The 10% who get their news from a chatbot
On 16 June the Reuters Institute at the University of Oxford published its Digital News Report 2026, the largest survey of how we get our news: 97,520 people interviewed online in 48 countries.1 Two numbers need to be read together. Trust in news has fallen to 37%, the lowest level since the institute began measuring it in 2015. And 10% of respondents use a chatbot such as ChatGPT, Gemini or Perplexity for news every week, up from 7% in 2025.
Ten per cent sounds small, and indeed only 1% name AI as their main source of news. But the growth is driven by the young: among 18 to 24-year-olds the share rises to 17%, three times that of the over-55s (5%).2 Those who use it are also mostly the people most interested in news, in other words the readers newspapers care about most. Only 20% of people trust news from a chatbot, but among those who use one the share rises to 44%.
The most quoted figure is about clicks. In the 27 countries where the question was asked, only 4% of respondents say they always or often open the original source when they read a chatbot's answer. From search the figure is 19%, from social media 17%.2 It needs careful reading, because the gap mostly reflects how few people still use chatbots for news. Among those who do, 42% say they often click through. They say so: these are survey answers, and the Reuters Institute itself warns that what we report and what we do can differ.
What browsing data shows
To find out what we actually do, you have to look at traffic. In March 2025 the Pew Research Center tracked the browsing of 900 American adults.3 Of their Google searches, 18% showed an AI-written summary at the top. When the summary was there, they clicked on a traditional result in 8% of visits; when it was not, in 15%. They clicked on links inside the summary in 1% of cases. And 26% of the time their browsing ended there, against 16% on pages without a summary.
How often people click on Google, with and without the AI summary
According to Similarweb, between May 2024 and May 2025 the share of news searches that ended without any click to a news site rose from 56% to 69%.4 Chartbeat, which measures the audience of thousands of news outlets, recorded a 34% drop in pageviews from Google Search between December 2024 and December 2025, and a 16% drop in those from Google Discover.5 Traffic from ChatGPT more than tripled in a year, but all chatbots together still account for under 1% of pageviews. The publishers surveyed by the Reuters Institute expect visits from search to fall by a further 43% within three years.1
The time that used to end up on a newspaper's page now stays inside the answer. Whoever writes it has taken the place of both the newsstand and the front page.
The Italian case
In Italy trust in news stands at 32%, four points down on 2025, and only 8% pay for online news.6 Chatbots are used for news less than elsewhere: 6% of respondents use them every week, one of the lowest figures among the 48 countries (the lowest is the UK's, at 4%). Just 16% trust what comes from a chatbot.7
The trouble for Italian newspapers, though, began before AI. According to AGCOM, the communications regulator, 430 million copies of daily papers were sold in 2025 across print and digital, about 1.6 million a day: 8.1% fewer than in 2024 and 30.6% fewer than in 2021.8 In the first quarter of 2026 the fall continued (-7.6%), and the most-read sites lost audience online too: in March la Repubblica's website had 28.8 million unique users (-8.9% in a year), Corriere della Sera's 26.6 million (-8%).9
Revenue tells the same story: in four years newspaper and magazine revenue fell by 18%, while online advertising grew by more than half and went almost entirely to the platforms.9
Revenue in Italy, in billions of euros
Deals, complaints and courts
Some publishers chose to negotiate. In September 2024 GEDI, then the owner of la Repubblica and La Stampa, signed an agreement with OpenAI to bring its content into ChatGPT's answers. On 27 November 2024 Italy's data protection authority, the Garante, warned it that newspaper archives contain personal data, some of it sensitive, which cannot be handed over to train a model without the safeguards required by the GDPR.10
In October 2025 FIEG, the Italian newspaper publishers' federation, filed a complaint with AGCOM against AI Overviews, Google's summaries, accusing them of cutting traffic to news sites and therefore advertising revenue.6 On 30 April 2026, after hearing Google and the publishers, AGCOM asked the European Commission to assess whether AI Overviews and AI Mode breach the Digital Services Act.11 In Brussels an antitrust investigation had already been open since 9 December 2025: the Commission wants to know whether Google uses publishers' content without fair payment and without giving them a real chance to refuse, since refusing risks disappearing from search.12
On 3 June the UK Competition and Markets Authority required Google to give publishers effective tools to stop their content powering AI features in search.13 Google built the option into Search Console and since 31 August has offered it to sites worldwide. It says the choice does not affect ranking in ordinary results, but warns that sites opting out receive no traffic or impressions from AI answers.14 In the United States, by contrast, on 30 September federal judge Amit Mehta dismissed the antitrust suits brought by Penske Media and Chegg against AI Overviews: that Google sends traffic to those who let it crawl their sites is, he wrote, "an expectation", not an agreement.15
Licensing deals exist, but they are few and concentrated. Reporters Without Borders counted 230 made public between 2023 and May 2026, 71% of them with outlets in the US, UK, Canada, Australia and New Zealand.16
How far the answer can be trusted
If the reader stops at the answer, what it says matters. Between 24 May and 10 June 2025 journalists from 22 public broadcasters in 18 countries, Italy's Rai among them, put current-affairs questions to ChatGPT, Copilot, Gemini and Perplexity in 14 languages and assessed 2,709 answers. Of these, 45% had at least one significant issue.17 The most common concerned sources (31%): they were missing, wrong, or did not say what the answer attributed to them. One in five contained factual errors, including invented or outdated details. Gemini did worse than the others, with significant issues in 76% of its answers.
