Something broke in July. Not visibly, not with an announcement banner, but if you run a branded TikTok account you probably felt it. Videos that would’ve cleared 40,000 views six months ago are stalling at 2,000. Formats that used to work like clockwork just… Don’t.
There’s a reason. TikTok’s US algorithm is mid-retraining, and Oracle is the one holding the wheel.
Since the finalized ownership deal put a new US joint venture in charge of the platform, Oracle has been rebuilding the recommendation engine from the ground up, training it exclusively on US user data rather than inheriting ByteDance’s global model. That’s not a small technical footnote. It means the signals that used to predict virality, the sounds, the hashtags, the posting windows, are being relearned in real time. Nobody outside Oracle’s data team knows exactly what the new model rewards yet. But the early patterns are starting to show, and they matter for exactly the kind of brands reading this.
What’s Actually Changing Under the Hood
The short version: distribution is in flux, and flux favors specificity.
Generic national content, the kind built for maximum reach across the widest possible audience, is getting buried more often than it used to. Broad lifestyle clips, generic “top 10” listicles, anything trying to speak to everyone at once. That stuff performed fine under the old model because ByteDance’s global engine was tuned for scale above all else.
Oracle’s version looks different. Early testing from marketers tracking watch-time-to-completion ratios suggests the new model is leaning harder into relevance signals tied to location, niche interest, and repeat engagement from smaller, denser audience pockets. Hootsuite’s 2026 breakdown of the algorithm notes that search behavior and saves are now weighted more heavily than raw view counts, a shift that rewards content built for a specific searcher rather than a passive scroller.
That’s a meaningful pivot. It means the accounts pulling ahead right now aren’t the ones chasing trend sounds. They’re the ones publishing narrow, useful, locally-flavored content that a specific person is actively searching for.
Here’s where it gets interesting for anyone doing regional or state-specific content work. Geo-targeted explainers, the kind that answer a very specific question for a very specific audience, are quietly outperforming the generic stuff across multiple verticals right now. State-level explainers, like coverage of the online casino market according to Metrotimes, are exactly the kind of geo-targeted content the reset seems to reward. It’s not flashy. It doesn’t chase a trending sound. But it answers a real question a real person in a real place is typing into a search bar, and that’s precisely the behavior Oracle’s retrained model appears built to surface.
Gambling involves risk. If gambling coverage or advertising in your feed is becoming a concern, resources like BeGambleAware exist for a reason, and it’s worth knowing they’re there.
Why Follower Count Stopped Being the Metric That Matters
Brands that built their entire strategy around follower count are the ones getting hit hardest by this transition. That number was always a vanity signal more than a functional one, but under the old model it at least correlated loosely with reach. Under Oracle’s retrained system, it barely correlates at all.
Watch time. Completion rate. Shares to private chat, not public reposts. Saves. Those are the signals doing the heavy lifting now. A video with 800 views and a 92% completion rate is outperforming a video with 40,000 views and a 12% completion rate, and that gap is only widening as the model matures.
This isn’t a totally new idea. Our own breakdown of why follower counts no longer predict reach covered the early signs of this shift back before the Oracle transition even finished rolling out. What’s changed is the scale of the effect. It used to be a soft trend. Now it’s the entire operating logic of the platform.
Brands still optimizing for follower growth as the north star metric are optimizing for a number the algorithm barely looks at anymore.
The Format Shift Nobody’s Talking About Enough
Length matters differently now too. For a while, the conventional wisdom was that shorter clips under 15 seconds always won. That’s no longer reliably true.
Retention-weighted models tend to favor content that holds attention across its full runtime, regardless of whether that runtime is 12 seconds or 90. A tight, information-dense 60-second explainer that gets watched to completion can outperform a punchy 8-second clip that gets skipped after three seconds. The metric isn’t length. It’s whether the viewer stays.
That rewards a specific kind of content discipline. No padding. No slow builds. Get to the point, deliver the payoff, and don’t waste the viewer’s attention on setup they didn’t ask for.
Brands producing tutorial content, explainer content, or anything with a clear informational payoff are adapting faster than brands built around aesthetic or lifestyle content, because informational formats are naturally structured around delivering value quickly.
What Regional and Niche Brands Should Actually Do Right Now
The instinct during any algorithm shakeup is to panic and start posting more, faster, hoping volume compensates for uncertainty. That’s usually the wrong move.
Better instinct: get narrower. Lean into whatever specific, local, or niche angle your brand already owns, rather than trying to widen your appeal to compensate for lower reach. The TechCrunch coverage of the finalized US TikTok deal makes clear this is a US-specific data retraining effort, not a global one, which means American regional and state-level content has an unusually good structural position right now. The algorithm is learning American user behavior in isolation for the first time. Content that’s distinctly, specifically American in its framing has a training-data advantage that won’t last forever.
Three things worth doing this month if you manage a brand account:
None of this is permanent. Algorithms retrain, then stabilize, then get nudged again. But right now, in September 2026, the window favors the specific over the broad, and brands that move on that early tend to keep the advantage even after the model settles.
Frequently Asked Questions
Why is TikTok’s algorithm changing so much this year? The platform’s US operations moved to a new joint venture, and Oracle is retraining the recommendation engine using only US user data instead of inheriting the previous global model. That retraining is still ongoing, which is why performance has felt inconsistent since mid-2026.
Does posting more often help during an algorithm transition? Not reliably. Volume without relevance tends to get buried under a retention-weighted model. Fewer, more targeted posts that hold attention to completion generally outperform a higher volume of generic content right now.
Is follower count still a useful growth metric? It’s a weak signal at best under the current model. Watch time, completion rate, and saves matter far more for actual distribution. Brands still chasing follower count as a primary KPI are measuring the wrong thing.
Should regional or niche brands change their content strategy right now? Most should lean further into their existing niche rather than broadening it. Geo-specific and narrowly targeted content appears to be getting rewarded under the retrained model, at least based on early performance patterns since the transition began.
How long will this transition period last? There’s no official timeline, but algorithm retraining of this scale typically takes several months to fully stabilize. Brands adapting early to the new signals are likely to keep whatever advantage they build once things settle.
