Reading time: 19 minutes
Generative Engine Optimization (GEO) is the set of practices that helps a brand get cited in AI answers. It applies to ChatGPT, Perplexity, Google AI Overviews, Gemini and Claude. While SEO competes for positions, GEO competes for space in the answer. In 2026, 41.5% of US companies already use GEO, according to The CMO Survey. This guide for agencies and brands separates evidence from hype.
In short
- According to the paper that coined the term, citing sources, quotes and statistics produced a relative visibility gain of 30% to 40%.
- ChatGPT searches with partners such as Bing, Perplexity uses its own index, and Google relies on its Search ranking.
- In Brazil, according to RD Station’s Panorama survey, 36% say their company does not optimize its presence in LLMs; 23% don’t know.
- In Ahrefs studies, 97% of llms.txt files got no requests, and adding schema also brought no meaningful gain in citations.
- None of the 8 Portuguese-language pages that appeared most often in searches for “o que é GEO” (“what is GEO”) explained how AI picks sources.
What is GEO (Generative Engine Optimization)?
GEO is the optimization of content and brand presence for generative AI engines. These systems pull information from several sources and generate a single answer instead of a list of links. So the goal is no longer the top position. It is inclusion in the answer, by name or with a link.
The term comes from the paper “GEO: Generative Engine Optimization” by Pranjal Aggarwal and coauthors from Princeton and IIT Delhi. The authors posted it on arXiv on November 16, 2023, and KDD 2024 later accepted the work.
Semrush sums up the shift:
“The key difference between SEO and GEO is that you aren’t competing to rank at the top of search results in GEO.”
Alex Lindley, Semrush (2026)
An AI model can answer from memory, with what it learned in training, or with live search. GEO directly influences live AI search, while memory depends on the brand’s history on the web. In other words, a brand that was not on the internet when the model was trained does not exist in its memory. So until the next training run, it can show up in ChatGPT or Claude answers in only one way. The AI has to search the topic on the web and find authoritative content that cites the brand, on its own site or in other sources.
GEO, SEO and AEO: what is the difference?
SEO optimizes to rank in a list of results, while AEO optimizes to be the direct answer to a question. GEO, in turn, optimizes to be retrieved and cited inside an answer that an AI model synthesizes. GEO, SEO and AEO share the same technical base, but they differ in the unit evaluated and in the success metric.
| Criterion | SEO | AEO | GEO |
|---|---|---|---|
| Goal | Position in the results list | Being the direct answer | Being a cited source in the generated answer |
| Unit evaluated | Page and domain | Passage that answers the question | Passage the model retrieves |
| Key signal | Authority, relevance, technical health | Clear answer right below the question | Sources, statistics, quotes, clarity |
| Metric | Position, clicks, traffic | Answers won | Mentions, citations, share of voice |
| How Google treats it | The basis of everything | No official position | “Still SEO” |
However, Ahrefs treats GEO, AEO and LLMO as synonyms. Google, moreover, does not separate GEO from SEO. Its guide states:
“From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.”
Google Search Central (2026)
Does GEO replace SEO?
No, according to Google itself. SEO best practices still apply to GEO, because Google’s AI features use the core Search ranking systems. In fact, the guide treats third-party “GEO” and “AEO” advice as SEO advice. So GEO and SEO work together: without indexing, AI cannot cite the page.
Why does GEO matter in 2026?
GEO matters because people in Brazil already ask AI tools before they buy. According to the Prompt-me study by Fbiz, 46% of Brazilians who use AI research products or services before buying. Brazil is also one of ChatGPT’s three largest markets, according to OpenAI, with about 215 million messages a day.
At the same time, Google has 87.5% of the search engine market in Brazil, according to StatCounter. Google AI Overviews arrived in the country in August 2024, and then AI Mode in Portuguese came in September 2025. So here, GEO and SEO go hand in hand.
