A growing share of your future customers no longer type a query into Google: they ask an assistant a question. "Which tool should I use to manage my invoicing?", "Which agency should build my application?". The answer cites three or four names. If yours isn't among them, you don't exist in that conversation.
In short: GEO (Generative Engine Optimization) is about making your company understandable, verifiable and citable by language models. In practice, it rests on four pillars: a site that AI crawlers can read, content that directly answers specific questions, a consistent identity everywhere on the web, and mentions on credible third-party sources. The last pillar carries the most weight, and it takes the longest to build.
How an AI decides who to cite
An assistant answers in two ways, often combined.
From its memory. The model was trained on a huge quantity of text. A company mentioned often, consistently, across varied sources leaves a trace there. This memory is only updated with each new version of the model.
From a live search. For recent or commercial questions, the assistant runs web searches, reads a few pages and synthesises. In this case, it relies on existing search engines. Ranking well therefore remains the foundation: a page that is invisible in classic search engines has little chance of being read by an AI.
The practical consequence: GEO doesn't replace SEO, it adds to it. A solid technical foundation remains the prerequisite.
What changes compared with classic SEO
| Classic SEO | GEO | |
|---|---|---|
| Goal | Appear in a list of links | Be cited in a written answer |
| Basic unit | The page | The passage: a paragraph, a table, a definition |
| Type of query | Short keywords | Long, conversational questions |
| Dominant signal | Inbound links | Brand mentions on credible sources |
| Measurement | Rankings and clicks | Presence and accuracy in the answers |
The most important change is the basic unit. An AI doesn't extract a page, it extracts a passage. Every section of your content must be understandable on its own, taken out of its context.
Pillar 1: a site that AI crawlers can read
- Allow the crawlers. GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended respect the robots.txt file. Check that they are not blocked, including by your hosting provider or your web application firewall.
- Serve the content as HTML. Most of these crawlers don't execute JavaScript. Content loaded only in the browser is invisible to them. Server-side rendering, built into a framework like Next.js, solves the problem.
- Structure with Schema.org data. Organization, Article, FAQPage, Product, Review: this markup explicitly describes who you are and what the page contains.
- Take care of the hierarchy. A single main title, descriptive subheadings, lists and tables.
A word about the llms.txt file, often presented as essential: it is a proposed convention for summarising a site for AIs. It is cheap to set up, but no major provider has confirmed using it systematically. Worth doing, without expecting miracles.
Pillar 2: content designed to be cited
The passages most often picked up in generated answers share a few characteristics.
- The answer comes first. Two or three sentences that answer the question, then the detail. No three-paragraph introduction.
- The subheadings are real questions, worded the way your customers ask them.
- The claims are precise. A figure, a range, a duration, a condition. "Between €25,000 and €60,000 for an MVP" is citable. "It depends on many factors" is not.
- Comparisons are in tables. Models extract tabular data very well.
- Terms are defined. A clear one-sentence definition has a strong chance of being picked up.
- The content is dated and signed. An identified author and a publication and update date strengthen credibility.
The most useful content is what your competitors don't dare to publish: your prices, your limits, the cases where you are not the right choice, honest comparisons.
Pillar 3: a consistent identity
A language model reasons in entities. For it to correctly associate your name with your business, the information has to be identical everywhere.
- The same name, written the same way, on your site, LinkedIn, your directory profiles and your legal notice.
- A stable one-sentence description: who you are, what you do, for whom, where.
- A factual "About" page and team page, with real people.
- Organization markup that links your site to your official profiles.
If your site says "development agency", LinkedIn says "product studio" and a directory says "IT services company", the model will hesitate over what you are, and when in doubt will cite someone else.
Pillar 4: being cited somewhere other than your own site
This is the most decisive factor. When an assistant searches for "best mobile development agencies in Paris", it mostly reads comparisons, directories, press articles and forum discussions. It cites the names that keep coming up.
The actions that count:
- complete profiles and customer reviews on the platforms in your sector
- a presence in existing comparisons and rankings
- guest articles, interviews and podcasts in your field
- case studies published and shared by your clients
- useful contributions in the communities where your customers are
This is public relations work as much as SEO work. There is no technical shortcut to replace it.
Measuring your visibility in AIs
- Write twenty to thirty questions that your customers really ask.
- Put them every month to ChatGPT, Claude, Perplexity, Gemini and Google's AI Overviews.
- Note whether you are cited, in what position, with what description, and which sources are displayed.
- Track the traffic coming from these assistants in your analytics.
The sources displayed are the most valuable information: they tell you exactly where you need to be present.
What doesn't work
- Hidden text aimed at AIs. It gets detected, and it is risky for your classic search rankings.
- Mass production of generated articles. Content with no new information has no reason to be cited.
- Fake mentions and fake reviews. A legal and reputational risk, for zero effect in the medium term.
- Expecting results in two weeks. The effects on live search appear within a few weeks to a few months; entering the models' memory takes longer.
Frequently asked questions
What is GEO?
GEO, for Generative Engine Optimization, refers to all the practices aimed at making a brand appear in the answers of artificial intelligence assistants such as ChatGPT, Claude, Perplexity or Gemini. It is also called AEO or AI search optimisation.
Does GEO replace SEO?
No. Assistants rely heavily on search engines to find their sources. Good organic search ranking remains the foundation, to which GEO adds work on content structure and brand awareness.
How long does it take to appear in ChatGPT's answers?
For answers that rely on a web search, a few weeks to a few months after publishing relevant content and earning third-party mentions. For the model's memory, you have to wait for the following versions.
Is the llms.txt file necessary?
It is useful and simple to set up, but not decisive. No major AI provider has confirmed using it systematically. Content quality and external mentions weigh far more.
Should you block or allow AI crawlers?
If your goal is to be recommended, allow them. Blocking the assistants' search crawlers prevents your pages from being read and cited in their answers.
👉 Want to know how your company appears in AI answers? Discover our SEO support or let's talk.