French publishers now have their first full September traffic numbers since Google launched AI Overviews in France.
They are not good.
Visits to major French news websites fell 21.8% on average in September 2026 compared with September 2025, according to figures reported by Stratégies from ACPM data.
That is a sharper fall than the 15.3% average decline reported for August.
It is also tempting to look at the timing and blame Google’s AI summaries.
That would go further than the data allows.
Google launched AI Overviews and AI Mode in France on 22 July. The publisher decline accelerated after that date, but September also includes changes in Google Discover, different news cycles, seasonality and large differences between individual publishers.
ACPM itself cautions that these traffic figures cannot isolate the impact of AI Overviews.
The useful story is therefore not “AI Overviews caused a 21.8% traffic loss.”
It is that French publishers now have enough evidence to start separating several different traffic problems that were previously being discussed as one.
The September declines are large, but they are not uniform
The year-over-year website declines reported for September vary sharply by publisher.
| Publisher | September website visits vs September 2025 |
|---|---|
| Le Parisien | -43.5% |
| Libération | -34.4% |
| Franceinfo | -32.9% |
| Le Monde | -22.9% |
| Le Figaro | -22.7% |
| Ouest-France | -9.1% |
| Les Echos | -1.5% |
| Average reported decline | -21.8% |
That spread should immediately make publishers cautious about a single explanation.
If one Search feature were producing the same effect everywhere, we would expect exposure and losses to line up much more neatly.
They do not.
The differences can reflect traffic mix, dependence on Discover, query intent, editorial portfolio, direct readership, app usage and how much evergreen information each publisher produces.
September traffic also remains substantial in absolute terms.
ACPM recorded approximately:
| News website | September 2026 visits |
|---|---|
| Ouest-France | 175.9 million |
| Le Figaro | 131.2 million |
| Franceinfo | 117.7 million |
| Le Monde | 102.4 million |
| Le Parisien | 52.1 million |
These numbers describe scale, not health.
A site can still serve more than 100 million visits while losing a meaningful share of its prior-year audience.
NEMO’s Web vs App Exposure Gap is 8.9 percentage points
There is another useful number in the September data.
News websites were down 21.8% on average.
Apps from the same media group were down 12.9% on average.
The difference is:
21.8% minus 12.9% = 8.9 percentage points
NEMO calls this the Web vs App Exposure Gap.
It is a diagnostic number, not an AI Overview attribution number.
Apps are much less directly exposed to changes on Google’s web search results page. A reader who opens a publisher’s app is already inside a direct distribution channel.
That makes the 8.9-point difference worth investigating.
It does not mean AI Overviews caused an 8.9-point loss.
Other web-specific acquisition channels can move at the same time, especially Discover.
The gap simply tells an editor or audience team where to look first.
If website visits fall much faster than app visits, investigate acquisition channels before assuming the audience has lost interest in the publication itself.
Radio France shows why Search and Discover must be separated
Radio France provides a more direct example.
Laurent Frisch, its digital director, said traffic to the Radio France website from Google was down 25% year over year during the first three weeks of September.
That is much closer to the question publishers actually want answered because it looks specifically at Google-originated traffic.
Frisch also said evergreen knowledge and educational content was among the most affected.
That fits a plausible AI Overview risk pattern.
A user asking a factual question may get enough information directly in the search result and have less reason to click through to an explainer.
But Radio France added another important detail: Discover also fell sharply.
That prevents the 25% Google decline from becoming clean evidence of a 25% AI Overview effect.
Google Search and Google Discover may both appear under a broad idea of “Google traffic,” but they are different products with different user behaviour.
Any publisher analysis that combines them loses the information needed to diagnose what actually changed.
There is separate evidence that AI Overview exposure can reduce CTR
The ACPM numbers do not isolate causation, but another French dataset gives us a stronger signal.
Ahrefs analysed 963 domains using aggregated Google Search Console data around the 22 July launch.
It compared 28 days before launch with the first nine days after.
Domains where more than 20% of tracked queries triggered an AI Overview recorded a 23.1% median decline in click-through rate.
The low-exposure control group, where fewer than 2% of queries triggered AI Overviews, saw CTR rise by 2.3% over the same period.
Ahrefs calculated a control-adjusted difference of 24.8%.
That is meaningful evidence that AI Overview exposure can reduce organic CTR.
It still does not prove that the French news industry’s 21.8% September traffic decline was caused by AI Overviews.
The datasets measure different things.
ACPM measures publisher visits across websites.
Ahrefs measures Search Console performance across a panel of domains, grouped by AI Overview exposure.
Those findings can support each other without being treated as interchangeable.
Publishers should run four splits before assigning a cause
A top-line website traffic number is now too blunt for this problem.
The September data becomes much more useful when publishers divide it four ways.
1. Split website and app
Start with the same comparison ACPM has made visible.
Record:
- website visits;
- app visits;
- year-over-year change for each;
- the difference between the two.
A large Web vs App Exposure Gap points toward acquisition or web-distribution pressure rather than a uniform drop in audience interest.
It does not tell you which web channel is responsible.
That comes next.
