Culture

How Netflix Uses AI: A Verified Case Study from Personalization to Production (2026)

TL;DR. According to the annual reports Netflix files with the US Securities and Exchange Commission (SEC), the company’s revenue rose from $15.79 billion in 2018 to $45.18 billion in 2025; its operating margin climbed from 10% to 29.5%. Over the same period technology and development spending rose from $1.22 billion to $3.39 billion, yet its share of revenue is printed as 8% in both reports. A 2015 paper by Netflix employees says the recommender system influences the choice for about 80% of hours streamed and is thought to save more than $1 billion a year; an independent review notes that how these figures were calculated is not explained. In a 2026 Netflix paper, the improvement delivered by a new ranker built on a large language model on the primary long-term metric is 0.006%. The CEO move: pick a single metric that measures retention, not clicks. The student move: run one small, controlled experiment every week and write down the result.

Related reading: the attention economy’s new weapons, trust calibration, the signal-to-noise crisis for separating company statements from verified data, how LLMs actually work for the logic of language models, and the AI productivity paradox for why small gains rarely show up in company results.

Why was this article rewritten, and what changed?

The first version of this page, dated 2019, was a short unsourced roundup describing personalized cover images. Three things changed in the rewrite.

First, the numbers were tied to primary sources. The old text said “In 2018 alone, it gained 8.8 million new members”. According to Netflix’s January 2019 shareholder letter, 8.8 million (1.5 million in the US, 7.3 million international) is the paid net membership addition for the fourth quarter of 2018 only. The annual report prints paid net membership additions for the whole of 2018 as 28,615 thousand, or about 28.6 million. The 139.3 million figure in the old text, by contrast, is consistent with the 139,259 thousand paid memberships in the report.

Second, claims that could not be verified were removed. No support was found, in any primary source opened for this research, for the claims that Netflix “acquired a simulation technique of its own”, that the algorithm “recognizes the entire movie”, takes “the scene with the best contrast” and inserts “the movie’s name and company logo into the scene”, or that member data was entrusted to the “Credentials Sharing Insight” software of a company called Synamedia.

Third, data for 2025 and 2026 was added: the latest annual report, shareholder letters, two 2026 papers by Netflix researchers, and independent studies that test the company’s narrative. Company figures are reported for context; this article is for general information and is not investment advice.

How did Netflix change from 2018 to 2025, in numbers?

The table below sets the line items the old article relied on side by side for fiscal years 2018, 2024 and 2025, as printed in the reports. The reason for the 2024 column is this: in its April 2024 shareholder letter, Netflix announced that from the first quarter of 2025 it would stop reporting quarterly membership numbers and average revenue per membership. The 2025 annual report also states in a footnote that reporting of membership numbers has ended and that the focus is on revenue and operating margin as the primary financial metrics. The last official year-end membership count therefore belongs to 2024.

Table 1. Figures reported by Netflix (company statement, from audited annual reports, as printed)

Line item 2018 2024 2025
Revenue (USD thousands) 15,794,341 39,000,966 45,183,036
Operating income (USD thousands) 1,605,226 10,417,614 13,326,603
Operating margin 10% 26.7% 29.5%
Technology and development expense (USD thousands) 1,221,814 2,925,295 3,391,390
Technology and development as a share of revenue (as printed) 8% 8% 8%
Marketing expense (USD thousands; “sales and marketing” in 2024-2025) 2,369,469 2,917,554 3,301,306
Paid memberships at end of period (thousands) 139,259 301,626 Not reported
Paid net membership additions during the year (thousands) 28,615 41,350 Not reported
Average monthly revenue per paying membership (USD) 10.31 11.70 Not reported

Source: Netflix, Inc. Form 10-K, fiscal years 2018, 2024 and 2025 (SEC EDGAR). The 2024 operating margin is the 26.7% value printed in the 2025 report; it was rounded to 27% in the 2024 report.