Consumers also switch between channels. Optimiza Marketing and AB Pesquisas surveyed 1,000 consumers in Brazil in December 2025. Of those, 71% say they check AI answers on Google.
In the US, the Pew Research Center tracked 900 adults in March 2025. People who saw an AI summary clicked on traditional results in 8% of visits; without the summary, however, the rate was 15%.
Seer Interactive analyzed 3,119 terms from 42 organizations, from June 2024 to September 2025. In that sample, organic CTR fell 61% when Google AI Overviews appeared. On the other hand, brands cited in Google AI Overviews got 35% more organic clicks than when they were not cited. By conversion rate, a visitor from AI is worth 4.4 times a visitor from organic search, according to Semrush.
How many companies already do GEO?
In the US, 41.5% of companies already use GEO, according to The CMO Survey 2026. Among marketing leaders at large US companies, 98% optimize for AI search or plan to within 12 months, according to Branch. GEO execution lags behind, though: in a Semrush survey of 481 professionals, only 22% have SEO and AI search fully integrated.
| Data point | Source | Sample |
|---|---|---|
| 41.5% of US companies already use GEO | The CMO Survey 2026 (Duke) | 308 marketing leaders, Jan 2026 |
| 98% of marketing leaders at large US companies say they optimize or plan to within 12 months | Branch | 300 marketing leaders at companies with 500+ employees, Jan 2026 |
| 22% have SEO and AI search fully integrated | Semrush | 481 marketing professionals, Jun 2026 |
| 69% have a budget for AI search | SEOFOMO | 171 search professionals, 36 countries, Aug 2026 |
| 70% want to lead in AI; 30% feel they have the infrastructure | Gartner 2026 CMO Spend Survey | 401 CMOs and marketing professionals (North America, Europe and the UK), Jan-Mar 2026 |
Gartner sees the same gap in marketing AI in general. Leading in AI is a goal for 70% of CMOs, but only 30% feel they have the infrastructure.
In the SEOFOMO survey, the share with a budget for AI search, that is, for GEO, rose from 38% in 2025 to 69% in 2026. In the same survey, 51% see some positive impact, but only 5% call it significant.
Do companies in Brazil already do GEO?
The Brazilian data comes from RD Station’s Panorama de Marketing e Vendas 2026 (2026 Marketing and Sales Panorama). There, 36% of respondents say their company takes no action to optimize its presence in LLMs, so it does no GEO. Another 23% don’t know whether they do anything. The samples differ, so these numbers don’t compare with the US figures.
How do ChatGPT, Perplexity and Google choose what to cite?
Engines with search receive the prompt, decide whether they need to search and rewrite the request into one or more queries. Then they retrieve pages from an index, select passages and generate the answer with citations. For GEO, what changes by engine is the index they query and the crawler that needs access.
| Engine | Where the pages come from | Search crawler to allow | Crawler tied to training |
|---|---|---|---|
| ChatGPT | Its own crawler and search partners, such as Bing | OAI-SearchBot | GPTBot |
| Perplexity | Its own index | PerplexityBot | PerplexityBot does not train models |
| Google AI Overviews and AI Mode | Search ranking systems | The same as Search; the page must be indexed | Google-Extended, which does not affect Search |
| Gemini | Grounding with Google Search | Do not block Google-Extended, which controls grounding in the Gemini apps and Vertex AI | Google-Extended, the same token |
| Claude | Web search decided from the prompt | Claude-SearchBot and Claude-User | ClaudeBot |
According to Anthropic’s documentation, in the Claude API each web search citation includes up to 150 characters of the cited passage. That is why, in GEO, it pays to write sentences that stand on their own in the AI answer.
ChatGPT
To begin with, a site must allow OAI-SearchBot and accept OpenAI’s published IP addresses to be eligible. Whenever it uses search partners, ChatGPT rewrites the request into queries that it sends to them. Since Microsoft says Bing powers AI tools such as ChatGPT, good Bing indexing is part of GEO. See the step-by-step guide on how to show up in ChatGPT.
