Dave LumAI here. I have been feeding TikTok videos on a fairly regular basis, and lately the algorithm has developed a fascinating new response to my efforts: absolutely nothing. Not bad views. Not disappointing views. Zero. The sort of number normally reserved for my chances of understanding why social-media algorithms do anything.
This is weird.
What makes it interesting is not simply that some videos performed badly. Badly would be 37 views and one pity-like from somebody who may have sneezed on the screen.
These are clusters of literal zeroes.
Then suddenly one video gets hundreds of views.
Then zero.
Zero.
Zero.
Another one escapes.
Then back to the void.
That is very different from the normal pattern of “the audience did not particularly care for this one.” It looks much more like certain videos are simply not getting normal distribution in the first place.
So let us tackle the important questions before the algorithm notices we are talking about it.
Did something change between September 15 and 18?
Maybe, but I cannot find evidence of a big public TikTok announcement saying, “Hello creators, on September 17 we changed the recommendation algorithm and good luck with that.”
The most interesting official breadcrumb is a September 15 TikTok Newsroom post about AI-generated content. TikTok said it was testing improved detection systems aimed at accounts dedicated to posting AI-generated spam.
That immediately got my attention because I publish plenty of AI-assisted visual content.
But there is an important wrinkle.
TikTok described those tests as particularly focused on areas such as politics and current events, financial advice, and medical information. Also, versions of this initiative had appeared in other regional TikTok announcements earlier in the summer.
So September 15 is not a smoking gun.
It is more like a suspicious-looking toaster standing near the crime scene.
There is also a larger 2026 change happening in the United States. The TikTok USDS Joint Venture says it now retrains, tests, and updates the U.S. recommendation algorithm using U.S. user data. That does not explain a September 17 cliff either, but it does mean the machinery underneath American TikTok is genuinely being worked on.
So my answer is:
I found no confirmed algorithm change specifically between September 15 and 18, but there absolutely has been ongoing work around recommendations, AI detection, moderation, and the U.S. algorithm during 2026.
And the timing is interesting enough that I would keep my eyebrow raised.
Does the desktop app versus the web app make a difference?
Officially, I cannot find anything saying TikTok gives better distribution to a video because it arrived through one upload doorway instead of another.
TikTok supports creator tools through the app and browser, including TikTok Studio, and its own explanation of how TikTok recommends content says things such as watch behavior, interactions, content information, location, time, and device information can influence recommendations.
But there is no documented “browser penalty.”
That said, software bugs do not read documentation.
It is entirely possible for one upload path to encounter different processing, metadata handling, encoding, moderation queues, or some delightfully obscure bug involving a server named Kevin.
So I would absolutely test it.
Not because I think the algorithm secretly hates Chrome.
Because controlled experiments are considerably cheaper than sacrificing a goat.
There is another clue hiding in my video lengths
This one surprised me.
TikTok’s Community Guidelines on originality and recommendation eligibility specifically say that low-quality content can be ineligible for the For You feed, and one of the examples they give is extremely short clips.
Look back at my recording.
Several videos are only around six seconds long.
That does not mean every six-second video automatically goes directly into algorithmic witness protection. Clearly some short videos perform perfectly well.
But it gives me something concrete to test.
And concrete is much nicer than “maybe Mercury is in retrograde.”
Zero views is not the same problem as low views
This distinction matters.
If a video gets 73 views and stops, TikTok probably distributed it somewhere and the early signals were not strong enough to keep pushing it.
That happens.
It hurts a little, but we put on our grown-up pants and continue.
Zero is different.
TikTok explains in its content review and eligibility guidance that content can be reviewed when it is uploaded and can also be deemed ineligible for the For You feed.
That means the first question for a zero-view video should not necessarily be:
“Was my opening hook terrible?”
It should be:
“Did this thing ever get properly released into the recommendation system?”
That is a much more useful mystery.
Here is exactly how I would troubleshoot it
- Check the account itself. Go into TikTok Studio and run the Account Check. If TikTok thinks there is a posting or account problem, I want to know before changing twelve unrelated things.
- Open each zero-view video individually. Go into its analytics and look for any notice saying it is ineligible for the For You feed. If there is one, read the reason and appeal it when appropriate. A video can exist perfectly happily on your profile while recommendation distribution is another matter entirely.
- Label the AI content. For videos that are substantially AI-generated, use TikTok’s AI-generated content setting. TikTok’s AI-generated content guidance specifically says using the label by itself does not hurt distribution as long as the content follows the rules. That removes one unnecessary ambiguity from the experiment.
- Retire the six-second test clips for a week. I would make the next batch roughly 15 to 25 seconds, with meaningful motion throughout, a strong opening second, useful on-screen text, and perhaps narration or another obvious layer of original editing. The goal is not to appease a mystical algorithm deity. The goal is to move well away from anything a classifier might mistake for low-effort, automated, or nearly static content.
- Test the upload path scientifically. Post three fresh videos through the normal app and three comparable fresh videos through the browser or desktop workflow. Do not upload the identical file six times. Keep the topics, length, quality, and posting schedule reasonably similar. Then compare views, For You traffic, eligibility notices, and processing time after 1 hour and 24 hours.
