Where a publisher's figures need the most care

I build churn models to decide what to do, not merely to rank subscribers by risk. At Vesterlund our models estimate two things for each digital subscriber: the probability of cancelling in the next ninety days, and the change in that probability if we act, whether through a reminder of a newsletter they no longer open, a pause, a cheaper plan or nothing at all. The second estimate comes from randomised holdout groups refreshed every quarter, because a model trained only on past campaigns learns our habits rather than readers' intentions. I also report a measure our board now sees every quarter: the share of paying subscribers who have read nothing in ninety days. In 2023 it was 12 per cent; in mid-2026 it is 6 per cent, partly because readers came back and partly because we let silent subscribers leave easily. Retention that depends on readers forgetting to cancel is a liability carried at the value of an asset.

Most of our readers arrive without logging in, and under the ePrivacy rules and the GDPR we need their consent before setting non-essential cookies or using their browsing for advertising. In 2022 I replaced thirteen separate consent banners with one, gave refusal the same prominence as acceptance on the first screen and wrote each purpose in plain Danish rather than as a vendor list. Acceptance fell, and so did revenue from open programmatic auctions. We rebuilt on data readers give us knowingly: a free login for three articles a week, newsletter preferences and stated interests. Logged-in reading rose from 22 to 61 per cent of page views between 2021 and 2025, and advertisers now buy segments built from declared interests and on-site reading that we can explain to any reader who asks. I describe the rules in general terms and leave their interpretation to our counsel and our data protection officer.

A recommender in news or broadcasting decides what a reader sees next, which makes it editorial whether or not anyone calls it that. My practice is to separate what an algorithm may optimise from what it may never touch, and to write the second list with editors. At Fjellvang Kringkasting that meant a reserved third of the streaming front page and rules protecting children's and minority-language programmes. At Vesterlund it means the editors' front-page order is never reshuffled by a model; personalisation runs only in a clearly labelled module beneath it, which readers can switch off. We publish a plain account of the signals the model uses. The EU Digital Services Act requires online platforms to explain the main parameters of their recommender systems; whether that duty reaches a news site is a question for lawyers, and we adopted the practice regardless because readers asked for it.

I began my career weighting survey panels, and audience measurement remains the part of my work I care most about. Digital figures look precise because they come from logs, but a browser is not a person: one reader may appear as four browsers, and refused or blocked cookies remove others entirely. Since 2022 I have sat on the board of Nordmål, the joint company through which Danish publishers and broadcasters fund independent audience measurement, and on its methodology committee. I supported its move from counting browsers to estimating people from a recruited panel combined with site logs, although our own reported reach fell by about a third. Inside Vesterlund I am now replacing page views in advertising sales with attentive time, defined in a published specification and checked against raw logs. Measurement earns trust when the seller's own analysts cannot quietly change it between campaigns.

Six years selling aggregated insights from a mobile network taught me where anonymisation ends. Location records are among the most identifying data any organisation holds; a few time-stamped points can single out most individuals even after names are removed. At Brosund Mobil we reported only counts of fifty devices or more per grid square and hour, suppressed or aggregated everything below that floor, and refused to sell movement paths, individual dwell times or repeat-visit measures. With every delivery we stated the share of cells we had suppressed, so that a planner would not mistake a blank square for an empty one. I apply the same discipline to the far smaller location signals a publisher holds: subscribers' delivery postcodes and the approximate location of app readers who allow it. We use postcodes for deliveries and regional editions, and nothing finer for advertising.