Netflix chooses the artwork each member sees
Netflix posts describe testing several images per title and then choosing the image shown to each member from their viewing history. The performance claims are Netflix's own. In 2018 users alleged racial targeting; Netflix denied it.
For a psychology course
Separate the documented system or audit from an inference about human behavior. Identify the objective, observation and comparison in this record, then ask what additional evidence would be needed for a causal claim about people.
This record was transferred from Popular Psychology with its original claims, sources and reading limits. Its historical review date does not establish that the current platform works the same way.
- Case type
- Practice
- Context
- Consumer
- Organizations
- Netflix
What happened
Netflix published two blog posts on artwork. An earlier post described A/B tests that showed members different images for the same title. A post reported by Dezeen in December 2017 described choosing the image for each member individually. Netflix said it relies on the data to decide which signals to use, treats the member as the context, and saw a significant lift in its core metrics. No figures were given in the sources opened.
The psychological idea
The earlier post begins from Netflix's own observation that if it does not capture a member's attention within 90 seconds, the member will likely lose interest. Its stated goals were to help members find a story faster, raise engagement, and avoid misrepresenting titles. A later Netflix talk abstract frames artwork personalization as a contextual multi-armed bandit problem. Netflix's 2015 paper adds that algorithms decide which image out of several versions to use.
In practice
In the A/B tests Netflix reports that it significantly raised view share for titles tested with several images, and that expressive faces and recognisable or polarizing characters tended to do well. Dezeen's report of the 2017 post gives an example: a member who mostly watches romance may see Good Will Hunting artwork with the lead and his love interest, while a comedy viewer may see Robin Williams.
Why it mattered
The posts document a company treating the thumbnail as something to test and tailor per viewer, on the stated premise that attention must be captured within 90 seconds. A public dispute followed in 2018 about whether images were targeted by race.
Limits and qualifications
All performance claims are Netflix's. In October 2018 users said posters made Black actors in minor roles look like leads. Netflix said reports that it uses demographics were untrue, that it does not ask members for race, gender or ethnicity, and that it uses viewing history. The complaints were user observations, not a systematic study, and none of the sources opened is an independent audit of the claim either way.
Evidence behind the claims
A Netflix Technology Blog post, as republished by a university site, says that if you do not capture a member's attention within 90 seconds the member will likely lose interest, and sets goals of finding a story faster, raising engagement and avoiding misrepresenting titles.
Netflix reports it significantly raised view share of titles tested with several artwork variants, and that expressive faces and recognisable or polarizing characters tended to perform well.
Dezeen on 20 December 2017 reported that Netflix said it relies on data rather than hand-derived rules to pick signals, that the member is the context, and that core metrics saw a significant lift with no numbers given; Dezeen's example was Good Will Hunting artwork showing a love interest to romance viewers and Robin Williams to comedy viewers.
Netflix's 2015 paper says evidence selection algorithms decide which image out of several versions to use for a recommendation.
Users including Stacia L. Brown said Netflix showed Black actors as leads on some film posters despite minor roles, for example Like Father, and other users said the Love Actually poster suggested a love story between two characters.
Netflix responded that reports of it looking at demographics for artwork were untrue, that it does not ask members for race, gender or ethnicity, and that personalization relies only on a member's viewing history.
A QCon San Francisco talk abstract by Netflix's Justin Basilico frames artwork personalization on the homepage as a contextual multi-armed bandit problem.
People involved
- Justin Basilico. Netflix machine learning director, speaker on artwork personalisation at QCon San Francisco
- Stacia L. Brown. User whose 2018 observations on posters were quoted by The Fader
Course connections
- OpenStax Psychology 2e, section 2.1: Why Is Research Important?Context. The section discusses this topic.
Opened the page: it advises asking about a source's expertise and what it might gain, which applies to a company describing its own tests. It does not discuss A/B testing.
Course relevance is not evidence for the historical or corporate claims.
Sources and inspection
Inspection statements below preserve the original reading scope. A source marked as not verified supports no claim by itself.
- [5]Basilico, J. (2018, August 21). Qcon san francisco 2021. QCon San Francisco. https://archive.qconsf.com/node/110208Basilico, J. QCon San Francisco. 2018. Company document. Company document.
Abstract or publisher record only. Read only the conference talk abstract page. The page header shows a 2021 conference banner; the talk appears to be from the 2018 event, which was not confirmed.
- [1]Blog, N. T. (n.d.). Selecting the best artwork for videos through A/B testing (republished Netflix technology blog post). Republished by BINUS University; Original on Netflixtechblog.Com (Not Opened). https://international.binus.ac.id/graphic-design/2021/07/25/selecting-the-best-artwork-for-videos-through-a-b-testing/Netflix Technology Blog. Republished by BINUS University; original on netflixtechblog.com (not opened). Undated. Company document. Company document.
Company document read. Read a republished copy of the Netflix post on a university site as returned by the fetch tool; the site credits Netflix Technology Blog and shows no date. netflixtechblog.com returned HTTP 403, so the original was not compared. Follow-up correction: publication date is unconfirmed, so the year field is null; a footer, plan or prior researcher’s assumption is not publication evidence.
- [4]Gomez-Uribe, C., & Hunt, N. (2015). The Netflix recommender system: Algorithms, business value, and innovation. ACM Transactions on Management Information Systems, 6(4), Article 13. https://gwern.net/doc/reinforcement-learning/exploration/2015-gomezuribe.pdfGomez-Uribe, C; Hunt, N. ACM Transactions on Management Information Systems. 2015. Journal article. Company document.
Company document read. Read pages 1 to 10 of the paper in the gwern.net PDF copy, including section 2.7 on evidence selection.
- [3]Renshaw, D. (2018, October 22). Netflix users suggest the service is targeting black viewers with inaccurate movie posters. The Fader. https://www.thefader.com/2018/10/22/netflix-target-users-race. The Fader. 2018. News. Secondary.
Reference article read. Read The Fader report of 22 October 2018 on user complaints and Netflix's response, as returned by the fetch tool. No independent audit was read.
- [2]Yalcinkaya, G. (2017, December 20). Netflix targets different film artworks to suit users’ viewing habits. Dezeen. https://www.dezeen.com/?p=1165745. Dezeen. 2017. News. Secondary.
Reference article read. Read the Dezeen report dated 20 December 2017 as returned by the fetch tool; the quoted Netflix wording and the Good Will Hunting example are from that report. It gives no figures.
Access note: Netflix's own 2016 and 2017 blog posts returned HTTP 403. The 2016 post is read through a republished copy; the 2017 post only through Dezeen. A lift figure seen in one report was not used.
Original independent review record
Claude review agent, independent of the research agent. 2026-10-21. Verdict: pass-with-edits.
- Changed the QCon talk claim from level page to abstract, because only the abstract was read.
- Removed 'engineers' as the blog authors and 'engineering director' for Basilico; the sources opened give only Netflix and a machine learning director title.
- Described Stacia L. Brown as a user quoted by The Fader, not as a writer, which the opened page does not state.
- Corrected the Fader claim: Brown's example was Like Father; the Love Actually complaint came from other users and concerned a suggested love story.
- Attributed the Good Will Hunting example to Dezeen and put it in a claim.
- Stopped implying the 2017 post used the bandit approach; only a later talk abstract names it.
- Replaced 'members decide within seconds' with Netflix's stated 90-second premise, and 'fed a dispute' with a non-causal 'followed'.
- Replaced 'no audit was found' with a statement about the sources opened.
- Codex audit: replaced an unverified publication-year placeholder with null; the actual historical milestone date remains unchanged.