We read the only peer-reviewed GEO study so you don't have to

If you want ChatGPT, Perplexity or Google's AI answers to cite your website, the evidence says to do three things: quote credible people, include real statistics, and cite your sources. That is the headline finding of GEO: Generative Engine Optimization, the study from researchers at Princeton, IIT Delhi, Georgia Tech and the Allen Institute for AI, peer-reviewed and published at KDD 2024. The authors found those three tactics lifted a website's visibility in AI-generated answers by 30 to 40 per cent, while the classic SEO trick of keyword stuffing did essentially nothing.

Most of what you will read about "GEO" or "AI SEO" is vendors guessing. This paper is the rare exception: a controlled experiment with published methods and numbers. So we read all twelve pages, checked the tables, and here is what it actually says, in plain English. (The practical playbook built on these findings is our guide to getting cited by ChatGPT.)

"AI search does not reward the loudest page. It rewards the best-evidenced one."

Dean Cookson, founder, Operosus

What did the researchers actually test?

The team built a benchmark called GEO-bench: 10,000 real and realistic search queries drawn from nine sources, including anonymised Bing and Google query logs, across 25 domains from health to law to games. For each query, they pulled the top five Google results, rewrote one of those pages using a specific tactic, then had a generative engine answer the query from those sources. They measured whether the rewritten page got more space and more prominent placement in the AI's answer.

Nine tactics were tested. Some added material to the page: statistics, quotations from credible sources, citations. Others changed style only: simpler language, more fluent writing, a more authoritative tone, technical terms, unique words, or extra keywords.

Two important caveats before the results. The answers were generated with GPT-3.5, which is now an old model, and the engines have moved on since 2024. And it is one study. It remains the best evidence available, but treat the percentages as directional, not gospel.

What actually got pages cited?

Three tactics stood out, and all three involve adding substance rather than polishing style:

• Adding quotations from relevant, credible sources was the single best performer in the main experiment.
• Adding statistics, replacing vague qualitative claims with concrete numbers, came a close second.
• Citing sources for the claims on the page came third.

Together these produced relative visibility improvements of 30 to 40 per cent on the study's position-adjusted word count metric, which measures how much of the AI's answer is built from your page and how early it appears.

Style was not worthless. Making the writing more fluent and easier to understand delivered improvements in the 15 to 30 per cent range. Generative engines reward clear writing, just less than they reward evidence.

Crucially, the researchers validated the findings outside the lab. On Perplexity.ai, a real deployed AI search engine, the same tactics produced visibility improvements of up to 37 per cent, with quotation addition again leading.

What failed?

Keyword stuffing, the backbone of bad SEO for two decades, offered little to no improvement in the main experiment and performed 10 per cent worse than baseline on Perplexity. A language model is not matching keywords; it is deciding which source material is worth building an answer from. Repeating your target phrase does not make your page more useful to it.

More surprisingly, an authoritative, persuasive tone showed no significant improvement on average. Sounding confident is not the same as being citable. The model wants verifiable substance: numbers, named sources, direct quotes.

Hence the one-line summary at the top of this piece: the loudest page loses to the best-evidenced one.

Why is this good news for smaller firms?

This is the finding we think UK SMB owners should pay most attention to. The researchers looked at how the benefits were distributed by search ranking. Pages ranked fifth in the underlying search results gained the most: adding citations lifted fifth-ranked sites' visibility by 115.1 per cent, while the average top-ranked site actually lost around 30 per cent of its share of the answer.

Traditional search rankings are heavily influenced by backlinks and domain authority, which large companies accumulate over years and small firms struggle to match. A generative engine reads the actual content. If your page carries better evidence than the incumbent's, the model can prefer it even though Google ranks you fifth. The authors describe this as levelling the playing field for small content creators, and the numbers in their Table 2 back that up.

There is one more practical wrinkle: the best tactic varies by topic. Statistics worked hardest in law, government and opinion-style queries. Quotations did best for history, explanations and people-focused content. If you write in one niche, it is worth matching the tactic to the territory.

What should you change on your site this week?

You do not need new tools to act on this. You need an editorial habit:

1. Put numbers in. Replace "many of our clients see faster turnaround" with a real, sourced figure, from your own data or a named third party. Never invent one; fabricated statistics are a reputational risk with humans and machines alike, and the web already has a sludge problem without your help.
2. Quote real people. A named expert, a customer, a regulator, a published author. Attributed quotations are the strongest single signal the study found.
3. Link to your evidence. Cite primary sources inline, the way this article links to the paper itself rather than to someone's summary of it.
4. Write clearly. Shorter sentences and plainer structure earned a measurable lift on their own.
5. Do it on the pages that answer questions. The study optimised pages that already sat in the top five results for a query. Start with your pages that already rank, where the uplift compounds.

This is also how we build at Operosus. Contentwell, our content system, drafts articles with sourced statistics and primary-source links as a requirement rather than an afterthought, because the evidence says that is what gets content cited. The same principle runs through Bidwell, where tender responses live or die on verifiable claims, and through the client content systems we build: every number traceable to a source, every quote real.

Should you hire a "GEO agency" off the back of this?

Be careful. The study has been widely misquoted, with the 40 per cent figure attached to all sorts of claims it never made. If an agency pitches you guaranteed AI rankings, proprietary GEO scores, or keyword-led tactics rebadged for ChatGPT, the only peer-reviewed evidence we have points the other way. The same scepticism applies to the llms.txt file currently being sold as an AI visibility fix: we checked the log data and it is a dud.

The honest version is less glamorous and more useful: make your pages genuinely better evidenced than your competitors' pages, in clear English, on the topics your buyers ask about. That has always been good marketing. The Princeton study's contribution is showing that the machines now reading the web reward it too, and measurably so.