August 29, 2026
The Princeton GEO study tested nine content optimization techniques across 10,000 real search queries. Here is what actually improved AI visibility, what did not, and how to apply the findings.
Ask ChatGPT or Perplexity a question your buyers actually type, and the source that gets named in the answer usually is not whichever page ranks highest on Google. It is the page that gave the model something concrete to point back to: a named source, a specific number, a quote from someone with real standing on the subject. That is not a hunch. It is the core result of the Princeton GEO study, the most rigorous academic research published so far on what actually earns a citation inside a generative engine's answer, and it found that a handful of deliberate, testable edits can lift a source's visibility by 30 to 40 percent, while several habits most content teams still treat as best practice moved almost nothing at all.
This piece walks through what the researchers actually tested across roughly 10,000 real queries, which of the nine techniques they measured worked and which quietly failed, why citations, statistics, and quotations carry so much weight with a generative engine, why chasing keyword density alone falls flat, and what all of it means for building an AEO content strategy your team can act on this quarter instead of one borrowed wholesale from traditional SEO.
The Princeton GEO study found that specific, testable changes to a page, not broad rewrites, produced a 30 to 40 percent gain in visibility inside AI generated answers, while a few common tactics did close to nothing. The paper is titled "GEO: Generative Engine Optimization," written by researchers at Princeton University working with collaborators at Georgia Tech and the Allen Institute for AI, and it was published at KDD 2024, one of the most selective research venues in the field. It remains the largest controlled study of what actually moves the needle when a generative engine decides which sources to pull into its answer.
Before this research, "visibility" in AI search was mostly a feeling. A brand either seemed to show up in ChatGPT or Perplexity answers or it did not, with no consistent way to measure why. The Princeton team fixed that by introducing two concrete metrics: Position Adjusted Word Count, which weights how many words from a given source actually appear in the generated answer and where, and Subjective Impression, a separate score for how prominently a source reads within that answer. Together, those two numbers turned AI content visibility from a guess into something a team can actually test, move, and report on.
That measurement framework is arguably the paper's most important contribution, since every technique level finding below only means anything because the researchers could prove it moved a real, repeatable number rather than relying on anecdote.
The study's core contribution is a benchmark called GEO-bench: roughly 10,000 real queries spanning about eight broad topic domains, drawn from nine underlying data sources, each paired with the actual web sources a generative engine would pull from to answer it. That scale is what separates this from the smaller, more anecdotal tests that circulate around most AEO content strategy advice, and it is a large part of why the findings have held up as the reference point for the field ever since.
The testing process itself followed a consistent, repeatable sequence:
Start from an existing web source already paired with a GEO-bench query
Apply exactly one of nine content modifications to that source, leaving everything else unchanged
Feed the modified content into a generative engine built to closely mimic how Bing Chat synthesizes an answer
Score the resulting answer using Position Adjusted Word Count and Subjective Impression
Compare that score against the unmodified baseline, across the full 10,000 query set
Take the modifications that performed best and validate them a second time on Perplexity, a live generative engine, to confirm the results held up outside a controlled simulation
That last step matters more than it might look. Plenty of AI search research stops at a lab simulation. Validating the strongest techniques on a real, publicly available engine is what turns this from an academic exercise into something a content team can actually trust and build an AEO content optimization process around.
Researchers tested nine distinct content modifications against the GEO-bench baseline. Five produced a real, repeatable gain in visibility. Four did not.
The five techniques that consistently improved AI visibility:
Cite Sources: a 30 to 40 percent improvement in Position Adjusted Word Count, with a far larger jump for content that was not already ranking near the top of traditional search
Quotation Addition: a 30 to 40 percent improvement, strongest in domains built around explanation and lived experience
Statistics Addition: a 30 to 40 percent improvement, strongest in fact heavy domains and even larger when paired with citations
Fluency Optimization: a 15 to 30 percent improvement from rewriting for clarity and natural flow without changing the underlying claims
Authoritative Language: a 10 to 20 percent improvement from stating claims directly instead of hedging them
The four techniques that produced little to no benefit, and in some cases a slight decline once validated on Perplexity:
Keyword Stuffing: no meaningful gain, and a small negative effect in the live validation
Content Padding: simply adding more words did not translate into more visibility
Oversimplified Language: stripping content down to be easier to read showed a small or negative effect
Persuasive Language Without Evidence: confident sounding claims with nothing backing them performed worse than plainly stated, sourced ones
The next two sections go deeper on why the winners worked and why the most familiar loser, keyword optimization, failed as badly as it did.
All three of the strongest techniques share one property: they give a model something specific and attributable to work with, rather than asking it to trust a general assertion. Citing sources works because it hands the model a verifiable claim it can point back to. The researchers were careful to note that the citation had to point to a credible third party, not to the same brand's own product pages, since a self referential citation carries none of the same trust signal.
