What Each Trade Costs
Forty-five of the fifty biggest news sites in the United States lost traffic last month, against the same month a year earlier.
0:00 / 10:38
What Each Trade Costs
The week's two most useful numbers sit on opposite sides of one problem. In September, 45 of the fifty biggest news sites in the United States lost traffic against the same month a year earlier — nine in ten, up from eight in ten in August, with Newsweek down 61 percent for its sixth straight month at the bottom. 1
And in a closed-door session at the Digiday Publishing Summit, one executive said his team keeps modelling a paywall and the arithmetic never works out; he has not found a locally focused publisher that made it pay. 2
No route leads back to the traffic that went away. What newsrooms did this week amounts to four trades: what they let a machine read, what they build that a machine cannot make, what they license and on what terms, and who pays for the whole thing. Every one of them has a price, and the price is the part nobody publishes.
The machine reads the archive
Category: Editorial Workflow / AI — immediate opportunity.
About 150 journalists, researchers and technologists met at The New York Times Building late last month for a forum on AI and investigative reporting, hosted by the paper and the nonprofit Hacks/Hackers. 3
The case that brought them there was the Epstein files. On 30 January the Department of Justice released more than 3.5 million pages, plus tens of thousands of images and video clips. Duy Nguyen, who leads AI science research at the Times, calculated that printing the PDFs would build a stack as tall as the Empire State Building, and that one reporter would need eleven years to read them. His team built an internal chatbot — the Epstein Files Engine — on a free, open-source, self-hosted platform called LibreChat. More than 100 journalists at the paper asked it more than 5,000 questions, and it fed at least 20 published stories. They followed it with an archive bot for the paper's own morgue, then opened News Agent in July, which the team calls the front door to AI at the Times: any reporter there can build a small search agent of their own.
What to take from it: the capability is not the model, it is the corpus you already own. A reporter at Semafor built the same kind of tool from a Google spreadsheet and a short script that watches his inbox; when he tags an email as a new source, the script writes the name, the beat and a one-line summary into his source sheet, then each morning scans the headlines and suggests a topic and a person to call. He says some suggestions are useless and others make him email someone he had forgotten about. The boundary came from Afrooz Mosallaei of the Center for News, Technology and Innovation: AI is very good at generating leads and much less capable of establishing publishable evidence. Verification, source development and editorial judgment stayed with people.
The product a machine cannot make
Category: Product / Subscriber Conversion — immediate opportunity, on publisher-reported figures.
The Daily Mail's interactive storytelling strand, Deep Dive, began in 2024 and moved behind the paywall in May. Its team is now 18 people; ten were hired in the last two months, and the paper is recruiting two more data journalists. 4
A Deep Dive averages about 440,000 views, holds a reader for two minutes and 45 seconds, and brings in between 60 and 100 new subscribers. The piece the team talks about most is on baby names: a reader types in their own name and sees how common it was over 30 years. Across the US and the UK it has driven roughly 4,500 subscriptions, the most subscribed article the Mail has ever published. The format also sells: the Alzheimer's Society ran five pieces with the paper, and in the six days after the third, the charity's legal and financial advice pages were getting about 65 times their previous views.
What to take from it: the transferable unit is not the 18-person team, it is the one input. A local newsroom has data nobody else can assemble — school catchment boundaries, planning decisions by street, restaurant inspections, a season of high school results. Put one field the reader fills in at the centre of it, and the story stops being something to read and becomes something to use. Harry Lewis-Irlam, who runs the strand, says the days of reaching as many people as you can as quickly as you can are no longer as lucrative as they used to be. AI sits in the background of this work — the paper's innovation team used it to sift hundreds of thousands of Epstein files, doing what Lewis-Irlam calls the jobs nobody wants to do.
The terms you license on
Category: Revenue / AI — trend to watch.
Nine Publishing, which owns The Sydney Morning Herald, The Age and The Australian Financial Review, set out its licensing strategy in INMA this week. 5
Nine appointed a head of strategic licensing, signed a deal with Microsoft for Copilot, and blocks large language models from scraping its journalism at all. Its non-negotiables are worth copying into any contract: attribution to the original publication, no alteration or misrepresentation of editorial meaning, and a scope limited to supporting an accurate answer in the moment, with depth, context and the journalists staying behind the subscription. The company says licensing often produces no immediate visit, so it measures awareness, attribution and repeated exposure instead of clicks.
What to take from it: the institutional deal matters more to a local publisher than the Copilot deal. Nine is signing universities and similar bodies — a campus agreement pays a licence fee straight away, puts the journalism in front of people who read it habitually, and builds a subscriber pipeline. A university, a library system, a school district or a large employer is within reach of many local newsrooms, and unlike a platform, an institution does not change its algorithm one morning and stop sending anyone. You sell the access you already give a subscriber, to an organisation, at a price and on terms you set.
Who pays, and what it costs them
Category: Revenue / Editorial Workflow — trend to watch, from qualitative research.
Jocelyn McKinnon-Crowley at the University of Memphis spent four months inside a four-person nonprofit newsroom in a midsized Rust Belt city and interviewed staff at seven more nonprofit outlets. 6
Her finding is that the money moved to the front. Fundraising, grants and organisational chores became the top priority for the people who were supposed to be reporting — what she calls a double worker exploitation, where journalists feel obliged to do good work as journalists and obliged to do good work for a morally just organisation, with both obligations landing on the same person. One man held the titles of executive director and managing editor without being paid for the higher-level work, and also served as grant writer, fundraiser, board member, photographer, tech support and social media manager. His own reporting, by his own admission, was slipping through the cracks. McKinnon-Crowley argues funders end up taking decision-making power from journalists: they sit on the boards, shape the application forms and set the renewal cycle.
What to take from it: the recommendations are a funding checklist a newsroom can put in front of its own board — recurring money with fewer restrictions, positions funded for five to ten years rather than one, money for overhead, less paperwork, and funders staying out of coverage decisions. She also suggests newsrooms say out loud what they would treat as unacceptable interference, which is cheaper to agree before a grant is refused than after.
She has company in that week's research roundup: a study of 343 Chinese news consumers, in a country that has run AI news anchors since 2018, found about 5 percent rated an AI anchor very engaging or more satisfying than a human one, and roughly seven in ten would not say the machine was as good. 6
What each trade costs
Four trades, and none of them restores the old traffic. Let a machine read the archive and keep the verification with people — the work is choosing a corpus you already own, not a model. Build something that cannot be copied and accept that it will only ever reach a few people, because that is what makes it worth paying for. License for awareness rather than clicks, and only on terms you can enforce. And if somebody else funds the newsroom, expect them to have a say in what it does — so write the boundary down first.
The numbers to keep are 4,500 subscriptions from one interactive article, and at least 20 published stories from one internal search engine. Those are what the trades actually produced. Pick the one you can carry, and price it honestly before you start.
References
- 1Top 50 US news websites: 90% see traffic drop in September
pressgazette.co.uk
- 2
- 3
- 4
- 5
- 6
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.