The Princeton GEO Study: Benchmark & Findings Explained

The paper "GEO: Generative Engine Optimization" (Aggarwal, Dugan, et al., KDD 2024) is the first systematic, benchmarked study of how content modifications affect citation visibility in AI-generated answers. It introduced GEO-bench — a benchmark of 10,000 real search queries across 9 datasets — and tested 9 optimization strategies with quantified visibility results. As of September 29, 2026, it remains the most methodologically rigorous reference study in the GEO field — though the September 2026 replication now quantifies precisely how much of its signal survives, and directional findings are validated by Conductor, GrackerAI, BrightEdge, and Axis Intelligence.

The paper was published at KDD 2024 (ACM SIGKDD Conference on Knowledge Discovery and Data Mining), with the arXiv preprint 2311.09735 first posted in November 2023. The research team spanned Princeton University, IIT Delhi, and Georgia Tech — a cross-continental collaboration. In 2026, the paper continues to be cited on Princeton's research portal as foundational work in AI content optimization.

How GEO-bench measures AI visibility

GEO-bench is constructed from 10,000 real search queries drawn from 9 authentic datasets, including Google Search logs and Perplexity user queries. Each query is run through a generative search engine, and the research team analyzes which sources are cited in each AI answer and how prominently.

The core metric is position-adjusted word count: it counts how many words from a source appear in the AI-generated answer, with words appearing earlier in the answer weighted more heavily. This design distinguishes between "briefly mentioned" and "core reference" — a crucial distinction that simple binary citation metrics miss. The metric captures both whether you are cited and how prominently you are cited.

The 9 strategies: complete results table

The research team applied each content modification to a baseline page and compared visibility before and after on GEO-bench. Below are the complete results:

Source: Aggarwal et al., "GEO: Generative Engine Optimization," arXiv:2311.09735, KDD 2024. Visibility measured by position-adjusted word count on GEO-bench (10,000 queries × 9 datasets).

The most counterintuitive finding: lower rank, higher GEO returns

Perhaps the most strategically significant finding from the Princeton study is that GEO's impact is inversely correlated with traditional search rank. Pages ranked 5th in Google search achieved up to +115% visibility improvement after GEO optimization. Meanwhile, the 1st-ranked page experienced a 30% visibility decline.

This finding inverts the traditional SEO logic where top-ranking pages compound their advantage through authority signals. In AI search, the re-ranking models do not inherit Google's ranking conclusions — they evaluate content quality independently. BrightEdge's February 2026 analysis confirms this mechanism at scale: 62–83% of sources cited by Google AI Overviews come from outside the organic top 10. The implication is clear — pages that struggle in traditional SEO have the most to gain from GEO.

"For two decades, SEO was a winner-take-most game — high-authority domains with deep backlink profiles consumed the majority of traffic. GEO may break this pattern. If content quality is strong enough — original data, expert quotations, clear structure — a new domain can earn more exposure in AI answers than established authority sites."

2026 independent validations of the Princeton findings

Since publication, multiple independent studies have validated and extended the Princeton team's findings. Here is a summary of the key validation studies as of September 3, 2026:

The July 2026 critical survey: what 45 studies actually prove

A July 2026 critical survey (Martinez, arXiv:2607.14035) reviewed 45 GEO studies published between November 2023 and July 2026 — the first systematic audit of the field. Its uncomfortable but useful conclusion: GEO techniques reliably change how an already-retrieved page is cited, but no reviewed technique shows a stable, cross-platform causal effect on organic discoverability. In other words, the research shows how to be quoted better once an engine already has your page — not how to make the engine find you in the first place.

The survey also traces the famous "+40%" to its source. In Aggarwal et al. (KDD 2024), the metric is position-adjusted word count — the share of the generated answer attributed to a source, weighted by position. For the quotation tactic it rose from 19.3 to 27.2, about +41% in relative terms. That is a share-of-answer-text gain inside a fixed testbed, not a click gain and not a retrieval gain. The generalized claim that "GEO increases visibility by 40%" is explicitly listed among the claims the review rejects as unsupported.

