8.3 The Algorithm in Practice: What Actually Works
There is no shortage of material about platform algorithms and most of it is either unfalsifiable or outdated. This chapter covers the parts that are stable enough to plan against, and the specific claim worth retiring: the shadowban for being AI.
[AS OF 2026-10] Every platform's distribution system changes without notice. What follows is a description of mechanisms that have persisted, not a specification.
The hook: two seconds, three structures
The first two seconds determine whether the rest is seen, on every discovery-driven platform. Three openings work, and they work for different reasons.
The visual interruption. A frame or first beat that is unusual enough to stop a scroll: an unexpected composition, an object out of place, a striking close-up. The mechanism is pattern interruption. For a persona account, this is a framing decision made at generation time (5.2), not an editing decision.
The stated promise. A line of text or speech that names a specific payoff: what the viewer will know or see by the end. Specificity is what carries it — "three things" beats "some tips," and a concrete outcome beats a vague benefit.
The open question. A question the viewer wants answered, posed in the first beat. It works because an unresolved question is uncomfortable, and the resolution is a few seconds away.
What fails, reliably:
- A logo or a title card
- A slow fade or an establishing shot
- Context that arrives at second four
- A greeting with no substance
- Music-only openings where the reason to watch appears late
For a generated post, the hook is a production decision. Chapter 6.4 puts it in the shot list because a hook found in the edit is a hook that is weaker than the material allows. Generate the hook frame or the hook A CLIP-style text encoder that responds best to comma-separated fragments; used by older model families. deliberately, first, before the body of the piece exists.
Which metric decides, and where
The metrics platforms use are not the same as the metrics you see first, and different platforms weight them differently.
| Platform type | Primary signal | Secondary signal | What it means for you |
|---|---|---|---|
| Short video, discovery-driven | Completion / watch time | Shares, then saves | Length and pacing matter more than the first frame alone |
| Image feed, discovery-driven | Saves and shares | Comments | Utility and identity, not novelty |
| Search-driven (long tail) | Click-through and satisfaction | Watch time | Titles and thumbnails are the whole game |
| Text-first | Replies | Reposts | A position, not an image |
The consequence for format choice: on a platform where completion decides, a four-second video that is watched twice outperforms a fifteen-second video that is watched once. This is not a platform rule; it is arithmetic, and it is why chapter 6.1's short-clip rule aligns with distribution rather than fighting it.
The metric to stop watching: likes. They correlate weakly with distribution and are the easiest signal to obtain, which is why they are the one everyone tracks.
Length: when short, when long
Short wins when the content is a single idea, a visual, or a punchline. Roughly seven to fifteen seconds on short-video platforms, and the reason is the completion metric above.
Long wins when the content has a structure that needs establishing — a story with a turn, a demonstration with steps, a piece where the payoff requires setup. Roughly thirty to ninety seconds, and it wins where the platform weights total watch time rather than completion rate.
How to test it rather than guess: take one idea, produce it at both lengths, publish a few days apart, and compare completion and shares. Two iterations of that test tells you which length your audience and your platform reward, and it costs four posts. Most operators never run it.
One structural note specific to synthetic personas: longer pieces are also riskier, because The slow migration of a face away from the reference plate, invisible in one frame and obvious across a year. accumulates with duration (6.1). A ninety-second piece is typically assembled from short clips with cuts, which is a production cost. Length is therefore a production decision as well as a distribution one.
Cadence: minimum, optimum, maximum
From 2.6, with the distribution consequences made explicit.
| Level | Cadence | Consequence |
|---|---|---|
| Minimum | 3 posts/week | Maintains the account; growth is slow but real |
| Optimum | 5–7 posts/week, mixed formats | The cadence most operators report as sustainable and effective |
| Maximum | 2–3 posts/day | Diminishing returns and rising production cost |
| Beyond maximum | Continuous | The audience tunes out; quality falls first |
The cost of exceeding the maximum is not primarily quality. It is that at high cadence, every post competes with your own previous posts for the same audience's attention, and the platform's distribution of each post suffers. Posting more is not free reach; at some point it is cannibalization.
And the production constraint is real: at 7.1's time table, a video post costs twenty-five to thirty-five minutes of attention. Two video posts a day is not a cadence, it is a full-time job.
The rule: choose the highest cadence you can sustain for a year, not the highest you can sustain for a month. Cadence collapses are worse than a lower starting cadence, because the account's distribution expectations reset.
Recycling: one asset, five formats
The highest-return distribution practice, and it is nearly free.
