Industry
Part of Why no feed ordering is neutral: the constraints that force a feed
Feed algorithms risks: six to watch, two to plan for
Feed algorithms risks: concentration, proxy capture, opacity, volatility, the eligibility cliff and dependence — what a publisher controls and what to plan for.
Depending on a feed for your audience is a bet on a system you cannot see, run by a party whose objective is not yours. That is not a reason to avoid it; there is no alternative at scale.
It is a reason to name the feed algorithms risks precisely, decide which ones you are carrying, and put a control against each. What follows is a register, built for someone who publishes. The risks feeds pose to readers and to platforms are noted where they change what a publisher should do.
What to take away
- Six risks cover nearly everything that goes wrong for a publisher who lives inside a feed: concentration, volatility, opacity, proxy capture, the eligibility cliff and dependence. Four have cheap controls. The eligibility cliff and dependence do not, and those two are the ones to plan for.
- The early signal for most of these is visible in your own numbers weeks before the damage, if you know which number to watch.
- The single most effective control is boringa way to reach your audience that no feed sits between.
The register
| Risk | Mechanism | Early signal | Control |
|---|---|---|---|
| Concentration | Attention flows to what already has attention, and the feed amplifies it. A few items or accounts take most of the slots; everyone else fights for the rest | A widening gap between your best item and your median item | Judge yourself on the median, not the peak; plan cash flow on the median |
| Volatility | The system changes constantly and nothing about you needs to change for your numbers to | Month-on-month swings larger than anything you did explains | A control series: comparable accounts or items you did not touch, tracked in parallel |
| Opacity | You cannot see why anything happened, so you adopt explanations that feel right | You find yourself certain about a cause you cannot test | Work the checklist in order; refuse a ranking explanation until cheaper ones are ruled out |
| Proxy capture | The feed rewards a countable stand-in for quality; producing for the stand-in slowly replaces producing the thing | Your output drifts toward whatever scored last month, and you notice it in retrospect | Keep a written statement of what you make and for whom; read it before each item |
| Eligibility cliff | A policy or classification decision removes you from consideration entirely, not gradually | A drop that is a step, not a slope, and that hits every item at once | Know the rules; keep records of what you published and when; have a route to appeal |
| Dependence | The feed is the only path between you and your audience, and the path belongs to someone else | You cannot name how you would reach your audience if the feed vanished tomorrow | An owned channel: an address list, a subscription, a place you control |
The first four have controls you can put in place this week. The last two are structural: the eligibility cliff can only be prepared for, not prevented, and dependence can only be reduced by building something outside the feed, which takes years.
Feed risk register
- Concentrationattention flows to what already has attention
- Volatilitysystem changes, your numbers swing
- Opacityyou cannot see why anything happened
- Proxy capturecountable stand-in replaces quality
- Eligibility cliffpolicy removes you entirely
Why concentration is not a bug
The pattern in the first row, in which the already-visible get more visible, is not a design choice any platform made. It is what happens whenever exposure produces the evidence that earns exposure. The general form is called the Matthew effect, and it runs in citation counts and city sizes as well as feeds.
The feedback loop inside a recommender is the specific mechanism. For a publisher the consequence is that outcomes are far more skewed than effort, so a plan that assumes effort maps to outcome will be wrong in a predictable direction.
There is partial consolation. The systems that concentrate attention at the top also make a long tail reachable: a narrow interest with a small, findable audience is served better by inference than by a broadcast schedule.
The register's first row is a risk if you need scale, and an opportunity if you can live on a niche.
Proxy capture in slow motion
This is the row people underestimate, because it does no damage on any particular day. The feed pays you in the proxy: the click, the completion, the reaction. You learn, correctly, what moves the proxy. You produce more of it. Each step is rational.
Proxy capture over one year
Start
- What you make
- Your chosen work
- Audience
- Assembled around you
- Feed numbers
- Fine
- Second measure
- Kept from start
One year later
- What you make
- What scored last month
- Audience
- Assembled around proxy
- Feed numbers
- Fine
- Second measure
- Only signal of drift
A year later you are making something you would not have chosen, for an audience that assembled around the proxy, not you. The numbers are fine.
Every feed pays in a proxy and can do nothing else, for the reasons set out in what forces a feed to exist. Knowing it is structural helps: stop expecting the platform to fix it.
Control does not mean ignoring the numbers. Keep a second measure the feed cannot supply: replies that say something, people returning through a route you own, work you are glad to have made.
When the feed's measure and yours diverge, that gap is the signal. It is visible only if you kept the second measure from the start.
The risks that are not yours
Readers carry risks a publisher should know about, because they shape policy and policy shapes eligibility: narrowing of what they see, amplification of whatever produces strong reactions, and inference about them from behavior they did not think of as disclosure.
Platforms carry risks too: gaming, homogenization, and loss of trust. When a platform tightens a rule, it is usually responding to one of these, and the tightening arrives for publishers as an eligibility cliff — a ranking change at Instagram, a policy update at TikTok, a monetization rule at YouTube. That is a constraint rather than a score, as set out in how ranking inputs divide.
Watching the platform's stated concerns is a way of seeing cliffs coming. They are stated before they are enforced, and almost always in public.
Common questions
Which risk should I deal with first?
Dependence, because it takes longest and because it is the one that turns every other risk from a setback into a catastrophe. Start the owned channel now, however small.
Is diversifying across platforms the same as reducing dependence?
Only partly. Several feeds are several systems with the same structure, and a policy shift can land on all of them in one season. An owned channel is different in kind, not just in number.
How do I tell volatility from a real change?
With the control series. If comparable accounts moved the same way, it was the weather. If they did not, something happened to you, and the register's other rows are the list of what.
Can a small publisher do anything about opacity?
Not about the opacity itself. About its effect, yes: refuse explanations you cannot test, and prefer the boring ones. The habit costs nothing and prevents most of the expensive mistakes. It works best beside a register of platform changes, which tells you which boring explanation to check first.
