The short answer
There is exactly one academic paper behind GEO. GEO: Generative Engine Optimization by Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, presented at KDD 2024 in Barcelona and funded under NSF grant 2107048.
It is a good paper. It also does not say most of what it gets quoted as saying. The famous "up to 40%" is a relative gain on a metric the authors invented, measured on a generative engine the authors built themselves, and it never touched Google AI Overviews.
What GEO-bench actually is
GEO-bench is 10,000 queries, split 8,000 for training, 1,000 for validation and 1,000 for test. The reported results are on the test split, averaged over five random seeds. The authors published the code and data, which is more than any vendor selling GEO services has done.
The queries come from nine sources: MS MARCO, ORCAS-1, Natural Questions, AllSouls essay questions, LIMA, Davinci-Debate, Perplexity's Discover feed, the ELI5 subreddit, and a set generated by GPT-4 to fill gaps in the distribution. They span 25 domains. The mix is deliberately weighted to real query distribution: roughly 80% informational, 10% transactional, 10% navigational.
Each query is paired with the cleaned text of the top five Google results. Those five documents are the corpus the engine answers from.
What engine they measured
The main experiments did not run on a commercial product. The authors built a two-step engine: fetch the top five Google results, then have gpt-3.5-turbo write a response grounded in them. They sampled five responses per query at temperature 0.7 to reduce variance. Section 6 repeats a subset of the methods on Perplexity, which is a real deployed engine.
That is a reasonable research design. It is not a claim about how Google composes an AI Overview. Google's AI Overviews rolled out broadly in May 2024, after this work was done, and the paper never mentions them.
What "visibility" means in the paper
Two metrics, both invented for the study.
Position-Adjusted Word Count. Count the words in sentences that cite a source, as a share of the whole response, then discount each citation by how far down the response it appears. It measures how much of the answer is yours and how early it lands.
Subjective Impression. Seven aspects, including relevance, influence, uniqueness, position, count, likelihood of a click and diversity, each scored by GPT-3.5 using the G-Eval method. A language model grading a language model's output.
Neither metric is clicks, traffic, conversions, or even citation count. If you have been told GEO tactics deliver a 40% traffic lift, that is not a misreading of this paper. It is unrelated to it.
The results, with the baseline attached
Nine content edits, each applied to one randomly chosen source per query, held constant across methods. Absolute scores, baseline included. Higher is better.
| Method | Position-Adjusted Word Count | Subjective Impression |
|---|---|---|
| No optimization (baseline) | 19.3 | 19.3 |
| Quotation Addition | 27.2 | 24.7 |
| Statistics Addition | 25.2 | 23.7 |
| Fluency Optimization | 24.7 | 21.9 |
| Cite Sources | 24.6 | 21.9 |
| Technical Terms | 22.7 | 21.4 |
| Easy-to-Understand | 22.0 | 20.5 |
| Authoritative | 21.3 | 22.9 |
| Unique Words | 20.5 | 20.4 |
| Keyword Stuffing | 17.7 | 20.2 |
The paper's caption on that table: "The best methods improve upon baseline by 41% and 28% on Position-Adjusted Word Count and Subjective Impression respectively."
Two results in that table rarely make it into the posts citing this paper.
Keyword stuffing scored below doing nothing. 17.7 against a 19.3 baseline. The one tactic borrowed straight from 2005 SEO is the one tactic that made the source less visible. On Perplexity it was 10% worse than baseline.
Writing more authoritatively barely moved the objective metric. Authoritative reached 21.3 on Position-Adjusted Word Count, behind plain fluency editing. The authors note they found no significant improvement from persuasive tone and conclude generative engines are "already somewhat robust to such changes".
The finding that should bother you
Table 4 of the paper shows worked examples of the winning edits. Its caption reads: "Without adding any substantial new information, GEO methods significantly increase the visibility of the source content."
The Cite Sources example scored a 132.4% relative gain by appending, to a sentence about Swiss chocolate consumption, the phrase "According to a survey conducted by The International Chocolate Consumption Research Group". The Statistics Addition example gained 65.5% by inserting "with a staggering 70% increase in robotic involvement in the last decade" into a passage about robots and work.
