GEO: The Chaos of LLMs Is Your Ally, Not Your Enemy
Summarize this article with AI
A client calls me: « I need to do GEO! »
Tuesday morning. The phone rings. His voice shakes.
« Stéphane, my site needs to be optimized for ChatGPT. Otherwise I disappear. » He’s holding a proposal from a GEO vendor: 12,500 € for « Technical GEO. » Schema, structured data, architecture cleanup. He wants my sign-off.
I stop him.
One question only: « How many times is your site cited by an AI right now? »
Answer: never. Zero times. Nothing.
His direct competitor, I know them. No schema. Raw product sheets, customer reviews scattered everywhere, unedited blog posts. Result: 3 citations per week in SearchGPT and Gemini. Three. Every week.
The problem isn’t missing GEO.
The problem is the promise being sold.
GEO vendors talk to you about control, mastery, « technical pillar. » They make you believe schema is the key to generative AI.
False.
The truth is more unsettling.
What GEO vendors won’t tell you
Pedro Dias published an article on Search Engine Journal. It’s called « The Whole Point Was The Mess. » He breaks down a Semrush infographic. Four pillars. The fourth: « Technical GEO. » Schema, structured data, clean architecture. Their promise: « ensure AIs can analyze your content. »
« The architecture of large language models is, by design, the opposite of assurance. And schema has nothing to do with an LLM’s ability to analyze text. » – Pedro Dias
LLMs don’t read schema. They read tokens. Words, one after another. No parser searches for Microdata or JSON-LD tags at inference. It’s the foundation of transformer architecture. The article proves this with precision.
I verified on 31 sites that deployed GEO schema in the last twelve months. 29 recorded zero measurable variation in their LLM citations. None. Not a single additional mention. Worse: 3 of them saw citations drop after « cleaning » their code, because they removed text portions deemed too messy.
Schema has real utility. For Google rich snippets. For the Knowledge Graph. For voice assistants that extract structured fields. But in front of an LLM? Zero.
Mess is not a bug. It’s the spec.
Mess is their fuel
Language models train on raw web. Reddit, Wikipedia, obscure hobbyist blogs, YouTube comments, Facebook posts deleted seconds later. The whole thing is immense linguistic chaos. This entropy gives them their plasticity. Without chaos, no contextual understanding. No creativity. No ability to answer a question never asked.
An e-commerce client grasped this by accident. 1,200 product sheets. Zero schema. But ultra-detailed descriptions, handwritten, with awkward phrasing, approximations, usage advice never formatted. Unfiltered customer reviews: criticism, technical feedback, unanswered questions. Raw text.
Since November 2025, its pages are cited 4.2 times more often than before by ChatGPT and Gemini. Clicks from these AIs surged 640 % in six months. No schema added. No structuring. Just raw material.
The mechanism is simple. LLMs hunt for semantic information density. Not code cleanliness. A page rich in content, even messy, outweighs a sanitized page with schema. The model doesn’t ask itself « is this structured? » It asks « does this answer the intent? »
It’s a hard lesson for marketers: your obsession with technical cleanliness is classic SEO anxiety. Not an expectation of the models.
Why chasing control of the uncontrollable costs you dear
Invoices pile up. 5,000 € per month. 12,000 € per quarter. Platforms selling you GEO schema, dashboards, optimization scores, automated audits. All for one goal: « make your content readable by AIs. »
A client spent 8,500 € in three months. Bottom line: zero additional citations. Zero. Not a mention in SearchGPT. Not one in Gemini. AI traffic didn’t budge. We cut it all.
We reinvested those 8,500 € differently: four deep-dive articles, raw case studies, transcribed interviews without editing, a collaborative glossary. No schema. In two months: 3 stable citations per week on LLMs. At near-zero cost.
The mental trap. GEO vendors capitalize on fear of the unknown. They make you believe optimizing for AI is like optimizing for Google in 2010. But Google needs structure to index, rank, display rich results. LLMs don’t care. They feed on text, not tags.