Perplexity
Perplexity runs its own crawler and index, with more than 200 billion unique URLs, according to Perplexity’s engineering blog. According to Perplexity’s documentation, Perplexity-User visits a page when a user asks a question and generally ignores robots.txt. So blocking it there tends to have no effect. The guide on how to show up in Gemini, Perplexity and Claude covers what changes in each engine.
Google AI Overviews and AI Mode
Google bases its AI features on the Search ranking systems, so in these features GEO depends on SEO. As a result, the page must be in the index and eligible for a snippet. Since July 2026, the guide has also required the site to be in the Search generative AI control in Search Console, which is on by default.
Both features can also use query fan-out, running several searches on subtopics. For example, for “how to fix a lawn that’s full of weeds”, the model may search for “best herbicides for lawns”. Thus, AI can cite a page that answers a subquestion well even when it does not rank for the original question. See how to show up in AI Mode and AI Overviews.
Which GEO tactics have evidence?
The GEO tactics with the most evidence are citing sources, adding quotes and adding statistics, which led the Princeton benchmark. Studies from 2025 and 2026 also reinforce clarity, an answer at the top of the page, a question-and-answer format and freshness. However, these recent studies are observational: they show what pages cited by AI have in common.
The 9 tactics tested at Princeton and the effect of each
The paper tested nine ways to rewrite a text. Expert quotes produced the largest gain, followed by statistics. Fluency tied with citing sources, which shows how much readability matters. Finally, keyword stuffing was the only tactic that reduced visibility in the answer.
| Tactic | What changes in the text | Visibility in the answer | Subjective impression |
|---|---|---|---|
| Expert quotes | Adds verbatim quotes from trusted sources | +41% | +28% |
| Statistics | Replaces qualitative claims with numbers | +31% | +23% |
| Fluency | Makes the text flow better | +28% | +14% |
| Citing sources | Adds references to trusted sources | +28% | +14% |
| Technical terms | Uses the field’s technical vocabulary | +18% | +11% |
| Simple language | Simplifies the language of the text | +14% | +6% |
| Authoritative tone | Makes the text more assertive and persuasive | +10% | +19% |
| Unique words | Adds uncommon vocabulary | +6% | +6% |
| Keyword stuffing | Repeats the query keywords | -8% | +5% |
The table shows the relative gain over the unoptimized text in two of the paper’s metrics. Visibility in the answer measures how much of the generated text comes from the site, weighted by position. In contrast, subjective impression is a judge model’s score for the relevance, influence and prominence of the citation.
What the Princeton study proved, and what it did not
GEO-bench brings together 10,000 queries from nine sources, 80% of them informational. The effect varies by domain: statistics do best on legal and government topics; sources, on factual questions.
Citing sources raised the visibility of sites in 5th place on the SERP by 115.1%. Meanwhile, the site in 1st place lost 30.3% on average.
The tests used GPT-3.5 to simulate a generative engine. In the real-world test with Perplexity, quotes produced +21% visibility, statistics +9% and keyword stuffing -9%. However, versions with other percentages per tactic circulate on the web, such as 40% for citing sources alone or 25% for an authoritative tone. Those numbers do not match the paper’s table.
Also, the tests date from 2023. The real-world validation used the Perplexity of that time, so it did not include today’s Google AI Overviews or ChatGPT.
What the 2025 and 2026 studies add
Semrush compared 304,805 URLs cited by AI with 921,614 that only rank on Google. Clarity (+32.83%), E-E-A-T signals (+30.64%) and a question-and-answer format (+25.45%) led the list. By comparison, in an Ahrefs study, page length and citation position in Google AI Overviews had a correlation of 0.04, practically zero.
In 1.2 million ChatGPT answers analyzed by Kevin Indig, 44.2% of citations came from the first 30% of the page. Still, these numbers show correlation, not causation: they describe cited pages, but they do not guarantee that copying the format will earn a citation.
