Babylon Newsletter, Issue #12

The click — the thing all publishers built decades of business on — is going. Google AI Mode takes the question and answers it on the spot. The useful question is not how to win back traffic. Instead publishers need to define what "reach" means once the audience stops arriving. Read more about this in the "Talk of the Week" below.

Google's AI Mode passed one billion monthly users — a year after its US launch, with queries more than doubling every quarter since. AI Mode is the option to skip the list of blue links and let Google answer directly, usually with a few source links attached. In effect, Google is redefining what it is: search engine and encyclopedia in one go.

This is a usage story as much as a business one. With AI Mode, Google found product/market fit for a genuinely new product — the billion users are the proof — and the revenue is already following. Google's Search & Other revenue hit $60.4 billion in the first quarter of 2026, up 19% year over year — and CEO Sundar Pichai credited AI Overviews and AI Mode for the rising usage. AI Mode searches run about 3× longer than a traditional query, and follow-up turns are growing more than 40% month over month. Separately, Google is turning the ad itself into a conversation: its new "Business Agent" lets a searcher ask an ad about pricing or availability and get answers pulled from the advertiser's own site — the lead qualification that used to happen after the click now happens inside the ad.

Why care? This is already a big business sitting inside an even bigger one — and it is growing because of the AI shift, not despite it. $60.4 billion is a single quarter of Search alone, and the reason to click away from Google keeps shrinking as it answers more on the page itself. The reach publishers used to get from Google is quietly turning into revenue Google keeps.

Sources: Google I/O 2026 · Alphabet Q1 2026 earnings (29 April) · Google AI Mode one-year report (Mohan, 19 May) · Search Engine Land (26 May)

Television news joins the copyright fight over AI answers
28 May 2026 · Variety / CNN

What's new: CNN has sued Perplexity in New York federal court, alleging the AI answer engine scraped more than 17,000 of its stories, photos and videos — CNN's first AI copyright case and, it believes, the first brought by a television network.

Why care? It comes down to a clean disagreement. Perplexity says: "you can't copyright facts." CNN says: facts are free, but the reporting that found and published them is not — and that is what it claims was copied. It is now for US courts to decide who is right, and the ruling will help set the price of news as an AI input.

Reality check: The industry is split, not united. CNN, the New York Times and Dow Jones are suing, while Time, Gannett, Le Monde and Der Spiegel have signed deals with Perplexity — and Meta licensed CNN itself last December. "Sue or sign" is a bet, not a consensus.

The AI licensing market is being built by the firms that broke the old one
27 May 2026 · Open Markets Institute / Nieman Lab

What's new: A report from the Open Markets Institute, "Same Gatekeepers, New Tollbooths," finds publishers caught in a "double bind": the same Big Tech firms whose AI products strip their traffic now control the licensing infrastructure meant to replace that lost revenue — "occupying both sides of the value chain."

Why care? The report maps a three-tier market: a few large bilateral deals, an intermediary marketplace layer, and an uncompensated majority — most publishers and creators, including smaller, regional and non-English outlets — left outside any deal. The terms being normalised now will be hard to reverse later.

Reality check: This is an advocacy report (35+ interviews, policy asks like statutory licensing), not a neutral market study. One hard number from it: AI bots bypassing voluntary access restrictions quadrupled in six months, from 3.3% to 12.9%.

One story, many formats — Schibsted makes content liquid
NAMS 2026, 27.-28. May, Copenhagen · Juan Carlos Lopez Calvet (Schibsted)

What's new: At the Nordic AI in Media Summit (NAMS) in Copenhagen, Schibsted demonstrated its "Liquid Story Engine," which reshapes one piece of reporting into several formats — article, short video, social clip — and reports cutting video production from over an hour to under 15 minutes.

Why care? This is "liquid content" as an actual workflow, not a slide. If reach now fragments across surfaces you don't control, the unit of journalism can no longer be the fixed article — it has to be a story you can pour into whatever shape each surface needs. Built for video desks, framed as augmentation.

