Generative engine optimization (GEO). Nine rewrites measured by one study, a follow-up that cuts many of them down to size, and a sheet to test them on your own page.
GEO is sold as the new SEO, but the study that named it measured it on a test engine. Here is what held up, and the Excel sheet to try it.
Generative engine optimization is work on a site’s text so that it gets picked up and cited in the answers of generative engines. The term was coined in 2023 by the study of Aggarwal and others, presented at KDD 2024.
The study tried nine rewrites on 10,000 questions, with an engine built for the test. Quotations, statistics and sources raised visibility by up to 41%; keyword stuffing lowered it. A 2025 follow-up, C-SEO Bench, finds most of these methods ineffective.
Google says that for its AI features SEO still applies, and Microsoft Advertising calls it the foundation to build on. For ChatGPT, Copilot and AI Overviews the basics remain the index and checkable facts.
The Excel sheet can be downloaded below, with a worked example and a PDF of the example, with no email required.
Written by: Mattia Esposito, AI and marketing consultant, founder of Itria AI. From Puglia, based in Turin since 2023, he works remotely in Italy and abroad. LinkedIn profile.
The sheet for rewriting a page with the study’s techniques, to download
An Excel sheet with one row per paragraph: today’s text, the technique chosen, the source, the figure or quotation you add and the new text. A second sheet records ten test questions before and after, a third lists the nine techniques with the study’s numbers. The example is filled in for a made-up ceramics company.

| File | Contents | Link |
|---|---|---|
| Excelto fill in, in Italian | Rewrite, Test, Techniques, plus the two example sheets. Words and citations are counted automatically. | |
| PDFthe example, in Italian | The worked example, to read before you start. |
| Column | What you enter | Technique from the study |
|---|---|---|
| Paragraph before | The text as it is today, one paragraph per row. | Baseline, 19.3 |
| Technique | One choice from the drop-down. To know what worked, one at a time. | Table 1 |
| Source | The document or body the fact comes from, with a link to the exact page. | Cited sources, 24.6 |
| Statistic | The figure, the unit and where it comes from: a test, a register, a report. | Statistics, 25.2 |
| Quotation | Who said it, the sentence and where it was published. Never invented. | Quotations, 27.2 |
| Paragraph after | The new text, written fluently, without repeating keywords. | Fluent writing, 24.7 |
Before: Our porcelain stoneware is very hard-wearing and absorbs almost no water, so it’s perfect for outdoor use too.
After: Our porcelain stoneware absorbs 0.08% water in the ISO 10545-3 test (March 2026 batch). The EN 14411 standard places tiles below 0.5% in group BIa.
On 7 September 2026 I opened the websites of 143 Italian companies that also sell abroad and looked for the file that describes a company to AI assistants, llms.txt. This is how I put it that day: «Twenty-eight out of a hundred and forty-three have the file that explains the company to an assistant, and in twenty-four a plugin wrote it.» The full post.
The easiest file to add is also the one Google’s guide to AI features says you can ignore. My site already had it on 7 August, and ChatGPT still couldn’t find me. Afterwards I wrote pages, 78 in five weeks, each on one question, and impressions in Search Console went from 24 to 1,222.
My site’s figures show that the pages were read and cited, not which technique counted most. That’s why the sheet asks for one technique per row and the same questions before and after: so you can see what works on your page.
From one page to five. On the morning of 25 September 2026, for Claude’s web search, itria.io had a single page; by the evening, after asking Brave to recrawl seven, it had five. Before and after, with the limits of the data, in Brave and Claude.
The email address that wasn’t there. Cloudflare’s email obfuscation removed the address from the HTML: to an assistant, the contact page had no contact details. The same flaw turned up on 14 out of 121 producers’ sites, in the piece on the email address AI assistants can’t see.
The result. Between 11 October and 7 November 2026 Bing’s AI Performance report counted 268 citations of itria.io pages in Copilot, and Search Console 3,550 impressions in AI Overviews and AI Mode.
What generative engine optimization is
Generative engine optimization, GEO, is the set of changes to a site’s content that make it appear, quoted and cited, in the answers written by generative engines. The term was proposed in November 2023 by six researchers from Princeton and IIT Delhi, in a study later presented at KDD 2024, in Barcelona.