Two current facts that are not in the table come from shareholder letters. The January 2026 letter reports that the company crossed the 325 million paid memberships mark in the fourth quarter of 2025 and that advertising revenue grew more than 2.5x to over $1.5 billion. The latest letter, dated October 2026, gives third-quarter revenue as $12.9 billion and the operating margin as 34.6%; the company’s full-year 2026 forecast is $51.1 billion in revenue and a 31.5% operating margin. In the same letter the 2026 advertising revenue forecast is lowered from about $3 billion to about $2.7 billion. Forecasts are not realized results.

What can be calculated from these numbers?

Table 2. CEOtudent analysis: ratios derived from Netflix 10-K figures

Indicator 2018 2025 (unless stated otherwise) Change
Revenue $15.79 billion $45.18 billion 2.86x
Operating income $1.61 billion $13.33 billion 8.30x
Technology and development expense $1.22 billion $3.39 billion 2.78x
Marketing expense $2.37 billion $3.30 billion 1.39x
Technology and development / revenue 7.7% 7.5% -0.2 points
Marketing / revenue 15.0% 7.3% -7.7 points
Paid memberships at end of period 139.26 million 301.63 million (2024) 2.17x
Average monthly revenue per paying membership $10.31 $11.70 (2024) +13.5%
Annual technology and development expense per membership $8.77 $9.70 (2024) +10.5%

CEOtudent analysis, from Netflix Form 10-K figures (2018, 2024, 2025). Expense per membership is annual expense divided by the number of paid memberships at end of period. If the phrase “over 325 million” is taken as a lower bound, membership growth is at least 2.33x.

The technology share is flat, the marketing share has halved. Technology and development expense grew at almost the same pace as revenue (2.78x against 2.86x) and stayed in a band of 7.5-7.7% of revenue. Marketing expense grew only 1.39x; its share of revenue fell from 15.0% to 7.3%. In 2025 technology and development expense exceeded marketing expense by about $90 million. The reports draw no causal link between these two trends; but against the old article’s emphasis on “marketing intelligence”, what the data shows is the steadiness of investment inside the product.

Technology spending per membership is small and rising slowly. Technology and development expense, $8.77 per membership per year in 2018, is $9.70 in 2024. Personalization, testing infrastructure and the entire interface cost less than one month’s average revenue from one membership ($11.70). The real factor that makes AI investment cheap is scale.

The company changed its scoreboard. The decision to stop reporting membership numbers and focus on revenue and margin is perhaps the most instructive CEO move in this case: deciding which metric matters comes before the algorithm.

What does the recommender system actually optimize?

The most detailed primary source on Netflix’s recommender system is the paper published by Carlos Gomez-Uribe and Neil Hunt of Netflix in ACM Transactions on Management Information Systems in December 2015. The paper was written from inside the company and is more than ten years old; even so, most of the famous figures in circulation come from it.

80% and 20%. According to the paper, the recommender system is used on most screens of the product and in total influences the choice for about 80% of hours streamed on Netflix; the remaining 20% comes from search. The word to watch is “influences”.

$1 billion a year. The paper writes that monthly subscriber churn is in the low single-digit percentages, that personalization and recommendations have reduced this rate by several percentage points over the years, and that the two effects combined are thought to save the company more than $1 billion a year. The sentence begins with “we think”; it is an internal estimate, not an accounting line.

The real target is retention. The least quoted but most instructive part of the paper is the choice of metric. According to the authors, the main measurement target of changes to the recommendation algorithms is member retention. But because retention rates are already high, creating even a 0.1% (10 basis points) difference requires a very meaningful improvement. So the company also measures medium-term viewing time, which it has found to be strongly correlated with retention, and typically runs its A/B tests for 2-6 months.

A local win is not an overall win. The paper distinguishes three kinds of result: an engagement increase in the part of the product that was changed, an engagement increase across the whole product, and an increase in retention. The authors say they find multiple clear wins every year, but that overall engagement wins that do not affect retention are seen more often, and local wins that do not change total viewing at all are seen more often still; because a local win usually steals viewing from another part of the product.

