Mehdi Kabbaj

Posted on Jul 16

We scanned 474 French SMB websites for AI-readiness. The average score was 48/100.

#ai #data #seo #webdev

Here's a paradox we didn't expect to find so clearly in the data: French small businesses almost never block AI crawlers, but their sites give those crawlers almost nothing worth reading.

We run an AEO (AI Engine Optimization) scanner at Les Créavores, a French web agency, and we wanted real numbers instead of vibes before writing anything about "AI search readiness." So we pointed our audit engine — the same one behind our free scanner — at a sample of French SMB websites and measured what's actually there, structurally, for an AI agent or crawler to parse.

Methodology, briefly

474 sites analyzed (509 collected, 35 were not analyzable — dead links, JS-only shells that timed out, etc.)
10 sectors × 15 cities across France
Sample = top 10 organic Google local results per sector/city combo, collected via Brightdata
Collection date: 2026-07-13
Homepage-only audit — we did not crawl full sites. This measures structural readiness, not actual AI visibility or citation frequency. Those are different questions, and we're not claiming this dataset answers the second one.

The full dataset is public under CC-BY 4.0 with a DOI, links at the bottom if you want to dig into the raw numbers yourself.

The headline number

Average readiness score: 48/100 (median 49). 80.8% of sites scored below 60. Only 1.3% scored above 80 ("excellent" in our scoring band).

Sub-scores tell you where the points are actually leaking:

Sub-score	Average	Max
Crawlability	14.9	30
Structure	17.4	35
Trust	15.6	35

None of the three dimensions is close to saturated. This isn't one weak spot dragging down an otherwise solid score — it's a broad, even shortfall.

Agentic navigation is close to absent

We check three basic signals of "can an AI agent actually work with this site": a structured accessibility tree (semantic landmarks like <main>/<nav> plus alt text coverage), a usable llms.txt, and genuinely open access for agents (no blocked AI bots, no noindex, canonical and viewport present). Average: 1.2 out of 3 checks passed. Only 5.3% of sites pass all three.

That's the number that worries me most, honestly. Chatbots reading a homepage for a summary is one use case. Agents trying to complete a task (book something, get a quote, find a phone number) on a site is a different, harder bar, and almost nobody clears it.

llms.txt: present but mostly wrong

19.6% of sites have an llms.txt file. Sounds decent for an emerging convention — until you check conformance: only 13.3% are actually conformant to spec. So roughly a third of the sites that bothered to add the file got the format wrong enough that it likely isn't doing its job.

Nobody is blocking AI bots

This is the other half of the paradox. Only 1.3% of sites block AI crawlers in robots.txt. 85.9% are fully open access — no restrictions on any of the major AI user-agents at all.

So the door is wide open. GPTBot, ClaudeBot, PerplexityBot, whatever — they can all walk in. The problem isn't access. It's that once they're in, there's often nothing structured to cite.

Structured data and authorship

JSON-LD present: 74.7% (the one bright spot — most sites at least attempt some schema)
Identifiable author/byline: 17.9%
Basic accessibility markers: 19.2%

Three out of four sites ship some JSON-LD, but fewer than one in five have a clear "who wrote/verified this" signal, which matters for both traditional E-E-A-T and how confidently an LLM can attribute or trust a claim from the page.

Sector and geography spread

Best-performing sector: consulting (54.8 avg). Worst: restaurants (43.6). Best city: Paris (55.8). Worst: Rennes (43.8). The spread exists but it's not huge — nobody is close to acing this, sector or city doesn't change the fundamental picture much.

The traffic shift is already happening

Here's a smaller, more anecdotal data point, but it's real and it's ours: on our own site's GA4, over the last 7 days, AI assistants sent more sessions than Google organic search. Copilot (10) + Perplexity (3) + ChatGPT (2) = 15 sessions, versus 6 from Google organic. That's 2.5x.

It's one site, one week — not a trend line you should extrapolate a market shift from. But it's a live signal that this isn't a purely theoretical concern for small sites. Someone is already sending traffic through these channels, on top of a landscape where the sites in our sample are almost entirely unprepared for it.

What you can actually do about it

If you build or maintain SMB sites, the fixes here are unglamorous but concrete:

Ship a conformant llms.txt, not just a file that exists. Check it against spec — don't assume "I added one" means "it works."
Server-render your JSON-LD. Don't inject structured data via useEffect or client-only hydration; a lot of AI crawlers don't execute JS the way Googlebot increasingly does.
Add real author/reviewer bylines where you're making factual claims (pricing, expertise, credentials). It's cheap to add and it's one of the weakest scores in the whole dataset.
Design for agentic navigation, not just human eyeballs: clear headings tied to actual content sections, discoverable contact/booking actions, consistent semantic markup — the stuff that lets an agent figure out what a page is actually offering without guessing.

None of this requires blocking anything. The access problem basically doesn't exist in this sample. The readability problem is the whole story.

DEV Community