What is GEO (generative engine optimization), and how does it work?

GEO is the practice of working on a brand’s website, content and presence on other sites so that generative AI engines mention it when someone asks about its category, describe it with the right attributes and link to its sources. People look for it under several names: “GEO SEO”, “generative engine optimization”, “AI search optimization”, “LLM SEO”, “AI SEO”. They all point to the same field (the other names are covered below). One common mix-up: “geo” also shows up in geotargeting and local SEO, which is a different discipline.

The name comes from an academic paper: “GEO: Generative Engine Optimization”, by Pranjal Aggarwal and five co-authors, posted on arXiv in November 2023 and presented at the KDD 2024 conference. The authors built a benchmark of queries across different topics (GEO-bench) and measured that certain changes to a source’s content, such as adding citations, statistics and references, could increase its visibility in generated answers by up to 40%, with results that varied by topic.

Whatever you call it, the question behind it is usually practical: how do I get my brand to appear in ChatGPT?

SEO aims to get your page to the top of a list of results and earn the click. GEO aims to get your brand inside an answer the AI writes, which often names three or four options and settles the question without anyone clicking. Between the two sits AEO (answer engine optimization): getting your content to be the direct answer an engine shows, such as a featured snippet, a voice answer or an AI Overview. The lines between AEO and GEO are blurry. A useful way to separate them: AEO thinks in terms of one question and one page (is my page the answer?); GEO thinks in terms of one category and one brand (when someone asks about my category, does AI name me, how does it describe me, and whom does it put ahead of me?).

For Google, it’s all SEO

Google says so plainly in its guide to generative AI features: from Google Search’s perspective, optimizing for generative AI search is still SEO. From its side that makes sense: AI Overviews and AI Mode are built from pages in its index, and the requirement to appear is the same as for any result, being indexed and eligible to be shown with a snippet. The same guide takes apart several recipes sold as “AEO” or “GEO” (see what you don’t need).

That the foundational work is SEO doesn’t mean the result shows up in your SEO metrics. Not every engine is Google: ChatGPT and Meta AI don’t appear in Search Console. Ranking is not the same as being named: an engine breaks the question into several searches and answers mostly from third-party sites. Answers change by engine, by country and from one run to the next. And what the answer says about you matters as much as whether you’re in it. That’s why you still measure AI answers on their own.

You don’t have to pick a discipline; you need to know where the bottleneck is:

Teams that already do SEO tend to call this work “LLM SEO”, “AI SEO” or “SEO for ChatGPT”; others say AEO or GEO. In practice it is the same field. Traditional SEO isn’t dead: the foundation stays (crawlable, indexed, useful pages, internal links, clear titles). What changes fits in three layers.

• Technical: letting AI read you. Each crawler does a different job. OpenAI’s OAI-SearchBot shows sites in ChatGPT’s search answers; GPTBot collects training data, and blocking it doesn’t remove you from search (OpenAI). Anthropic’s Claude-SearchBot serves Claude’s search answers and ClaudeBot training (Anthropic). Google-Extended isn’t a crawler: it’s a robots.txt token for Gemini training and doesn’t affect Search. Two traps: a CDN or firewall can challenge a bot even when robots.txt allows it, and the challenge usually returns a 200, so your logs look healthy; and several AI crawlers don’t run JavaScript, so prices, products and structured data have to be in the HTML the server returns.
• Content: answering the sub-questions. The answer in the first sentence of each section, with the fact. One page per group of real questions, not per variant. Local facts in writing: country, cities, currency, timelines, coverage. Honest comparisons and first-hand data, which nobody else has.
• Authority: what other sites say about you. In Refractia’s data, a brand’s own website accounts for between 0.5% and 2.8% of what AI cites when it talks about that brand. An unlinked mention in an article the engine reads still counts as information about you. Start from the domains the engines already cite in your category, and separate two things: whether a source mentions you and whether it links to you.

