly tilted toward sources that started out ranked lower. For the citing-sources tactic specifically, sources that started in the fifth position of the underlying search results saw visibility gains of up to 115% once optimized—while sources that already sat in the top position sometimes saw little additional gain, or even a measured decrease, from the same tactic.
Read plainly, that's good news for a smaller or newer Thai business that doesn't rank first for its category: the paper's own headline finding is that GEO tactics "can boost visibility by up to 40%" across the full benchmark, but the real opportunity is concentrated in giving lower-ranked, less-established content a genuine way to compete for an AI citation—not in defending a position that's already dominant.
Picture this: Rueang Thong Herbal Tea earns a citation without topping the rankings
Picture a hypothetical business called Rueang Thong Herbal Tea, a small herbal tea brand selling online. It has never ranked first for "herbal tea for sleep Thailand"—a handful of larger wellness retailers with bigger content budgets consistently outrank it. Rather than trying to out-publish those competitors, the team rewrote its main product page around the Citation Triad: naming the specific herbal source and harvest region, quoting an actual customer's description of the taste rather than generic marketing copy, and stating the exact steeping time and temperature instead of "steep to taste." Within weeks, an AI search summary answering "what herbal tea helps with sleep in Thailand" cited Rueang Thong by name alongside two of the larger retailers—not because it had climbed the rankings, but because its page gave the generative engine something specific and attributable to work with that the higher-ranked pages didn't.
The Citation Triad Playbook
- Name Your Source — every factual claim on a key page should trace to somewhere specific: a named report, a dated study, an identifiable person. "Studies show" is not a citation.
- Quote a Real Person — a direct, attributed quote from a customer, expert, or team member gives an AI system something concrete to attribute, not just paraphrase.
- Anchor Every Claim in a Number — replace "fast delivery" with the actual average delivery time; replace "popular treatment" with how many bookings it gets a month.
- Skip the Keyword Padding — redirect the time your team spends on keyword-variation tweaks toward writing one more genuinely sourced, quoted, or statistic-backed paragraph instead.
A neighborhood driving school choosing what to publish on its pass-rate page faces the identical choice: repeat the phrase "high pass rate" five times across the page, or state the actual number once with a source and a date. The second one is what the data says earns a citation.
Evidence: what it all comes down to
If there's one word for what this study actually found separates a citable page from an invisible one, it's Evidence—not persuasion, not keyword density, just something specific enough that a reader or a machine can point to it and say "that's where this came from."
So here's the question worth sitting with: pick the page on your website you'd most want an AI system to quote. Does it currently name a source, quote a real person, or state an actual number—or does it just describe things in general terms and hope that's convincing enough?
FAQ
What is the "GEO" this study introduces?
Generative Engine Optimization—the researchers' term for the practice of adjusting content specifically to improve how AI systems like ChatGPT or Google's AI-powered search cite and summarize it, as distinct from traditional SEO, which optimizes for ranking position.
Is this the same study as the "1,885 pages" schema markup research?
No, they're separate studies asking different questions. We cover the schema markup study, and what it found about a different, widely-repeated claim, here.
Will rewriting my page with citations and statistics guarantee more AI mentions?
No study can promise that for a specific business. What this research shows is that, across a large controlled benchmark, citing sources, quoting real people, and using statistics produced meaningfully better results than the alternative tactics tested—including no change at all.
Does this mean my top-ranked pages don't need this work?
Not exactly, but the study's own data suggests the biggest upside is for content that isn't already dominant in traditional search rankings. A page that already ranks first may see a smaller effect, or occasionally none, from the same tactics.
Is keyword stuffing actively harmful, or just ineffective?
The study found it had little to no measurable effect on citation visibility—not that it actively damaged it. The finding is that the time spent on it would be better spent elsewhere.
Do I need an AI expert to apply these findings?
Not necessarily. The three top-performing tactics—naming a specific source, quoting a real person, and stating an actual number—are editorial habits a content team can apply directly, without any technical AI work required.
What this looks like with NAWA AI Websites
NAWA AI Websites turns complex services into short, useful, specific answers—the same content discipline this study found actually earns a citation, rather than generic AI-written summaries that repeat what competitors already say. We build the structured pages and internal linking that put a sourced, quoted, number-backed answer in front of both your customers and the AI systems summarizing your industry.
If you want a second opinion on which of your pages would benefit most from this kind of rewrite, talk to us about an AI website and content review.