What is GEO and how do you do it?

When people look for a product they increasingly ask an AI instead of clicking through ten links: "What is the best accounting software for small businesses?" The answer is one piece of text and a few source links. If your name is not in that text, ranking first on a results page means nothing to that user.

GEO deals with three things: the brand being named in the answer (mention), your site being listed among the answer's sources (citation), and the background searches the model runs on its own before answering (query fan-out). If the model does not find you in the results of those searches, it will not mention you either.

• —SEO: the target is the position on the results page; measured by clicks and impressions.
• —GEO: the target is the AI answer's text and sources; measured by mention rate, position, tone and share of citations.
• —They are not alternatives: AI search is mostly fed by search engine indexes, so good SEO is the ground GEO stands on.

Each assistant reads from a different index with different priorities, which is why different engines show different sites for the same question.

ChatGPT

Its web search relies on the Bing index. Industry measurements find very little overlap (6–8%) between the pages ChatGPT cites and Google's top 10; your site needs to be indexed in Bing, and OpenAI's crawlers (OAI-SearchBot, ChatGPT-User) must not be blocked.

Google AI Overviews / AI Mode

Read from Google's own index. Classic SEO signals and structured data (schema.org) matter most here.

Perplexity

Searches on every question and weighs freshness and community proof; in measurements a large share of its most-cited sources are community sites such as Reddit.

Gemini and Claude

With search on, they use their own search infrastructure; with search off, the answer comes from training data, where how often the brand appeared on trusted sites over the years decides.

Most GEO advice comes from experience; less of it has been measured. What is known:

• —Princeton's 2023 GEO study showed that adding statistics, quotations and sources to a text could raise visibility by up to 40%. A 2026 critical survey notes the gain holds only when the content is already in front of the model; relevance and retrievability decide.
• —Most of the pages ChatGPT cites most are Wikipedia, brand home pages and app store pages. In an analysis of 6.8 million citations, 86% came from sources a brand can manage (its own site, and list/directory pages).
• —Recently updated pages get cited more. In one industry measurement, content updated in the last 30 days received about 3.2× more citations in AI Overviews (a single source; read with care).
• —Asking the same model the same question again can change the answer and its sources noticeably. Deciding from a single try is misleading.
• —An llms.txt file is easy to add, but no major platform is confirmed to use it; 97% of llms.txt files across 137,000 domains received no requests at all.

So rather than one "trick", focus on the fundamentals and track the result with repeated measurement.

From the best-evidenced to the most experimental:

1. 01List the questions your customers actually ask AI ("best …", "… vs …", "alternative to …") and write a separate, clear answer page for each.
2. 02Put a short, definite answer that can be quoted on its own in the first paragraph of every page.
3. 03Give numbers, sources and named expert views; publish original data from your own measurements where you can.
4. 04Keep pages current and show the last-updated date.
5. 05Add structured data: Organization, Product or SoftwareApplication, FAQPage, Article.
6. 06Check robots.txt does not block AI crawlers (GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended) and that your content is readable without JavaScript.
7. 07Add the site to Bing Webmaster Tools and report changed pages with IndexNow; ChatGPT search uses the Bing index.
8. 08Be present on third-party sources: comparison and "best of" lists in your category, review sites, industry directories, Wikipedia/Wikidata, and real answers in communities such as Reddit.
9. 09Write your brand, product names and category the same way everywhere; AI matching you to the right entity depends on it.
10. 10Add llms.txt: it is cheap, just know its effect is unproven.

Measuring GEO means asking the questions your customers would ask to several AI engines, several times, and recording for each answer whether your brand appears, in what position and tone, and which sites are cited.

• —Mention rate: in what share of questions the brand appears.
• —Position and tone: where it ranks among recommendations, positive or negative.
• —Citation share: your site's and your competitors' share of the source list.
• —Source gap: in answers that mention a competitor but not you, which sites are cited? Those are the places to get into.
• —Background searches: what does the model search before answering? Write content for those searches.

One important distinction: an engine giving no answer, or a case that could not be measured, is not the same as "not mentioned" (zero). Mavzena measures by these rules; the details are on the Methodology page. How we measure

1. Aggarwal et al., GEO: Generative Engine Optimization (Princeton, 2023)
2. Optimizing Visibility in Generative Engines: A Critical Survey of GEO (2023-2026)
3. Profound — AI platform citation patterns
4. WhyIQ — What each AI engine wants before it cites you
5. Frase — What is Generative Engine Optimization (GEO)?
6. llms.txt: where it goes, who reads it