The practical takeaway is not that GEO is worthless — it is that indexing and retrieval are the real bottleneck, not formatting tricks. The Princeton benchmark remains the gold-standard method for measuring citation once a page is in the pool; the 2026 survey tells you not to confuse "more citable" with "more findable." This is exactly why the indexing workflow in our get-indexed-by-AI-search-engines guide matters as much as on-page optimization.

August 2026: the volatility evidence that reframes the +41%

Three weeks after that survey, the field got its sharpest natural experiment. Between July 18 and August 7, 2026, reddit.com held a steady 3.83% share of ChatGPT Search citations — one of the largest of any single domain. On August 14 it fell below 1%, and the August 14–17 average settled at 0.52%: an 86.4% relative drop in four days (Promptwatch). Two independent panels measured the same event: Qwairy recorded a 95% fall, and Peec AI recorded 88% — and found Reddit was not alone, with arXiv down 84% and YouTube down 78% over the same window.

The mechanism is what makes it a GEO lesson rather than a Reddit story. On August 8, ChatGPT's query fan-out changed: the share of fan-out queries using the site: operator jumped from 0.37% to 16.8% — roughly a 46× increase in a single day — while the average number of searches per response nearly doubled, from 1.08 to 1.83. The domain-scoped searches were added on top of the generic ones rather than replacing them. Crucially, Google's surfaces showed a different shape entirely: AI Overviews declined only 11.3% (2.37% → 2.10%) and AI Mode 30.5% (2.22% → 1.54%), both gradually, with no cliff. Reddit did nothing wrong, was being added to the S&P 500 that same week, and still lost almost its entire ChatGPT citation footprint in four days.

"Reddit's experience exposes GEO's original sin, the same one SEO has always had: you are building on ground you do not control. When a single model provider adjusts how it selects sources, your visibility can be rewritten overnight, with no notice and no explanation."

Read alongside the Princeton paper, this clarifies exactly what the +41% does and does not buy you. Princeton measures share of answer text, conditional on your page already being in the retrieved pool. The August collapse happened one stage earlier — at source selection, before any page-level optimization could matter. Both facts are true simultaneously, and confusing them is the most common strategic error in GEO today. The full breakdown is in our dedicated analysis, AI Citation Volatility 2026.

September 2026: the replication that finally measures the gap

The strongest 2026 evidence against over-reading the nine strategies is a September 7, 2026 arXiv preprint by Bajemon and Rochet. It did two things the original paper never did: it built a deterministic, query-blind GEO score from 11 text features (entropy, information and entity density, semantic coherence, self-containment, citation F1, statistic density, MMR, nDCG, redundancy and quotable density), and it re-tested the three headline interventions as volume-controlled paired edits. The result is the cleanest quantification yet of where page-level optimization actually sits.

Source: Bajemon & Rochet, arXiv preprint (September 7, 2026); Aggarwal, Dugan et al., "GEO: Generative Engine Optimization" (KDD 2024).

Two readings matter, and the difference between them is the whole lesson. The 0.114 → 0.37 jump shows that query context carries roughly three times more signal than page formatting — exactly the semantic-alignment share AgentGEO identified. The flat-to-negative intervention results do not prove the tactics are worthless; they prove that adding volume without adding relevance does nothing. A statistic bolted onto a passage that was never about that statistic does not improve query fit. Note also that the GEO score's own authors state the paper is a preprint, not peer reviewed, and that two of three confidence intervals include zero. The honest summary is narrower than either camp claims: page-level tactics are conditional, and query relevance is the condition.

