Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →A “Recommended for you” shelf is the visible result of a system predicting which items may interest you, selecting candidates, ranking them, and deciding how to display them. It is not mind-reading, and it is rarely just one formula. The signals and priorities differ by service, and public explanations describe only part of each company’s system.
What does a recommendation system do?
At a basic level, a recommendation system uses information about items and people’s past interactions to predict what may be relevant. Google for Developers describes recommendation models as finding similarities between items and interactions. That can surface something a person might not have thought to search for.
As an Amazon Associate I earn from qualifying purchases.
Some recommendations are personalized to a person, such as items on a homepage. Others are anchored to a particular item, such as related titles or videos. Services can combine both approaches, but they are not interchangeable: one asks what might suit this person, while the other asks what is related to this item.
Many systems can be understood through three broad stages. This is a useful general model, not a claim that every service uses the same architecture.
#1 Best Overall
1. Candidate generation
The system narrows a large catalog to a manageable set of possible items. Google’s instructional example says a YouTube candidate generator can reduce billions of videos to hundreds or thousands.
2. Scoring and ranking
A model estimates how suitable each candidate may be and orders them. The system can then select a smaller group for the page or shelf.
3. Re-ranking and constraints
Before display, additional rules or goals can change the order or remove items. Google’s overview gives examples such as excluding explicitly disliked items, emphasizing newer items, and supporting diversity or fairness. It notes, “Finally, the system must take into account additional constraints for the final ranking.”
Rank #2
What signals can shape a recommendation?
A signal is information the system can use when making a prediction. Different services disclose different signals, and a signal’s presence does not mean it has the same weight for every person or situation.
Netflix’s disclosed signals
Netflix says it may use a member’s interactions with the service, similarities with other members, title attributes, time of day, language preferences, device, and viewing duration. Its personalization can affect which rows appear, which titles are in a row, and the order of titles. Netflix also says newer engagement can outweigh older engagement, and that viewing, completion, and rating feedback update its predictions.
YouTube’s disclosed signals
YouTube lists watch and search history, subscriptions, likes, dislikes, and “Not interested” feedback among the signals that can shape recommendations. Its 2021 explanation also discusses clicks, watch time, surveys, sharing, likes, and dislikes, noting that the importance of signals can vary by viewer.
YouTube Help currently describes its system as learning from more than 80 billion pieces of information it calls signals. This is YouTube’s own description, not an independently audited count.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errorsWhy does the shelf—and not just the item order—change?
Personalization can shape the page itself. A service may choose which shelves to show, what titles or videos go in them, and their order. Netflix describes all three interface decisions as possible parts of personalization.
That distinction matters: a recommendation is not only a list of scores. The service has to turn predictions into a page, deciding which groups are relevant and how to arrange them. A 2015 Netflix research paper described multiple rankers—including Top-N, trending, Continue Watching, and video similarity—combined by a page-generation algorithm that considered row relevance and page diversity. That paper documents Netflix’s approach at that time; it should not be taken as a blueprint for its present system or for other services.
Rank #4
What are platforms trying to optimize?
Engagement is useful to a recommendation system, but it is not identical to satisfaction. A click does not prove that someone found a video valuable, and longer viewing is not automatically better in every context.
YouTube says clicks alone did not show whether viewers actually watched, so it added watch time to recommendations in 2012. Its 2021 post reports a 20% drop in views after that change. The same post describes user surveys intended to estimate “valued watchtime,” as well as sharing and feedback signals. For news and information content, YouTube says it considers information quality and context, and uses human evaluations and classifiers to identify authoritative or borderline content. These statements describe the company’s stated practices and priorities; they do not independently establish the effect of each measure.
Recommended Free Tools
YouTube VP of Engineering Cristos Goodrow wrote in 2021: “Our recommendation system is built on the simple principle of helping people find the videos they want to watch and that will give them value.” He also wrote: “That’s why providing more transparency isn’t as simple as listing a formula for recommendations, but involves understanding all the data that feeds into our system.”
Google for Developers says its page’s figures are 40% of app installs on Google Play coming from recommendations and 60% of YouTube watch time coming from recommendations. The page was last updated on August 25, 2025, but does not state the underlying measurement period, so these should not be read as newly measured 2025 results.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.Why can’t a company explain every recommendation with one formula?
A service may use multiple models, signals, stages, and interface rules, and those may change over time. Public descriptions offer useful evidence about what a company says it does, but they do not expose every proprietary model or prove how every recommendation was produced.
There is also a separate design question: explaining a recommendation to a user is not the same as generating the recommendation. A 2018 survey by Yongfeng Zhang and Xu Chen describes explainable-recommendation research as producing recommendations alongside explanations, and frames possible explanations through what, when, who, where, and why. A system can select an item and still face a distinct challenge in offering a clear, accurate account of why it appeared.
What controls can users use?
Controls vary by service. They can affect the information available to the system, but do not necessarily reveal or directly control every ranking decision.
Netflix
Netflix says members can search its catalog and may optionally choose favorite titles when getting started. Those options provide ways to find content or give the service initial preference information.
YouTube
YouTube says users can pause, edit, or delete search and watch history. Because its stated signal list includes both histories, managing them can change what information is available for future recommendations.
Quick Recap
What to keep in mind when you see a recommendation
- It is a prediction based on signals, not evidence that a service knows your thoughts.
- Similarity can refer to patterns among users, attributes of items, or a relationship to something you are currently viewing.
- The placement and order of shelves can be personalized along with the items in them.
- Signals such as clicks or viewing time are proxies for interest; YouTube says it supplements engagement signals with surveys and quality measures.
- Company explanations are partial and service-specific. A published account of one platform should not be assumed to describe another.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
