GEO France Guide: the manual to exist in AI

This guide is the practical manual for GEO — optimization for generative AI. Across eight chapters, it explains why your brand vanishes from ChatGPT, Perplexity and Gemini answers, how these engines choose their sources, and which levers to pull to become citable again. No ivory-tower theory: the method we apply on engagements, laid out plainly, written in full below. You don't have to download anything to read it.

Why a dedicated GEO guide

The topic exists mostly in English. The reference manuals are American, built for the US market, its queries, its sources. But AIs don't behave quite the same in every language: the corpora differ, the cited sources too, and brands compete against a different field.

This guide fills that gap. It's written by an agency that runs GEO audits every week, for B2B SaaS and SMEs. What we put in isn't what we read — it's what we've seen work, and what we've seen fail.

The goal: that you understand the mechanics well enough to decide with full knowledge. Whether you apply it yourself or delegate it, you'll know what's happening under the hood. The eight chapters that follow are the complete guide, not a summary — you can read it in one sitting or jump straight to the chapter that concerns you.

The table of contents — 8 chapters

Chapter 1 — Why your brand vanishes from AI

Your brand vanishes from AI because search has shifted surface, not because your site is bad. For twenty years, visibility came down to one question: where do you appear in Google's results? Today, a growing share of searches no longer goes through a results page. The user asks their question to ChatGPT, to Perplexity, to Google's AI Overview — and gets a synthesized answer, with a handful of cited sources. You're in it, or you're not. There's no page two in an AI answer: there are three or four sources, and the rest doesn't exist.

Why ranking first on Google is no longer enough

Ranking well on Google no longer guarantees being cited by an AI. The two signals partly overlap — a poorly ranked page has little chance of being cited — but AI citation adds its own demands. A page can be number one on Google and invisible in generative answers, because it isn't extractable, because it isn't structured, or because AI bots simply can't read it. SEO optimizes for a human to click a link. GEO optimizes for a machine to copy your sentence into its answer. These aren't the same winners.

What "not being cited" really costs

Not being cited by AIs costs a silent erosion, not a sharp drop. You don't see your traffic collapse overnight: you see your competitors cited in your place, in answers your prospects read before they even reach a traditional engine. The brand that answers the question inside the AI captures attention, perceived authority and often the decision. The others become invisible, with no alert, no red line in a report. That's the danger: GEO produces no pain signal until you measure it. You lose ground on queries you weren't tracking, against brands you hadn't identified as competitors.

The lead window, and why it's closing

The good news is that this ground is new. Very few French brands have understood and invested in it. The lead window is real: on most B2B queries in French, AIs still cite the first structured content they find, for lack of better. Whoever publishes citable content today holds a spot no one is contesting yet. But that window closes as serious players move in. The lead is taken now, not when the market matures — by then, it will be too late to be the first cited.

Chapter 2 — How LLMs work (no jargon)

An LLM is a language model that predicts the most likely next word from what it has learned. To answer a question, it doesn't consult a database of facts: it generates text that resembles what's expected. This nuance changes everything for GEO. An AI doesn't "know" your brand like a directory does: it reproduces what recurs often enough, consistently enough, in the sources it has seen or consults live. Optimizing for an AI means raising the probability that your name and your phrasings are the ones the model judges most plausible for a given question.

Trained knowledge versus live search

Two moments must be distinguished. The first is the knowledge frozen into the model at training time: if your brand appears there, it's because it was present, repeated and consistent in the public corpus when the model learned. The second is retrieval-augmented search (RAG): ChatGPT with search, Perplexity, AI Overviews fetch pages in real time, read them, then synthesize an answer while citing their sources. For GEO, this second mode is the most actionable: it's where your page today can be read, extracted and cited tomorrow, without waiting for a future retraining.

How an AI chooses its sources

An AI in search mode chooses its sources on three cumulative criteria: relevance (does the page answer the query precisely?), extractability (can it isolate a clear, self-standing passage?), and trust (is the source credible, consistent, cross-checked elsewhere?). A highly relevant but bot-unreadable page won't be cited. A perfectly structured but off-topic page won't either. GEO works all three at once. That's why good content on a site technically closed to bots returns nothing — and why citable content on an obscure site can break through if the trust signals are there.

