The system is optimising something, and it is not your work

What a ranking system is actually doing

A recommender does not assess whether a piece is good. It predicts, for a given viewer, how likely they are to produce some measurable behaviour (watch to the end, watch again, react, comment, share, not scroll away) and ranks by that prediction.

That measurable behaviour is a proxy for the thing the platform wants, which is time and return visits, which is itself a proxy for advertising revenue. There are several layers of substitution between the number being maximised and anything a person would recognise as value, and each layer is a place where the two can come apart.

Goodhart’s law, and why it is not cynicism

The economist Charles Goodhart observed that a measure used as a target stops being a good measure. The mechanism is mundane: any proxy correlates with the real thing across the range of behaviour that existed when it was chosen, and optimising hard against the proxy pushes you outside that range, into the region where the correlation breaks.

Concretely:

• Completion rate correlates with satisfying pieces, and also with pieces that are short, or looped so the counter never completes, or structured to withhold the answer until the final second.
• Comments correlate with pieces people care about, and also with pieces people are angry at. An argument in the replies is indistinguishable, from a metric’s point of view, from a conversation.
• Watch time correlates with good, and also with slow, padded and artificially delayed.

None of these optimisations require anybody to be dishonest. They happen by selection: the creators who accidentally do them get distributed more, so more people see and copy them, and within a year the format has drifted.

The division that keeps you honest

A workable line, and one you can actually hold under pressure:

Optimise the form. Never optimise the promise.

Form is everything about delivery that does not change what you said: the aspect ratio, the safe areas, the opening frame, the caption, the pacing, cutting a slack ten seconds, putting the specific noun first in the title. All of this makes it easier for the right person to find and follow you, and none of it costs the audience anything.

The promise is what the piece claims it will give. Inflating it (a cover showing something that does not happen, a title asking a question the piece never answers, an opening implying a revelation you do not have) buys a click by spending trust, and trust is the thing you cannot buy back.

The reason to hold this line is not moral, although it is also that. It is that the proxy is measured and trust is not, so the damage is invisible in exactly the place you would look for it.

Nobody knows the algorithm, including the people saying they do

Ranking functions are not published, they change frequently, they are personalised per viewer, and large parts of them are learned rather than written. What circulates as algorithm knowledge is folklore assembled from small samples by people who saw one thing work once.

This matters because the folklore is generated by survivorship. You hear the account of somebody whose video did well and who has an explanation for it. You do not hear from the thousands who did the same thing and did not, because they did not get a platform to explain it from. Any rule derived this way is untestable by you, because you cannot run the counterfactual.

What you can rely on are the durable, stated things: the piece has to work in the shape and conditions the surface imposes, the opening has to pay, and a viewer who leaves immediately after arriving is a signal every system on the market is designed to notice.

Distribution you do not own is a structural risk

A platform can change its ranking, its interface, its terms or its mind overnight, without notice, without explanation and without appeal. Accounts are suspended by automated systems with no usable route to a human. This is not a rare event; it is a normal one.

The defence is unglamorous and it works: own something. An email list, a website, a phone number, a client relationship: anything where reaching your audience does not require somebody else’s permission. Use the platforms for reach, and move the people who care onto something you control. A newsletter of two hundred people you can reach directly is worth more than twenty thousand followers on a service that can turn them off.

The limitation

None of this is an argument for ignoring distribution. Work nobody sees helps nobody, and the platforms are where the audience is, and refusing to learn how they present your work is a choice to be worse at your job.

The position that holds up: learn the mechanics thoroughly, meet the format’s real constraints, and keep the promise fixed. Then judge yourself against something the platform is not measuring: whether anybody came back, whether the client got the outcome in the brief, whether the piece did the thing you wrote down at the start.