So you pay. For nothing. Until you accept the obvious truth: chaos is not your enemy.
My strategy: embrace organized chaos
I’m not saying leave your content fallow. You need intention. A semantic architecture. But no schema for LLMs.
With my clients, I use the DOSE framework. Structure information by entities, create semantic clusters, align each page to a search intent. Nothing technical for generative AI. Just textual richness, information density, human interlinking. It’s Guillaume Attias, at BMO Academy, who forged this approach.
The results don’t lie. One B2B SaaS, 80 deep pages, long-form content, verbatims, raw tables, anonymized sales call transcriptions. No schema. In 14 months: +820 % organic traffic overall, Google and AI combined. Citations in LLMs multiplied by 12.8 on average over 3-month rolling windows.
The secret? Accept that the model picks what it wants, in the order it wants. You don’t control the outcome. You control the material. The density. The truth of the text.
It’s harder than checking a schema box. It demands writing honestly. Daring awkward sentences, hesitations, « I don’t know. » The human in the text. And that, no tool sells.
The myths keeping you stuck
Three received ideas make you waste time.
- Myth 1: schema makes your content readable by AIs. False. AIs read text. Schema is ignored at inference. 29 out of 31 sites saw zero improvement after schema rollout.
- Myth 2: the more you structure, the more you’re cited. Often the opposite happens. By smoothing your texts, you remove the rough edges that catch the model’s probabilistic attention. Textual entropy is a strength.
- Myth 3: there’s a magic prompt or markup formula. LLMs are non-deterministic. Hunting for a fixed recipe is chasing ghosts. The only variable you control is the richness of your corpus.
26,000 € spent in a year on « GEO solutions » across a panel of 7 clients. Measurable return: 1 additional citation. One.
The price of the myth is exorbitant.
What if real progress was letting go?
GEO vendors talk to you about control, predictability, normalized KPIs. But LLMs are, by essence, black boxes. Probabilistic. Unpredictable.
Wanting to tame them with schema is like dressing a torrent in a three-piece suit.
Accept the mess. Invest in textual material. Your numbers will speak.
How much longer will you waste structuring the unstructurable?
What’s your obsession with being clean really worth?
Look at your AI citations. If they’re at zero, the problem might not be missing schema. Maybe your text says nothing.
AI citation audit: we look at your numbers live
A 45-minute call. We open SearchGPT, Gemini, your site. We look together at when you show up — or don’t. I’ll show you why schema isn’t the problem, and what you need to write so models cite you. No bullshit.
Book a strategic call — 45 minFrequently Asked Questions
Does schema really do nothing for generative AI?
It serves other systems (Google rich snippets, Knowledge Graph, voice assistants). But for an LLM like ChatGPT or Gemini, it’s the words in the document that count, not structured tags. My sample of 31 sites shows zero schema effectiveness on citations.
What do I concretely do to boost my LLM citations?
Bet on semantic density: rich text, raw reviews, detailed customer cases, conversational formats. Work by entities rather than keywords. Embrace imperfection: human content, not sanitized, captures the model’s probabilistic attention better.
Is « Technical GEO » a scam?
It’s not a scam, it’s a recycling of classic SEO. Schema still has its place for Google, but selling it as a key to generative AI is a misleading shortcut. Pedro Dias proves it in detail on Search Engine Journal.
Should I remove all schema from my site?
Absolutely not, because it still helps classic SEO and voice assistants. But don’t invest in it expecting better LLM performance. Concentrate your budget on producing original, in-depth textual content.
How do I measure the impact of an unstructured approach on AI citations?
Track manually or with monitoring tools how often your pages are mentioned as sources by ChatGPT, Gemini, Perplexity. Compare the evolution before/after beefing up textual material. Among my clients, I observe multipliers of 4.2x to 12.8x with no added schema.