What can be verified about personalized cover images?

This was the subject of the old article. The existence of the mechanism is confirmed by primary sources; most of the details in the old text are not.

Verified. The 2015 paper writes that “evidence selection” algorithms also decide which image, out of several versions, to use to support a recommendation. An arXiv paper by Netflix researchers dated August 2026 describes the current system: each title has a set of candidate images, and the system algorithmically selects an image or video preview from that set for each member. The same paper states that artwork personalization was first formulated in 2017 as a contextual bandit problem over a fixed inventory of images per title.

What changed in 2026. According to the paper, the old models treated each image only as an identification number without looking at its content; because newly released titles had no interaction history, the system fell back on popularity. Once off-the-shelf image embeddings representing the content of the image were added, a single model began to serve all five artwork formats used on different devices and replaced five separately trained models. A small portion of member traffic is set aside for evaluation, and an exploration policy that picks at random among the candidates with a known probability distribution is applied to that portion. The system deliberately tries things at random in order to learn.

Not verifiable. The old article’s example of “graphics that will evoke action” for the action viewer and a more emotional cover for the viewer of emotionally heavy series spread through association with the Netflix team’s 2017 piece “Artwork Personalization at Netflix”; but because that text could not be opened in this research, its examples and figures were not used. The two primary sources that could be opened say the system selects among ready-made candidates; no source was found to support the account that the algorithm generates a new personalized cover and adds a logo.

An independent caution. The review by Dietmar Jannach and Michael Jugovac, published in the same journal in 2019, points out that artwork personalization creates a measurement problem: if the image is chosen to persuade the member that the content suits them, counting only “started watching” can be misleading, because that count also includes those who started and did not like it.

What changed in 2025-2026: language models and generative AI?

The primary sources of the last two years show four separate areas of use.

1. A large language model in recommendation ranking. The GenRec paper by Netflix researchers, dated August 2026, describes a trial move from the traditional ranker, which uses thousands of hand-engineered features, to a ranker built on a language model that expresses the member’s history in natural language. The online test in the paper ran for 4 weeks on about 10% of Netflix traffic. The result: a relative improvement of +0.115% on the short-term homepage engagement metric and +0.006% on the primary long-term metric. The authors describe the latter as “statistically significant at Netflix scale”: in a mature system even the newest architecture delivers a gain of this size, and it takes hundreds of millions of memberships to see it. The gain on the short-term metric is about 19 times the gain on the long-term one: metrics close to the click move more easily.

2. Search and discovery. The October 2025 shareholder letter writes that machine learning and AI have been used for many years in title recommendations as well as in production and promotion technology; the examples given for generative AI are conversational search, in beta testing, which lets members explore the catalog in natural language, and the localization of promotional materials into different languages.

3. Production. The same letter gives two examples: the de-aging of characters with the help of generative AI in the flashback scene that opens Happy Gilmore 2, and the use of generative tools by the producers of Billionaires’ Bunker for costume and set design before shooting. The letter also states that a production guidance document has been published for creative partners; because the guidance itself could not be opened, its content is not reported. The January 2026 letter adds the use of AI in subtitle localization.

4. Advertising. According to the October 2025 letter, the company uses AI to test new ad formats, to generate the most relevant ad creative and placement for the member, and to speed up media plans; the ad-supported plan is offered in 12 markets.

The company’s risk statement is also on the record: the 2025 annual report writes that the development and use of generative AI tools is subject to uncertain legal frameworks and that it is uncertain whether copyright and other intellectual property protection will be available for AI-generated material.

Which parts of the company narrative does independent research confirm, and which not?

Almost all of the sources above are Netflix’s own texts. Four studies from outside the company mark the limits of this narrative.