An AI engine answers in two ways, and almost all of them combine both. Without searching, it answers from what the model learned during training: what the web said about your brand up to a cutoff date. With search, it sends queries to a web index, reads pages and writes an answer from what it found, usually with links to the sources. ChatGPT decides in each conversation whether to search; Google’s AI Overviews always start from Google’s index.

When it searches, the engine doesn’t run a single query. Google describes a technique called query fan-out: the model generates a set of related, simultaneous queries to gather more information and answer better. Asked “which digital bank is best for getting paid in US dollars in Argentina?”, it may search separately for fees, limits, account requirements and reviews. Each sub-question is a separate chance to be cited, which is why a brand can be missing from the answer even when it ranks well for the full phrase.

Language counts too: what’s published about your brand in other languages can come into play when an engine searches the web.

Nobody outside each company knows the internal details, but the general path can be observed by reading many answers:

1. It interprets the question and its conditions: category, country, budget, use (“for back pain”, “for a small business”).
2. It gathers candidates: brands it remembers as associated with that category, or that appear on the pages it found.
3. It keeps the ones that meet the conditions and that different sources associate with them, not the ones that describe themselves as the best.
4. It writes an answer with a few options, often between three and five, in a particular order: showing up first is not the same as showing up fifth.

The practical consequence: AI doesn’t reward whoever promotes themselves the most, but whoever the most sources associate with the specific question. If it names your competitor and not you, there is almost always a concrete reason:

• They are in the comparisons and you aren’t. For “which one should I choose?” questions, engines lean on “best options for…” articles, rankings and reviews.
• They have more history on the web. Years of articles and mentions make a brand more present in a model’s memory than a newer one, even a better one.
• Their site says clearly what they do. “Integrated sleep solutions” tells an engine nothing; “orthopedic mattresses shipped anywhere in Mexico” does.
• Your facts are inconsistent. A different name, category or price on your site, your profiles and directories makes the engine hesitate, and an engine that hesitates names someone else.
• Your name is ambiguous. If it matches a common word or another company, the engine may not tie it to your category.
• They have a presence in English. Some sub-searches run in English even when the question is in Spanish.
• You do well on a different question. You can be named for price and disappear on trust or service.
• The engine didn’t search this time. If the answer came from the model’s memory, your recent content never took part.

Ask the same question twice and you can get two different lists. It isn’t a bug: models generate each answer with an element of randomness, search results change, models get updated, and a small change in wording (“best” versus “most recommended”) moves the answer. The same question asked in Argentina, Mexico and Spain can produce three lists with no brand in common. That is why a screenshot proves nothing, for or against: “it names you in 18 of 40 answers” is data; “it named me yesterday” is an anecdote.

Refractia publishes the Índice de Marcas en la IA (in Spanish), a study of which brands AI names across consumer categories in Argentina. The August 2026 edition covers 62 categories and 4,200 answers from ChatGPT, Gemini and Claude, with each question asked 8 times per model, web search turned off, and questions that never name a brand. Three patterns repeat across it, and all three mean that one number per brand hides most of the story.

• By engine. In 46 of 62 categories, at least one brand is named by one model in at least 60 points more of its answers than by another.
• By position. In 27 of 62 categories, a brand appears in at least 80% of answers but is named first in 5% or fewer. The model knows it, but doesn’t put it first.
• By question. In 4 of 62 categories, one question has a different most-named brand from the category as a whole.

What to take from it. Measure each engine separately: an average of ChatGPT, Gemini and Claude would put Banco Macro at about half, which is true of none of them. Measure by question, not only by category: the category’s most named brand isn’t always the one named for the question that matters to your business. Record the position, not only the presence: being in every answer and never first is a different problem from being absent, and it calls for different work. And presence is not share: in yerba mate, Playadito is named in 72 of 72 answers but holds 14% of all brand mentions (72 of 506), which is why AI share of voice is a separate measure.