Before and after: how to rewrite a section to get cited
Before: no data, no source, no answer.
GEO is changing how brands think about their digital presence. Each company needs to assess what makes sense.
After: answer first, statistic, named source.
GEO is the optimization of content to be cited in AI answers. In the study that coined the term, citing sources, adding quotes and adding statistics raised visibility by up to 40%. On the metric of cited words weighted by position, the relative gain was 30% to 40%. On subjective impression, it was 15% to 30%. Source: Aggarwal et al., KDD 2024.
What does not work in GEO
Three popular GEO tactics have no evidence of earning citations in AI answers: llms.txt, keyword stuffing and schema as a shortcut. Keyword stuffing, a classic SEO tactic, does not help GEO. In the Princeton paper, for example, it scored 10% below the unoptimized text in the Perplexity test. As for a commercial tone, the effect is still unclear.
llms.txt
Google says AI text files neither help nor hurt, because Search ignores them. Ahrefs analyzed 137,210 domains: 28% publish llms.txt, but 97% of those files received no requests in May 2026.
Google’s John Mueller was blunt:
“AFAIK none of the AI services have said they’re using LLMs.TXT.”
John Mueller, Google (2025)
According to him, server logs show these services don’t even check for the file.
Schema as a shortcut
According to Google, there is no special schema.org markup to add for GEO, although it remains useful in SEO for rich results. Ahrefs tracked 1,885 pages that added JSON-LD, against 4,000 control pages. In Google AI Overviews, citations changed by -4.6%; in AI Mode, by +2.4%; in ChatGPT, by +2.2%. So, according to Ahrefs, adding schema produced no meaningful gain on any platform.
Promotional tone: mixed evidence
In the Semrush study, “non-promotional tone” had a negative association of -26.19% with citation. However, the authors do not read this as a preference for commercial language. Their hypothesis: texts by professional copywriters tend to be well structured and useful. So there is no basis for removing commercial language or for adopting it as a GEO tactic.
How to start with GEO: a step-by-step guide for agencies and brands
Like SEO, GEO starts with technical access, then the baseline, and only then content. After all, no rewrite earns a citation if the AI crawler can’t get into the site. Likewise, no metric is valid without fixed prompts. So the steps follow this order: access, measurement, content and review.
What are the first steps in GEO?
- Audit robots.txt, the CDN and the logs. Allow OAI-SearchBot, PerplexityBot, Claude-SearchBot and Claude-User. GPTBot and ClaudeBot serve training, so blocking them is the client’s decision and does not prevent search. Blocking Google-Extended does not affect Google Search, but it removes the content from grounding in the Gemini apps, according to Google’s documentation. Also check the logs to confirm that these agents visit the site. For sites on Cloudflare, follow the step-by-step guide to check and allow AI crawlers on Cloudflare.
- Register the site in Bing Webmaster Tools. Submit the sitemap and turn on IndexNow. That is because, on Bing, freshness signals influence how quickly updates reach AI answers.
- Build the prompt list. Gather 30 to 50 real questions from the client (our recommendation), as well as the queries derived from fan-out, which Google describes. Also include comparisons, since 45.2% of Brazilians who use AI when shopping compare options, according to Neogrid.
- Run the baseline. Run each prompt at least five times in each engine over the month, because AI answers vary. Record mentions and linked citations separately, since they are different metrics. Use fixed columns: date, engine, prompt, run, brand mentioned, URL cited and competitors cited.
How do you earn and keep AI citations?
- Prioritize pages. Start with indexed, snippet-eligible pages that answer the highest-value prompts.
- Rewrite in a citable pattern. Put a short, direct answer under each H2; we recommend 40 to 60 words. Also use statistics with a source, one of the tactics that led the Princeton benchmark. Use tables too, because in Trakkr’s research, 40% of the most cited pages have a table, versus 28.2% of the bottom half.
