Media & Entertainment

The newsroom reports different reader numbers for May than the publisher does, and the subscription count in sales matches neither. That is rarely down to mistakes. Web analytics, app, streaming player, podcast hosting, newsletter and subscription system each count in their own way, and the CMS is the only one that knows which article belongs to which section.

For publishers, broadcasters and streaming providers, we build a metric model where editorial, product, sales and management see the same numbers. AI comes in where it takes work off editorial teams and the decision stays with people.

What is an active subscription?

The question sounds trivial. Do trials count, paused subscriptions, bundles? Does a cancellation take effect on the day it is declared or at the end of the contract? Depending on the answer, the same quarter looks different. So we write these definitions down first, together with editorial, sales and finance, and put them into one central model. The sources are loaded into a data warehouse on a regular schedule.

Churn only becomes meaningful when it is split by product, entry point and tenure. A subscription that came through a promotion behaves differently from one taken out after three weeks of reading.

Connecting usage and subscriptions

Whether a section brings subscriptions or only visits becomes visible only when usage can be tied to a subscription. That mostly works through the ID assigned at login. Without a login, the cross-channel view is limited, and the report should say so openly instead of extrapolating audiences.

From the CMS we need section, author, topic, format and publication time, otherwise usage cannot be traced back to content. Then more than the pageview becomes measurable: days visited per month, return rate, reading depth or listening time often say more about loyalty. And articles with many views do not necessarily lead to subscriptions. Streaming and audio often measure differently from web analytics, so their numbers can only be added up with care. Anyone who wants ad revenue per article also needs data from the ad server. For us that is one more source, not a focus.

Measuring with consent

In Germany, section 25 of the TDDDG governs access to end devices, and the GDPR applies on top. Audience measurement therefore usually depends on consent. Server-side tracking with clean consent management makes the data more dependable but does not replace consent. Part of the visits stays unmeasurable. We show that gap in the report. The changes planned in the EU’s Digital Omnibus are so far only a proposal and have not been adopted.

AI in the archive and the newsroom

In media, trust is the capital, so we proceed carefully. For tagging, the AI suggests topics, places and people for articles, images and videos, and editors confirm or correct before anything enters the archive. An archive search answers questions and names its source. Before that, it must be settled which collections may be searched, because photos, wire material and freelance contributions come with rights and licences. Technically, you can control what the search sees at all.

Churn predictions flag subscriptions with declining usage. They give probabilities, not certainty, and we test them on historical data before anyone works with the signals. Summaries and drafts go out only after editorial review. Wrong tags or faulty summaries in the archive damage credibility more than they save time. That is why spot checks with the newsroom, a documented approval path and a measured error rate come first, before a feature is used more widely.

What the EU AI Act means for newsrooms

The transparency duties under Article 50 have applied since August 2026. Deployers must disclose AI-generated text on matters of public interest. That does not apply if the text was reviewed by a human or is under editorial control and a person holds editorial responsibility. Deepfakes in image, audio and video follow separate labelling rules, and providers of generative systems must mark their output in a machine-readable way.

Article 4 adds that anyone deploying AI needs AI literacy in the team. After the most recent changes, the rule no longer prescribes a particular level.