GEO From the Source: What the Paper That Coined the Term Measured, and What Google and Bing Say

Picture a GEO proposal arriving at your company. On the second slide comes the big claim. A Princeton study proved up to 40% more visibility in AI answers. It has a university name and a round number.

The study exists, and it went through a serious conference. What almost nobody tells you is what it measured, on which engine, and under what conditions. This piece takes apart the paper that coined the acronym GEO. Then it sets beside it what Google and Bing write about optimizing for AI. Where the study and the platforms disagree, I show you.

A separate page, on LLMO, GEO, and AEO, covers the fight over the acronyms. The map of the whole subject is in SEO for AI. Here the subject is the evidence.

Where the acronym GEO came from

The acronym was born in an academic paper. “GEO: Generative Engine Optimization” went up on arXiv on November 16, 2023, signed by six authors. ¹ In the first version, they are affiliated with Princeton, Georgia Tech, the Allen Institute for AI, and IIT Delhi. ¹ The paper was accepted to KDD 2024, a data mining conference held in Barcelona. ¹

Aggarwal and his coauthors gave a category its name. A generative engine, in their definition, is a system that retrieves documents and writes an answer grounded in them. ¹ GEO would be the site owner’s work to win space in that answer. ¹

The starting point is a worry about the people who produce content. A ready-made answer spares the reader the visit to the site. And writers have “little to no control over when and how their content is displayed.” ¹

The engine the study built

Think of a test kitchen. Instead of tasting the dish at a restaurant, the researcher sets up a stove at home. She picks the ingredients, controls the heat, and measures the effect of each spice.

The study did that with search. For each question, the authors took the top five results from Google. ¹ They handed the text of those five sites to GPT-3.5-turbo, with instructions to answer and cite the sources. ¹ Then they rewrote one of the five sources and measured again. ¹ This is called a simulated engine.

AI did the rewriting too. A language model got the instruction to apply a recipe to the source text. ¹ There were nine recipes. ¹ Three added material: citing sources, adding a direct quotation, and repeating words from the question. The other six changed the form: authoritative tone, statistics in place of description, fluency, plain language, unusual words, and technical terms. ¹

The questions came from a benchmark the authors built, called GEO-bench. It holds 10,000 questions from nine sources. ¹ Some are real searches from Bing and Google, and some come from a Reddit forum. Others are Oxford essay questions, debate topics, and questions generated by GPT-4. ¹ Eight in ten questions are informational. ¹ The results came from the benchmark’s test split, with 1,000 questions. ¹

What the study called visibility

Picture a pizza divided among the five sites in each answer. Each site’s slice is how much of the AI’s text came from it. A sentence cited at the start of the answer weighs more than one cited at the end. ¹ This is called Position-Adjusted Word Count, the paper’s main metric.

The pizza has a fixed size. The scores of all the sources in one answer always add up to one. ¹ When one slice grows, another shrinks.

There was a second yardstick, called Subjective Impression. Here GPT-3.5 itself acted as the judge. It scored each citation on relevance, influence, and the chance of a click, among other facets. ¹

Notice what both yardsticks leave out. Neither one asks whether the site made it into the top five. The five were already there, chosen by Google.

The recipes that won and the ones that lost

Three recipes pulled ahead: citing sources, adding a direct quotation, and including statistics. In the test, they improved the main metric by 30% to 40%. ¹ The best one was the direct quotation. The rewritten source’s average slice rose from 19.3 to 27.2 points, about 41%. ¹ That is where the “up to 40%” on the slides comes from.

Stuffing the text with words from the question had the opposite effect. The oldest trick in SEO landed below the original text, at 17.7 points. ¹ The authoritative tone brought no significant gain, according to the authors. ¹ More fluent text and plainer language did help. ¹

In one analysis, the authors crossed the gain with the position each source held on Google. Citing sources raised the slice of the fifth-place site by 115.1%. The same recipe cut the top result’s slice by 30.3% on average. ¹ The authors read this as democratization, with the small site gaining ground. ¹ With a pizza of fixed size, that reading has another side. Someone paid for that slice.

The test on Perplexity

The paper has a section called “GEO in the Wild.” There the recipes were tested on Perplexity, a commercial engine. ¹ The decisive detail sits in the appendix. Perplexity did not let the authors choose the source pages. So they uploaded the source text as files and asked for answers based only on them. ¹ The test used 200 questions. ¹

On Perplexity, the direct quotation came out ahead again, 22% above the baseline on the main metric. ¹ The excess of question words came out 10% below. ¹ The result confirms the lab. And it is, once again, a test with the sources handed over. Perplexity’s search chose none of them.

The limits, told by the paper and by those who came after

The paper’s limitations section is short and honest. The authors write that they “didn’t evaluate how GEO methods affect search rankings.” ¹ They warn that the methods “may need to adapt over time as GEs evolve.” ¹ And they risk a prediction. With larger context windows, models would read more sources, and ranking would weigh less. ¹ The paper did not measure that prediction.

In July 2026, Olivier Martinez published a critical survey of 45 studies on GEO. It is a single-author preprint, without peer review. ² His conclusion about the founding paper is direct. The gains hold within the experiment, “conditional on a source already being present in a fixed context.” ² They establish neither organic discoverability nor durable traffic effects. ²

Put the two readings together and three points do not carry over to your site without adjustment. The engine dates from 2023, with GPT-3.5 and five sources, and the authors themselves warn that methods age. ¹ The tested source was already among the five handed to the model before the recipe. ^{1 2} And the yardsticks measure a relative slice of text and a score given by another AI. ¹ Clicks, sales, and citations with the brand’s name stayed out of the count.

