French Media AI Visibility Study: 100 Sites Audited, None Above a C

We ran 100 leading French media sites through our analysis engine. Each site was scored on the same 46 criteria we use for our own clients. The verdict is blunt: an average score of 46.3/100, and not a single site above a C. Between their command of Google SEO and their AI visibility — how ready they are to be cited by AI search engines — the gap reaches 35 points. Here are the numbers, the ranking, and what they say about the next battle for audience.

What does the study show, in four numbers?

In July 2026, the 87 French media sites we were able to audit score an average of 46.3/100. None reaches grade B. These sites are strong on Google search ranking (67.1/100), but that strength does not carry over to their AI visibility at all (32.5/100). In plain terms: they won the Google battle, but haven't yet started the AI-engine one. On July 12, 2026, the average classic-ranking score hit 67.1/100 — that's solid. The average AI visibility score, meanwhile, tops out at 32.5/100.

• 46.3/100: the average score across all sites (half do better than 47, half do worse), measured in July 2026.
• No site graded A or B in July 2026: 27 sites in C, 58 in D, 2 in F.
• 100% of the sites audited in 2026 lack complete Article structured data.
• 35 points of gap on average, in July 2026, between the search-ranking score (67.1) and the AI visibility score (32.5).

Put another way: the technical foundations are there. But almost nothing that AI engines look at to pick their sources is in place. These engines want passages that stand on their own, complete structured data, and facts tied to their source. None of that is ready.

Why such a big gap between search ranking and AI visibility?

Because everything built to please Google does not work for AI. Google ranks whole pages. AI engines, on the other hand, grab chunks of text. They look for specific things: answers that stand on their own, complete structured data, facts tied to their source. So a typical French outlet in our panel scores decently on classic ranking in 2026, but its AI visibility score is twice as low. This isn't a measurement fluke. It's the result of twenty years of work aimed at a single engine — Google, which ranks pages — applied unchanged to engines that carve out passages instead.

GEO (optimizing for AI engines) follows different rules, proven by research. According to the Princeton/ACM SIGKDD 2024 study on AI citations: a question-and-answer section gains 27% AI visibility, a passage that stands on its own at the start of a section is cited 2.3x more often, and a statistic tied to a named source is worth +40%. These are exactly the points almost our entire panel misses.

The one criterion 100% of media sites fail

Of 87 sites audited, all 87 fail the Article schema criterion. What is it? A small block of code, called JSON-LD, tucked inside the page to describe it to machines. At a minimum, this block must state the article's headline, its description, and its main topic. Many sites have a piece of it. But based on our readings, none fills it out completely on the pages we tested. It's the clearest finding of the study. This structured data helps AI engines understand a piece of content. Yet media outlets still treat it as optional.

Here's the rest, across the 84 sites where the content analysis could run all the way through:

Let's be honest about one thing: our audit covers 3 important pages per site, homepage included. These are mostly pages that gather links, not articles. Some individual articles surely do better. But those homepages are exactly the first pages AI bots run into.

The ranking: specialists ahead of generalists

The top of the ranking looks nothing like the audience leaderboard. The best-scoring sites are specialized outlets, not the big generalist media:

At the very bottom, two sites land the F grade, with AI visibility at zero on the pages we tested: Elle.fr (17/100) and Marianne (19/100). Big names like L'Équipe (39), 20 Minutes (41) and BFMTV (41) stay below the panel average. Their editorial weight isn't yet turning into signals AI engines know how to use.

Thirteen sites couldn't be audited at all: their protections blocked our bot. They are Le Figaro, Libération, Les Échos, Le Point, Ouest-France, CNews, Europe 1, La Nouvelle République, Madame Figaro, Biba, Pour la Science, L'Usine Digitale and L'Usine Nouvelle. Guarding your content against automated scraping is a defensible choice. But a site closed to bots it doesn't recognize risks being closed to AI bots too — the ones that read the web on behalf of these engines. And so risks never showing up in their answers.

"Media is the one sector where the gap is this fixable: they already own the structural edge generative engines are hunting for — editorial freshness — and all they're missing is the markup and structure to turn it into citations," says Nicolas Meridjen, founder of LightSpot and author of the study.

Why this is an audience problem, right now

AI engines have a soft spot for recent content, and that favors media. According to the Princeton/ACM SIGKDD 2024 study, 76% of the sources cited by Perplexity are less than 30 days old. And 65% of AI citations rest on content less than a year old. A newsroom that publishes dozens of fresh articles every day has exactly what these engines are after — as long as its pages are built well enough to be cited.

The stakes are real. Every AI answer that cites a competitor — or no outlet at all — is a visit lost for good. And that lost visit is invisible in the usual audience tools, because it produces no click. The more AI search takes ground from link-based search, the bigger this blind spot grows. This is today's missing link, and it's fixable. Structured data, a direct answer at the top of the article, source attribution: these are small editorial-organization jobs, not site rebuilds.

How we measured it

The method is the same as for any LightSpot audit.

1. A panel of 100 French media sites: national and regional press, television, radio, web-native outlets, and specialized sites.
2. An automated audit of each site on July 12, 2026: 3 important pages, 46 criteria (21 for search ranking, 25 for AI visibility), while respecting each outlet's robots.txt — the file that tells bots which pages they may visit.
3. Each criterion returns a verdict: pass, fail, or impossible to evaluate. Criteria that can't be evaluated are set aside, not counted against the site.
4. The 13 inaccessible sites are removed from the statistics. The published scores are the ones from the standard report, untouched.

What can media outlets fix this quarter?

Three projects are enough to close almost the entire gap. One: fully populate the Article schema markup, the code that describes an article to machines. Two: open every article with a clear 40-to-60-word answer that stands on its own. Three: make source attribution machine-readable — that is, tie every number to its source.

And the good news from the 2026 study is that these three points, missed by nearly everyone, are among the cheapest to fix.

Article schema markup lives in the page template, not in the writing. You do it once, and it covers the whole site. Opening every article with a 40-to-60-word answer is a simple editorial guideline, and it earns 2.3x more citations. Attributing numbers and quotes as a matter of course ("according to the INSEE," "per study X") is already a journalist's reflex. You just have to make it machine-readable by placing the source right next to the number. Three habits, one quarter: and the 35-point gap between search ranking and AI visibility starts to close.

What AI engines actually see

An AI doesn't "read" a site the way a human does. It explores the site with a bot (GPTBot for OpenAI, PerplexityBot for Perplexity, Google-Extended for Gemini). It pulls passages out of it. Then it builds an answer while citing two or three sources. According to the academic research that founded GEO (Aggarwal et al., ACM SIGKDD 2024), the choice of those sources depends less on the site's overall fame than on how easy each passage is to reuse: does the passage stand on its own, is it rich in facts, does it cite its sources?

That's why Article structured data and the guidance in Google's structured-data documentation weigh so heavily among our 46 criteria. They turn content that's readable by a human into content that's usable by a machine. In the same spirit, an llms.txt file placed at the root of the site tells AI engines which pages really matter. It's a signal that's still rare, including across our panel.

To see where your own site stands — media outlet or not — the LightSpot audit applies these 46 criteria to 3 pages for free and hands you the top 3 fixes to make first.