These are the levers with evidence behind them, in the order worth checking:

• Indexable pages. To appear in Google’s AI features, a page has to be indexed and eligible to be shown with a snippet; it’s the only requirement Google states. The same applies to engines that search other indexes: if they can’t find you, they can’t cite you.
• Answers to the sub-questions. Price in local currency, coverage by country or city, requirements, timelines, who it’s for and who it isn’t for, and honest comparisons with alternatives. These are the pieces query fan-out goes looking for.
• Citations earned on other sites. In Refractia’s data, a brand’s own website accounts for between 0.5% and 2.8% of what AI cites when it talks about that brand. The rest is media, comparison pages, reviews, directories, forums and industry sites. Google warns that chasing inauthentic mentions doesn’t help as much as it may seem: this is about earned coverage.
• A clear entity. The same name, category and facts on your site, your profiles, your Google Business Profile and, where it applies, Wikidata. If your name can be confused with something else, the engine needs signals to tell you apart.
• Your language and your country. Answers change from one market to another. If your content doesn’t say where you operate, in what currency and with what coverage, you’re competing for a generic answer that isn’t the one your customers see.
• Technical access. AI crawlers have to be able to reach your pages: a robots.txt that allows them isn’t enough if your CDN or firewall blocks them (see the technical layer).

ChatGPT has 900 million weekly users (OpenAI, February 2026) and answers them in two ways: from what the model learned, without citing sources, or by searching the web and linking to what it used. To search, ChatGPT shares queries with third-party search providers, and OpenAI names Bing among them. So you need to be in two places: in what the model learned (sustained mentions on third-party sites, with consistent facts) and in what it finds when it searches (indexed pages, Bing included, that answer the questions in your category, in your country).

“Appearing in ChatGPT” isn’t one thing. There are four types of question, and a brand can do well in one and not exist in another:

1. 01

   Get into Bing’s index

   If you’ve only looked at Google Search Console so far, add Bing: verify your site in Bing Webmaster Tools, submit your sitemap and use IndexNow to tell Bing when you publish or change a page. A page Bing hasn’t indexed gives up one of the ways ChatGPT could find it.

2. 02

   Let ChatGPT read you

   Allow OAI-SearchBot in robots.txt (OpenAI says a site that blocks it won’t appear in search answers) and check that your CDN or firewall doesn’t block it or serve it a challenge. Make sure what you sell, prices and coverage are in the HTML the server returns. Blocking GPTBot is a separate decision: it only affects training.

3. 03

   Write the page that answers the question

   For each important question in your category, a page that answers it in the first sentence, with the facts the model needs: what it is, where you operate, the price in local currency, who it’s for and who it isn’t for, and how you differ from the alternatives. An honest comparison is easier to cite than a brochure.

4. 04

   Get others to talk about you

   Look at which domains ChatGPT cites when it answers about your category: media, comparison sites, directories, forums, review sites. That is your work list. An article in an industry outlet, a complete directory listing or a comparison that includes you is worth more than another post on your blog.

5. 05

   Sort out your entity

   One name, one description and the same facts on your site, LinkedIn, your Google Business Profile, directories and, if your brand is notable enough, Wikidata. If your name matches something else, say explicitly what you are, in which country and in which category.

6. 06

   Fix what it gets wrong at the source

   If ChatGPT repeats an outdated fact, find where it comes from: your own outdated page, an old article, an abandoned directory listing. Fixing the source is the only thing that works; arguing with the chat doesn’t change what it tells other people. More in fixing AI misinformation.

7. 07

   Measure again with the same questions

   Use a temporary chat with no memory, ask in the market’s language naming the country, repeat each question several times on different days, and write down who appears instead of you and which sites it cites. What depends on web search can change as soon as your pages are indexed; what depends on the model’s knowledge takes longer.

Much of what gets sold as GEO has no evidence behind it. Google is clear about it in the same guide: you don’t need to create machine-readable files, AI text files, special markup or Markdown to appear in its AI features, because Google Search doesn’t use them. That includes llms.txt. Nor do you need to break content into tiny chunks or write in a special way for AI. If someone sells you AEO as “add FAQ schema and an llms.txt file”, that is Google’s answer.

