# Aftonbladet's subscription-prediction model: the 75% front-page sales lift, in Schibsted's own words

> Schibsted says an in-house machine learning model that picks Aftonbladet's front-page articles for non-subscribers lifted front-page subscription sales 75% in an A/B test against its previous models. The figure is Schibsted's own, and its scope is reported inconsistently, so this file caps below green.

- Verification status: pending
- Case type: deployment
- Provider: Schibsted Media & Tech (Curate team), in-house subscription-prediction model
- Client: Aftonbladet (Schibsted), media (named)
- Sector: media / SE / marketing
- Canonical URL: https://theinternetninja.com/stories/aftonbladet-ml-front-page-subscription-model-75pct-front-page-sales-2025/
- Source: The Internet Ninja (theinternetninja.com), independent verified-proof platform

## Outcomes

| Metric | Before | After |
| --- | --- | --- |
| Subscription sales from front-page articles, A/B test against previous models |  |  |

## Verification method

Media AI business-value case file. The 75% figure was extracted verbatim this session from two fetched sources saved to sources/: an INMA Ideas Blog piece written by three Schibsted staff (8 December 2025) and a WAN-IFRA report (29 August 2025). WAN-IFRA scopes the same figure two ways ('in some of the use cases' and 'an average of 75 percent'); both are shown, not merged. web.archive.org snapshots were not captured and are flagged as a gap. The figure is Schibsted's own, so no confirmation was sought and none would count; the file caps because no independent measurement of the A/B result exists.

## Full case file

## The problem

Aftonbladet's hardest conversion problem is the visitor it knows nothing about. Schibsted's team described "personalisation for anonymous users" as one of the biggest challenges in the space, meaning visitors "who aren't logged in and for whom we have little to no historical data" ([source](https://www.inma.org/blogs/ideas/post.cfm/aftonbladet-sees-75-increase-in-subscription-sales-with-front-page-ai-content-recommendations)). The front page is where those visitors land, and Schibsted product manager Christoph Schmitz called it "the most powerful tool for this growth" because it "engages users already on our platform, requires minimal external cost" ([source](https://wan-ifra.org/2025/08/how-schibsteds-ai-model-helped-boost-subscription-sales/)).

## What was built

Aftonbladet "developed a machine learning (ML) model designed to predict which articles are most likely to result in a subscription" and places the likeliest converters in automated front-page positions ([source](https://www.inma.org/blogs/ideas/post.cfm/aftonbladet-sees-75-increase-in-subscription-sales-with-front-page-ai-content-recommendations)). During development the team "tested <span class="kpi">158</span> data points for their impact on subscription sales, narrowing them down to around a dozen features" ([source](https://wan-ifra.org/2025/08/how-schibsteds-ai-model-helped-boost-subscription-sales/)). The recommendations are shown only to non-subscribers; logged-in subscribers get interest-based recommendations "served by other ML models and approaches" ([source](https://www.inma.org/blogs/ideas/post.cfm/aftonbladet-sees-75-increase-in-subscription-sales-with-front-page-ai-content-recommendations)).

## The outcome

Schibsted's own staff put one number on it: "When A/B tested against our previous models, our new approach led to a <span class="kpi">75%</span> increase in sales from front-page articles" ([source](https://www.inma.org/blogs/ideas/post.cfm/aftonbladet-sees-75-increase-in-subscription-sales-with-front-page-ai-content-recommendations)). WAN-IFRA carried the same figure but scoped it two different ways in one article: first as "a 75 percent increase in subscription sales from front-page articles compared to previous models, in some of the use cases" ([source](https://wan-ifra.org/2025/08/how-schibsteds-ai-model-helped-boost-subscription-sales/)), and later as "front-page sales increased by an average of <span class="kpi">75</span> percent" ([source](https://wan-ifra.org/2025/08/how-schibsteds-ai-model-helped-boost-subscription-sales/)). Those are not the same claim, a lift seen in some use cases is not an average, and TIN shows both rather than picking one ([source](https://wan-ifra.org/2025/08/how-schibsteds-ai-model-helped-boost-subscription-sales/)).

The weakest point is the source of the number itself: the only first-hand account is a blog post written by three Schibsted employees, and WAN-IFRA restates the company's result rather than measuring it ([source](https://www.inma.org/blogs/ideas/post.cfm/aftonbladet-sees-75-increase-in-subscription-sales-with-front-page-ai-content-recommendations)). No absolute sales figure, test period, sample size or significance level is disclosed in either source, so the 75% is a relative lift against an undisclosed baseline ([source](https://wan-ifra.org/2025/08/how-schibsteds-ai-model-helped-boost-subscription-sales/)).

## Why this caps below green

Every load-bearing figure on this page traces to Schibsted. WAN-IFRA is independent trade press, but it reports the company's A/B result and does not audit it, and its own two scopings of the figure disagree. Green never rests on the subject's own numbers, and no independent measurement of this test exists in the public record. The honest ceiling is corroborated.

## How this was verified

The 75% figure and its two scopings were read verbatim this session from the INMA Ideas Blog piece by Vipul Goswami, Jacob Welander and Christoph Schmitz (8 December 2025; the page returns 403 to direct fetch and was captured through a reader proxy) and from the WAN-IFRA report by Research Editor Neha Gupta (29 August 2025); both copies are saved to `sources/`. The system description (158 data points, about a dozen features, non-subscribers only) was confirmed against the same two sources. web.archive.org snapshots were not captured from the war-room and the gap is flagged. Method date: 2026-09-30.

## Related case files

- [TikTok's recommendation algorithm fined by Italy's competition authority](/stories/tiktok-agcm-french-scar-recommendation-algorithm-10m-fine-minors-2024/): the other side of algorithmic content ranking, an engagement-optimised recommender judged by a regulator rather than by its owner's A/B test.
- [The Chicago Sun-Times AI summer reading list](/stories/chicago-sun-times-ai-fabricated-summer-reading-list-10-of-15-fake-books-king-features-2025/): a newspaper's AI deployment documented from the outside, the opposite evidentiary position to a publisher reporting its own lift.
- [Ars Technica's retraction of an article with AI-fabricated quotations](/stories/ars-technica-retracts-article-ai-fabricated-quotations-scott-shambaugh-2026/): newsroom AI where the public record, not the publisher's own metric, settles what happened.

## Sources

1. INMA Ideas Blog, "Aftonbladet sees 75% increase in subscription sales with front page AI content recommendations", Vipul Goswami, Jacob Welander, Christoph Schmitz (Schibsted): 8 December 2025. https://www.inma.org/blogs/ideas/post.cfm/aftonbladet-sees-75-increase-in-subscription-sales-with-front-page-ai-content-recommendations **(Tier 1, first party: Schibsted staff stating their own A/B result; not an independent measurement. Captured via reader proxy, saved to sources/.)**
2. WAN-IFRA, "How Schibsted's AI model helped boost subscription sales", Neha Gupta, Research Editor, 29 August 2025. https://wan-ifra.org/2025/08/how-schibsteds-ai-model-helped-boost-subscription-sales/ **(Tier 2, strong secondary: independent news-industry association, but it restates Schibsted's figure rather than measuring it, and scopes it two inconsistent ways. HTML saved to sources/.)**