Quotation Addition works on a related but distinct mechanism. Attributing a specific claim to a named expert or organization functions as a portable credibility signal the model can lean on when deciding what to repeat. The effect was strongest in domains built around explanation, history, and lived experience, exactly the kind of content where an unattributed opinion reads as weaker than the same idea tied to a named source.
Statistics Addition works for a more structural reason. A specific number is a self contained, easily attributable unit of information, while a paragraph of narrative reasoning is not. Generative engines extract and synthesize claims rather than absorbing an argument the way a person reads one, and a sentence built around a concrete figure is simply easier to lift cleanly. The effect was strongest in fact heavy domains like law, policy, and anything opinion adjacent where a specific number carries real weight, and it compounded further when paired with a citation.
The common thread across all three, and really across everything covered in this generative engine optimization research, is verifiability. A generative engine has to decide what it is safe to repeat in a synthesized answer, and specific, attributable, checkable claims are simply a safer bet than confident but unsupported prose. That distinction is what separates genuine answer engine optimization content from a page that only repeats a target phrase enough times to satisfy a keyword tool.
This is the finding that should change the most habits. Keyword Stuffing was one of the nine conditions the Princeton team tested directly, not a side observation, and it produced no meaningful improvement in the primary evaluation and a small negative effect once validated on Perplexity. For a discipline built for two decades around keyword density, that is a significant, uncomfortable result.
Recommended Tool
Don't guess whether ChatGPT, Claude, Gemini, or Perplexity can access your content. Analyze your site in seconds.
Check AI Visibility Try Prompt Finder
No signup required • Instant results
The reason is structural rather than a quirk of one study. A classic search algorithm once leaned heavily on how often and where a term appeared on a page. A generative engine does not work that way. It is synthesizing an answer from the underlying claims a page makes, not counting how many times a phrase repeats, so packing a page with a target phrase gives the model nothing new to attribute or trust. It just makes the page harder to read, which the fluency findings above suggest actively works against a page rather than for it.
None of this means keywords are irrelevant. Clear, consistent topical language still helps a retrieval system find your page in the first place, which is a separate and necessary step before optimization can matter at all. What the research rules out is treating density and repetition as the goal itself. That is the core distinction behind real AEO content optimization: use language that makes a page findable, then spend the actual optimization effort on citations, statistics, and quotations, the three techniques the data shows genuinely move a generative engine's decision about what to trust.
Pulled together, the nine results point to one underlying principle: generative engines reward verifiable specificity over volume or confidence. Every technique that worked gave the model something concrete to attribute. Every technique that failed either added bulk without substance, stripped out the very specificity that made content trustworthy, or asked the model to take a claim on faith. That single distinction is worth more to an AI search content strategy than any individual tactic on the list.
The domain by domain variation is worth taking seriously too. Quotations carried the most weight in explanation and history heavy domains, statistics carried the most weight in fact heavy ones, and the paper is explicit that no single technique dominates every category equally. That argues against treating this list as a fixed checklist applied identically everywhere, and for treating it as a toolkit a team tests against its own category and its own most important prompts.
The finding most worth sitting with, though, is one that rarely gets top billing in shorter summaries of this research: pages sitting further down a traditional results page gained far more from these techniques than pages that were already comfortably in the top few spots. The Cite Sources method alone produced close to a 115 percent visibility increase for sources ranked around fifth in Google, against a barely noticeable change, and in some cases a slight decline, for sources already sitting at the very top. Generative engines appear to weigh the strength of an individual claim more than the reputation of the whole domain behind it, which means a smaller or newer brand willing to do this work has a genuinely realistic path to AI content visibility that traditional ranking rarely offered against an established competitor.
Turning this research into a working GEO content strategy does not require a full content rebuild. It requires a specific, ordered set of edits applied to the pages that already matter most.
Audit your highest traffic and highest intent pages for how many claims are actually backed by a citation, a named quotation, or a specific statistic
Add citations and statistics first, since the research shows these produce the largest and most consistent gains, especially in combination
Layer in one or two expert quotations per page, attributed by name and title rather than left as an unnamed opinion
Edit for fluency and directness rather than simplicity for its own sake, keeping precise or technical language wherever it is actually accurate
Stop treating keyword density as a target in itself, and reserve keyword focus for making a page findable rather than for the content quality work above
Prioritize this work on pages that are not already dominant in traditional search, since that is exactly where the research suggests the largest gains are sitting unclaimed
Re run the same buyer facing prompts across the AI engines your audience actually uses and track whether citation and mention rates actually move, rather than assuming a change worked
That last step is where a clear framework for AI search visibility metrics and KPIs earns its place, since none of the work above means much without a consistent way to confirm it actually changed anything. A dedicated AI visibility platform is what turns this from a one time content sprint into an ongoing AEO content strategy, tracking citation and mention rates across ChatGPT, Gemini, and Perplexity as each page gets updated, so a team can see exactly which edits earned a citation and which still need work.