2. Split Search and Discover
Do not report “Google traffic” as one number.
Separate at least:
- Google Search;
- Google Discover;
- Google News if material;
- other Google referrals.
For Search, use Search Console clicks and impressions.
For Discover, use the Discover performance report where available.
The distinction is critical in France because publishers have already reported significant Discover volatility alongside the AI Overview rollout.
If Search is stable and Discover collapses, AI Overviews are a weak explanation.
If Discover is stable while Search CTR drops sharply, the AI Search hypothesis becomes more credible.
3. Split evergreen content from news-cycle content
Radio France’s observation about evergreen material gives publishers another useful test.
Create content groups such as:
- explainers;
- definitions;
- health and science information;
- biographies;
- guides;
- service journalism;
- breaking news;
- live coverage;
- investigations;
- interviews;
- exclusives.
Then compare Search clicks and CTR by group.
Pages that answer stable informational questions are more likely to face direct-answer competition than a live political story, an exclusive interview or original investigation.
The sitewide average can hide that difference.
4. Split high AI Overview exposure from low exposure
The strongest test is a cohort comparison.
Take queries or pages with similar prior traffic and divide them by AI Overview exposure.
For example:
High exposure More than 20% of tracked queries regularly show an AI Overview.
Low exposure Few or none of the tracked queries show one.
Then compare:
- impressions;
- clicks;
- CTR;
- average position;
- Discover traffic;
- page type;
- conversions or subscriptions.
If rankings and impressions remain reasonably stable while CTR deteriorates much more heavily in the high-exposure group, the evidence becomes more interesting.
That still requires care, but it is far stronger than comparing July site traffic with September site traffic and assigning the entire difference to AI.
Use rankings and impressions as controls, not conclusions
One of the easiest mistakes in an AI Overview traffic audit is to look only at lost clicks.
Suppose a publisher loses 25% of Search clicks.
That could happen because:
- rankings fell;
- query demand declined;
- fewer pages were indexed;
- AI Overviews reduced CTR;
- another SERP feature took attention;
- seasonality changed;
- the publisher lost Discover traffic that was incorrectly grouped with Search.
Check impressions and average position alongside clicks and CTR.
If impressions and rankings remain relatively stable but CTR falls, the case for a search-result interface effect gets stronger.
If impressions collapse as well, demand or visibility may be part of the story.
No single metric answers the question.
Measure revenue before declaring a traffic crisis
Publishers should also resist treating every lost visit as equally valuable.
A lost evergreen information click may reduce advertising inventory.
A lost branded subscription query may affect paid conversion much more heavily.
A lost Discover visit may have a different commercial value again.
Add business outcomes to the analysis:
- advertising revenue per session;
- subscription starts;
- registrations;
- newsletter sign-ups;
- returning users;
- direct visits following first discovery.
Le Monde has previously said original journalism, analysis, interviews and exclusives behave differently from commodity information.
Whether that pattern holds across the wider French market is measurable.
Traffic loss matters.
Revenue loss is the stronger business signal.
A reproducible publisher worksheet
For each major content cohort, record:
| Field | Example |
|---|---|
| Surface | Web / App |
| Google source | Search / Discover |
| Content type | Evergreen / Breaking / Original |
| AI Overview exposure | High / Low |
| September impressions | Current value |
| September clicks | Current value |
| CTR change YoY | Percentage |
| Average position change | Position difference |
| Visits change YoY | Percentage |
| Subscription or revenue change | Business result |
Do not compare one high-exposure health explainer with one low-exposure breaking-news story and call that a controlled test.
Use groups of comparable pages or queries.
Repeat the analysis monthly.
The most useful evidence will come from patterns that persist.
What the September data actually tells us
We can now say several things with confidence.
French news websites had a difficult September.
The reported average year-over-year website decline was 21.8%.
Apps fell less, at 12.9%, creating an 8.9 percentage-point Web vs App Exposure Gap.
Radio France separately reported a 25% decline in Google-originated website traffic during the first three weeks of September.
Ahrefs has measured a substantial CTR difference between French domains with high and low AI Overview exposure.
And Google launched AI Overviews in France less than three months ago.
Those facts belong in the same investigation.
They should not be collapsed into the claim that AI Overviews caused 21.8% of French publisher traffic to disappear.
The stronger conclusion is more useful anyway.
French publishers now have a live market where AI Overview exposure, Search traffic, Discover traffic, app usage and content type can be compared against each other.
That gives the industry a chance to move from anecdotes to attribution.
The next question is not whether publisher traffic is falling.
It is which part of the decline survives once Search, Discover, apps, content type and AI Overview exposure are separated.
That is the number NEMO will watch next.
Sources
- ACPM, September 2026 unified public website rankings.
- ACPM, September 2026 application rankings.
- Stratégies, 9 October 2026 analysis of ACPM news-site traffic changes.
- Google France, AI Overviews and AI Mode launch announcement, 22 July 2026.
- Stratégies interview with Radio France digital director Laurent Frisch, 1 October 2026.
- Ahrefs, French AI Overviews CTR study based on 963 domains.