What Google Says to Ignore for AEO/GEO (and What You Should Really Do)
Summarize this article with AI
37 000 organic sessions. 14 months ago, the same site had 4 000.
37 000 organic sessions.
14 months ago, the same site had 4 000.
No Google Ads. No bought backlinks.
Just a quiet overhaul of semantic architecture.
A client calls me on a Monday morning.
He’s shell-shocked.
His e-commerce site, 1 200 references of French designer furniture, isn’t showing in AI Overviews.
Worse: he spent 4 700 € with an agency to fix it.
The agency sold him a full AEO package: LLMS.txt files, content chunking, exotic JSON-LD tags, Markdown pages reserved for bots.
Result after 5 months: zero impressions in generative answers.
Not a single one.
I can hear the shame in his voice on the call.
Like he’d wasted his budget on empty promises.
I tell him: « Stop everything. Let’s go back to basics. We’re rebuilding your silos. »
He agrees.
We audit 945 pages.
We group product families into clusters.
We rework every H1, every title, every meta description to reflect one search intent.
We restructure internal navigation to strengthen co-occurrence links.
In 14 months, organic sessions jump from 4 000 to 37 000.
And AI Overviews? He now captures 47 distinct generative queries.
Some place him at the top of the snippet before even the paid ads.
The secret wasn’t in an LLMS.txt file.
It was in the architecture of his HTML pages—already understood by Google.
Google le dit clairement : pas besoin de fichiers spéciaux. Les chiffres le confirment : 87% des pages présentes dans AI Overviews n’ont aucun fichier llms.txt. La majorité silencieuse a raison.
Les pages qui apparaissent dans AI Overviews utilisent-elles llms.txt ?
Seulement 13% des sites captant des réponses génératives ont un fichier llms.txt
What Google officially tells you to ignore
Yesterday, Google published an official guide for publishers who want to improve their presence in generative answers.
A thick one.
Yet three lines are enough to sweep away most emerging practices.
The guide says it plainly:
- No need to create machine-readable files like LLMS.txt.
- No need to « chunk » your content into small pieces so the AI understands it better.
- No need to add special markup or convert your pages to Markdown reserved for bots.
Translation: everything some agencies charge you 300 to 5 000 € for has zero technical impact on your visibility in AI Overviews.
Excerpt from Google’s guide: « You don’t need to create new machine readable files, AI text files, markup, or Markdown to appear in generative AI search. »
Google explains it can crawl, index, and understand many file types, including your standard HTML pages. That doesn’t mean an LLMS.txt file gets special treatment. It’s just indexed like any other content.
Yet I watch entrepreneurs waste 3 weeks every week on Markdown files for AI while their competitors optimize their thematic silos.
The worst part? The guide also clarifies another point few SEOs mention: Google’s systems analyze the entire page. Not chunks. Not pieces.
File format hardly matters.
Only information architecture counts.
Semantic architecture: the only fuel for generative answers
A site appears in AI Overviews not because it sent Google an LLMS.txt file.
It’s because its pages are structured as silos.
I’ve been building silos since 2016.
650 clients.
1 300 silos delivered.
The principle is simple: organize your pages into thematic clusters, create content pillars, link every page to its parent and neighbors. The algorithm then grasps the site’s domain of expertise.
No external file needed.
In March 2025, a bakery equipment site, 2 100 pages, 70% orphaned product sheets, contacted me.
Their goal: appear in generative snippets on queries like « best professional bakery mixer. »
We audited the pages. We built 12 silos.
Each silo answered one precise intent.
Result: in 4 months, impressions from AI Overviews across the entire site jumped +610%.
No line of Markdown.
No LLMS.txt file.
Of the 45 sites I audited in 2025, none capturing generative answers had implemented a special file. Their strength: coherent internal linking, sharp standard tags, silo structure.
The DOSE framework I apply—Demonstrate, Optimize, Structure, Evaluate—rests on one principle: the AI doesn’t hunt for chunks of text. It hunts for complete, reliable, well-architected pages.
Everything you do to strengthen your silos also serves AEO.
With no extra effort.
Le site de mobilier design a doublé son trafic en 14 mois, sans aucun fichier exotique. Voici l’évolution chiffrée des indicateurs clés avant et après la restructuration en silos.
Avant / Après : l’effet de l’architecture sémantique
De 4 000 à 37 000 sessions organiques, et 47 requêtes capturées dans AI Overviews
+820% impressions in AI Overviews in 3 months. Without one Markdown file.
Back to the bakery equipment site.
Before intervention: not a single impression in AI Overviews across 45 queries in their market.
