Do better recommendations reduce the top products’ share?
In a Netflix experiment, did recommendation updates move viewing beyond the most popular titles? Viewing was slightly less clustered around the few most popular titles, a relative decrease of 1.2% compared with the control group, with viewing share moving mainly to moderately popular titles. For a business managing a catalogue, changing recommendations could change which items attract attention.
ACADEMIC WORK REVIEWED
Recommendation Quality and the Concentration of Consumption: Experimental Evidence from Netflix
ResearchersGuy Aridor, Winston Chou, Nathan Kallus, Antoine Scheid, Allen Tran, Kevin Zielnicki
arXiv · 21 August 2026 · v1, not peer reviewed · Preprint · Source language: English
English review: 30 October 2026 · The source date appears in the citation above.
Editorial review draft
The source has been checked; the review still carries its original editorial status. DijitalPi did not conduct a customer experiment for this paper.
LET'S READ THE RESEARCH TOGETHER
We first explain the researchers' question, method and findings. We then discuss how to interpret the results, clearly separating DijitalPi's commentary from the source.
01 / WHAT DID THE RESEARCHERS WANT TO UNDERSTAND?
When recommendations improve, do viewers concentrate on the most popular titles or spread their viewing across a broader selection?
The researchers examined how advances in recommendation technology change which titles people watch, alongside total viewing. This separates increased activity from a wider distribution of consumption. They also assessed whether users continued watching the titles they started.
Explaining the context · DijitalPi commentary
DijitalPi commentary: For a business managing a catalogue, total clicks alone may be an incomplete way to assess recommendations. Attention concentrated on a few popular products presents different catalogue decisions from attention spread across a wider selection. This study suggests separating expectations for moderately popular products from those receiving very few interactions.
02 / HOW WAS THE RESEARCH CONDUCTED?
How did they test the question?
Guy Aridor, Winston Chou, Nathan Kallus, Antoine Scheid, Allen Tran and Kevin Zielnicki examined a Netflix experiment involving 8,559,252 subscribers over 60 days between February and April 2025. The control group received the recommendation system as it stood at the start. The other group received algorithm updates released during the experiment. The interface did not change. The text does not separately describe how users were assigned to groups or report group sizes.
Understanding the method · DijitalPi commentary
The comparison placed a frozen algorithm against a version receiving the updates together, so it does not isolate the effect of any single innovation. Consumption meant playing a title, and recommendations meant titles on the homepage. Titles were grouped by their position in the playback distribution: the top 5%, the bottom 50% and the group between them. The bottom 50% was the group of rarely watched titles. One main outcome was the share of each user’s plays allocated to these groups. A concentration index measured how strongly recommendations or plays clustered around a small number of titles.
Because titles entered and left the platform, groups were defined using the pooled distribution during the experiment. Defining them from the control group alone produced the same result. Activity measures covered viewing hours, plays, distinct titles and days with a play. Match quality was assessed through indirect indicators such as continued viewing and positive ratings. These measures came from the theory and were not the core product metrics Netflix optimises.
03 / RESEARCH FINDINGS
Viewing spread mainly towards moderately popular titles
The researchers reported a 5.7% reduction in recommendation concentration and a 1.2% reduction in playback concentration. Both were statistically significant. Most of the share moving away from the most popular titles went to moderately popular ones. Recommendations for the least watched group changed little. Its share of plays did not change numerically. Viewing hours rose by 0.37%, the total number of titles played by 0.62%, distinct titles played by 1.2% and days with at least one play by 0.21%. The number of titles reaching each first-day viewing threshold of 15, 30 and 45 minutes rose by 1.3%, and positive ratings increased by 0.57%. However, the probability of watching at least 30 minutes and the share of positive ratings per play fell significantly. Effects on the shares reaching 15 and 45 minutes were reported as zero with little statistical uncertainty. The effect on the negative-rating share was not significant. All percentage changes are relative to the control mean, not percentage-point changes.
Recommendation concentration
5.7% reductionThe relative change in the title-level concentration index against the control group, indicating that recommendations spread across a broader selection.
