There's an enormous folk literature about what recommendation algorithms want. Post at specific times. Never leave the app. Use certain words, avoid others. Reply to every comment within the first hour.
Some of this is right. A lot of it is superstition — patterns inferred from small samples by people with a strong incentive to believe they've found something.
What's actually known publicly is more limited and more useful than the folklore.
What these systems optimise for
The general shape is not secret. Platforms have published papers and documentation describing their recommender architectures at a high level, and the objectives are broadly consistent.
They predict the probability that a given user will engage with a given item, across several kinds of engagement — watch time, completion, likes, shares, comments, follows — and combine those predictions into a ranking.
Different actions carry different weights, and those weights change. A share is generally worth more than a like, because it predicts distribution. A follow is worth a great deal, because it predicts future sessions.
Crucially, the optimisation target is usually something like long-term user retention, not immediate engagement. That distinction matters, because it means content that generates engagement while making people enjoy the platform less can be down-weighted.
The prediction is personal, not global
The most common misunderstanding. People talk about "the algorithm" as though there's one ranking of content.
There isn't. The system is predicting engagement for a specific user, based on that user's history. A piece of content isn't good or bad in the system's terms; it's a good or bad match for particular people.
Which means the practical question isn't "will the algorithm like this" but "for whom is this the best thing they could see right now". Content that's an excellent match for a small, well-defined audience often performs better than content that's a mediocre match for a large one.
This is the single most useful reframing available and it contradicts most generic advice, which pushes towards broad appeal.
Where the folklore goes wrong
Posting times. Mattered a great deal in chronological feeds. Matters much less in recommendation-driven ones, where content is surfaced when the system thinks a user will engage, potentially days after posting. Still relevant on platforms with a chronological component; largely superstition elsewhere.
Hashtags. Their role has diminished substantially as content understanding improved. Systems now analyse the content itself — audio, visuals, text on screen, captions — and infer topic. Hashtags remain useful as a signal and are not the lever people treat them as.
External links kill reach. Widely believed. There's some logic to it, since a link takes the user off-platform and off-platform time doesn't serve retention. The magnitude is likely smaller than the folklore claims and varies enormously by platform.
Deleting underperforming posts. The theory that poor performers drag down an account. There's no public basis for this and it's an example of a superstition that's cost people a lot of archived work.
What genuinely does matter
Retention within the first few seconds. The most reliably confirmed factor for short video. If people leave immediately, nothing else compensates. This is why the opening has become so aggressive across every platform.
Completion rate. Shorter content completes more often, which is a large part of why everything has got shorter.
Rewatches and shares. Both are strong positive signals and both are hard to manufacture.
Consistency. Regular posting improves the system's model of your content and audience. Not a mystical reward for effort — just more data.
Negative signals. Not hidden, not discussed. "Not interested", hiding, reporting. These carry real weight and content that provokes them can be suppressed even while generating engagement.
The thing that's genuinely unknowable
Weights change constantly. Platforms run continuous experiments, and different users are in different experimental groups at any moment.
Which means a creator observing their own results is observing a system that is both changing and not identical for everybody. Patterns inferred from personal data have a short shelf life and may never have been general.
That's the honest core of it. Anyone claiming certain knowledge of current ranking behaviour is either working on the system or guessing.
What I'd actually do
Make something specific enough that a defined group of people will genuinely want it, rather than something broad enough that nobody objects. Recommendation systems are good at finding audiences for specific things and bad at doing anything with generic ones.
Get the opening right, because that's the one factor everybody agrees on.
Post consistently and stop reading the analytics daily, because daily variation is noise.
And treat every confident claim about algorithm mechanics, including the ones above, as provisional. That's not a cop-out; it's the accurate epistemic position about a system nobody outside can observe directly.
Why the folklore persists
Worth asking why superstition is so durable here, because the answer is structural rather than a comment on anybody's intelligence.
Creators receive noisy feedback on a small number of attempts, with no control group and no way to run the same content twice under different conditions. That is precisely the environment in which humans generate causal explanations for random variation — the same conditions that produce superstitions in sport and gambling.
Add a strong emotional stake, a community that shares theories, and a system that genuinely does change, and you get a stable body of folk knowledge that is partly correct, partly obsolete and partly invented, with no reliable way to sort the categories.