The business value figures cannot be audited. The 2019 review by Jannach and Jugovac surveys the published findings on the business value of recommender systems. On Netflix’s $1 billion estimate, the authors note that unfortunately it is not explained in detail how these figures were calculated. Another of their cautions: when almost everything on the page is a recommendation, users are inclined to pick what is recommended simply because it is there; adoption rates can therefore make the system’s value look larger than it is.

The feedback loop can homogenize. The 2018 study by Allison Chaney, Brandon Stewart and Barbara Engelhardt of Princeton University shows that training recommender systems on data from users who have already been exposed to recommendations creates a feedback loop, and that in simulations this loop homogenizes user behavior without increasing utility. The finding rests on simulation, not on Netflix data. The random exploration traffic described in Netflix’s 2026 paper is a mechanism set up precisely to reduce this kind of bias.

The “too much choice paralyzes” argument is weak. The 2015 paper opens by saying that people are surprisingly bad at choosing between many options. The meta-analysis by Benjamin Scheibehenne, Rainer Greifeneder and Peter Todd, published in the Journal of Consumer Research in 2010, however, combined 63 conditions from 50 experiments and found the mean effect size to be virtually zero and the variance between studies to be considerable. Choice overload does harm under some conditions and not under others; it is not a universal law.

The controlled experiment is sound as a method. The guide by Ron Kohavi and colleagues, published in Data Mining and Knowledge Discovery in 2009, regards the controlled experiment as the best scientific design for establishing causality between a change and user behavior, and advises teams to listen to the user instead of “the highest paid person’s opinion”. Netflix’s method is consistent with this literature; the size of the results, however, is a company statement.

Famous Netflix claims: which are verified, which are company statements, which are myth?

Table 3. CEOtudent analysis: a source scorecard for the claims circulating about Netflix’s use of AI

Claim What the primary text says Status
“80% of viewing comes from recommendations” The recommender system influences the choice for about 80% of hours streamed; the remaining 20% comes from search (2015 paper) Company statement, dated 2015; with the word “influences”
“The recommender system earns $1 billion a year” Personalization and recommendations combined are thought to save more than $1 billion a year (2015 paper) Company estimate; an independent review notes the method is not explained
“The Netflix Prize was $1 million and asked for a 10% improvement” Announced in 2006; $1 million to whoever improved the accuracy of the existing system by 10% (Amatriain, SIGKDD Explorations) Verified (published paper by a company employee)
“The prize-winning algorithm was never used” The two best algorithms of the 107-algorithm blend that won the first progress prize with an 8.43% improvement were put into production; the additional methods in the final blend that won the grand prize were not worth the engineering effort Half true: the final blend was not used, two of its components were
“Cover images change from person to person” The image is selected algorithmically from a candidate set for each member (2015 and 2026 papers) Verified (technical papers by company authors)
“The algorithm scans the movie, generates a personalized cover and adds a logo” The sources opened describe only selection among ready-made candidates Could not be verified; removed from the old article
“New AI models radically improve recommendations” The ranker built on a language model delivered a +0.006% improvement on the primary metric (2026 paper) Company data; the improvement is real but very small

CEOtudent analysis. “Verified” means the claim is written in the primary text; it does not mean the figure has been independently audited.

The lesson of the Netflix Prize row is clear: winning on a competition metric and making a product change that works are not the same thing.

How does a one-person business build the Netflix loop?

Netflix’s scale cannot be copied; a newsletter writer or freelancer with a few hundred customers will never be able to see a difference of 0.006%. What can be copied is the order of decisions: deciding first what will be measured, testing the change under control, separating the local win from the overall win.