One example comes from Rai itself. Asked what NATO does, Copilot gave as sources a video from 2022 and an article from 2014. For a news outlet the risk is twofold: it loses the click and can lose its reputation too. BBC research published alongside the study found that 42% of British adults would trust the original source less if an AI summary contained errors.
Advertising moves into the answer
On 9 February 2026 OpenAI began showing ads in ChatGPT to American users on the free and Go plans. The ads are kept apart from the answer, labelled as sponsored and excluded from conversations about health and politics.18 Since 24 August advertising has reached 31 European markets, Italy included, for now without personalised ads in the European Economic Area.19 Google works on even bigger numbers: it says AI Overviews has more than 2.5 billion monthly users and AI Mode over a billion.14
For twenty years someone searched, saw ten links and chose. Brands paid to appear among those links, newspapers lived on the visits. Now the answer arrives ready-made and the space that counts is inside it. For a company the question becomes whether it gets named, with what information and next to which ad. It is the same shift I described for shopping in When machines start to choose.
What I think
I do not believe the clicks will come back. On 15 October Google announced more links and more room for sources in its AI answers, but did not say how much traffic they will bring.20 That is good news, and it remains a concession that whoever grants it can take back. A newspaper that builds its accounts on visits from search is making a bet that depends on someone else.
I see two things to do. The first is a direct relationship with readers: newsletters, apps, subscriptions, events. In a country where 8% pay, it is a long road, but it is the only one that does not depend on someone else's algorithm. The second is bargaining power. Now that opting out of AI answers is possible, a publisher has something to put on the table. Alone it counts for little, as that 71% of deals signed in the English-speaking world shows. I think Italian publishers should negotiate together, as RSF urges, rather than one at a time and in private.
I work in media, and for brands I see a specific risk. Whoever writes the answer decides which sources to cite and which companies to name, and now also sells the advertising that appears next to it. OpenAI says the two systems are separate; I would like someone outside to be able to check. In the meantime a company should measure, besides clicks, how it appears in answers, and correct the wrong information circulating about it.
When copying is cheap, the world changes.
Seventh law of innovation, from Un perché alla volta (translated)
An answer that summarises ten articles reproduces, almost for free, the work of the people who wrote them. At that point the problem is no longer finding information but choosing whom to trust. To choose, you need to see where it comes from, as I also wrote about generated content in SynthID and the rest.
How to keep the sources in view
- SourcesAsk where it comes fromAsk the chatbot which sources it used and open at least one. In the EBU and BBC study almost a third of answers had problems right there.
- DatesCheck when it was writtenAnswers can rest on articles that are years old. If the cited source is from 2014, it is not recent news.
- Breaking newsGo to the outletFor stories still unfolding Google often shows no AI summary and points to the articles. The app or website of a paper you know is still the shortest route.
- HabitChoose your own sourcesA newsletter, an app, a subscription. A direct link with two or three outlets you trust brings you the news from the people who checked it.
Anyone who runs a website, meanwhile, has a new figure to watch. Since the end of August Search Console has shown how often pages appear in Google's AI answers and in which countries.14 That is the number to start from before deciding whether to stay in those answers or leave them.