- Post one absurdly ordinary control video. Shoot 15 seconds of something completely original with the phone camera. A tree moving. Coffee being poured. A squirrel committing whatever crime squirrels are planning today. No AI. No imported artwork. No elaborate editing. If that gets normal distribution while the generated visual videos stay at zero, we have narrowed the problem dramatically. If the ordinary phone video also gets zero, I would start looking much harder at the account or distribution pipeline.
- If zeroes continue, report the pattern instead of saying “shadowban.” Give TikTok specific post IDs, dates, screenshots, and the useful facts: public post, zero views after 24 hours, account check clean, no visible eligibility warning. “Something is broken in distribution” gives support considerably more to work with than “the algorithm hates me,” even though the second version may be emotionally more satisfying.
What I would not do
I would not start deleting and reposting everything.
I would not create three emergency replacement accounts.
I would not change the captions, hashtags, video length, upload device, posting time, music, account type, editing software, and phase of the moon simultaneously.
Because then if something works, we learn absolutely nothing.
This is where social-media troubleshooting starts looking suspiciously like software troubleshooting.
Change one variable.
Observe.
Write it down.
Repeat.
Yes, I know. We have somehow turned posting a ten-second video into a laboratory experiment.
Welcome to 2026.
The AI angle is the part I would watch most closely
TikTok is plainly paying more attention to AI-generated material.
That does not mean TikTok hates AI art.
Its own documentation talks extensively about allowing AI-generated content when it follows the rules and is appropriately disclosed.
The more interesting question is whether automated systems can reliably distinguish:
“This creator deliberately made an original AI-assisted artwork and animated it into a video.”
from:
“This account is mass-producing synthetic filler at industrial velocity.”
Those look very different to a human.
To a classifier examining metadata, posting frequency, video length, visual similarity, editing patterns, and repeated structure?
Maybe not always.
That is the hypothesis I would test before deciding my audience suddenly became 100 percent allergic to everything I make.
And apparently I am not the only person seeing the zero-view phenomenon. Other creators were reporting clusters of zero-view posts during September, including people saying their accounts showed no obvious violations.
Anecdotes are not proof.
But when enough canaries begin complaining about the coal mine, I at least check whether the coal mine has Wi-Fi.
So what do I think is actually happening?
My best read of the evidence is that this does not look like ordinary poor performance.
There are too many abrupt zeroes mixed with videos that still manage to receive hundreds of views.
That points me toward some combination of:
processing delays,
For You eligibility,
automated content classification,
AI/spam detection,
very short video formats,
or an account-level distribution problem.
The desktop-versus-web question is absolutely worth testing, but I would treat it as a possible technical bug rather than a known ranking rule.
Most importantly, I would not rewrite my entire content strategy yet.
First I want data.
If you have been seeing the same thing, comment and tell me when it started, whether your videos are AI-assisted, how long they are, and whether you uploaded through the app or browser. That might actually help us spot a pattern TikTok has not documented.
And follow @davelumai because apparently we are now doing algorithm archaeology together.
At least until the next upload gets zero views and the algorithm pretends it has never met me.
Art Prompt (Tonalism): Create a quiet autumn night landscape in the refined spirit of early American Tonalism, with a broad meadow disappearing into velvety dusk beneath a low, luminous sky. Place a sparse line of tall, slender trees along the horizon, their trunks and branches softened by blue-gray atmospheric haze. Use deep olive, muted umber, smoky violet, charcoal blue, tarnished gold, and faint silver-green, allowing neighboring tones to dissolve gently into one another rather than forming hard outlines. Let a pale moon glow through thin drifting clouds while subtle light settles across distant grass and a narrow reflective stream. Keep the composition restrained and horizontal, with abundant negative space, delicate layered brushwork, softened edges, and an almost musical rhythm between dark trees and open sky. The mood should feel hushed, introspective, slightly mysterious, and deeply serene, with the landscape appearing to emerge gradually from darkness. Museum-quality oil painting, subtle surface texture, no readable text, no logos, no recognizable people, family-friendly.
Video Prompt: Begin with a sudden gust sweeping through a dark autumn meadow, sending dry leaves spinning rapidly across the foreground as tall grasses bend in rolling waves. Reveal slender trees silhouetted against a luminous blue-gray night sky while thin clouds race past a pale moon and silver mist pours slowly between distant trunks. Let the camera glide low through the grass, rise smoothly above the field, and briefly follow a stream catching broken reflections of moonlight. Introduce quick flickers of fireflies, drifting seed heads, rippling water, and leaves crossing close to the lens to create constant layered motion without losing the quiet, painterly atmosphere. Keep the palette restrained to deep olive, muted umber, smoky violet, charcoal blue, tarnished gold, and silver-green, with softened Tonalist edges, atmospheric depth, subtle oil-painted texture, elegant natural movement, no readable text, no logos, no recognizable people, family-friendly.
Songs: “Autumn Sweater” — Yo La Tengo; “Under the Milky Way” — The Church.