How Does the TikTok Algorithm Work? A Builder’s View

The TikTok algorithm is a recommendation system that picks the next video for each viewer from a large pool of clips, instead of showing a list of accounts the viewer follows. TikTok has published an explainer on For You recommendations. It groups the inputs into what the viewer does, what the video contains, and the device and account settings. This post separates what TikTok documents from what is standard practice in recommender systems, then turns it into something a founder can build and tune.

If you are planning your own short video network, the feed is the part to understand first. A white-label TikTok clone ships with a working feed, but you still decide how it behaves. The sections below cover the signals, the test-and-widen loop, cold start, the levers worth keeping, and the mistakes that sink small platforms.

What a "for you" feed is, and why it replaced the follow list

A classic social feed is a filter on a graph. You follow accounts, and the feed shows what those accounts posted, newest first or lightly ranked. A "for you" feed reverses the order of operations. The system looks at all eligible videos, predicts which ones this viewer is likely to want, and fills the screen with them. The follow graph becomes one input among several.

That design has a business consequence. In a follow-list product, a new creator with no followers is invisible, and growth depends on being discovered somewhere else. In a recommendation product, the first upload is already in the running. TikTok has said, in its published explainer that neither follower count nor whether an account has had previous high-performing videos is a direct factor in its recommendation system. Instagram describes separate ranking systems for each part of its app, Reels included, and YouTube and Snapchat publish similar explainers (see the Instagram ranking explainer).

For a creator, this is the promise that makes posting worthwhile. For an operator, it is the reason the feed is the product. If the feed is dull, creators leave because nobody sees their work, and viewers leave because nothing surprises them.

The two feeds are different jobs

Most short video apps keep both. The For You feed is for discovery and is ranked. The Following feed is for loyalty and is mostly a chronological or lightly sorted view of accounts the viewer chose. Keep them separate in your own app. Mixing them confuses creators, who cannot tell whether a view came from a follower or a stranger, and it weakens your own analytics.

The signals a feed like this reads

TikTok's published explainer on For You recommendations, first published in June 2020, describes three families of input. What follows is TikTok's own description, not something an outsider can test. Read the current explainer on TikTok's own site before you rely on any detail, because the company can revise it at any time.

Two further points, again as the explainer is quoted, matter for builders. First, signals are weighted. TikTok has described an example in which finishing a longer video from start to end is a strong indicator of interest, while the viewer and creator being in the same country is a weak one. Second, the explainer describes new users as starting from the interests they pick at sign-up, with the system adjusting as they show what they do not want to see.

The explainer is also quoted as describing filtering: TikTok has said it does not recommend duplicated content, content the viewer has already seen, or content it considers spam, and that a video that was just uploaded or is under review may not yet be eligible for recommendation. That second sentence is worth remembering when you design moderation: eligibility is a gate that sits in front of ranking.

What is documented and what is not

TikTok does not publish the model, the exact weights, the number of stages or the size of any test audience. Anything on the web that claims to give a specific watch-time percentage for "going viral" is a guess, not a company statement. Here is how we label the rest of this post.

• Attributed to TikTok's published explainer: the three signal families, weighting of strong over weak signals, follower count not being a direct factor, diversity in the feed, filtering of duplicates and spam, and eligibility for new or reviewed videos.
• General recommender practice: candidate generation followed by ranking, exploration of new items, and the use of completion and rewatch as quality signals. YouTube, Instagram and Snapchat each describe similar ideas in their own help pages.
• How our product behaves: the specific test-and-widen loop, the weight sliders and the admin levers described below. These are facts about our software, not about TikTok.

Other companies publish comparable explainers, and reading them side by side is the fastest way to see what is common. YouTube says its system looks at viewing behavior, likes, dislikes, subscriptions and feedback, including satisfaction surveys (how YouTube recommendations work). Instagram says Reels ranking leans on surveyed entertainment value and predicts reshares, full watches, likes and visits to the audio page. Snapchat lists country, language, age, engagement history and hidden or reported content for Spotlight (Spotlight ranking). The shared pattern is behavior first, content description second, context last.

Signal table: what each signal tells the system, and how an operator can weight it

The table below is a builder's reading of signals in general. The right-hand column is advice for operators of their own app, not a description of TikTok's settings. What TikTok itself uses is limited to the families quoted above.

The loop: candidate pool, ranking, small test audience, widening

This section describes the general design of a short video feed, and what our product does. It is not a claim about TikTok's internals.

A feed of this kind runs in two broad steps. The first picks a candidate pool: a few hundred clips out of everything eligible, chosen cheaply by interest match, freshness, region and language. The second ranks that pool for the viewer with a model that predicts how likely they are to watch, finish, replay, like or share each clip. The top of the ranked list goes on screen. Each swipe feeds new behavior back in, so the next batch is better informed.

Test, read the response, widen

New videos have no history, so the system has to create some. The usual answer is a staged release:

1. The clip passes eligibility checks (moderation, duplicate and spam filters, rights flags).
2. It is shown to a small group whose interests plausibly match the topic, sound and tags.
3. The system reads the response: how many finished it, replayed it, shared it, followed the creator, or swiped away in the first seconds.
4. Strong response widens the audience to the next ring of similar viewers. Weak response stops the spread.
5. The process repeats until the response weakens or the clip stops being new.

Our own product works this way. The ranking engine shows a new clip to a limited group first and reads completion, replays and shares, and the operator can tune how generous the test is. The effect is that a first-time creator can be seen and that one lucky upload cannot flood the feed on the strength of a single metric.

