GEO: how to get cited by ChatGPT and AI assistants (the new SEO)

Your future customers no longer always type “best plumber in Lyon” into a search engine. They ask an assistant: “which plumber near me can I trust?” — and the assistant answers with two or three names. If you are not one of them, you do not exist for that person. That is the whole point of GEO, generative engine optimization: becoming the answer the AI cites, rather than merely the link it offers you to follow.

GEO, AEO, SEO: what are we talking about

SEO places you in a list of links. GEO — sometimes called AEO, answer engine optimization — gets you cited inside the answer an assistant writes: ChatGPT, Gemini, Perplexity, Copilot, or the generated summaries that now sit on top of so many results pages.

The difference is brutal, and it comes down to a single number. A results page shows ten links, plus more as you scroll. An assistant’s answer cites two or three. Visibility no longer spreads out, it concentrates — and the third name cited gets more attention than the third blue link ever did.

What has actually changed

Two shifts happened at the same time, and they are often confused. The first is the arrival of generated summaries at the top of the results: the engine answers for itself, and the click becomes optional. The second runs deeper: a share of searches simply no longer goes through an engine at all. They start in a conversation, continue with follow-up questions, and end with a supplier’s name — without a single results page ever being displayed.

It is that second shift that counts, because it is invisible in your analytics. A drop in organic traffic gets noticed; a conversation in which your competitor was cited and you were not leaves no trace whatsoever. Companies that discover GEO usually discover it through a customer’s remark — “I asked ChatGPT and it gave me your name” — or, more often, through the absence of one.

How models decide who to cite

An assistant does not rank the way an engine does. It looks for the answer that is most reusable as it stands, and four properties make the difference.

Structural clarity comes first. A heading that poses a question, a clear answer in the first two sentences, then the detail: that is a format a model can extract without getting it wrong. A page that opens with three paragraphs of scene-setting before reaching the point will be read, but not quoted.

Then come the verifiable facts. A figure, a timeframe, a definition, a comparison table — that is what a model quotes willingly, because it is what it can reuse without rephrasing at its own risk. Vagueness, on the other hand, is never picked up: “fast turnaround” cannot be quoted, “on site within 48 hours on weekdays” can.

Consistency across the web plays the role links used to play in SEO. Your name, your activity, your address and your description must be identical on your site, on your business listing, in professional directories and on your social profiles. A model cross-checks; a contradiction between two sources does not make it choose, it makes it give up and move on to the next one.

Finally, third-party mentions. What the web says about you carries more weight than what you say about yourself, and models lean heavily on the sources already present in their data or reachable on the fly: reviews, articles, forums, sector directories, comparison sites.

The six-step action plan

The technical detail almost everyone forgets

The assistants that browse read your site with crawlers of their own, and those crawlers can be blocked. Plenty of sites block them without knowing it, through a robots.txt file copied over years ago or an application firewall rule that treats any unknown agent as a threat.

Checking takes five minutes and comes before everything else: producing perfectly quotable content behind a closed door achieves nothing. In the same spirit, an llms.txt file at the root — a plain-text page describing what the site does and where to find what — costs an hour and removes a great deal of ambiguity for a model discovering your domain.

Server-side rendering matters too. Content that only appears once JavaScript has run is read by modern search engines, but not always by assistant agents, which often work from raw HTML. If your sales argument sits in a tab that only fills on click, it is entirely possible that no model has ever seen it.

Measuring, without kidding yourself

There is no Search Console equivalent for assistants. Measurement is therefore done by hand, and it is less tedious than it sounds: draw up a list of ten to fifteen questions your customers genuinely ask, put them to three or four assistants every month, and note for each one whether you are cited, who is cited in your place, and what the cited page says.

Two precautions. First, an assistant’s answer is not deterministic: the same question can produce two different answers ten minutes apart, so an isolated absence proves nothing and it is the trend over several months that tells you something. Second, turn off personalisation or test from a fresh session, otherwise what you are mainly measuring is your own account history.

A second, quieter indicator is worth tracking: referral traffic coming from the assistants’ domains. The volume is low, but its quality is generally above average, because someone who clicks through from a generated answer already has their information and is looking to go further.

The mistake that makes you invisible

Writing to look good rather than to answer. A page that circles the subject, with no figures, no structure, no direct answer, is unusable for a model: it will read it and move on to the next. The competitor cited in your place is not necessarily the biggest or the best ranked — it is the one whose page contained the sentence the model could lift as it stood.

The good news is that this constraint pulls in the same direction as your human readers’. Nobody has ever complained about a page getting straight to the point.

Where to start

Three things, in this order. Check that the assistants’ crawlers can read your site. Pick the five most important questions in your market and give each one its own page, with a clear answer in the first two sentences and at least one hard figure. Then test, take notes, and start again the following month.

Our GEO page sets out the full method and what we put in place for our clients; the checklist goes through the points to verify one by one, on a single sheet. For an initial picture, the SEO audit analyses the structure of your pages free of charge, and our pricing shows what ongoing support costs. If you would rather first find out where you stand, get in touch: we will look together at what the assistants answer today to your market’s questions.