The other side of the coin: defending the pipeline from malicious GEO

There is a symmetry the practitioner literature usually misses. If the nine tactics are weak levers for earning citations, they are demonstrably strong levers for manufacturing them — and by September 2026 the academic response had arrived. A September 2, 2026 arXiv paper, GEO Defender (Li, Shao, Lin, Guan, Zhou & Shi, arXiv:2609.02964), builds a two-stage defence against adversarial GEO rewrites and reports reducing the average attack success rate from 50.32% to 6.20% while retaining 94.12% of benign-evidence use.

The finding that should worry every content team is not the defence number — it is why four off-the-shelf safety systems failed. The authors note that attack documents remain factually consistent with their originals, which defeats fact-verification and perplexity filtering, and that the features they amplify are the same ones that characterise high-quality benign content. In other words: the very signals Princeton measured as quality are indistinguishable from the signals of manipulation. That is a structural property of the tactic set, not a bug in one paper.

Practical application: a 5-step content checklist from the Princeton study

Based on the Princeton findings and their 2026 validations, here is a prioritized checklist for content optimization:

1. 1.
   Add expert quotations (+41% lift) — Interview domain experts with full attribution. Use blockquote formatting with named sources. GrackerAI 2026 reports that attributed quotes increase AI citation probability by 3.7×, the single largest structured data multiplier.
2. 2.
   Replace vague claims with statistics (+33% lift) — Every assertion should include a verifiable number with source and date. Authoritas 2026 found that pages with original statistical data achieve +156% higher citation probability in Google AI Overviews.
3. 3.
   Optimize readability (+29% lift) — Remove redundancy, ensure clear subject-verb structure, limit sentences to 25 words. The fluency + statistics combination outperforms individual strategies by +5.5% (Princeton paper).
4. 4.
   Cite authoritative sources (+28% lift) — Link every factual claim to academic papers, official statistics, or recognized industry reports. University of Toronto 2025 research (arXiv:2509.08919) found that brands cited on third-party platforms receive 6.5× more AI citations than those appearing only on owned domains.
5. 5.
   Stop keyword stuffing (−8% penalty) — AI engines use semantic embedding matching, not keyword counting. BrightEdge 2026 data confirms natural writing outperforms keyword-optimized content for AI citation. Use synonyms and related entities instead.

September 2026: the failure-mode taxonomy that reorders the nine tactics

The most useful thing published about the Princeton tactics in 2026 is not a replication — it is a study asking a different question. Instead of "which edit produces the biggest lift in a controlled test," the AgentGEO paper asked: when content fails to get cited at all, what is actually going wrong? The answer reframes the whole nine-strategy list.

Source: AgentGEO academic taxonomy of AI citation failure modes (2026). Read alongside the Princeton nine-strategy list, which is a content-quality intervention set.

Put the two side by side and the mismatch is obvious. Statistics addition and source citation are content-quality interventions, so they can only ever operate inside the 27.1% slice — and even there they are one sub-tactic, not the category. The single largest reason content is skipped is semantic alignment: the page answers a question nobody is currently asking, in the wrong framing, or with data that has gone stale. No number bolted onto a paragraph fixes that.

The same 2026 literature also clarifies where Princeton's testbed sits in the pipeline. Sequenced, four decisions happen before any page-level tactic can matter:

1. 1.
   Does the engine search at all? Profound found Claude invokes live web search for only 36.6% of tested prompts. For the rest there is no citation opportunity to win, regardless of how well the page is optimized.
2. 2.
   Which sources get selected from what is retrieved? DEJAN measured OpenAI selecting Reddit as a source in just 0.61% of retrieved candidates after the August 2026 change — a selection decision, not a content-quality one.
3. 3.
   Why does retrieved content fail to get cited? This is AgentGEO's stage, and it is dominated by semantic alignment at 62.2%.
4. 4.
   Was the recommendation already set? Seer Interactive's ghost-citation work found a roughly 5× citation-rate lift once a brand is already the chosen recommendation — meaning some citation decisions were effectively settled before retrieval began.