Table 4. CEOtudent editorial framework: translating the Netflix loop into a one-person business in six steps

Step Netflix equivalent One-person business equivalent CEO decision Student habit
1. Choose the metric The main target is retention; viewing time is its proxy “After 90 days, are they still a customer, still reading?” Clicks, likes and open rates are not the main metric Regularly check the proxy metric’s relationship to the real target
2. Measure Every impression and view is recorded A single table: who came, what they bought, when they left Close the unmeasured channel or make it measurable Update the table on the same day every week
3. Run a small experiment A/B test with a control cell One variable, two versions, an expectation written in advance One experiment at a time Record the result even if the expectation does not hold
4. Personalize the presentation Same title, different cover image Same product, a different headline, example and intro for 2-3 audiences Change the shop window, not the product Note which audience responds to which framing
5. Separate the local win A rise in one row can steal from another row Clicks on one email can rise while total sales stay the same Decide on the total result In every experiment ask “where did this gain come from?”
6. Set aside a share for exploration Random selection for a small portion of traffic Set aside a small portion of effort for trials the data does not suggest Fix the exploration budget in advance Keep surprising results in a separate list

CEOtudent editorial framework. The “Netflix equivalent” column is based on the 2015 and 2026 papers; the other columns are editorial interpretation, not research findings.

Why is retention worth more than a click? A sample calculation. The numbers below are not Netflix data; they are a hypothetical example. In a newsletter with 1,000 subscribers, if monthly churn is 5%, then after 12 months, assuming no new subscribers arrive, 540 subscribers remain; the average subscription lifetime is 20 months. If monthly churn falls to 4%, 613 subscribers remain after 12 months and the average lifetime rises to 25 months. A one-point difference amounts to about 72 subscribers in a year (using unrounded values). That is why the order of experiments starts with the changes that could affect churn.

Weekly checklist

  • Is the main metric written in a single sentence, and is it not a click-type metric?
  • Does this week’s experiment change a single variable?
  • Were the expected result and the “success” threshold written down before the experiment started?
  • Will the result be read against the main metric, or only against a local indicator?
  • Could the gain have been stolen from another channel or product?
  • Is there at least one trial the data does not suggest in this month’s plan?

What does this case not mean?

AI has not been proven to be the cause of the growth. The reports do not apportion the growth in revenue and profit among content investment, price increases, the crackdown on account sharing, the ad-supported plan and recommendation technology. The fact that technology spending has stayed constant as a share of revenue is a correlation.

Technology and development expense is not an “AI budget”. The annual report defines this line as the payroll and related expenses of the technology personnel responsible for improving the service; the testing, maintenance and modification of the interface, the recommendations and the infrastructure all sit in this line. The share of AI within it is not reported separately.

A company-authored paper is not independent evidence. Most of the descriptions of mechanisms in this article rest on texts written by Netflix employees. Their being published makes the method visible; it does not make the figures auditable.

Frequently asked questions

Where exactly does Netflix use AI?
According to primary sources, in four areas: title recommendations and ranking, cover image and preview selection, production and localization, and advertising. The October 2025 letter states that machine learning has been used for many years in recommendations and in production and promotion technology, and that generative AI is a new opportunity.

Is it true that 80% of viewing comes from recommendations?
The figure comes from the 2015 paper by Netflix employees, and its exact wording is that the recommender system influences the choice for about 80% of hours streamed, with the remaining 20% coming from search. An independent review cautions that on a page where almost everything is a recommendation, ratios of this kind can overstate the system’s value.

Does Netflix show everyone a different cover image?
Yes, the mechanism is verified: each title has a prepared set of candidate images and the system selects one of them for each member. According to the 2026 paper, a single model serves five different artwork formats. No source was found indicating that the system generates new personalized images.

Can a small business run A/B tests?
It can, but in a small sample only large effects can be detected. Big changes such as the offer, the price or the onboarding flow should be tested; the result should be read against retention or total sales, not clicks.

Why did Netflix stop disclosing membership numbers?
In its April 2024 letter the company explained that, because of new revenue lines such as advertising and different price tiers, the business impact of each membership had come to vary widely, so from the first quarter of 2025 it would not report quarterly membership numbers and revenue per membership, and would announce major milestones instead. The most recently announced milestone is 325 million paid memberships.

Sources

The Netflix reports and letters (1-8) are company statements.

