Platform trend research is an argument, not a measurement

You do not need to know a single line of anyone's ranking code to predict some of what an engagement-optimized feed will drift toward. The prediction follows from three facts that are true of every such system, whoever built it.

The overview of platform trends sets out the vocabulary. This page works through the argument, states what it does and does not license you to conclude, and separates the pressures that are structural from the ones that are somebody's design choice.

What to take away

The three premises

First, the system optimizes something measurable. It has to, because it is trained on data, and only measurable things become data.

Second, what it measures is measured on what it showed. Nobody collects a response to an item that was never displayed.

Third, what gets shown is chosen by the system itself.

Put those together and you have a closed loop. Today's choices generate tomorrow's training data, and tomorrow's choices are made by a model fitted to it. It is the loop described in the background on recommendation systems, and everything below follows from it rather than from any particular formula.

What the loop pushes toward

The measurable thing is a stand-in for the thing anyone actually wants. Attention, completion and repeat visits correlate with value and are not value. Optimize the stand-in hard enough and the two come apart, which is the whole content of Goodhart's law: a measure that becomes a target stops being a good measure.

Fast responses beat slow ones. If the objective is scored over a short window, anything whose payoff arrives later is invisible to the optimiser. A piece that changes how someone works next year produces no signal at all inside a seven day window, and it loses to a piece that produces a reaction in ninety seconds.

The earliest moments of an interaction get disproportionate weight, because that is where the measurable decision happens. Whatever determines whether a person stays past the first few seconds is doing more work than anything that comes later, regardless of what the rest is worth.

Variance is rewarded when the objective is a sum. A format that usually does nothing but occasionally does enormously well can beat a format that reliably does moderately well, because the total is dominated by the tail. That is a mathematical property of summing, not a preference anyone expressed.

Exposure manufactures its own evidence. An item shown gets a response, and that response justifies showing it again. An item never retrieved stays unproven forever.

This leaves new and unusual material structurally disadvantaged. The same handful of things can occupy a feed for a long time. The map of ranking stages sets out which stage of retrieval and scoring each pressure acts on.

The forces pushing the other way

The right column is the part that gets left out of most accounts. A feed that fully surrendered to the left column would become unpleasant, and an unpleasant feed loses the audience the business depends on. So platforms build counterweights, and the shape you see is the outcome of a fight rather than the endpoint of a slide.

The narrowing effect discussed as filter bubbles is exactly this fight. How strong the counterweight is at any moment is not something an outsider can measure, so the practical move is to notice when it shifts, which is what a weekly trend watch is for.

What the argument does not give you

It does not name a format, a length, an opening style or a posting frequency. Anyone who moves from these premises to a specific tactic has added an assumption they did not tell you about, and the assumption is where the error will be.

It also does not tell you the strength of any pressure. Structural arguments give direction and say nothing about magnitude, so a pressure that is real can still be small enough to ignore next to whether your work is any good.

What it does give you is a filter for claims. A claim that a system will reward something with no measurable near-term response is fighting the loop and needs an explanation. A claim that unusual work is structurally harder to get started is consistent with the loop and needs less.

A claim that a format keeps winning after everyone has adopted it ignores what saturation does to a viral format, which is the most common way a confident tactic expires.

Common questions

Does this mean engagement optimization is bad?

It means it is a tradeoff with known failure modes, which is true of every objective anyone has proposed. The alternatives have their own drifts. What is unreasonable is treating any single objective as neutral.

If the drift is structural, is there anything I can do?

Yes, and it is unglamorous. Work out which pressure applies to what you make. Decide whether you are willing to fight it. If not, pick a distribution route where it does not apply.

Fighting a structural pressure is legitimate, but it should be a choice, not an accident. Moving route carries its own bill: the four categories of a platform move price it.

Why do specific formats seem to win and then stop winning?

Because the counterweights move, the population adapts, and a format that everyone adopts stops being distinctive. Any account of why a format worked that leaves out the fact that it was scarce is missing the main term.

Can platforms stop the drift completely?

Not while they optimize something measurable, which is to say not at all. They can measure better things, add counterweights and change the window. Each of those moves the drift rather than removing it, and each has its own side effects.