Working · 2026-08SecondaryGrowth opsAnalytics

Contextual bandits to personalize creative

Use a bandit that also takes user attributes into account to choose which version of an image, headline or offer each user sees, learning online instead of waiting for a batch model and an A/B test.

Also called: contextual bandit, personalized artwork

Claim and evidence

The claimAs we suspected, the personalization worked and generated a significant lift in our core metrics.

Netflix · reforge.com · 2020-10-02

What the numbers show

No published number for this tactic has passed our word-for-word check yet, so there is no chart.

Why it works

Diagram, not data
Read the user'scontextServe the likelybest variantReward deeperengagementModel updates,keeps exploring
Read theuser'scontextServe thelikely bestvariantReward deeperengagementModelupdates,keepsexploring

Different users respond to different versions of the same content, so one global winner leaves value on the table. A contextual bandit keeps a little controlled randomness to keep learning while serving each context its likely best variant, and quality-of-engagement metrics stop it from learning clickbait.

How to run it

  1. Prepare a varied pool of creative for each item
  2. Choose context features (user history, device, page)
  3. Reward deeper engagement, not just clicks
  4. Evaluate candidate models offline on logged data before going live
  5. Confirm the lift with an A/B test against an unpersonalized bandit

Before / after

Before
Featured
Open
Same image and headline for everyone
After
Featured
Chosen for you
Open
Each user sees the version likely to fit

Generalised wireframe of the pattern, not a copy of any company's screen.

Ways to run it

Design options to test, not promised results.

  1. Option APer-user artwork

    Choose which image of the same item each member sees based on their history, as Netflix does with title artwork.

  2. Option BOffline replay first

    Evaluate candidate models on logged data before going live, then confirm with an A/B test.

  3. Option CEngagement-quality reward

    Reward deeper engagement rather than clicks so the model does not learn clickbait.

  4. Option DStable pages only

    Use it on pages whose content does not change often, such as the home page.

Dated examples

  1. —

    Netflix

    Netflix moved from finding one best artwork per title with bandits to choosing artwork per member with contextual bandits, evaluated offline by replay and then A/B tested against unpersonalized bandits.

    Netflix reported a significant lift in core metrics; no figure given in the summary.

    “As we suspected, the personalization worked and generated a significant lift in our core metrics.”
    ✓ quote checked word-for-word on the page

    reforge.com · 2020-10-02

Does it still work?

Workingas of 2026-08

Optimizely's current docs offer contextual bandits for pages whose content changes rarely; the Netflix case is older and was read through a Reforge summary because the Netflix blog blocked direct fetching. source

Risks

  • Needs large traffic and engineering for low-latency serving
  • Can learn clickbait if rewarded on clicks
  • Changing creative a user already knows can hurt recognisability

Sources

  1. reforge.com · Reforge Blog · 2020-10-02
    “We also look at the quality of engagement to avoid learning a model that recommends “clickbait” images”
    ✓ quote checked word-for-word on the page
  2. support.optimizely.com · Optimizely Experimentation Strategy docs · 2026-08-05
    “Contextual bandits are ideal to be run on pages where the content is not changed too frequently, such as the homepage”
    ✓ quote checked word-for-word on the page

Change log

  • 2026-11-20 — harvest v0.1 (pass2-expansion-ops)