  1. Netflix, Inc. Form 10-K, fiscal year ended 31 December 2018. SEC EDGAR, 29 January 2019.
  2. Netflix, Inc. Form 10-K, fiscal year ended 31 December 2024. SEC EDGAR, 27 January 2025.
  3. Netflix, Inc. Form 10-K, fiscal year ended 31 December 2025. SEC EDGAR, 23 January 2026.
  4. Netflix, Inc. Fourth quarter 2018 shareholder letter (exhibit to Form 8-K). SEC EDGAR, 17 January 2019.
  5. Netflix, Inc. First quarter 2024 shareholder letter (exhibit to Form 8-K). SEC EDGAR, 18 April 2024.
  6. Netflix, Inc. Third quarter 2025 shareholder letter (exhibit to Form 8-K). SEC EDGAR, 21 October 2025.
  7. Netflix, Inc. Fourth quarter 2025 shareholder letter (exhibit to Form 8-K). SEC EDGAR, 20 January 2026.
  8. Netflix, Inc. Third quarter 2026 shareholder letter (exhibit to Form 8-K). SEC EDGAR, 20 October 2026.
  9. Gomez-Uribe, C. A. and Hunt, N. The Netflix Recommender System: Algorithms, Business Value, and Innovation. ACM Transactions on Management Information Systems, vol. 6, no. 4, article 13, December 2015. Authors are Netflix employees.
  10. Amatriain, X. Mining Large Streams of User Data for Personalized Recommendations. SIGKDD Explorations, vol. 14, no. 2. Author is a Netflix employee.
  11. Kong, E. Y., Deshpande, A., Yan, B. et al. Multimedia Asset Personalization via Multimodal Embeddings at Netflix. arXiv preprint 2608.18322, 18 August 2026. Authors are Netflix employees; unreviewed preprint.
  12. Li, Y., Sehgal, S., Rao, A., Houthooft, R. et al. GenRec: An LLM-Backed Recommendation Ranker at Netflix. arXiv preprint 2608.10257, version 2, 21 August 2026. Authors are Netflix employees; unreviewed preprint.
  13. Jannach, D. and Jugovac, M. Measuring the Business Value of Recommender Systems. ACM Transactions on Management Information Systems, vol. 10, no. 4, article 1, December 2019 (author copy, arXiv 1908.08328).
  14. Chaney, A. J. B., Stewart, B. M. and Engelhardt, B. E. How Algorithmic Confounding in Recommendation Systems Increases Homogeneity and Decreases Utility. arXiv 1710.11214, version 2, 27 November 2018.
  15. Scheibehenne, B., Greifeneder, R. and Todd, P. M. Can There Ever Be Too Many Options? A Meta-Analytic Review of Choice Overload. Journal of Consumer Research, 2010 (author copy).
  16. Kohavi, R., Longbotham, R., Sommerfield, D. and Henne, R. M. Controlled Experiments on the Web: Survey and Practical Guide. Data Mining and Knowledge Discovery, vol. 18, 2009, pp. 140-181.

Table 2 and Table 3 are CEOtudent analysis; Table 4 and the weekly checklist are a CEOtudent editorial framework. Table 1 reports the figures printed in company reports; the 1,000-subscriber newsletter example is hypothetical. Unverifiable claims removed from the old version: that Netflix acquired “a simulation technique of its own”; that the algorithm scans the movie, generates a personalized cover and adds a logo; the cover examples shown to viewers of action and emotional series; that member data was entrusted to Synamedia’s Credentials Sharing Insight software; that the most important reason for the rise was marketing intelligence. Corrected figure: “8.8 million new members in 2018” (fourth quarter only). Not used because they could not be opened: the 2017 artwork post and the foundation model post on the Netflix technology blog, the generative AI guidance for production partners, the “10 times faster” visual effects remark attributed to an earnings call, and the 10.06% value often cited for the Netflix Prize.


This content was compiled with the support of AI following in-depth research, then written and prepared for publication by the CEOtudent editorial team.

This post is also available in: Türkçe

Benzer içerikler