GEO Expert — GEO explained
Your organisation already has a reputation with AI. You didn't write it.
ChatGPT, Claude, Perplexity and Google's generative summaries are already answering questions about your association, your network, your town — with you or without you. GEO (Generative Engine Optimization) is the work that makes that answer the right one.
GEO, SEO, AEO: three words, three targets
SEO (Search Engine Optimization) targets a classic search engine: a page that gets indexed, that ranks, that someone clicks. GEO (Generative Engine Optimization) targets a generative agent: it doesn't return a list of links, it composes an answer, and cites — or doesn't cite — its sources inside the text. AEO (Answer Engine Optimization) is often used as a synonym for GEO; where a distinction exists, it concerns the direct answer boxes of classic engines (featured snippets, Google's AI Overviews) rather than purely conversational agents.
The difference that matters in practice: an SEO click is measurable in your analytics. A GEO citation may generate no click at all — the answer was enough for the reader — and still be the reason they chose your organisation over another. You cannot manage what you do not measure; that is precisely why a dedicated probe exists (see below).
How an agent discovers a source
A classic search engine crawls the web continuously and builds an index it consults at query time. A generative agent combines several mechanisms, not all of them visible from the outside:
- Training What the model learned once and for all, up to a cut-off date — a frozen fact, never updated between two versions of the model.
- Live retrieval Many agents (ChatGPT, Claude, Copilot, Perplexity, Google's summaries) run a real web search at the moment of the question, then summarise what they find — this is where crawlers DEDICATED to AI come in, distinct from classic SEO crawlers.
- The provider's own index Some agents (Mistral, for instance) rely on an index they control themselves: being absent from THAT index is not fixed by working on Google.
A direct consequence: a site invisible to GPTBot can remain
perfectly indexed by Google, and the other way round. Blocking those
crawlers by accident — in a robots.txt copied from a generic
template, or through an over-broad WAF rule — is one of the most frequent
causes of total absence, and the simplest to fix once seen.
The technical fundamentals, common to every agent
Before any content strategy, a technical floor decides WHETHER an agent can read the page at all.
- robots.txt, by name
Explicitly allow or refuse each AI crawler (
GPTBot,ClaudeBot,PerplexityBot,Google-Extended,CCBot…) — a genericDisallow: /copied from another site blocks some of them without anyone having decided it. - llms.txt A file at the root that tells an agent explicitly where the structured content of the site lives — a deliberate signal, not yet universally read by every agent, but already verifiable and increasingly consulted.
- Structured data (schema.org)
Event,Place,LocalBusiness,FAQPagemarkup states unambiguously WHAT each piece of information is — an event date, an opening time — where an agent would otherwise have to guess from free text. - The HTML basics everyone forgets
<title>, the meta description, the canonical link, thelangattribute, a sitemap kept current with its last-modified date — each of these is routinely missing on a page that is otherwise visually polished.
What each engine wants
Each agent favours a different signal — knowing that hierarchy keeps you from working for the one that matters least to your audience.
Leans heavily on existing SEO ranking: solid classic indexing stays the foundation, GEO is added on top, it does not replace it.
Requires clear attribution: an entity name and a source identifiable without ambiguity, not a text that assumes the reader already knows who or what is being discussed.
Insists on verifiability: a dated, attributed figure pointing at a source that genuinely exists — an unsourced claim is more likely to be set aside than cautiously rephrased.
Values attributable and recent sources: content that states explicitly where each number comes from is picked up more readily than an unsourced synthesis.
Asks for freshness: a current sitemap and IndexNow submission count for more here than on other engines.
Depends on its own index: your site's accessibility to ITS specific crawlers is the condition, independently of classic SEO or of your presence on other agents.
What we measured on our own ground
We are not selling an imported theory: Alliance MANA runs an open directory readable by AI agents (manafrance.org) and a territorial application (mana.bzh), and we measure their real traffic in our own database, not through a third-party tool.
Measured directly in the production database on 20 October 2026 — to be recounted before being quoted elsewhere: a living figure is only true on the day it is read.
- > 8.6 M
- AI agent visits per month across our two sites, measured in our own traffic table.
- > 6.9 M
- Places recorded from OpenStreetMap, made legible through structuring.
- llms.txt
- Published and readable at alliancemana.com/llms.txt — not a promise.
- 0
- AI crawlers blocked by accident — verified, not assumed.
Those two sites are French, and so is the ground we measured. We say so rather than implying a worldwide figure: what transfers across borders is the method, not the audience we happen to have.
Writing content worth citing
- The inverted pyramid The direct answer at the top of the section, the development after: an agent that summarises often reads the first sentences first, sometimes without going further down.
- Attributed data A source, a date, a scope (which population, which exact territory) and its stated limits: a figure without those four things is a figure a careful agent sets aside rather than misattributes.
- Consistent entities The same organisation name, written the same way, on the site and on the external sources that mention it — an AI links entities, not just character strings.
- Genuinely original content A tested method, a dataset of your own, an explicit comparison: what exists nowhere else is what an agent has reason to cite rather than paraphrase from a better-known source.
The mistakes that keep coming back
- Confusing a link with a call
An
<a href>to a third-party site is not a request the page makes; mixing the two up invents a problem that doesn't exist, or hides one that does. - A robots.txt copied without reading it A template found online sometimes blocks AI crawlers by default — the fix is usually one line, not a project.
- Measuring only afterwards The prerequisite to any correction is knowing what an agent ALREADY says about you — without that diagnosis, you correct at random.
- Believing one check holds forever
A site read and judged sound one day can regress by adding a
script, a
preconnect, or by changing arobots.txtwithout re-checking it — GEO is monitored, it is not ticked off once.
Frequently asked questions
Does GEO replace SEO?
No. SEO is still what makes a page exist for a classic search engine. GEO is added on top: it aims for a generative agent to understand, cite and correctly attribute the information, often without a single click being counted.
Do we have to rewrite the whole site for GEO?
No. GEO structures and republishes what already exists — it does not replace a site, it makes it legible from the outside to machines that do not read it the way a human does.
Is an llms.txt file enough?
No, but it is a real, verifiable signal rather than a cosmetic one: it has to come with a robots.txt that does not exclude AI crawlers, with structured data, and with content that answers directly, at the top of the page.
Does this work outside France?
The method does not depend on a country. What changes from one territory to another is the set of sources an agent reads — national registers, open-data portals, sector directories. Identifying those sources for your country is part of the audit, and we say so when a source we rely on in France has no equivalent where you are.
Where does our data go?
Your content and your data stay on a European chain. The measurement prompts, by the very nature of the product, are sent to the engines being tested — that is what measuring a public answer means. The audit includes a sovereignty sheet listing where processing happens, which engines are queried and what they receive, retention periods, and the providers involved.