"Princeton tells you what a page-level edit is worth once the page has cleared every earlier stage. AgentGEO tells you that most pages never clear those stages. Both are correct, and reading only the first one is why so much GEO effort lands on the wrong 10% of the problem."

None of this makes the nine tactics wrong — it makes them conditional. Applied to a page that is retrieved, semantically aligned, and in the consideration set, statistics (+33%), quotations (+41%), and citations (+28%) remain the best-measured levers available. Applied to a page that fails stage 1 or stage 3, they change nothing. Diagnose the stage first, then spend the tactic.

Two further 2026 data points belong in that diagnosis. A ConvertMate benchmark across 12,500 queries and 8,000 domains found 83% of AI Overview citations go to pages outside the organic top 10, with AI search traffic converting at roughly 4.4× organic — the upside for pages that do clear retrieval is real. And Ahrefs' 75,000-brand analysis found unlinked brand mentions correlating 0.656–0.709 with AI visibility while raw backlinks sat near zero, which is the off-site counterpart to the same lesson: being retrieved and being named are upstream of being well-formatted.

October 2026: the trust-tier finding that explains why the same tactic lifts one page and not another

The single most useful 2026 addition to the Princeton picture is not another lift percentage. It is the discovery that identical optimization produces radically different results depending on where the page already sits in the engine's trust hierarchy.

A 2026 preprint analyzing 21,143 valid citations separated citation presence from citation absorption — the first rigorous attempt to measure not just whether a page was listed, but whether the generated answer actually used its content. Two results matter:

The transparency caveat is important and belongs next to the numbers: this is a preprint, the influence construct is a proxy rather than a measured business outcome, and it does not establish that a citation causes a visit or a sale. Read it as a re-weighting of the Princeton tactics, not a refutation.

"The controlled KDD 2024 study tested nine optimization methods on a 10,000-query benchmark and reported visibility gains of up to 40%. These historical experimental results do not establish current-engine performance."

That disciplinary point is worth adopting directly: several widely repeated GEO statistics (an $886 million market size, a 4.4× conversion claim, a 2.8× citation uplift from heading hierarchy, 80% list usage) were withdrawn in that review because no traceable primary denominator could be found. The Princeton numbers survive the same scrutiny because the study publishes its benchmark, its query sources and its evaluation prompt. Prefer the primary source over the restatement — and expect the restated figures you see in competitor roundups to keep drifting.

The operational reading for a GEO program in Q4 2026 is therefore a three-stage filter, not a nine-item checklist. First, establish that the page absorbs — is it in the trusted tier for this topic, or is it a page that gets listed and ignored? Second, apply the Princeton levers only where absorption is plausible: statistics, named-source citations and quotations are evidence equipment, and evidence equipment on a page nobody quotes is inventory. Third, rebuild the off-site record, because 82% of commercial-intent citations resolve to third-party sources that must decide to name you before you can be absorbed.

Study limitations

While the Princeton study is the methodological gold standard in GEO, several limitations merit attention:

• ▸ The experiment used a single generative engine (an open-source GPT-3.5 system), not production ChatGPT Search or Perplexity. Different AI engines may exhibit different citation behaviors.
• ▸ The 9 strategies were tested individually, not in combination (except the fluency + statistics pair). In practice, multiple strategies interact, and the combined effects may differ from individual lifts.
• ▸ 10,000 queries, while substantial, may not represent highly niche or industry-specific search intents. Results may vary by vertical.
• ▸ The benchmark predates the ChatGPT Search launch (October 2024), Google AI Mode (March 2025), and the 2026 shifts in market dynamics. Subsequent validations address these gaps but do not replicate the controlled experimental design.

The directional conclusions should be treated as reliable guidance rather than precise prediction formulas. The convergence of independent validations — Conductor, GrackerAI, BrightEdge, Authoritas, Axis Intelligence — provides confidence in the overall framework while acknowledging that exact lift percentages vary by platform, vertical, and content type.

Frequently asked questions