Your website ranks first on Google for a key search term. Yet when a potential customer asks ChatGPT or Perplexity exactly that question, a competitor’s name appears in the answer. It happens. And it’s going to happen more and more often.
The problem is that, for years, online visibility has been measured in search rankings and clicks. This model is changing right before our eyes. According to several studies published in 2026, which all reach the same conclusion, 68% of Google searches end without a single click to an external site. For queries where an AI summary is displayed, this figure rises to 83%. A high ranking no longer guarantees visibility.
It was against this backdrop that the GEO, or Generative Engine Optimization. The term was formally defined in 2023 in a research paper published by researchers from Princeton, Georgia Tech and the Allen Institute for AI. The basic idea is simple to articulate, but it changes a great deal in practice: it is no longer simply a matter of being ranked in a list of links, but of being mentioned in an AI-generated response.
SEO and GEO: different rules of the game
Traditional SEO optimises a page to appear in Google’s search results. GEO optimises content so that a language model, such as ChatGPT, Perplexity or Google’s AI Overviews, recognises it as a reliable source and cites it in its responses.
The fundamental difference Here’s how it works: Google directs traffic to a URL. A language model, meanwhile, quotes a passage. The user receives the answer directly within the conversation, without necessarily visiting your site. Your visibility no longer depends solely on clicks; it depends on being mentioned.
That’s no reason to give up on SEO. The two approaches are complementary : SEO ensures your visibility in traditional search results, whilst GEO increases your chances of being featured in AI summaries. By 2026, it will make more sense to talk about SEO + GEO rather than choosing between the two.
What AI systems look for when selecting their sources
Generative models do not simply cite any old content. Several factors come into play.
The clarity of the answer. AI systems favour content that answers a question directly in the first paragraph, without a three-line introduction. A «short answer followed by further explanation» format is much easier to extract than a continuous, unstructured text.
Verifiable data. Figures, dates, sources cited: quantifiable elements are widely incorporated into generative models. Content that makes claims without providing evidence is less reliable than content based on facts.
The perceived authority of the source. Wikipedia is the top source cited by ChatGPT. Reddit is the top source for Perplexity. This is no coincidence: these platforms have accumulated years of structured content, frequently updated and widely cited elsewhere. For a professional website, this means that external mentions, testimonials and appearances in other publications contribute to the authority that the models attribute to it.
The freshness of the content. When it comes to searches related to current events or evolving trends (including digital technology), a recently updated article has the edge over outdated content.
Three practical steps to get started
Restructure your existing content. Reread your articles and service pages with this question in mind: could an AI extract a clear answer from the first 100 words? If not, rephrase the introduction. Add an FAQ section at the end of your articles. Generative models process this format very effectively.
Add a file called llms.txt to your website. This file, placed in the root directory of your website, provides AI crawlers (GPTBot, ClaudeBot, PerplexityBot) with a structured map of your priority content. By April 2026, more than 780 major sites will already have them, including Cloudflare, Stripe and Zapier. For a WordPress site, plugins such as Rank Math allow you to generate it automatically. It is not an official standard, but its adoption is growing rapidly and the cost of implementation is minimal.
Build your online presence beyond your website. Brand search volume is the best predictor of AI citations, with a correlation measured by AirOps in 2025. The more your name is mentioned in other publications, forums, the press or leading directories, the more search engines recognise you and reference you. This ties in with what SEO has long referred to as «off-site» factors: backlinks, brand mentions and presence in industry discussions.
What this means for the self-employed or very small businesses
Major brands are already included in the model training datasets. They exist within the data. For a consultant, a service provider or an SME, the question is a different one: how can you ensure you feature in the AI responses to the queries your clients actually ask?
The answer lies in specialisation. Highly specific content on a niche topic (such as WordPress site redesigns for B2B coaches, or local SEO for alternative medicine practices) is more likely to be cited than general content on «how to do SEO properly». Search engines look for sources that offer something distinct, not a rehash of what already exists in abundance.
Ultimately, this is consistent with what common sense has long advised: write for a specific audience, on a subject you know well, using real-life examples. The difference is that the audience has changed. It now includes language models that read your site to answer questions from your future customers.
Sources
- Aggarwal et al., «GEO: Generative Engine Optimisation», Princeton / Georgia Tech / Allen Institute for AI / IIT Delhi, arXiv:2311.09735, 2023
- PingPrime, «Decline in organic traffic in 2026: causes, figures and solutions», May 2026
- Michelle Blanc, «Robots.txt and llms.txt: the two files that determine whether AI systems can cite you», May 2026
The term ‘GEO’ has been around since late 2023, but it has taken a while to filter through into French-language discussions. It’s starting to become a key topic within agencies, especially since Google’s AI Overviews have started appearing for queries we’d worked hard on. What the article doesn’t mention yet is that measuring your GEO visibility is much more complicated than measuring your SEO rankings. There isn’t yet a consumer-facing tool that does this properly.
As a web editor, this section on content structure really strikes a chord with me. The format of a short answer followed by further detail is exactly what is sometimes referred to as the ‘inverted pyramid’ in journalism. What’s new is that the audience is no longer just human. It really changes the way we think about the introduction to an article.