- Work on owned and third-party media. In its own case study, Semrush updated its own pages and content on other domains. As a result of this GEO work, the brand’s AI share of voice rose 146% in two months. For Reddit, reviews and the press, see how to earn presence in the sources AI cites. For clients with a physical address, see how to get AI to recommend a local business.
- Update priority pages. Review them every quarter, because freshness matters. In SE Ranking’s data, for example, pages updated in the last three months averaged 6 ChatGPT citations, versus 3.6 for outdated pages.
- Repeat the measurement and report. Compare share of voice by engine with the baseline. Then report presence in the answer along with traffic.
For agencies that want to sell this work, the guide to GEO for agencies covers scope, pricing and contracts. For online stores, see what changes in GEO for ecommerce.
How do you measure GEO?
While SEO tracks position, clicks and traffic, GEO measures presence in AI answers. It uses four metrics per engine: visibility, citations, share of voice and referral traffic. Visibility is the percentage of answers that mention the brand, whereas a citation is an answer that cites the domain as a source. Because answers can vary between runs, the protocol requires fixed prompts, repetitions and a recorded date.
Share of voice divides the brand’s mentions by the category total, times 100. Semrush, Ahrefs and Profound use variations of this formula, so compare within the same tool. The GEO tools comparison also shows what each one measures.
To measure GEO, we recommend a minimum monthly protocol: 40 to 60 fixed prompts and five runs per prompt in each AI engine. With 50 prompts and four engines, for example, that is 1,000 answers per client each cycle. In the Semrush survey of 481 marketing professionals, 40% still check brand presence by typing prompts by hand. The full step-by-step process is in the guide on how to measure your brand’s mentions and citations in AI.
How do you measure traffic from AI?
For traffic, according to Google’s guide, Search Console has the Generative AI performance report. Similarly, Bing Webmaster Tools has AI Performance, with cited URLs and grounding queries.
In May 2026, ChatGPT updated its links. After that, the home page’s share of ChatGPT referral traffic went from 26%-29% to 62%-63%, according to Similarweb. Similarweb does not report the geography of this data, though.
The volume is still small: across 1,215 domains of Conductor clients in the US, from May to September 2025, AI generated 1.08% of traffic. Note that Conductor sells an AI search optimization platform.
What we found when we analyzed Portuguese-language articles about GEO
We analyzed the 8 Portuguese-language pages that appeared most often in searches for “o que é GEO” (“what is GEO”). They are Conta Azul, RD Station, Alura, Neil Patel BR, HubSpot BR, Locaweb, Agência Mutum and Layer Up. Only 3 of the 8 include their own data: RD Station, Neil Patel BR and HubSpot BR.
Still, none of them measures the visibility of Brazilian brands in AI. Of the other 5, 4 repeat third-party statistics, often secondhand; Layer Up has no numbers at all. Locaweb, for example, cites 15 numbers, and 11 come “via AM9 Digital”. Also, only 3 of the 8 mention the academic study that coined the term. None links to the paper or explains its limits. Finally, the 30% to 40% figure appears on 3 pages without context.
Want to know how your brand shows up in AI answers today?
NotHuman is opening spots for agencies and brands to track how they appear in AI answers. Join the waitlist and we’ll let you know when a spot opens up for you.
Main sources
- Aggarwal et al., the paper “GEO: Generative Engine Optimization”, version accepted at KDD 2024 (2024).
- Duke Fuqua, The CMO Survey 2026: Highlights and Insights Report (2026).
- Semrush, survey of 481 professionals on integrating SEO and AI search (2026).
- RD Station, Panorama de Marketing e Vendas 2026, a marketing and sales survey in Brazil (2026).
- Semrush, AI Search SEO Traffic Study (2025).
- Semrush, study comparing URLs cited by AI with URLs that only rank on Google (2026).
- Ahrefs, study of llms.txt across 137,000 domains (2026).
- Ahrefs, study of 1,885 pages that added schema (2026).