After 3 months: 47 queries captured, some at position zero.
Category pages for « Professional Mixers », « Industrial Laminators », and « Bakery Accessories » started generating snippets directly in generative answers.
Average click-through rate from these snippets: 4.2%.
Impressions for the affected pages (not just snippets, but standard blue links shown under generative answers) rose +820% versus the previous quarter.
What did we do differently?
- Restructured the 12 silos.
- Rewrote title tags to include concrete questions (« Which professional mixer to choose? »).
- Added a table of contents at the top of each pillar page.
- Cross-linked blog posts and product sheets, no external file.
The beauty of the system: these pages have been live for months already. Google didn’t need an extra signal to push them forward in AI. It simply recognized their thematic authority.
It’s the architecture that speaks.
Not the format.
Why AEO specialists sell you the opposite of the solution
The SEO market loves new names.
« AEO », « GEO », « LLM Optimization »… Every buzzword spawns new offers.
The problem? Most sell a band-aid on a wooden leg.
Creating an LLMS.txt file takes 2 hours. Auditing silos takes 40. Selling the first brings in 300 € quick. Selling the second demands pedagogy, analysis, time. So some choose commercial shortcuts.
But when Google itself says it’s useless, case closed.
What I see with my clients: sites capturing AI Overviews are often those who’ve never heard of LLMS.txt. Their leaders invested in content quality and structure.
The irony? Sites with hundreds of hours sunk into Markdown files for AI end up behind raw product sheets—but well-linked ones.
Google’s guide clarifies one precise point: crawl and indexation haven’t changed. Generative AI draws from the standard index.
Your efforts focus on what makes your pages indexable and authoritative: content, relevance, linking.
Your AEO checklist in 3 steps (without spending 1 € on exotic files)
Here are the three checks I run on every e-commerce audit.
1. Assess the depth of your silos.
Look at a category page. How many clicks away are its sub-pages? If a product sheet is more than 3 clicks from the homepage, it’s too deep.
2. Make sure every page answers one precise question.
AI Overviews love question/answer formats. Rewrite your H1s and product titles to include intent. « Desk lamp » becomes « What desk lamp works best for a designer’s office? ». This shift, with solid content, makes your page eligible for AI Overviews.
3. Forget LLMS.txt.
If you want to optimize crawl, a solid XML sitemap does the job. LLMS.txt files are a distraction. Your time is worth more.
A personal tip: When an e-commerce owner asks me where to start, I always say the same thing: take your best category page, map its silo, list its 20 related pages, check anchor consistency. In 2 hours, you’ve laid the foundation for a working silo. And no Markdown file will ever match that.
Are you sure your site doesn’t need LLMS.txt to capture AI?
The killer question.
I ask it on every discovery call.
And often, silence settles in.
Because deep down, many entrepreneurs have been conditioned to think a new acquisition channel requires an extra technical layer.
Reality is the opposite: the work you do for human visitors is exactly what Google expects for its generative answers.
A well-structured site, pages that answer questions, logical links between them, expertise. That’s what the AI asks for.
So I leave you with one simple question:
When did you last verify your internal linking actually made sense?
I’m not selling you the method. I’m showing you the pages.
A live audit of your e-commerce site in 45 minutes
I’m not selling you an AEO package. I take your site, show you on screen what’s broken in your silos, and explain what to fix to capture generative answers. No LLMS.txt file.
Book a strategic call — 45 minFrequently Asked Questions
What is an LLMS.txt file and why does Google say it does nothing for AEO?
An LLMS.txt file lists a site’s URLs in Markdown format. AIs often use it. Google gives it no special treatment for generative answers. It indexes it like any other file, with no particular weight.
Should I delete my existing LLMS.txt file?
No. You can keep it. No penalty risk. But if you want to show up in AI Overviews, spending time or money on it is a waste. Focus on your HTML page architecture.
Can chunking content improve my visibility in generative answers?
Google’s guide says chunking is useless. Google understands a complete page. Cutting your content into small blocks does nothing for AEO and can hurt your visitors’ reading expérience.
How can I tell if my site is well-structured for AI Overviews?
I see many sites forget to check silo depth (less than 3 clicks), the presence of explicit questions in their H1s and meta descriptions, and linking quality between pages of the same topic. A quick semantic audit of your best category pages is enough to get clarity.
Should I use special JSON-LD tags to appear in generative answers?
Standard structured markup (Schema.org, FAQ, Breadcrumb) is useful for other rich result types. For AI Overviews, Google enforces no special AEO markup. The AI uses only your HTML page content to generate an answer.