What Google says about optimizing for generative AI

Google published a guide to the AI features in Search, last updated July 10, 2026. ³ In it, Google names the acronyms AEO and GEO. And it closes the matter this way: “From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.” ³

The same guide lists what you can ignore for Google Search. Chopping your content into tiny pieces for the AI is not a requirement. ³ Nobody needs to write in a specific way for generative AI search. ³ An llms.txt file and special markup for AI change nothing in Search. ³ And chasing inauthentic mentions across the web helps less than it seems. ³

Another Google page covers third-party tools and advice. It mentions tools sold as “AEO” or “GEO” that promise improvements in AI experiences. ⁴ It says “Google doesn’t evaluate third-party services.” It adds that “using a service or tool doesn’t guarantee ranking success.” ⁴

A limit comes attached, and I repeat it every time I cite that sentence. This is Google talking about its own system, and Google has a stake in the matter. The sentence does not cover ChatGPT, Perplexity, or Copilot. The path in order, with Google’s documentation gate by gate, is in how to show up in AI answers.

What Bing documents

Bing went further on one point. It uses the acronym GEO inside its own Webmaster Guidelines. ⁵ The first item carries a plain title, “SEO Fundamentals Still Apply to Grounding and AI Experiences.” ⁵ The text says “Bing and Copilot search experiences rely on the same core crawling, indexing, and ranking foundation as traditional search.” ⁵

Further down, the page defines the term: “Generative Engine Optimization (GEO) focuses on content eligibility for grounding and reference in AI responses.” ⁵ The next sentence puts SEO and GEO side by side: “SEO does not guarantee rankings or traffic, and GEO does not guarantee grounding or citations in AI experiences.” ⁵

The page includes a list of abuses. “Content should be created for users, not to manipulate ranking systems or trigger citations or AI responses.” ⁵ Attempts to manipulate the language models behind Bing or Copilot may lead to reduced visibility or removal. ⁵ Lightly rewriting someone else’s content “without new insight or perspective is considered scraping.” ⁵

What Bing asks for instead fits in three lines. “Facts and definitions are explicit.” “Key statements do not rely on implied content.” Content that “stands on its own.” ⁵ The page carries no update date, so the note records the day it was read. Your site’s citations in Copilot show up in Bing’s AI Performance report.

Where the study and the platforms disagree

The first disagreement is one of principle. The paper states that traditional SEO methods “are not directly applicable to Generative Engines,” and that new techniques are needed. ¹ Google calls the work SEO. ³ Bing says the fundamentals of SEO still apply. ⁵

This is where almost nobody looked. The paper’s own engine pulled its sources from Google’s top five. ¹ In the test that coined the acronym, getting into the answer depended on ranking in search. The recipe only kicked in after that door. That reading is mine, and it comes from section 3.1 of the paper.

The second disagreement is about writing. In the paper, rewriting the source changed its slice of the answer. ¹ Google says nobody needs to write in a specific way for AI. ³ Bing condemns language engineered to trigger citations. ⁵

There is agreement too, more than the market admits. Keyword stuffing failed in the lab, failed on Perplexity, and sits on Bing’s list of abuses. ^{1 5} The winning recipes were citing sources, bringing numbers, and quoting experts. They resemble the content that goes beyond common knowledge, which Google asks for. ³ They also resemble the explicit facts Bing asks for. ⁵ In my reading of the three sources together, what survived the test is what makes a text easier to check.

What the search for “geo seo” showed in Brazil

I measured the search “geo seo” on September 14, 2026, twice, in Brazil, where we are based. Once in Semrush, in its Brazil database. Once live, on Google in Brazil. ⁶ The term gets 390 searches a month in that database. ⁶ The top ten in Semrush’s snapshot are about Generative Engine Optimization. Live, so were all nine web results. ⁶ With “seo” next to it, the acronym is free of contamination.

The version with “marketing” belongs to someone else. For “geo marketing,” the top ten in Semrush are about location-based marketing. ⁷ That measurement is on the page about search volume traps. Even under “geo seo,” Google’s automatic translation of one English result called GEO “otimização geográfica,” geographic optimization. ⁶

The AI Overview at the top of the live page, written in Portuguese, said GEO uses short sentences, structured data, and verifiable facts. ⁶ It also recommended lists, tables, and summaries that are easy to summarize. ⁶ Compare that with Google’s own guide. It says chopping content is not a requirement, and that structured data is not required for generative AI search. ³ Google’s AI in Search repeated what the market writes. The company’s own manual says something else.

I read the top results, all of them in Portuguese. Alura, a Brazilian online tech school, credits the study to Princeton and Georgia Tech, the affiliations in the first version. It repeats the range of 30% to 40%. ⁸ The text does not describe the engine or the five sources. ⁸ Conversion, a Brazilian SEO agency and the top live result, treats GEO as a subfield of SEO. ⁹ Neither one links to the paper or to Google’s or Bing’s documentation. ^{8 9}

The right question for the next GEO proposal

From now on, when someone shows you the “up to 40%,” you know where it came from. It came from an engine built in 2023, with the sources already chosen by Google. And it measured the text slice of one source out of five. ¹

Ask the seller three questions. On which engine was the number measured? What was measured: a slice of text, a citation with the brand’s name, or a visit? Was my page among the sources consulted before the recipe? If nobody can answer the third, the 2023 number says nothing about your site. And even Bing, which uses the acronym in its guidelines, promises no citations. ⁵

The first step is physical and takes ten minutes. Open the paper on arXiv and read section 3.1, which is half a page long. It describes the whole engine. Then open the proposal you received and check whether it tells the same story.