What LLMs favor

LLMs favor clear, structured, self-contained text. They readily reuse a definition phrased "X is…", an answer given upfront rather than buried after three paragraphs of introduction, a list or a table rather than a long block, and a passage that makes sense on its own, lifted out of context. They distrust vagueness, filler, unsourced claims, and pages where the information is drowned. Unlike old-school SEO, you don't win by repeating a keyword: you win by being the cleanest, most direct and most reusable phrasing of an answer.

Chapter 3 — The 5 technical levers

The five technical levers of GEO are the foundation: without them, the best content stays invisible to AIs. They are bot accessibility, structured data, speed and rendering, page structure, and the llms.txt file. We handle them first because they condition everything else: an AI bot that can't load your page will never read the citable passage you crafted. The ordering rule is simple — make the page machine-readable before polishing what the machine reads in it.

Lever 1 — AI bot accessibility

Bot accessibility is the most neglected lever and the most costly when it's missing. Many sites unknowingly block GPTBot, PerplexityBot, ClaudeBot or Google-Extended, either in their robots.txt or via a firewall or CDN that returns a 403 to AI crawlers. The result: the page is perfect, but no AI can read it. The basic check consists of inspecting your robots.txt, explicitly allowing the AI bots you want to be cited by, and examining your server logs to spot requests from these bots refused by your infrastructure.

Lever 2 — Structured data (schema)

Structured data tells the machine, plainly, who you are and what the page is about. Organization or Person markup establishes your identity; Article declares the author, date and title; FAQPage exposes questions and answers directly reusable by an AI in an answer; BreadcrumbList places the page within your hierarchy. This isn't an SEO gadget: it's a layer of meaning that LLMs read first to understand your page without having to interpret it. Valid schema, free of blocking errors, increases machine trust. Broken schema is useless, even harmful.

Lever 3 — Speed and JavaScript-free rendering

Speed and rendering condition what the bot actually sees. Most AI crawlers don't render heavy JavaScript: if your main content is injected client-side, they read only a blank page. The citable content must be present in the raw HTML, served quickly, on mobile as on desktop. A crawler that times out or finds only empty tags leaves with nothing. The concrete test: display your page's raw source (without executing the JS) and check that your main answer appears there as text.

Lever 4 — Page structure

Page structure guides extraction. A clean heading hierarchy — a single H1, H2s and H3s that pose clear questions or assertions — lets the AI cut your page into passages and match each section to an intent. Lists, comparison tables, boxed definitions are formats engines reuse easily. A well-structured page isn't just pleasant to read: it's segmentable, therefore citable block by block. Conversely, a wall of text with no headings forces the machine to guess, and it would rather look elsewhere.

Lever 5 — The llms.txt file

The llms.txt file is a text file placed at the site's root that guides AIs toward your key content, like a sitemap designed for language models. Few French sites have one today, which makes it a first-mover advantage. It replaces no other lever, but it signals to generative engines which pages matter and how to understand them. Paired with a clean text or markdown version of your important pages, it eases extraction and marks your site as deliberately open and organized for AIs.

Chapter 4 — Optimizing your content for citation

Optimizing content for citation means writing passages an AI can copy as-is into its answer. The central rule: each section must be liftable out of the page and remain understandable on its own. An AI doesn't extract your whole article — it takes a piece. If that piece depends on what precedes it ("as we saw above"), it becomes unusable. The editorial work of GEO consists of building self-contained, dense and direct blocks the machine can lift without breaking anything.

The direct answer upfront

The direct answer upfront is the most rewarding reflex. Give the answer to the section's question in its first sentence, before any context. AIs extract what answers, not what introduces. A paragraph that starts with "In a world where…" will be ignored in favor of one that starts with "GEO is…". This inversion — conclusion first, justification after — is the opposite of the school essay, and it's exactly what generative engines reward. Every H2 and H3 should be followed by a clear assertion in the first line.

Self-contained passages of 130 to 170 words

A citable passage runs between 130 and 170 words and is understandable out of context. That's the size LLMs reuse most readily: long enough to be complete, short enough to fit in an answer. Each passage handles a single idea, with no reference to another paragraph, no pronoun whose referent is elsewhere. Break your 600-word blocks into self-contained chunks under clear subheadings. The test: copy a paragraph alone, read it without the rest of the page — if it stands up and answers a question, it's citable. If not, rewrite it so it does.

