The first thing to understand is that "the algorithm" isn't one thing. It's at least seven nested systems making decisions in cascade · and the part most posts on LinkedIn obsess over is the second-to-last step. By the time a video gets to the part everyone writes about, the harder calls have already been made.
Over four months we sat down with eleven people who've worked inside it: former product managers, two engineers who built parts of the recommendation system, six creators with 10M+ followers, and three growth heads at brands you've definitely seen on a For You page this month. Most asked not to be named. What follows is a synthesis of what they all agreed on · including the parts that contradict the popular narrative.
This is going to take seventeen minutes and it'll change how you brief your next campaign. If you're going to read one of our pieces this year, make it this one.
Section 01The four signals that actually matter.
If you've read anything about the TikTok algorithm, you've seen a list of "ranking signals" · usually pulled from a leaked engineering doc from 2022. That list is real but it's also misleading: it tells you what the system can see, not what actually moves the needle.
The four signals that actually determine whether your post gets distribution, in order of weight:
- Completion rate, but only when the video is under 21 seconds.
- Re-watch rate · how many viewers loop the video at least once.
- Comment-to-view ratio in the first 90 minutes.
- Share-to-view ratio across the first 24 hours.
You'll notice likes are not on this list. Three of the engineers we spoke to confirmed that likes are heavily de-weighted in the actual ranking model · they're useful as a counter-fraud signal but they barely move distribution. This contradicts about half the LinkedIn content on this topic.
Likes are vanity. Comments and saves are work · and the system rewards work.Anonymous, former TikTok PM
Why completion rate is more nuanced than you think.
Everyone tells you to hook viewers in the first three seconds. That's right but it's not the whole story. The completion rate metric is computed in weighted segments: the first 3 seconds count for ~40% of the score, seconds 3–10 count for ~35%, and the remaining time scales down sharply.
This is why a 12-second video with 90% completion will out-distribute a 60-second video with 60% completion, even if both have the same total watch time. The system isn't optimising for time-on-platform per video · it's optimising for perceived quality density.
We tested this on 14 brand pages.
Across 600 posts, we cut every video at the moment retention dropped below 50%. Average reach went up 2.7×. Eleven of the 14 brands saw their best month ever. Two saw no change. One declined · turns out their content was retention-strong throughout, and we'd cut the wrong frames.
Section 02Why "watch time" is the most misread metric in social.
Watch time appears in every TikTok dashboard and every algorithm explainer. It is also, by itself, almost useless as a creative metric. Here's why.
The system computes something called relative watch time: your video's watch time compared to videos of similar length, posted in the last 48 hours, in similar categories. A 17-second beauty TikTok with 14 seconds of average watch time will outperform a 30-second one with 22 seconds · because the comparison set is different.
This means you can have worse absolute numbers and better distribution if you change format. Brands routinely make the wrong call here because their dashboard shows raw watch time, not relative.
The 21-second cliff.
Engineers we spoke to all referenced a "21-second cliff" · a hard internal threshold above which videos are evaluated against a different cohort with much higher retention requirements. Below 21 seconds, you're competing against meme replies and casual posts. Above it, you're competing against documentary edits and scripted content.
For most brands, that's the wrong fight. The math says: if you're not absolutely confident the longer cut earns its length, cut it under 21 seconds. We checked our own data: 78% of our top-performing brand work this year clocks in between 14 and 19 seconds.
Section 03The cold-start window: 90 minutes that decide everything.
When you upload a video, TikTok runs a "cold start" experiment. It pushes your video to a small, deliberately-mixed seed audience · typically 200–500 viewers · and watches what happens in the first 90 minutes.
That's where comments matter most. Comments in the first 90 minutes are weighted approximately 5× higher than comments after 24 hours. This is why "first comment wins" is real, and why brand pages that respond to early commenters within minutes consistently out-distribute brand pages that don't.
It's also why prompts matter. A video that ends with "what's your take?" will out-distribute the same video without that line, almost every time. The hook gets you in; the prompt keeps the experiment going.
We don't write captions last anymore. We write the comment-bait line first, then build the video around it.Mia Chen · creative director · attn:seeker
Section 04What this actually means for brand pages.
If you've read this far, you can probably already see the implications. Three things, in order of importance:
- Stop publishing on a "content calendar." Publish when you can publish well, with a comment-bait line ready and a team that can respond inside 90 minutes. One great post a week beats five mediocre ones every time.
- Cut harder. If a video isn't earning its length, cut it. The 21-second cliff is real and most brand work belongs under it.
- Reply early or don't bother. Comment replies in the first 90 minutes are worth more than 24 hours of replies later. If your social team can't be on for the first 90 minutes after publish, schedule your publish for when they can.
None of this is hacky. It's just understanding what the system is actually rewarding · which turns out to be a pretty fair set of behaviours: quality, attention, conversation. The work is in being honest about whether your content earns those.
If you want to dig deeper, we go into the data live with three engineers and two creators on next week's Stay Curious podcast. And if you want our full benchmark dataset (600 posts, 14 brands, 12 weeks of A/B testing), drop us a note · we share it with newsletter subscribers on request.
Stanley