Nearly everything written about how TikTok’s algorithm works describes a system that no longer serves American users. The US recommendation model was licensed from ByteDance, retrained from the ground up on US-only data, and completed its transition during the second quarter of 2026.
The interface looks identical. The machinery behind it is not the same machinery, and the practical consequences are more specific than most coverage has managed.
Key Takeaways
- The US For You algorithm was retrained on US-only data and now runs on Oracle Cloud.
- Cross-border trend spillover is the clearest casualty of a narrower training set.
- Core ranking signals did not change: completion and watch time first, then saves and shares.
- Likes rank last among engagement signals, which makes bought likes close to worthless.
- Follower-first distribution means inactive purchased followers damage the initial test.
What Happened
The transfer of TikTok’s US operations closed on 22 January 2026, creating TikTok USDS Joint Venture LLC. Oracle, Silver Lake and Abu Dhabi’s MGX lead the investor group, existing ByteDance investors hold a further tranche, and ByteDance retains 19.9%, deliberately below the statutory ceiling.
The algorithm was the hard part of the negotiation, and the resolution was neither a sale nor a handover. Oracle received a licensed copy and rebuilt it, because the 2024 divest-or-ban legislation prohibits cooperation between ByteDance and the new owners on the operation of a content recommendation algorithm. Retraining was not a design preference. It was a legal requirement.
Oracle Cloud now hosts the US recommendation model, its training pipeline and the content moderation stack, with a majority-American board overseeing data protection, moderation, software assurance and the retraining programme. Roughly 170 million US users sit on the other side of it.
What a Narrower Training Set Actually Does
A recommendation model reflects the data it learned from, so the consequential change is not political. It is statistical.
The clearest effect is on trend spillover. TikTok’s global model let a sound or format gain traction in Southeast Asia or Latin America and arrive in American feeds carrying momentum it had already earned elsewhere. A US-only model has no visibility into that. Trends that would have crossed over now surface later, differently, or not at all.
Niche categories feel this most. Content that reached critical mass by aggregating small pockets of interest across many countries no longer has those pockets to draw on, so discovery cycles in specialised verticals run slower.
Forrester’s analysis of the divestiture put the expectation plainly, that retraining on US data means the experience will feel different and users will notice. The firm’s assessment also flagged the open question nobody can yet answer, which is how far US TikTok diverges from the international product now that two models optimise for two audiences.
The Signals That Did Not Change
This is the reassuring half, and it is where planning should sit.
Through the transition and after it, the ranking hierarchy held. Completion rate and watch time carry the most weight. Saves and shares come next. Likes rank below both. Follower count is not a ranking input, which is the structural feature that made TikTok distinctive in the first place and survived the rebuild intact.
Retraining changed which content the model has learned to associate with those outcomes. It did not change what the model is optimising for. A video that people finish and send to someone performs under the new system for the same reason it performed under the old one.
Why Purchased Engagement Fails Twice
Services selling TikTok followers, likes and views advertise themselves as working with the algorithm. The ordering above explains why they cannot.
Likes are the weakest of the engagement signals, so buying them purchases the least valuable input available. Nothing purchasable touches completion rate or watch time, because those require a person to actually watch.
Follower-first distribution makes it worse rather than neutral. New videos are tested on existing followers before wider release, which means an account padded with accounts that never watch anything is handing the system a test audience guaranteed to produce a poor completion rate. The purchased followers actively suppress the distribution the account was buying them to get.
There is a policy layer on top. TikTok’s Integrity and Authenticity guidelines prohibit accounts that mislead or manipulate the platform, and separately prohibit the trade of services that artificially boost engagement or trick the recommendation system. In the US, buying fake indicators of social media influence also violates the FTC’s rule at 16 CFR Part 465. The enforcement mechanics behind that are worth understanding before rather than after.
What TikTok Does Not Do
One claim circulates widely enough to need correcting: that TikTok reads facial reactions through the camera to gauge responses and rank content.
It does not. The signals TikTok uses are behavioural and documented: whether you finished the video, replayed it, saved it, shared it, followed the creator, or scrolled past. Those are recorded interactions with the app, not observations of you. The distinction matters because the fabricated version makes the system sound both more sophisticated and more invasive than it is, and it displaces the real explanation, which is that completion rate does most of the work.
Similar caution applies to claims about AI detecting emotional tone in audio to predict virality. Machine learning is genuinely central to how TikTok ranks content. The specific mechanisms described in most articles about it are invented.
Planning Around the Uncertainty
Four practical positions, given a model that is still settling.
Hold your benchmarks loosely. Performance data from before January 2026 describes a different system, and comparisons across that boundary carry noise that looks like signal.
Stop importing trends from international TikTok on the assumption they will cross over. That pipeline is narrower than it was, and formats that look inevitable elsewhere may simply not arrive.
Optimise the first three seconds and the last one. Completion rate remains the dominant signal, and it is the one thing entirely inside your control.
Treat US and international audiences as separate problems if you serve both, because two models optimising for two datasets will keep diverging.
Conclusion
The rebuild was substantial and the fundamentals survived it. What changed is the pool of behaviour the model learned from, which affects what surfaces and when rather than what gets rewarded.
Make videos people finish. Everything else in this category, including every service selling engagement, is an attempt to avoid that requirement, and the retrained model is no more susceptible to it than the original was.
Read Next
- How to Avoid a TikTok Ban: What Actually Triggers Enforcement
- Instagram Just Made Reposting Unrecommendable. Here’s What Counts as Original Now.
- X Marketing in 2026: The Algorithm Is Public, So What Are You Paying For?
