Generative AI could in principle feature a great diversity of sources in output responding to news-related queries.
But do various generative AI products actually do this?
In early 2026, Roa Powell and Carsten Jung published a piece of research for the Institute for Public Policy suggesting maybe not – based on analysis of responses to a sample of 100 hypothetical UK news queries submitted to four different AI tools (ChatGPT, Perplexity, Gemini, and Google AI overview), they found that these tools draw “on a narrow range of prominent news brands”.
As a benchmark for interpreting their results, I have created the slide below.
The figures in red are my calculations of the Herfindahl–Hirschman Index (HHI) (a common measure of market concentration) for each of the four AI tools based on the share of reference that goes to each of the top ten news brands (as reported by Powell and Jung).
The figures in black are taken from this piece of work I did with Richard Fletcher included for comparison, namely similar HHI calculations based on historic data from the UK from 2017 for online news accessed directly, via search, news aggregators, or various kinds of social media, as well as, for further comparison, HHI figures for television viewing and weekly print newspaper circulation at the time.
The concentration of attention, with a few brands accounting for a very large share of reference, is far greater for the AI tools than any other form of access.
While the underlying data is not like-for-like comparison (the bulk of the 2017 data is based on passive tracking of actual UK users, the 2026 AI tool data is generated by prompting), I think the figures are still interesting and thought-provoking. (And not necessarily unique to the UK – Nikos Smyrnaios and Olivier Koch has published a piece of work, based on a somewhat different methodology, suggesting results for France that are also about the 2,000 bar for a highly concentrated market.)
It’s not just that Powell and Jung are right to stress that a narrow range of prominent news brands (some of whom have commercial deals with the AI companies) loom very large in AI output.
It is also that the concentration in question, measured here in terms of share of reference, is far, far greater than any of the different kinds of access we analyzed based on the 2017 data – even more concentrated than direct access, which itself was significantly more concentrated than any other kind of access at the time.
We already know that AI tools generate far fewer referrals than established platforms do – worrying enough for a multitude of publishers competing to connect with the public.
It also seems the (comparatively fewer) referrals they do drive are highly concentrated amongst a select few publishers, much more so than has been the case for search, social, or news aggregators.
(Underneath the topline, there is some diversity from tool to tool (just as we found different outlets doing well via different platforms), as Powell and Jung writes, their findings suggest “each AI tool prioritizing news brands in different ways, in each case foregrounding a distinct selection of news outlets compared with those that are currently most popular across the UK” – go read their report for more details.)
rasmuskleisnielsen.net 