AI Visibility Live preview
Check whether AI engines can reach, read and recommend your pages.
For teams still early in this process, a beginner friendly walkthrough of answer engine optimization strategies and frameworks is a useful foundation to build on before layering these specific findings on top, and our broader guide to content strategy for AI visibility covers how to decide which pages deserve this treatment first if you are working with a large content library rather than a handful of flagship pages. Teams further along who want the fuller organizational picture, beyond this one study, can start with our complete guide to building an AEO strategy , and our practical breakdown of structuring a site for AEO covers the heading and FAQ schema work that pairs naturally with the citation heavy editing this research recommends.
It is the research paper "GEO: Generative Engine Optimization," published at KDD 2024 by a team based at Princeton University, with contributors from Georgia Tech and the Allen Institute for AI also credited on the paper. It introduced GEO-bench, a benchmark of roughly 10,000 real queries, along with two new metrics, Position Adjusted Word Count and Subjective Impression, to measure how visible a source is inside a generative engine's answer. It remains the largest controlled study to date of which content changes actually improve that visibility.
Researchers tested nine distinct content modifications. Five produced a real, repeatable improvement in visibility: citing sources, adding quotations, adding statistics, optimizing for fluency, and using more authoritative language. The other four, keyword stuffing, content padding, oversimplified language, and persuasive language without evidence, failed to move the needle in a consistent way, and a couple even slipped slightly once the researchers checked them against a live engine.
No. The study tested keyword stuffing directly as one of its nine conditions and found no meaningful improvement in the primary evaluation, along with a small negative effect once the finding was validated on Perplexity. Generative engines synthesize answers from the claims a page makes rather than counting how often a phrase repeats, so repeating a phrase more often does not hand the model any new claim it can point back to.
Generative engines appear to evaluate content largely at the level of the individual claim rather than the overall reputation of the domain behind it. A page that was not already dominant in traditional search has more room to gain when its claims become more citable, while a page that was already trusted and frequently cited has less headroom left to gain. The gap the researchers actually measured makes the point well: the Cite Sources method alone lifted visibility by close to 115 percent for a source sitting around fifth place on Google, while a source already at the very top saw almost no additional movement.
Start with the pages that already get the most traffic or carry the most buying intent, and add specific, sourced statistics and citations to credible third parties before anything else. The study found these two techniques produced the largest and most consistent gains of the nine tested, particularly when used together, which makes them the highest value starting point before layering in quotations or fluency editing.
What the Princeton GEO Study Reveals About Content Visibility in AI Generated Answers
How Researchers Tested Content Optimization Across 10,000 Search Queries
Which Content Optimization Techniques Had the Biggest Impact on AI Visibility?
Why Citations, Statistics, and Quotations Make Content More Valuable to Generative Engines
Why Traditional Keyword Optimization Alone Isn't Enough for AI Search Visibility
What the Research Teaches Us About Creating Content for Generative Engines
How to Apply the Princeton GEO Findings to Your AEO Content Strategy
Frequently Asked Questions
Summarize with ChatGPT Summarize with Claude Summarize with Perplexity
S
Satvik Mishra
Co Founder of Verseodin
Satvik Mishra is the Co Founder of Verseodin, an AI visibility platform that tracks brand citations across ChatGPT, Gemini, Claude, and Perplexity. He writes about generative engine optimization strategy and what actually works for brands trying to earn visibility in AI powered search.
Newsletter
Get the latest AI Visibility insights, GEO research, product updates, and SEO strategies delivered straight to your inbox.
Continue improving your AI visibility.
Discover what users are asking AI about your industry and uncover prompt opportunities.
Open Prompt Finder
Analyze whether AI search engines can discover and cite your website.
Run Free Report
Tutorials 12 min read
A step by step method for finding where AI answers cite competitors and not you: collect comparable answers, check every cited page, confirm the cause, rank the gaps, and verify the fix.
Read article
Tutorials 12 min read
A practical method for measuring AI share of voice: pick tracked competitors, freeze a prompt set, run clean repeat samples, and combine ChatGPT, Gemini, and Perplexity scores without hiding the gaps.
Read article
Tutorials 12 min read
Ten criteria for judging an AEO insights platform, from prompt tracking and brand mention analysis to bot data, citation sources, share of voice, and platform coverage, plus a way to choose.
Read article
Want more AI SEO insights?
Monthly research on AI search, GEO and citation trends. No noise.
[ partner program ]
Join our partner network, whether you're an agency, consultant, or reseller. Leave your email and we'll reach out with details.
Revenue share
Earn a cut of every referral you bring on, renewing month over month.
Co-marketing
Joint webinars, case studies, and content with the Verseodin team.
Priority support
Direct line to our team plus early access to new features.
AEO Content Strategy: What the Princeton GEO Study Reveals About AI Search Visibility | VerseOdin