Reality check: The time saving and the "high quality" claim are Schibsted's own, presented at a conference, with no independent benchmark. Speed is measured; quality is asserted.

For several decades now, reach meant one number going up: pageviews, uniques, referrals from search and social. The platforms sent the crowd, publishers counted it, advertisers paid for it. That number is now leaking, and the figure at the top of this issue is why.

But "reach" was always two things wearing one name. There is the volume — the casual, low-intent arrival who reads one piece and leaves. And there is the connection — the reader who comes back, logs in, pays, trusts the masthead. The old metrics could not tell them apart, because the platforms delivered both in the same stream.

AI separates them. The casual arrival is exactly what an answer engine absorbs: a fact, a summary, a click that no longer needs to happen. At the World News Media Congress in Marseille, A. G. Sulzberger, the publisher of The New York Times, described the AI companies as "consolidating their outsize control over our data and our attention" — taking the audience and the work at once. The volume side of reach is being captured and resold.

The connection side cannot be. A media researcher at the University of Copenhagen put it plainly at the Nordic AI in Media Summit: content gets commoditised, connection does not. The summary of your article is now a commodity any model can produce. The relationship with the reader who chose you is not.

This is the quiet good news under a grim number. Audiences are shrinking — but the part that is shrinking is the part that was worth the least: the drive-by traffic that never converted and barely paid. What is left is harder to win and worth far more.

The publishers acting on it are changing what they measure. A media strategist at the same summit, Florent Daudens, urged them to "own the demand signal" — to capture the agents asking on a reader's behalf, and know who is on the other end. Be the destination, not the supplier feeding someone else's answer.

None of this brings the old traffic back. The shift is from a business built partially on volume to one built on connection the media brand has to earn. Fewer people, holding on tighter.

A site built for agents — Hacks/Hackers rebuilt its site to serve AI agents directly, pointing them to an MCP server and RSS feed instead of letting them scrape the HTML. A small, concrete template for the agent-readable web. → LINK

The Economist's two-track web — The publisher is testing agent-readable versions of its marketing and B2B pages — outside the paywall for now — preparing for "two versions of the web," one for humans, one for machines. → LINK

Mistral goes industrial — The French lab signed Airbus (five years, defence and space) and BMW (crash simulation) for a physics-aware AI stack, pitched as a European, sovereignty-minded alternative to US providers. The sovereignty argument, now with anchor customers. → LINK

TAUS Massively Multilingual AI · 3–5 June 2026 · Rome ·
The language-technology industry's working session on multilingual AI.
https://www.taus.net/events

DW Global Media Forum · 23–24 June 2026 · Bonn · International media conference at the World Conference Center Bonn, with AI and the future of journalism across the agenda.
https://dw.com/gmf

"How do we get cited by AI?" is now a real distribution question, and most of the answers on offer are either vague or sell schema tricks that do not work. The evidence points somewhere more specific.

It is not a markup problem. Google's own guidance is that no special schema, file, or tag earns a place in AI Mode — engines read the words on the page, not the code around them, which makes most "answer-engine optimisation" a shortcut to nowhere. What they read for is structure and trust: a clear question answered directly in the opening lines, in plain text rather than buried under carousels and feature art, from a source the model already has reason to believe. It is the same format publishers from The Economist to Hacks/Hackers are now rebuilding their pages into. And the uncomfortable part is the last one — the work that earns a citation is the same work that earned readers: original reporting, stated plainly. The clarity that earns the citation is the same clarity a machine can copy — your newest reader is one that takes the whole thing.

I am Mirko Lorenz. I work on language technology projects at Deutsche Welle in Germany.

Three projects you will hear about in this newsletter:

I cover all three with the same critical lens applied to competitors.

AI use: I use Claude (Anthropic) for research and to edit this newsletter, based on refined and specific prompts. My goal is to understand where the AI performs and where it fails. I learn something every week. Responsibility for stated facts, names, and links is entirely mine.

babylon-newsletter.com · Every Tuesday

7,000 languages. AI works for 20.