Loading
Skip to content
Miracle Minds AI
  • Try “HIPAA”, “model risk”, “FinOps” or “agentic”.

Personalization and recommendation

A recommender pays off when a controlled test shows it moves retention, not when offline accuracy improves. Netflix's own team found offline results were not as predictive of A/B outcomes as they wanted, and the data you train on is shaped by what you already recommended.

Part of AI for Telecom, Media, Entertainment & Gaming. 8 sources, checked .

Share of Netflix hours streamed where the recommender influenced the choice (Netflix, 2015)
About 80%
Share of Netflix hours streamed where the recommender influenced the choice (Netflix, 2015)
ACM Transactions on Management Information Systems, Vol. 6, No. 4, Article 13, 2015(opens in a new tab)
Netflix's own estimate of yearly savings from personalization and recommendation, attributed to lower churn (2015)
$1B+
Netflix's own estimate of yearly savings from personalization and recommendation, attributed to lower churn (2015)
ACM Transactions on Management Information Systems, Vol. 6, No. 4, Article 13, 2015(opens in a new tab)
Time a typical Netflix member spends choosing before losing interest, per Netflix consumer research (2015)
60 to 90 s
Time a typical Netflix member spends choosing before losing interest, per Netflix consumer research (2015)
ACM Transactions on Management Information Systems, Vol. 6, No. 4, Article 13, 2015(opens in a new tab)

The problem

Viewing has moved to streaming, and the choice screen is now where sessions are won or lost. Streaming took 49.6% of US TV viewing time in August 2026, with YouTube alone at 15.8% and Netflix at 7.7%.1 Every one of those services competes for the same evening, and for the same few minutes on a home screen.

Netflix described the stakes in 2015. Its consumer research found a typical member loses interest after perhaps 60 to 90 seconds of choosing, having reviewed 10 to 20 titles. The recommender influenced choice for about 80% of hours streamed, and Netflix estimated that personalization and recommendations saved it more than $1B per year through lower churn.2 Those are Netflix's own figures, but they show where the value sits: retention, not clicks.

Measurement is where most programs go wrong. Netflix relied on offline experiments to choose what to test but did not find them as predictive of A/B test outcomes as it wanted.2 Research presented at ACM RecSys 2018 showed that training on data from users who were already shown recommendations creates a feedback loop that homogenizes behavior without increasing utility.3 A model can look better every quarter while the catalog people actually see gets narrower.

The rules have caught up. Under the EU Digital Services Act, platforms must explain the main parameters of their recommender systems, and very large platforms must offer at least one option not based on profiling.4 The EDPB's 2026 guidelines add that both options should be presented equally and that profiling must stop while the non-profiling option is active.5 In the US, viewing history and carrier customer data carry their own consent rules.6,7

What we build

  1. 01

    Retention-targeted ranking

    Ranks rows and titles on predicted medium-term engagement and renewal rather than click-through, with business rules for new releases, rights windows and editorial picks.

  2. 02

    Experimentation platform with long-run holdouts

    A/B testing with guardrail metrics and a standing holdout group that never gets the new model, so you can measure the cumulative effect on churn, not just the first week.

  3. 03

    Consent-aware offers for telecom customers

    Next-best-offer models for plans and add-ons that check each customer's CPNI approval status before the data is used, and record which approval covered each decision.

  4. 04

    Non-profiling mode and plain-language explanations

    A ranking built on popularity, recency and editorial signals for users who opt out of profiling, plus a short 'why this' explanation tied to the main parameters you disclose.

  5. 05

    Personalized artwork with provenance

    Generates and tests artwork and trailer variants per audience segment, with C2PA Content Credentials attached so the record of how each asset was made travels with the file.

Where it goes wrong

The failure modes we design against from the first week, and the control for each.

  • Shipping on offline accuracy

    Offline metrics assume members would have behaved the same way under the new algorithm, which is why Netflix found them weak predictors of live results.2 Use offline tests only to shortlist, and make release decisions on controlled online tests measured against retention.

  • Feedback loops that shrink the catalog

    Training on interactions your own recommender caused reinforces what it already shows and homogenizes behavior.3 Reserve an exploration budget, log the probability each item was shown, and track catalog coverage and diversity next to engagement.

