The Princeton dividend is boring

The most useful research result in AI visibility was presented at a data-mining conference in Barcelona in August 2024, and almost nobody selling AI visibility talks about it. Not because it is obscure. It is peer-reviewed, free to read, and built on ten thousand queries. It goes unmentioned because it is bad for the mystery, and the mystery is what most of this industry sells. What the paper actually found is a spec. And the spec is boring.

This is a research brief on a single study: GEO: Generative Engine Optimization, by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande (a collaboration spanning IIT Delhi and Princeton), presented at KDD 2024. To our knowledge it remains the only peer-reviewed, large-scale, controlled measurement of what moves a source's visibility inside AI-generated answers. One paper is a thin foundation for an industry. It is also the best foundation anyone has, and honesty starts with saying both things out loud.

Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan & Deshpande, "GEO: Generative Engine Optimization," KDD 2024, Barcelona (arXiv:2311.09735).

Section OneTen thousand queries, five competitors, one answer

The experiment is clean enough to describe in a paragraph. The researchers built a benchmark, GEO-bench: 10,000 queries drawn from nine source datasets spanning multiple domains, everything from everyday questions to specialist ones. For each query they took the top five Google results as the competing sources, and had a generative engine (an AI system that reads sources and composes a cited answer) write its response. Five sources go in. One answer comes out. Every sentence of that answer leans on somebody.

Then the intervention: take one of the five sources, rewrite it with a single tactic (add quotations, add statistics, add citations, smooth the prose, stuff keywords), leave everything else untouched, and re-run. Measure how much of the answer is now attributed to the rewritten source. The paper's primary metric, Position-Adjusted Word Count, counts the words the engine attributes to you, weighted by how early in the answer they appear. A second metric, Subjective Impression, rates how prominently and favorably the source comes across. Nine tactics, tested one at a time, against the same untouched baseline.

The abstract's headline claim is stated plainly: "GEO can boost visibility by up to 40% in generative engine responses." The interesting part is which tactics carried that number, and which one fell on its face.

Aggarwal et al., KDD 2024, abstract and benchmark section (arXiv:2311.09735).

Section TwoThe boring winners

Ranked by lift in attributed visibility against the untouched baseline, on the paper's controlled benchmark: adding quotations lifted position-adjusted attributed visibility by 42.6 percent (and subjective impression by 28.0 percent), the strongest single tactic in the study. Adding statistics lifted it 32.8 percent (22.8 percent on impression). Fluency optimization, simply making the text easier to parse, lifted it 28.7 percent. Citing sources lifted it 27.7 percent. Further down: technical terms +18.5 percent, easy-to-understand wording +13.8 percent, an authoritative tone +11.8 percent on attribution (though a notable +18.7 percent on impression), unique wording a mere +6.2 percent.

Aggarwal et al., KDD 2024, per-strategy results, PAWC and Subjective Impression vs. baseline (arXiv:2311.09735).

The best-performing tactic in the age of machine answers is the oldest move in journalism: quote somebody, and say who said it.

Look at the top four as a set. Quotations. Statistics. Fluent prose. Named references. This is not a growth hack. It is the grading rubric from a decent journalism course, rediscovered by a benchmark. The tactics that won are the unfashionable ones: receipts, named sources, verifiable numbers. The margin was not subtle.

Section ThreeThe tactic that went negative

One tactic in the study did not merely lag the field. Keyword stuffing (loading the text with more mentions of the query's terms, the reflex trained into a generation of website owners by two decades of search-engine optimization) reduced attributed visibility by 8.7 percent, landing below the untouched baseline. On the impression metric it managed +4.7 percent, the weakest showing in the study. The move that once dragged pages up a ranked list actively repelled the machine that writes answers.

Aggarwal et al., KDD 2024, keyword-stuffing result: −8.7% PAWC vs. baseline, +4.7% SI (arXiv:2311.09735).

Keyword stuffing didn't merely underperform. It fell below doing nothing at all.

The paper also tested combinations, and the finding compounds the point. In the authors' words, "the best combination (Fluency Optimization and Statistics Addition) outperform[s] any single GEO strategy by more than 5.5%." Readable prose carrying verifiable numbers beat every individual tactic, including the quotation champion. The moves stack, and they stack in the direction of substance.

Section FourRead the fine print

Now the part a research brief owes you. These numbers come from a controlled benchmark, not from the open market. Each query had exactly five competing sources, and visibility there is zero-sum: when the rewritten source gains attributed words, the other four must lose them. On the open web your competition is not four documents; relative gains in a five-source arena can look larger than what any single business would measure in the wild. The study was run on the engines and the web of 2023 and 2024, and generative engines have been rebuilt several times since. And a lift in attributed visibility is not a revenue line: the paper measures how much of the answer cites you, not what that citation is worth.

None of that weakens the paper. It defines what the paper is: the one peer-reviewed data point in a field otherwise run on vendor decks and vibes. A study, not a guarantee, and still the only playbook with a methods section.

Section FiveThe dividend: a spec, not a mystery

Here is what the study pays out, and why we call it a dividend. AI visibility, as sold, is a mystery: opaque models, secret rankings, a consultant between you and the machine. AI visibility, as measured, is a content specification. Give the machine quotable, attributable matter. Quote real people by name. State numbers with units and origins. Cite what you cite. Write prose a parser doesn't stumble over. Stop repeating your keywords at it. That is the entire peer-reviewed playbook, and every item on it is checkable by reading your own page.

A repeatable spec is worth more than a secret, because a spec compounds. It can be applied to every page you ship, checked in review, taught to a writer, or to an agent, in an afternoon. The mystery, by contrast, renews monthly and invoices accordingly.

You cannot game your way into a machine's answer. You can only become the easiest thing in the room to cite.

And a note you may have already noticed: this article is written to the spec it describes. Direct quotations from the paper, with attribution. Verified statistics, each with its window and its caveat. A named source with a live link. Fluent, parseable prose. No keyword got stuffed in the making of this brief. We publish this way not as a flourish but because we are our own first client. Every method we describe runs on our own pages before it is offered to anyone else's.

In ClosingBoring is a strategy

Everything we publish reduces to one claim: machines can only act on what they can clearly read. That holds for the machines answering your customers' questions outside your company, and now for the agents doing the work inside it. The Princeton study is the cleanest evidence yet for the outside half. The engine did not reward the loudest source or the most optimized one. It rewarded the source that arrived legible: quoted, sourced, numbered, parseable. Visibility went to whoever was easiest to verify.

That is good news for small businesses, if they act on it. The winning moves require no budget, no tooling, and no consultant. They require only the discipline to write receipts-first, which most competitors will skip precisely because it is boring. If you want to know how visible your own pages are to the machines that now answer for you, you can measure it at /signal-index/, or write to us at /contact/.