A generative engine, in the study’s definition, first retrieves a few sources from the web and then has a language model write an answer that cites them. This is the retrieval-augmented generation pattern, explained on the page about RAG or fine-tuning.
Microsoft Advertising, in its GEO guide of 6 January 2026, describes it as optimising content for generative search environments so that it is «discoverable, trustworthy, and authoritative». The technical terms are in the LLM glossary. The rules don’t change between Italy and the English-speaking world; the study’s benchmark, though, is in English.
The Aggarwal study: how it was measured
The study built a benchmark of 10,000 questions, GEO-bench, taken from nine different collections and split into 25 domains, from art to health. 80% are informational questions, 10% transactional and 10% navigational, as in real searches.
The engine used is neither ChatGPT nor Gemini. For each question it takes Google’s top five sources and has gpt-3.5-turbo write the answer, five times, to reduce chance. Then a single source is rewritten with one technique and the space it gains in the answer is measured.
There are two measures. Position-adjusted word count counts how many of the source’s words end up in the answer and how high. Subjective impression combines seven judgements made by GPT-3.5, including the likelihood that a reader would click.
The nine GEO techniques and how much they counted
Quotations raised visibility from 19.3 to 27.2, 41% more, and are the study’s best technique. Statistics and more fluent writing follow. In brackets is the subjective impression, on the same baseline.
| Technique (name in the study) | What you do to the page | Visibility (impression) |
|---|---|---|
| None | The page as it is. | 19.3 (19.3) |
| QuotationsQuotation Addition | Sentences attributed to a named person or organisation. | 27.2 (24.7) |
| StatisticsStatistics Addition | A figure with its origin instead of an adjective. | 25.2 (23.7) |
| More fluent writingFluency Optimization | The same content, written more smoothly. | 24.7 (21.9) |
| Cited sourcesCite Sources | The document every fact comes from. | 24.6 (21.9) |
| Technical termsTechnical Terms | The exact name instead of the generic one. | 22.7 (21.4) |
| Easier to understandEasy-to-Understand | Simpler sentences, less jargon. | 22.0 (20.5) |
| Persuasive toneAuthoritative | The same text in a more assertive tone. | 21.3 (22.9) |
| Unusual wordsUnique Words | Rarer, more varied vocabulary. | 20.5 (20.4) |
| Keyword stuffingKeyword Stuffing | The search words repeated several times. | 17.7 (20.2) |
Two things stand out in this table. SEO’s oldest technique, keyword stuffing, is the only one below the baseline. A persuasive tone, which you might expect a model that follows instructions to reward, does not improve things significantly, according to the authors.
Among pairs of the four best techniques, tested on 200 questions, fluent writing plus statistics wins: more than 5.5% above any single technique. Cited sources perform far better in a pair than on their own.
On a real engine, Perplexity, with 200 questions, quotations brought 22% more and keyword stuffing went 10% below the baseline. This was the study’s only check on a real product, in 2024.
Who gains most: position and sector
The gain depends on where the page starts: with cited sources, the site ranked fifth on Google gained 115% visibility and the first lost 30% on average. Sources all rewritten together. This is the study’s table 2, which simulates a world where everyone does GEO.
| Technique | Site ranked first on Google | Site ranked fifth on Google |
|---|---|---|
| Cited sources | Loses 30.3% visibility. | +115.1% |
| Quotations | Loses 22.9%. | +99.7% |
| Statistics | Loses 20.6%. | +97.9% |
| Persuasive tone | Loses 6.0%. | +6.1% |
| More fluent writing | Loses 2.0%. | +2.2% |
The authors draw an optimistic reading: sites without links or fame can win back space. But the engine read only five sources, and anyone outside Google’s top five never got into the game.
Sector matters too. In table 3, statistics perform best in law and government questions, debates and opinions; quotations in people and society, explanations and history; cited sources in factual statements.
The critical follow-up: C-SEO Bench
In 2025 a group of researchers built C-SEO Bench, a broader benchmark to test these methods, and found that most of them don’t work. The paper, accepted at NeurIPS 2025 in the datasets track, tests two tasks, question answering and product recommendation, with three domains each.
The abstract is blunt: «most current C-SEO methods are not only largely ineffective but also frequently have a negative impact on document ranking». They often push the document down. Classic SEO strategies, the ones that lift the source within the model’s context, work better.