  • Viewing history leaking to ad partners

    Under the VPPA, information identifying a person as having requested specific video materials is protected, consent must be distinct and separate from other terms, and courts can award liquidated damages of $2,500.6 Inventory every pixel, SDK and data feed that carries viewing events and gate them on consent.

  • Nudging users toward the profiled feed

    The EDPB says very large platforms should present profiling and non-profiling options equally, not nudge users toward profiling, and honor the choice across sessions.5 Design review should cover the choice screen, not only the model.

Rules that apply

  • Platforms must explain the main parameters of their recommender systems in clear terms, and very large platforms must offer at least one recommendation option not based on profiling.

  • Video Privacy Protection Act, 18 U.S.C. 2710

    Primary text on law.cornell.edu(opens in a new tab)

    A video service may not knowingly disclose information identifying what a consumer requested or watched without informed, written consent given separately from other terms, for up to two years or until withdrawn.

  • FCC CPNI rules, 47 CFR 64.2007

    Primary text on ecfr.gov(opens in a new tab)

    Carriers may use customer proprietary network information to market communications-related services only with opt-out or opt-in approval, must keep approval records for at least one year, and need opt-in approval for most other uses.

  • Defines a signed manifest of assertions about how an asset was made, so generated or edited artwork and promos carry verifiable provenance.

How to measure it

  • Churn and renewal rate against a long-run holdout
  • Share of viewing hours started from recommendations versus search
  • Browse abandonment: sessions that end without a play
  • Catalog coverage: share of titles that receive meaningful viewing
  • Consent coverage for every data flow that carries viewing or CPNI data

Questions buyers ask

How do we prove our recommender actually reduces churn?

Keep a randomized holdout group that gets a simple baseline, such as popularity ranking, for months rather than weeks, and compare cancellation and renewal rates against the group on the full model. Pair that with shorter A/B tests on engagement for individual changes. Offline accuracy tells you which ideas to test, not which to ship. Report the holdout result to finance in the same terms as any other retention lever.

Do we need a non-personalized option for our users?

If you are designated a very large online platform or search engine under the EU Digital Services Act, yes: you must offer at least one recommender option not based on profiling. Every online platform in scope must explain the main parameters behind its recommendations. Even outside the EU, a non-profiled mode is a useful baseline for measurement and a fallback when consent is missing.

Can we share viewing data with advertising partners?

In the US, the Video Privacy Protection Act limits disclosure of information that identifies what a person watched. Sharing generally needs informed, written consent that is separate from your other terms and lasts up to two years or until withdrawn. Many exposures come from pixels and SDKs rather than deliberate deals, so map every data flow that carries viewing events before you build on them.

Sources

Every figure on this page comes from the sources below. We checked each link on .

  1. 1.TV Viewing Cools Off in Nielsen's August Gauge Reports as Sports Take a Breather and School Resumes(opens in a new tab)Nielsen, October 22, 2026 · nielsen.com
  2. 2.The Netflix Recommender System: Algorithms, Business Value, and Innovation(opens in a new tab)ACM Transactions on Management Information Systems, Vol. 6, No. 4, Article 13, December 2015 · doi.org
  3. 3.How Algorithmic Confounding in Recommendation Systems Increases Homogeneity and Decreases Utility(opens in a new tab)Proceedings of the 12th ACM Conference on Recommender Systems (RecSys 2018), September 27, 2018 · arxiv.org
  4. 4.User rights under the Digital Services Act(opens in a new tab)European Commission, July 17, 2026 · digital-strategy.ec.europa.eu
  5. 5.Guidelines 3/2025 on the interplay between the DSA and the GDPR, Version 2.0(opens in a new tab)European Data Protection Board, September 17, 2026 · edpb.europa.eu
  6. 6.18 U.S. Code 2710: Wrongful disclosure of video tape rental or sale records(opens in a new tab)Legal Information Institute, Cornell Law School · law.cornell.edu
  7. 7.47 CFR 64.2007: Approval required for use of customer proprietary network information(opens in a new tab)Electronic Code of Federal Regulations, 2026 · ecfr.gov
  8. 8.Content Credentials: C2PA Technical Specification, Version 2.5(opens in a new tab)Coalition for Content Provenance and Authenticity (C2PA), October 2026 · spec.c2pa.org