From one photo session or one generated One production run, sized from the survival rate rather than from ambition., with the assets already in hand:
- The still as a feed post
- The still with motion — a subtle push or parallax — as a Reel or Short (6.1's substitute)
- A carousel of three to five related stills with a caption that carries the narrative
- A cover frame for the video with the strong image, plus a text overlay
- A pin on the search-driven image platform, with a title written for search rather than for a feed
Five outputs from one production cycle. The marginal cost of the second through fifth is minutes, and the reach surfaces are different audiences.
The limit worth respecting: recycling is re-formatting, not re-posting. The same image posted four times to the same feed is repetition, and the audience notices (5.3's rolling exclusion). Recycling means the same underlying asset reaching different surfaces, which is why the secondary platform choice in 8.1 is about format rather than about mirroring.
"Shadowban for AI": what is actually happening
A silent shadowban applied because content is AI-generated is largely a myth. What exists is narrower, more specific, and easier to avoid.
What does not exist, or is not evidenced: a hidden penalty applied to an account because its images were generated. [UNVERIFIED] The claim circulates widely and is rarely supported by anything beyond a reach decline that has another explanation — the 35% year-over-year fall in platform reach, a change in format, a saturated niche, or an ordinary weak run of posts.
What does exist:
Disclosure enforcement. Platforms require AI-generated content to be labeled, and unlabeled synthetic content that is detected or reported is subject to restriction. This is not a shadowban; it is a rule with a penalty, and it is documented.
Impersonation enforcement. Presenting a synthetic persona as a real, identifiable person — or copying a real person's appearance — is prohibited and enforced. This is the rule that ends accounts.
Spam and engagement-manipulation enforcement. Automated following, commenting, and mass DMing are detected and penalized. Chapter 7.3's table covers what crosses this line.
Account-level signals unrelated to content. Chapter 8.5's finding: the ban signal follows the device and the network, not the content. A new account created on a flagged device can start deprioritized without any content violation.
Reach volatility. Ordinary distribution variance. Some posts travel and some do not, and a bad week is not evidence of a penalty.
How to diagnose a reach decline, in order:
- Did the format change? A shift from video to stills, or a change in length, resets the signal.
- Did the disclosure posture change or lapse? Check the last twenty posts.
- Did the cadence change? A gap or an overshoot affects distribution.
- Is it a platform-wide decline? Check whether comparable accounts report the same.
- Has this account been created or operated on a shared device or network? See 8.5.
Only after those five does "the algorithm is punishing AI" become a candidate explanation — and even then, the actionable version is about labeling and impersonation, not about the images being generated.
Reading a post's numbers
Four numbers per post, and what each tells you when it is the odd one out.
| If this is high and the rest is low | The likely cause |
|---|---|
| Impressions, low completion | The hook worked, the body did not |
| Impressions, low saves and shares | Interesting but not useful or passable |
| Saves, low impressions | Useful but not discoverable — a title or caption problem |
| Shares, low profile visits | The content travelled without the account travelling with it |
| Profile visits, low follows | The profile under-converts (8.2) |
The diagnostic order: start from the metric that is furthest below the others, find the row, apply the cause. This replaces the usual practice of looking at reach alone and concluding either that the account is broken or that nothing can be learned.
One number to record that most operators do not: the ratio of shares to saves. Saves indicate utility to the individual; shares indicate social value. Content that saves well and shares poorly is reference material; content that shares well and saves poorly is entertainment. Both are viable, and they lead to different content plans (2.6).
Where this goes wrong
No hook. The rest of the piece is not seen.
Optimizing for likes. The weakest available signal.
A default length. Untested, and probably wrong for your platform.
A cadence you cannot sustain for a year. Collapse resets the account's expectations.
Posting the same asset repeatedly to the same feed. Repetition, not recycling.
Blaming a shadowban for a reach decline. It prevents finding the real cause.
Unlabeled synthetic content. The documented rule, with a documented penalty.
Automated engagement. The enforceable violation, and the one that ends accounts.
The gate
- Does every post have a hook that is legible in two seconds, generated deliberately?
- Do you know which metric decides distribution on your primary platform?
- Have you tested two lengths on the same idea?
- Is your cadence the highest you can sustain for a year?
- Do you produce five formats from one production cycle?
- Do you have a five-step diagnosis before concluding that you have been penalized?
What you leave with
Three hook structures with the mechanisms behind them and a list of reliable failures, a platform-by-platform table of deciding metrics with the format consequence, a length test that costs four posts, cadence levels with the cannibalization cost of exceeding the maximum, the five-format recycling cycle and its limit, and a clear account of what AI-related enforcement does and does not exist — with a five-step diagnosis to run before blaming the algorithm.