There is a second, less comfortable result: the more players adopt the same techniques, the smaller the gain. The authors describe it as a crowded, zero-sum game.
A University of Toronto preprint from September 2025 adds that the AI engines tested strongly prefer earned media, meaning authoritative third-party sources, over brand content and social media. The paper is a preprint, not yet peer-reviewed.
SEO and GEO: what stays the same and what changes
SEO gets a page into the index and up the rankings; GEO works on what a model takes from a page it has already found. Both paths go through the same index, which is why the second depends on the first: what many call SEO for AI, or GEO SEO, starts with SEO. The diagram below shows where each one acts.
| Point | SEO | GEO |
|---|---|---|
| Where the game is played | A ranking of links. | A written answer, with a few cited sources. |
| To get in | Open crawlers and an indexed page. | The same: no index, no citation. |
| What gets chosen | The whole page. | The passage the model picks up. |
| What is measured | Position, impressions, clicks. | Citations and words picked up in the answer. |
| Keywords | They match the page to the search. | Repeating them goes below baseline: 17.7 against 19.3. |
| Authority | Inbound links and domain history. | Sources, figures and quotations in the text; citations from third parties. |
| Who gains | Those already at the top. | In the study, the fifth-ranked site more than the first. |
| Where to read it | Search Console, Bing Webmaster Tools. | The same dashboards: the AI report and AI Performance. |
Does GEO replace SEO? On the documents read, no. AI Overviews and AI Mode draw on Google’s index, Copilot on Bing’s, and the rules explained in the guide on how to get a website indexed on Google apply. That’s why SEO is still relevant in 2026.
Google’s position: it’s still SEO
Google Search Central, in its 15 May 2026 guide to generative AI features, names AEO and GEO and brings them back to SEO. This is the sentence: «From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.»
The same guide, updated on 10 July 2026, lists tactics Google Search says can be ignored: chunking text into blocks for AI, creating files such as llms.txt, chasing inauthentic mentions. Instead it asks for original content that offers something beyond common knowledge. llms.txt has its own page: what llms.txt is and who reads it.
Keyword variants matter less, Google explains, because its AI features break a question down into several related searches. The guide’s limit is clear: it applies to Google Search, not to ChatGPT, Claude or Perplexity. For AI Overviews the practical rules are on the page about how to appear in AI Overviews.
Microsoft’s position: SEO as the foundation
Microsoft Advertising, in its January 2026 guide «From Discovery to Influence», writes that traditional SEO remains essential because assistants search the web in real time. GEO is built on top, with product data, crawled pages and information beyond the site.
The document is a guide for advertisers, aimed at online shops, not Bing technical documentation. To show up in Copilot, Bing’s index is what counts, with the registration explained on the page about Bing Webmaster Tools and IndexNow.
AEO, answer engine optimization, and GEO: are they the same thing?
No official body and no peer-reviewed study gives a shared definition that separates AEO from GEO. Google names them together, answer engine optimization and generative engine optimization, as terms you may come across, and treats both as SEO. The 2023 paper uses only GEO.
Does GEO really exist? As a research term, yes: it has a peer-reviewed study, a follow-up and a Microsoft definition. As a discipline separate from SEO, the official documents read say no. Acronyms such as LLMO have no definitions in official or academic sources.
GEO strategies that stand up to the sources
The study, its critical follow-up and the engines agree on five strategies, and the first three are SEO. Open the doors to search crawlers (how to show up in ChatGPT). Be in Google’s and Bing’s indexes. Put the text in the HTML, readable without scripts.
Then write checkable facts: figures with their origin, sources linked to the document, attributed quotations. They are the study’s best techniques and also the original content Google asks for. Finally, get cited by others: third-party sources, trade press, associations, which according to the Toronto work count for more than the brand’s own site.
To avoid: keyword stuffing, buying mentions, inserting text written to manipulate the model. The GEO study separates its techniques from the adversarial sequences described by Kumar and Lakkaraju of Harvard, which can shift a model’s recommendations.
Tools to measure GEO
The free tools to measure GEO are three, and they sit in dashboards a site already uses. Search Console has a report on performance in Google’s generative AI features. Bing Webmaster Tools has the AI Performance report, which counts citations in Copilot. ChatGPT adds utm_source=chatgpt.com to the links it shows.