Both edits were written by a language model told to add citations and statistics. The paper does not claim that either the survey or the 70% figure is real, and its own caption says no substantial new information was added. What the metric rewards is text that looks sourced, whether or not it is.
So the honest translation of "add statistics and citations to gain visibility" is: real statistics and real citations, or you are publishing fabrications to move a machine's attention. We are not going to pretend that distinction is a footnote. It is the whole difference between a research finding and a bad instruction.
Who the gains actually went to
Table 2 breaks the results down by where the optimised source already ranked in Google, in the condition where every source in the set gets optimised at once.
| Method | Rank 1 | Rank 2 | Rank 4 | Rank 5 |
|---|---|---|---|---|
| Cite Sources | -30.3 | 2.5 | 15.5 | 115.1 |
| Quotation Addition | -22.9 | -7.0 | 25.1 | 99.7 |
| Statistics Addition | -20.6 | -3.9 | 10.0 | 97.9 |
Figures are relative change in visibility, as a percentage. Rank 3 is not reported.
The pattern is consistent. Sites sitting fifth roughly doubled their share of the answer. Sites already sitting first lost between a fifth and a third of theirs. The paper frames this as good news for small creators, and on its own terms it is. It also means the headline number is not addressed to you if you already rank first, which is exactly the reader most likely to be sold a GEO retainer.
What happened on a real engine
Section 6 runs a subset on Perplexity. Baseline 24.1 on Position-Adjusted Word Count and 24.7 on Subjective Impression.
| Method | Position-Adjusted Word Count | Subjective Impression |
|---|---|---|
| No optimization | 24.1 | 24.7 |
| Quotation Addition | 29.1 | 32.1 |
| Statistics Addition | 26.2 | 33.9 |
| Keyword Stuffing | 21.9 | 28.1 |
The paper reports Quotation Addition improving the objective metric by 22% here, against 41% for the best method on the authors' own engine. Keyword stuffing again came in below baseline. The direction of every result held on the deployed engine, and the size roughly halved.
The limitations, in the authors' words
Section 9 is short and worth reading in full. It says the methods "may need to adapt over time as GEs evolve, mirroring the evolution of SEO", and that the query distribution will drift. It then states, "owing to the black-box nature of search engine algorithms, we didn't evaluate how GEO methods affect search rankings".
That last clause matters commercially. Nobody tested whether these edits cost you organic position while gaining answer share. The authors argue the changes are textual rather than structural and so unlikely to hurt rankings, and they are careful to call that an expectation rather than a result.
What survives for a real site
Four things, none of them exotic, and none of them a separate discipline. Google's own AI features documentation says there are "no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary", which is worth holding next to any GEO proposal you are quoted for.
- Quote and cite accurately, and name the source in the sentence. The strongest effect in the paper, and the only version of it that is safe to run is the honest one. If you are adding schema alongside it, Article schema's required and recommended fields are the part Google actually reads.
- Edit for fluency. Plain readability editing beat writing more authoritatively on the objective metric.
- Do not stuff keywords. Measured worse than doing nothing on both engines tested.
- Get retrieved before you get clever. Every result in this paper starts from a source already in the top five. If a crawler cannot fetch or render your page, none of it applies. Check what your robots.txt is telling each bot with the AI crawler list and, if you serve agents directly, the well-known files they look for.
Table 2 also makes a point the paper does not dwell on. Ranking first and being cited are separate outcomes, and here they moved in opposite directions. We pulled that apart in ranking first does not mean cited.
What we measured
We ask Google buyer-intent questions live and read back the AI Overview with its citation URLs. Bright Data lists that call at $0.0015, or $1.50 per thousand, which is list price rather than a figure we diffed against our own balance. What the rest of the data costs is broken down in what SEO data actually costs.
A nine question scan of stripe.com cost $0.0135 and showed stripe.com cited in 44% of the questions Google answered. Nine questions is a small sample and should be read as a shape, not a rate. On our commercial buyer-intent set, every question returned an AI Overview, well above the 15 to 20 percent the provider documentation suggested. Query type explains that, not a change in Google.
SEOBuilder is our product and it asks seven answer engines, Perplexity among them. Published measurements are still rare, though, and the paper above is one of the few places anyone has put a Perplexity number next to the method that produced it, which is part of why it deserves reading rather than quoting. The full method for doing this yourself is in how to measure AI visibility.