I remain sceptical about the figure of 68% for no-click searches. This statistic is being bandied about a lot at the moment, but it groups together very different types of queries: weather, quick calculations, definitions. For queries with a commercial intent or those seeking in-depth information, the zero-click rate is much lower. That doesn’t mean GEO is of no interest, but the figure on its own is a little misleading.
@Théo: That’s a fair point. For transactional queries and long-tail informational queries, clicks still account for the majority. But the underlying trend is real, and above all: even if the zero-click rate is 40% for your key queries rather than 68%, this represents a significant loss of visibility that Google’s ranking alone can no longer compensate for.
To conclude the point I raised at the start: I still maintain that the figures on zero-click searches need to be put into context. However, the general direction outlined in the article strikes me as correct. Optimising to be cited and optimising to rank are not mutually exclusive, and starting to think about GEO now, whilst competition in this area is still low, is probably the right time.
A specific question about the llms.txt file: I installed Rank Math a few months ago. Is the generation of the llms.txt file included in the free version or only in the Pro version? I haven’t been able to find the option in my dashboard.
@Amandine: In Rank Math Pro, yes. In the free version, you can create the file manually and upload it to the root directory of your server via FTP or your hosting provider’s file manager. The structure is simple: a text file with a few lines pointing to your priority content. There’s a free online generator at llmstxt.org if you’d like to start with a template.
@Thomas, regarding llms.txt: I’ve found the generator – it really is easy to create. I uploaded it this morning. I don’t know if it’ll make much difference in the short term, but the cost of setting it up is really minimal, as the article says. We might as well do it.
What interests me most about this article is the link between AI-generated citations and off-site brand mentions. In practice, what does this actually mean for a freelance consultant who doesn’t have a PR budget? Niche podcasts, contributions to LinkedIn groups, guest articles? Are these formats actually picked up by the models’ crawlers?
@Karima: The models’ training data includes the indexable web, so articles on other sites are included, yes. Audio podcasts do not, unless they are transcribed. LinkedIn discussions have partial and variable indexability. The most reliable factor remains text-based publications on domains that themselves have good authority: sector-specific media, leading directories, and long-established blogs.
Technically speaking, is there a difference in how GPTBot, ClaudeBot and PerplexityBot handle the llms.txt file? When it comes to robots.txt, each crawler has its own rules. I wonder if the same applies here, or if the file is interpreted more consistently.
@Florent: llms.txt is not an official standard, so there are no binding specifications. In practice, crawlers from different platforms interpret it differently, and some do not read it at all yet. It is more a gesture of goodwill and a navigation aid than a strictly enforced instruction. The robots.txt file remains the only standard that is legally binding for crawlers.
I’ve been testing this since this morning: I asked ChatGPT the question my patients ask me most often, namely how to find a therapist specialising in brief therapy in Lyon. ChatGPT listed three names. Mine wasn’t among them. Yet I’m on the first page of Google for that search term. The article explains exactly why. It’s a bit unsettling to see it happen in real time.
@Isabelle: This simple test you’ve described should be the first thing every professional does tonight. Ask ChatGPT or Perplexity the questions your clients ask you and see if you come up. The result is often a reality check, but it’s the best starting point for understanding where you really stand.
I ran the same test on my personal finance blog. For search queries I’d put a lot of work into – such as savings accounts in 2026 or the PEA for beginners – Perplexity lists two or three sites that I know well and which are genuine authorities in the niche. My blog doesn’t appear. What’s interesting is that the sites listed all have one thing in common: they have a lot of backlinks from forums and press articles. It’s not just about good content.
On the subject of the FAQs at the end of the article: I started adding them to my content three months ago, initially for Google’s rich snippets. Since then, two clients have told me that an AI assistant quoted my answer when they were searching for information on the topic. Whether this is a coincidence or not is hard to say, but the FAQ format does seem to be well received.
What strikes me about this article is that the GEO recommendations are almost identical, point by point, to what good SEO practice was already advising: provide clear answers, cite your sources, specialise, and build off-site authority. This is perhaps reassuring: there’s no need to reinvent the wheel, just to do better what we were already doing only halfway.
@Victor: That’s exactly right, and it’s a point that the best articles on GEO highlight. The difference lies not in the nature of the actions but in the reason behind them. Previously, we structured an article so that a human could read it easily. Now, we also structure it so that a model can extract the information accurately. The purpose of the content has changed, but not necessarily the rules of writing.
To be honest, I understand the general principle, but I’m not sure how to apply it to my bookshop. Could someone give me a concrete example of what a well-structured FAQ section might look like for a small specialist shop? I’m finding it hard to imagine what questions my customers might ask an AI about Occitan regional literature.
@Paul-Antoine: Think about the questions your customers ask you on the phone or in-store. Which Occitan authors should a beginner read first? What’s the difference between Occitan literature in the original language and in translation? Any books you’d recommend to understand the history of the Languedoc? These are exactly the sort of questions someone might ask ChatGPT – and they’re areas where you have genuine expertise that no one else has shared online.