None of the three says why a page gets cited. For that you need an experiment: same questions, same engine, before and after, repeated several times because answers vary. The Test sheet does exactly this. How to build it for a real company is on the page about presence on search engines and AI assistants.
Questions and answers
What is generative engine optimization, in short?
It is work on a site’s content so that it appears, cited, in the answers written by generative engines such as ChatGPT, Copilot, Perplexity or Google’s AI Overviews. The name comes from a Princeton and IIT Delhi study presented at KDD 2024, which measured nine ways of rewriting a page.
Does GEO replace SEO?
No. Generative engines first look for sources in an index, and without indexing a page cannot be cited. Google says optimising for its AI features is still SEO; Microsoft Advertising calls SEO the foundation to build on. GEO works on the last stretch, where a model chooses what to cite.
Which GEO techniques work?
In the original study, quotations, statistics and cited sources performed best: up to 41% more visibility on an engine built for the test. Keyword stuffing fell below the baseline. A 2025 follow-up, C-SEO Bench, finds most methods of this kind ineffective.
What is the difference between GEO and AEO?
AEO stands for answer engine optimization, GEO for generative engine optimization. No official body or peer-reviewed study gives a shared definition that separates them; Google names them together and treats both as SEO. GEO has a formal definition in the KDD 2024 study and one from Microsoft Advertising.
Does SEO still matter in 2026?
Yes. AI Overviews and AI Mode draw on Google’s index, Copilot on Bing’s, and ChatGPT search uses a crawler you have to let in. The C-SEO Bench follow-up finds strategies that improve a source’s rank, classic SEO, more effective than rewrites aimed only at the model.
Does generative engine optimization actually work?
The evidence is mixed. In the Aggarwal study, on 10,000 queries, quotations raised a source’s visibility from 19.3 to 27.2 and keyword stuffing lowered it; on Perplexity, with 200 queries, quotations brought 22% more.
In 2025 C-SEO Bench, accepted at NeurIPS, found that most of these methods do not work and that classic SEO strategies work better. What holds is the foundation: pages engines can read, facts with sources, attributed quotations, and third-party sources that talk about you.
Notes on sources
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan, Deshpande, GEO: Generative Engine Optimization, KDD 2024 (doi 10.1145/3637528.3671900), CC BY 4.0 licence. Reread on the v3 PDF on 10 November 2026: section 3 (engine and GEO-bench), table 1 (the nine techniques), table 2 (by position), table 3 (by sector), figure 4 (pairs), table 5 (Perplexity).
- Puerto, Gubri, Green, Oh, Yun, C-SEO Bench: Does Conversational SEO Work?, NeurIPS 2025 Datasets and Benchmarks, CC BY 4.0 licence: abstract read on 10 November 2026, source of the quotation.
- Chen, Wang, Chen, Koudas, Generative Engine Optimization: How to Dominate AI Search, University of Toronto, preprint of 10 September 2025: abstract read on 10 November 2026.
- Google Search Central, Optimizing your website for generative AI features on Google Search, 15 May 2026, updated 10 July 2026, and AI features and your website, content licensed under CC BY 4.0: source of the quotation.
- Microsoft Advertising, From Discovery to Influence: A Guide to GEO, 6 January 2026, with the 16-page PDF the definition comes from.
- Bing Webmaster Blog, Introducing AI Performance in Bing Webmaster Tools, 10 February 2026. OpenAI, Overview of OpenAI Crawlers, for utm_source=chatgpt.com.
- Kumar and Lakkaraju, Manipulating Large Language Models to Increase Product Visibility, Harvard, 2024: cited in the GEO study as adversarial work.
- The itria.io figures come from Search Console (generative AI report) and Bing Webmaster Tools, window 11 October to 7 November 2026. The llms.txt count is from 7 September 2026, across 143 sites.
Techniques are tested on a page. Deciding which questions to own is the work before. Positioning is built starting from the market.
Mattia Esposito, who wrote this guide, also works with Itria on positioning: in-depth market analysis to find where competitors are exposed and take that ground. He did it on his own site first. In August, asked who had founded itria.io, ChatGPT couldn’t find him. From 11 October to 7 November 2026 itria.io pages appeared 13,803 times in Google and Bing results and 3,818 times in AI answers: Google’s AI Overviews and AI Mode, and Microsoft Copilot.