Definitions and heading format

Self-contained definitions and question-headings are two formats AIs love. Define your key terms on the spot, in the "X is…" style, without sending readers to another page: that's exactly the structure an AI reuses to explain a concept. On the heading side, phrase your H2s/H3s like your users' real questions ("How do you measure AI visibility?") or as sharp assertions. The machine matches question to answer: a heading that poses the question and a first paragraph that answers it directly form the ideal pair. Lists and tables round out the arsenal for comparisons and steps.

Banning filler

Filler is the enemy of citability. Every hollow paragraph between a question and its answer dilutes the passage and lowers your chances of being extracted. Cut the empty transitions, the "it's important to note that," the introductions that say nothing. An effective GEO page is dense: useful information on every line, zero padding. This demand meets a deeper principle — you don't write to fill a page, you write to answer a question as cleanly as possible. Content that's thin in signal but fat in words is precisely what AIs avoid citing.

Chapter 5 — Off-site presence

Off-site presence is the set of signals about your brand that exist somewhere other than on your site — and AIs cross them before citing you. An AI trusts a brand mentioned, discussed and cross-checked across several independent sources more than a brand that only talks about itself on its own domain. Your site says what you want people to believe; the rest of the web says what's true. Off-site GEO consists of building, honestly, that third-party presence engines use as proof.

Why your site isn't enough

Your site isn't enough because it's both judge and party. An AI that has to decide whether your brand is a reference on a topic seeks outside confirmation: does your name appear on sites you don't control, in contexts consistent with what you claim? Without those third-party mentions, your expertise is just a self-declaration. With them, it becomes a cross-checked fact. It's the transposition to AIs of a simple principle: reputation is built in other people's mouths, not in your own.

Reddit, forums and YouTube

Reddit, forums and YouTube are sources AIs draw on heavily for community signal and video. When your topic is discussed where users genuinely exchange — not by spamming, but by participating usefully — you widen the surface where an AI can encounter your name. A video that covers your topic adds a source generative engines increasingly tap. The idea isn't to be everywhere artificially, but to exist where the conversation happens, with real contributions that hold up under scrutiny.

Press mentions and quality links

Press mentions and inbound links from credible sites remain authority signals AIs partly inherit from Google. A citation in a sector media outlet, a guest article on a reference site, a link from a recognized source: all of it builds the layer of trust engines cross. Link building hasn't disappeared with the arrival of AIs — it has changed purpose. You no longer aim only to pass "juice" to a page, but to inscribe the brand in a network of credible sources the models recognize as reliable.

Brand consistency (NAP and entity)

Brand consistency is the simplest off-site signal and the most often neglected. Your name, address, contact details and description must be identical everywhere: site, directories, social profiles, listings, press mentions. An inconsistency — two spellings, two addresses, two contradictory descriptions — sows doubt about your identity and lowers machine trust. A consistent entity, linked to its official profiles via Person/Organization schema, lets the AI cross-check "this really is the same brand" at every encounter. Consistency goes unnoticed, but its absence is paid for.

Chapter 6 — Measuring your AI visibility

Measuring your AI visibility means regularly checking whether generative engines cite you, on which queries, and how it evolves. Without that measurement, you steer blind: you know neither where you're winning nor where competitors are taking your place. The difficulty with GEO, as said, is that it sends no spontaneous pain signal. Measurement is therefore what turns GEO from a bet into a steerable lever. It rests on four practices: testing citations, tracking AI referral traffic, benchmarking competitors, and doing it on a regular cadence.

Testing whether AIs cite you

Testing citations means putting your business queries to ChatGPT, Perplexity and Gemini, and noting who's cited. List your ten to twenty priority queries — the ones that matter for your business, not a vague average — and query each engine on them, regularly. Are you in it? In what spot, in what context? Aren't you? Who holds the ground? This manual test, run methodically and logged, already gives 80% of the diagnosis. Per-query granularity is what guides action: you don't optimize "AI visibility" in general, you work a specific query where you're absent.

Tracking AI referral traffic

AI referral traffic exists and is measured in GA4. When a user clicks the source cited in an answer, the visit arrives with an identifiable referrer (the domains of generative engines). By segmenting that traffic, you see which pages AIs actually send to, and which convert. It's not yet a volume comparable to classic SEO, but it's an early, qualified indicator: a visitor coming from an AI answer has often already been pre-sold by the citation. Tracking this segment means spotting early the pages that are breaking through and those that deserve reinforcing.

