News audiences are falling everywhere. The interesting number is not the size of the drop, it is the gap between the channels that absorb it and the ones that escape it. That gap describes a shift reaching well beyond the press.
A global decline, not a French accident
Over the year to November 2025, traffic sent by Google to publishers fell by a third worldwide. Thirty-eight per cent in the United States, seventeen per cent in Europe. The data comes from Chartbeat, reported by the Reuters Institute.
The mechanism is documented. Google's AI-generated summaries appear on roughly 30% of queries and cut click-through by around 35%. The answer is consumed inside the results page, and the link becomes optional.
France got the feature later, on 22 July 2026, and therefore took the hit in a more concentrated form. Between September 2025 and September 2026, across forty national and regional news brands, visits fell from 2.14 to 1.70 billion. Twenty per cent less in a year, according to the ACPM ranking analysed by The Media Leader.
The gap that tells the story
Within that same panel, websites lose 22.9% and apps 9.4%. Thirteen points of difference, identical content, same period.
Title by title, the fault line holds. Ici, formerly France Bleu, loses 55.8% on its website while its app gains 21.8%. La Croix is down 6.3% on the site and ends at +2.1% overall thanks to the app. Radio France follows exactly the same pattern.
The United Kingdom gives the same signal through another door. Carly Steven, SEO director at the Daily Mail, sums it up in a sentence: most of the title's search traffic comes from branded queries. In other words, even inside Google, what holds up is what the reader was already looking for by name.
First useful question: what share of the audience arrives because the brand is known, and what share because an algorithm decided so?
Three business models, one dependency that never changed
The recent history of the press is usually told as a run of technological disruptions it handled badly. Reading it through the acquisition channel is more revealing.
First model, the newsstand and the print subscription. Second, the digital shift, where the audience is resold to advertisers. Third, the return to subscriptions once advertising stopped being enough.
Three different ways of making money. But the last two rested on the same thing: a reader brought in by a third party. Google for the most part, social networks for the rest. What changed was what the press was selling. Not how it found its customer.
Language models therefore do not break a revenue model. They cut the acquisition channel shared by the last two.
The real question is not “what new model for the press”. It is: who owns their reader relationship, and who was renting it.
Licensing does not protect what people think it protects
The industry's instinctive answer is the content licence negotiated with AI companies. A report published in April 2026 by the Center for Journalism & Liberty and the Open Markets Institute makes it possible to measure the actual effect.
The amounts first. News Corp is reported to have signed around $250 million over five years with OpenAI. The New York Times around $20 million a year with Amazon. Dotdash Meredith at least $16 million a year with OpenAI. Thomson Reuters discloses $33 million in AI licensing revenue year to date. Deal terms remain confidential; these are the figures reported in the press.
Against that, the report puts annual publisher revenue lost to zero-click behaviour at around $2 billion. But the decisive number is not that one.
Signing did not protect the traffic. Licensed publishers in fact lost more in percentage points, because they started higher.
A licence does not replace an acquisition channel, and it does not preserve one either. It offsets a fraction of the loss, for a while, and funds the tool doing the disintermediation.
The report describes the market as an oligopsony: a handful of buyers facing thousands of fragmented publishers. Around thirty-five deals on OpenAI's side, eight on Microsoft's marketplace, roughly twenty with Perplexity. Local and regional titles are very largely excluded.
Second useful question: a licence buys time. How much time, and to build what during that time?
The four positions actually on the table
Selling a place in the answer rather than an impression. Jérémy Parola, head of digital at Reworld Media, argues that scarcer clicks can push CPMs up on premium environments, and that media sites become structured databases capable of influencing brand visibility inside assistants. What is sold changes in nature.
Rebuilding leverage over access. Charging for every crawler visit, whitelisting known bots and blocking the rest by default. These levers have been seriously discussed in France since the start of the year. None has produced a public figure so far.
Falling back on the brand. This is the Daily Mail's position, betting on distinctive journalism as the best algorithmic defence, and the Telegraph's, whose SEO lead Harry Clarkson-Bennett considers subscription publishers structurally better protected.
Treating search as a channel at the end of its cycle. Stuart Forrest, global audience director at Bauer Media Group, describes managing for efficiency rather than growth, shifting effort to Discover, social and email. The Reuters Institute measures a net score of −25 on the effort publishers intend to devote to Google in 2026.
That last move is not risk-free. Consultant Barry Adams points out that disengaging from search can accelerate the very decline it claims to manage.
A fifth route: selling access to the data, not the article
The four positions above try to defend an existing model. A fifth consists of changing it, and the trade press is better placed to attempt it.
Le Moniteur, published by Infopro Digital, launched Moniteur+ in beta in late January 2026 and officially on 24 February. A subscriber-only assistant built on more than 18,000 construction projects referenced and qualified by Le Moniteur, AMC and Les Cahiers Techniques du Bâtiment. It compares projects, details the parties involved, timelines, materials and costs, and links back to sources.
Danielle Attias, deputy managing director for the group's Construction and Local Government media, notes that early usage went to comparing projects, checking costs and identifying which players to bring in. Infopro has announced an equivalent for retail on LSA.
What is sold is no longer an article, it is an answer to a question at work.
The reader no longer comes to read, they come to check something before deciding. The B2B press starts with an advantage the consumer press does not have: its reader is identified, logged in, and its database exists nowhere else. The relationship was never rented from Google.
The move has a social cost, and it is public. In May 2026, Infopro newsrooms went on strike against a plan to cut nineteen sub-editor roles, partly replaced by five senior editing positions working with AI across twenty-six titles. The unions point to a 27% EBITDA margin on the perimeter concerned in 2024.
Both facts belong to the same file. A publisher moving to a service model starts valuing its database above its editorial output, and the chain that feeds that database becomes a cost to optimise.
Third useful question: when the product becomes the answer, what still gets paid for, the database or the work that fills it?
What these numbers do not say
September 2025 was an exceptionally heavy news month, which raises the French comparison base. The ACPM ranking measures visits by platform, not their origin, which rules out isolating the precise effect of search. Google disputes any significant effect from its feature, and an anti-spam update landed over the same period.
Not everyone sees the same thing. Bauer reports a single-digit decline, far from the double digits in the Reuters report. In France, La Nouvelle République gains 8% and Paris Match 2.2%.
The thirteen-point gap between websites and apps, on the other hand, survives every one of those objections. It is the most solid data point in the file.
A mechanism that reaches beyond the press
Any company whose acquisition runs through a platform it does not control is exposed to the same shift. Online retail dependent on Google Shopping or Amazon. The software vendor getting half its sign-ups from search. Marketplaces buying their demand on Meta.
The press gets there first because its product is text, and text is precisely what models are good at reproducing. The others will follow at different speeds, as assistants learn to compare, book and buy on the user's behalf.
Three questions arise now for any executive, whatever the sector.
- What proportion of revenue depends on an acquisition channel we do not own?
- What is left of demand if that channel closes, and how long does it take to build something else?
- Is the product distinct enough to be searched for by name, or does it only exist inside an intermediary's answer?
Thirteen points separate websites from apps. That is the price of a direct relationship, and it is paid before you need it.