Why completion and rewatch usually beat raw likes

A like is cheap. A viewer can tap it on reflex, or tap it for a creator they already know. Finishing a clip, or watching it twice, costs the viewer time, and time is the one thing every viewer has a fixed amount of. That is why most vertical feeds lean on completion, rewatch and share as their strongest quality evidence, and fits TikTok's quoted example of a strong signal, finishing a longer video. It also explains why a ten-second clip with a strong ending can outrun a three-minute clip that loses people at second twenty.

Cold start: a new user and a new creator

A recommender learns from behavior, and on day one there is none. Cold start has two sides, and each needs its own plan.

A new viewer

TikTok's explainer is quoted as saying new users start from the interests they express at sign-up. Our product does the same: a short interest picker seeds the feed, then each swipe teaches the ranking. A practical checklist for your own launch:

• Offer a short interest and language picker at sign-up, and make it skippable.
• Seed the first screens with clips that are broadly liked inside the chosen categories, not random uploads.
• Keep a browse tab organized by category, hashtag and creator, so viewers can hunt for content instead of waiting for the feed to learn.
• Use region and language as soft filters until you have behavior to replace them.
• Mix a share of trending clips with new creators' uploads from the first session, so you gather signal on fresh content straight away.

The first session, step by step

For a new viewer, the first five minutes decide whether they come back. This is the sequence we recommend, and our product follows it in spirit:

1. Ask for language and a few interests. Two or three taps, skippable.
2. Build the first batch from broad appeal inside those interests. Use clips with strong completion across many viewers, not random uploads.
3. Insert one or two fresh clips from new creators in every batch. This is the exploration share, and it is how new uploads get a test audience.
4. Read the first ten swipes. Completion, replays and quick skips begin to separate the viewer's real tastes from the picks they made.
5. Shift weight from declared interests to observed behavior. By the end of the session, behavior should lead.
6. Keep a small share of surprise. Diversity stops the feed narrowing too early, and it gathers signal on categories the viewer did not pick.

A new creator

A creator with no audience depends entirely on the test step. Make it visible. Show the creator views, completion and where viewers dropped off within a day, so a weak result teaches something. If new creators see three views and no explanation, they stop posting, and your supply dries up. Our guide to attracting creators to a new short video platform covers the supply side, and the cold start problem post explains why viewers and creators have to be grown together.

Controls an operator should keep

A feed you own is a set of dials. A rented feed is a black box. The table lists the levers worth keeping, what each does, and the risk of misusing it. Our product exposes most of these in the admin panel: adjustable weighting between recency, engagement and creator diversity, curated categories, hashtags and featured trends, region-specific recommendations and region-wise availability.

The discipline that matters more than any single lever is patience. Change one weighting, give it a few days of traffic, and compare watch time and return visits before touching anything else. If one creator dominates, raise diversity. If viewers drop off quickly, check whether onboarding interests are steering them toward content you do not have.

Keep a not-interested action in the app from the first release. TikTok's Creator Rewards terms refer to viewers marking videos as not interested, which confirms that this signal exists in its system. It is also the cleanest negative label you can collect.

Operator tuning checklist

1. Write down the single outcome you want to improve, such as return visits on day seven or median watch time per session.
2. Record a baseline over at least a week of normal traffic before you change anything.
3. Change one weighting, one category rule or one region setting.
4. Leave it for several days of traffic, and do not compare a weekday with a weekend.
5. Look at the outcome, then at side effects: creator concentration, new-creator reach, reports per thousand views.
6. Keep the change, reverse it, or adjust by half the original step.
7. Log the date, the setting and the result in a place your team can search.
8. Tell creators in plain language what changed when it affects their reach.

What our product ships for this

Our short video platform includes a behavior-based For You feed. Ranking starts from the interests chosen at sign-up, then adapts to watch time, replays, likes and shares. The weighting between recency, engagement and creator diversity can be adjusted, and region-specific recommendations are part of the admin controls. Geography-based trending sections are available with our platform. Moderation runs before distribution, with automated flags and a review queue. The full list is on the page for TikTok clone features, including the feed and discovery tools.

We do not claim the engine matches TikTok's. A company with thousands of engineers and years of data trains a different system. What you get is a working, tunable feed from day one, which is the right starting point for a regional or niche network, where owning the rules matters more than matching a global model.

Glossary: the terms used in this post

Mistakes small platforms make with feeds

1. Ranking before there is content. A model cannot rank six clips. Until each viewer has a few hundred plausible candidates, use editorial picks, categories and trending lists, and turn on personalization gradually.
2. No moderation gate before distribution. If a clip reaches the feed before anyone checks it, the harm is done by the time a report arrives. TikTok has said new or under-review videos may not be eligible yet. Build the same gate, and see our guide to content moderation for a short video app.
3. Opaque rules for creators. Creators will invent theories if you say nothing. Publish a plain description of what the feed looks at, in the style of the explainers above, and update it when you change weights.
4. Retuning every day. Daily changes make results impossible to read. Change one thing, wait, measure.
5. Letting one metric run the feed. Completion alone favors loops, likes alone favor bait, and followers alone bring back the follow-list problem. Blend signals and watch return visits as the check.
6. Copying a global feed for a local market. A single city or language needs different diversity and freshness settings than a global network.

What to decide next

Pick your audience before your ranking. A narrow audience, such as one language, one city or one hobby, lets a few hundred creators make a feed feel full, and a feed that feels full is when ranking starts to matter. Then decide which levers from the table you will expose to your team, and who is allowed to change them.

If you want a working feed without building a ranking team, read the step-by-step plan in how to start a short video app, and look at the TikTok clone script to see the product. For the cost side, the TikTok clone development cost page separates the one-time price from the running costs of video storage, delivery and moderation.