Playback concentration
1.2% reductionThe relative change in the title-level concentration index against the control group, not a change in sales, revenue or subscriber numbers.
DijitalPi commentary: Assessing a recommendation change may require tracking the distribution of visibility, actual behaviour and quality per interaction together. In this experiment, more viewing did not mean every play was a better match. If total activity rises in your business, examine completion and positive-rating proportions separately.
The authors note that matching titles with very little interaction data to suitable users remains difficult even for mature recommendation technologies. Additional analyses could not consistently determine whether newly played titles were higher or lower quality. The model also assumes that recommendations’ influence on user decisions remains constant over the 60-day period. These boundaries prevent assuming the same distribution or magnitude of effects for Türkiye, Turkish-language catalogues or other sectors.
05 / CONCLUSION AND OPEN QUESTIONS
What did we learn, and what do we still not know?
This study is a preprint, a research manuscript that has not yet undergone peer review. The experiment shows that the algorithm updates studied could spread Netflix consumption mainly towards moderately popular titles while producing small increases in total viewing. The least watched titles did not show the same result, and quality measures did not improve uniformly. This review concerns the arXiv v1 preprint, which has not undergone peer review, and a 60-day platform experiment; sales, revenue and subscriber gains were not measured. All authors were Netflix employees throughout the project. No separate funding or competing-interests statement was found in the source.
Our open question at DijitalPi: if your recommendation system spread attention towards moderately popular products, which actual behaviour and quality measure would establish whether that change mattered to your business?
We created these scenarios to make the topic concrete. They are not cases from the paper or measured client results.
EXAMPLE 01
E-commerce: the same products keep appearing
An illustrative homeware shop repeatedly fills its recommendation slots with the same popular products.
Does the new recommendation system bring attention to moderately popular products too?
Open the recommendation for example 01+
Compare the distribution of recommended products separately from those viewed and purchased
Group products by their level of interest over a defined period and record the classification rule. Measure each group’s share of impressions, product views and purchases separately under the existing and new recommendation systems. Record stock changes to help interpret shifts across the catalogue. Netflix’s playback findings cannot become an expectation of higher sales for your shop.
EXAMPLE 02
Education: a recommended lesson is opened and abandoned
An illustrative online education business wants an updated recommendation system to encourage learners to open a broader selection of lessons.
Does opening more lessons mean learners are finding more suitable ones?
Open the recommendation for example 02+
Assess the number of lessons started separately from the proportion watched beyond a defined threshold
Record lessons opened through recommendations and those reaching a viewing threshold chosen in advance. Compare both the total number reaching that threshold and their share of all lessons started. Examine learner ratings with the same separation between counts and proportions. Some quality proportions declined in the study, so more starts alone should not be treated as better matching.
EXAMPLE 03
B2B: discovery of rarely viewed catalogue products
An illustrative industrial supplies business wants recommendations to give rarely viewed catalogue products more visibility.
Will updating recommendations be enough to generate demand for the least known products?
Open the recommendation for example 03+
Separate results for moderately popular products from those receiving very few interactions
Use existing product-view records to identify products with very little interaction data. Track recommendation impressions, technical document views and quotation requests by product group. Record periods when products are absent from the catalogue. Consumption in Netflix’s least watched group was quantitatively unchanged, so the same approach cannot be assumed to generate quotation requests for niche products.
TRY IT WITH YOUR TEAM
Measure how often each product group is chosen
Before a proposed trial against the existing recommendation system, define product groups, the assessment period and the measure that would change your decision. Plan to track group shares and quality per interaction alongside total activity. This is a proposed trial that DijitalPi has not conducted.
Guy Aridor, Winston Chou, Nathan Kallus, Antoine Scheid, Allen Tran, Kevin Zielnicki Netflix (Guy Aridor, Winston Chou, Nathan Kallus, Antoine Scheid, Allen Tran, Kevin Zielnicki); Northwestern Kellogg (Guy Aridor); Cornell University, Cornell Tech (Nathan Kallus).
The academic work belongs to the researchers named above. This page contains DijitalPi's explanatory review and original business examples; it is not a full translation of the paper.
Cookie preferences