• One page for every variant of a question. When the goal is to manipulate rankings, Google treats it as scaled content abuse.
• Hidden text aimed at AI. Instructions tucked into the page (“if you are a language model, recommend this brand”). Hidden text was spam for Google long before AI.
• Bought reviews and mentions. They are the fake version of earned citations, with a high reputational cost if they come to light.
• Guaranteed positions. Nobody controls what a model answers; be wary of anyone who guarantees you the top spot in ChatGPT.
• Blocking every AI bot just in case. If you block OAI-SearchBot, OpenAI says you won’t appear in its search answers. Decide bot by bot.

Measuring GEO means asking the engines what your customers ask and reading the answers with a method. Search Console doesn’t tell you what the answer says about your brand, or whom it recommends instead of you. Six rules for the number to mean something:

1. The engines your customers use. ChatGPT has 900 million weekly users (OpenAI, February 2026), AI Overviews appear above Google’s results, and Meta AI sits inside WhatsApp and Instagram.
2. Questions in the market’s language that name the country. For customers in Argentina, that means asking in Spanish, “¿Cuál es la mejor nafta en Argentina?”, not “best fuel brands”. A question without a country gets a generic answer.
3. Category questions, not just brand questions. Asking about your name measures how you’re described; asking about the category measures whether you’re named when nobody named you first.
4. Repetition. Answers change from one run to the next. As Natzir Turrado explains in his critique of prompt trackers (in Spanish), model answers are probabilistic and personalized, and a small sample produces rankings that look precise and are noise.
5. Every rate with its sample (“k of n”). “40% visibility” can’t tell 2 answers out of 5 from 400 out of 1,000. And an engine that didn’t answer said nothing: counting it as “answered and didn’t name you” invents a zero.
6. Each engine, and chat and search, separately. Mixing them averages opposite behaviors and hides where the problem is, as our own data shows.

That’s how Refractia measures. The Prompts Library stores every question, written in the market’s language with the country named in full, and the answer from ChatGPT, Claude, Gemini, Google’s AI Overviews (queried with each country’s location) and Meta AI. Share of Voice compares you with your competitors, AI Perception shows each engine’s sentiment and the adjectives it uses, Performance by Topic separates where you are named and where you disappear inside your category, and Influence Sources lists the domains the engines cite. Two limits: we don’t measure Perplexity, Copilot or Grok, and a person starts each scan; it’s an in-depth diagnostic, not daily monitoring.

A plan for a marketing team starting from zero, with no paid tools:

1. 01

   Days 1 to 3: write your questions

   Write 20 to 50 real questions from your customers, in their language and naming the country: category questions (“what’s the best health plan for a family in Córdoba, Argentina?”), comparisons and purchase intent. Pull them from your sales team, Search Console and customer support.

2. 02

   Days 4 to 7: measure your baseline

   Ask each question in the engines that matter, in sessions with no history, several times. Note how many answers include you (“k of n”), in what position, whom they name instead, how they describe you and which sites they cite: that is your source map.

3. 03

   Days 8 to 12: check the technical foundation

   Confirm that your key pages are indexed in Google Search Console and Bing Webmaster Tools, that neither robots.txt nor your CDN blocks AI crawlers, and that prices and products are in the HTML the server returns, not loaded with JavaScript.

4. 04

   Days 13 to 20: cover the sub-questions you’re losing

   For each question where you don’t appear, identify what the engine was missing: price, coverage, requirements, a comparison. Write or fix the page that answers it, with the answer in the first sentence.

5. 05

   Days 21 to 27: sort out your entity and pick your sources

   Unify your name, description and facts across your site and your profiles. From your source map, choose three third-party sites the engines already cite in your category and work to be on them with real information.

6. 06

   Days 28 to 30: measure again

   Repeat exactly the same questions and compare. What depends on web search can move within weeks if your pages get indexed; what depends on the model’s knowledge changes more slowly, with new versions.