Getting Cited by ChatGPT and AI: The 4 Real Levers

The StartingUp Summer, continued. Taking stock of a dense year also means moving from theory to practice: after the questions AI made obsolete and defining GEO, here's the question everyone asks us: how do you show up in ChatGPT and get cited by AI?

"How do we get ChatGPT to talk about us?" The question deserves better than the two answers already flooding the market: the fatalists' "it's a black box, there's nothing we can do", and the miracle-recipe sellers' "just hand us your budget". Reality sits between the two: there's no button, but there are mechanisms. They're known, documented, and you can act on each of them.

A brand shows up in the answers of ChatGPT, Claude or Perplexity through four paths: the model's memory, live search, the citability of its content, and the technical accessibility of its site. This article reviews each one, then lands on the tricky question: measurement. Promising nothing: here are the levers available, and what we actually know about them.

When a generative engine answers you, its text combines two sources: what the model learned once and for all during training, and what it just read by querying the web at the exact moment of your question. Two distinct doors in, with very different levers and timelines.

Models learn from vast corpora of text: press, Wikipedia, forums, review sites, comparisons, technical documentation. A brand present enough in these texts eventually becomes part of the model's representation of an industry: it knows the brand without needing to search. This is the deepest, and least steerable, lever there is. Training cycles span months, no action produces a fast effect on it, and nothing gets corrected on demand: wrong information the model learned stays there until the next generation.

In practice, you feed this corpus through means marketing has known for a long time: press relations, contributions to specialized media, presence in independent comparisons, genuine participation in the forums and communities where your industry gets discussed. We wrote this while defining GEO: the link used to be negotiated, the mention has to be earned. There's no shortcut for this path, and that's precisely what gives it its value.

The second door is much faster. ChatGPT, Claude and Perplexity no longer answer from memory alone: for any somewhat factual or recent question, they query a search index, read the top-ranked pages and compose their answer from what they find there. As we write this, Claude relies on Brave Search's index, Perplexity built its own, and OpenAI feeds its own index with a dedicated robot. The architectures differ, the principle is the same: behind every "live" answer, there's a search engine.

The consequence is counterintuitive for anyone quick to bury SEO: SEO remains the front door. A page absent from the top results is never read by the generative engine, so never cited. One key nuance: Google is no longer the sole judge. Your visibility on Bing or Brave Search, long neglected, now weighs on what millions of assistants say about you.

Being read isn't enough: the model also needs something to pick up. GEO's founding study already measured a visibility gain of roughly 30 to 40% for content enriched with statistics, citations and sources. Three years later, field observations converge on the same profile of citable content:

• Dated, sourced data. A precise figure, with its date and origin attached, gets picked up in one sentence while crediting you. A cautious generality gives nothing to grab onto.
• Clean definitions. A concept defined in one or two self-contained sentences, understandable out of context, is exactly the format a model can extract and cite.
• Answer pages. A real customer question as the title, a direct answer in the first paragraph, details afterward. The opposite of editorial teasing that makes the reader wait for the answer at the bottom of the page.
• Signed stances. An argued, owned opinion stands out from interchangeable content, for human readers as much as for machines trying to figure out who claims what.

None of this is a trick for robots: these are simply the qualities of good content. Generative engines just reward them more bluntly than Google ever did.

Generative engines read the web with their own robots, and every publisher operates several, each with a distinct role: gathering training data, feeding a search index, or visiting a page on a user's request. Each is controlled separately in your robots.txt file:

On Google's side, the logic differs: the classic Googlebot feeds the engine's generative answers, while the Google-Extended directive only controls whether your content is used to train Gemini. Either way, granularity is yours to control: you can, for instance, refuse training while staying present in search:

Two checks are essential before any strategy. First: review your robots.txt and your CDN or firewall configuration, since some services offer, sometimes by default, a blanket block on AI robots. Blocking training is a legitimate choice; disappearing from answers by accident isn't. Second: make sure these robots actually see your content. Most of them read the HTML as the server sends it, and don't execute JavaScript, or execute it poorly. A server-rendered site delivers its full text to them; an application rendered only in the browser hands them a mostly empty shell. It's one of the reasons, alongside performance and classic SEO, why we build our sites with server-side rendering.

There's no Search Console for LLMs yet, and anyone selling you a guaranteed "ChatGPT ranking" is selling you hot air. What does exist is a hand-crafted, honest method: list ten to twenty questions your customers actually ask, put them to ChatGPT, Claude and Perplexity every month, and note what comes back. Are you mentioned? Cited as a source? Is what's said about your offer accurate? Which pages, which competitors show up? Repeat each question several times and vary the phrasing, because answers fluctuate from one session to the next: you're after a statistical snapshot, not a single instant. A spreadsheet and some consistency are enough to get started.

A market of tracking tools is emerging to automate this listening: platforms specialized in brand mentions within AI answers, "generative visibility" add-ons bolted onto established SEO suites. We follow them with interest and caution: the sector is young, methodologies aren't settled, and the numbers they produce compare poorly from one tool to another. For a small or mid-sized business, the manual protocol described above remains, in our view, the best starting point: it costs a few hours a month and gets you to actually read what the machines say about you.

None of the four paths described here guarantees a citation, and no one can honestly promise otherwise: generative engines remain opaque, changing, and different from one another. But these levers have two merits. They're cumulative: citable content, on a robot-readable site, well ranked in the indexes, backed by real brand awareness, multiplies its chances at every step. And they're regret-free: each one also improves your classic SEO and, above all, what you offer your human readers.

That might be the most reassuring lesson of this emerging discipline: GEO doesn't reward some new form of trickery. It rewards, faster and more visibly than before, what the web has always rewarded: being a source others have good reason to cite. The work can start today; it's right now that the representations are being formed.