Benchmarking your competitors

Competitive benchmarking answers a simple question: who's cited in your place, and why? On each of your priority queries, identify the brands AIs cite. Go look at their pages: what do they have that you don't? A more direct answer, better structure, stronger off-site presence, better-laid schema? The target is defined by comparison. GEO being new ground, the brand cited today isn't always your historical commercial competitor — sometimes it's a player that simply published the right citable content before you. To be identified, to be overtaken.

How often to measure

GEO measurement is steered as a monthly review, not a one-off setting. Corpora evolve, engines change, your content and your competitors' shift. Once a month, run your priority queries again, compare to the previous measurement, note the gains and losses, and decide on corrections. GEO isn't a project you finish: it's a process you maintain. This monthly cadence is the minimum to see a trend and react before a competitor settles durably on a query that was yours.

Chapter 7 — The 90-day action plan

The 90-day plan unfolds GEO in three 30-day phases, in the right order: first make the site readable by AIs, then make it citable, then inscribe it in its ecosystem and steer it. You don't do everything at once. You tackle what unlocks the rest — technical access — before polishing the content, and you only measure seriously once there's something to measure. Here's the realistic sequence we apply on engagements.

Phase 1 (days 1-30) — Diagnosis and technical foundation

Phase 2 (days 31-60) — Content and citability

Phase 3 (days 61-90) — Off-site and steering

Chapter 8 — Real-world brand cases

The cases that follow are typical situations, observed in the field and presented generically. They illustrate the patterns that recur — what unlocks AI visibility, and what makes it stall — without naming a brand. The goal isn't the anecdote, but the pattern: recognizing your own situation in one of these cases and knowing which lever to pull.

The B2B SaaS first on Google, absent from AIs

A B2B SaaS vendor ranked at the top on its Google queries but never appeared in ChatGPT or Perplexity answers. The diagnosis revealed two cumulative causes: its main content was rendered in client-side JavaScript, invisible to AI bots, and its pages, though rich, buried the answer after long introductions. The fix addressed rendering (content served in raw HTML) and rewriting into direct answers with self-contained passages. The pattern is classic: excellent in SEO, absent in GEO, for lack of extractability. Being good on Google wasn't enough — it had to be readable and directly citable.

The industrial SME invisible for lack of off-site

An industrial SME had a clean, well-structured site but stayed absent from AI answers in its sector, where other players were systematically cited. The page wasn't at fault: it was the third-party presence that was missing. The brand barely existed off its domain — few mentions, no participation in the sector's discussions, an approximate entity consistency across its various profiles. The work consisted of locking down brand consistency and building a real off-site presence. The pattern: a good site isn't enough when AIs find nothing to cross-check elsewhere.

The solid content blocked by robots.txt

A brand had invested in quality editorial content, structured and well written, without seeing any return in AI visibility. The cause was singular and brutal: its robots.txt and its CDN blocked the main AI bots, which had never been able to read a single page. The perfect content was locked away. Unblocking access changed everything by putting the editorial work back in play. The pattern is the most frustrating and the most common: months of content effort canceled by a configuration file. That's why the access audit is always the first step — an invisible wall makes everything else useless.

The first mover that grabs the query

On a niche query still little worked in French, a brand was the first to publish citable content — direct answer, self-contained passages, clean schema, open access — and settled in as a source cited by several engines, for lack of serious competition. The pattern illustrates the lead window from chapter 1: on queries no one has yet optimized for AIs, the first to do the work correctly takes the spot. The lesson isn't "publish anything fast," but "identify the still-vacant queries and be the first to answer them cleanly."

Get the guide in PDF

You've just read the full guide — all eight chapters are on this page. If you want the PDF version to keep, print or circulate internally, tell us where to send it. First name, email, your role and your company size — so we can then send you what actually applies to you, not a generic newsletter.

Get the full guide →

Zéro spam. Désinscription en 1 clic.

To move from the guide to action, two direct complements. The 40-point checklist puts chapter 3 into practice. And if you'd rather learn in small daily doses, the 30-day ChatGPT masterclass has you apply one action a day for five days. To delegate the implementation, our GEO agency page lays it all out.