Guides
Why no feed ordering is neutral: the constraints that force a feed
An argument that no feed order is neutral: the five constraints that force one, why every objective is a proxy, and what to check first when reach drops.
A feed is not a design choice that could have gone another way. It comes from three conditions: more material than slots, no query, and a fraction of a second to answer.
Change any one and the feed stops being necessary. Leave them all and a feed shaped thing appears, whoever builds it.
That is the useful thing to know, because it is stable. The implementations change constantly and no outside description of one survives a quarter. The constraints do not move.
What to take away
- A feed is forced by five constraintsbounded slots, unbounded supply, no explicit query, a hard latency budget, and an objective that has to be measurable.
- Those constraints rule out a neutral order. Any order is a choice, including reverse chronological, which is a choice to rank by arrival time.
- When your reach drops, an algorithm change is the last hypothesis worth testing, not the first, because everything above it on the list is cheaper to check and more often the answer.
The constraints that make a feed a feed
| Constraint | Why it cannot be removed | What it forces |
|---|---|---|
| A bounded number of slots | A screen holds what it holds, and attention is finite | Selection. Something is left out of every response |
| Unbounded supply | More is published than any person could see | Ordering, because leaving it out is not enough |
| No explicit query | The reader arrived without asking for anything in particular | Inference of intent from behavior rather than from words |
| A latency budget | The response has to arrive before the reader loses interest | A cheap first pass, because an expensive model cannot see everything |
| A measurable objective | You cannot optimize toward something you cannot count | A proxy, because the thing actually wanted is rarely countable |
None of this is peculiar to social products. The same five constraints define the older field of information retrieval, the source of most of this vocabulary. Every argument about feeds is really an argument about one of these five rows.
Five constraints of a feed
Constraint
- Bounded slots
- Selection
- Unbounded supply
- Ordering
- No explicit query
- Inferred intent
- Latency budget
- Cheap first pass
- Measurable objective
- A proxy
What it forces
- Bounded slots
- Unbounded supply
- No explicit query
- Latency budget
- Measurable objective
A complaint that a feed hides things concerns the first two rows, and one that it guesses wrong concerns the third. A complaint that it favors whatever is quick and loud concerns the fifth.
There is no neutral order
A machine with a fixed number of pockets must put every card somewhere, and the sorting rule is chosen before the first card arrives. The common demand is a feed that shows everything, in order, without interference.
That request is coherent only when supply is small enough to fit the slots. Above that point, showing everything is impossible, so the question is not whether to select but on what basis.
Reverse chronological is not the absence of ranking; it ranks by one signal, arrival time, weighted at one hundred percent.
The results are predictable: frequent posters take the most slots, timing outweighs content, and anything posted while the reader slept is gone.
These are not failures; the ordering works as specified.
The honest framing is that every feed encodes a preference. Naming the preference is a fair thing to ask for. Removing it is not available.
Every objective is a proxy
A ranking system optimizes something countable. Nobody can count whether a reader was well served, so the system counts something correlated with it: a click, a dwell time, a completion, a return visit, a follow. Each of those is a stand in, and the general result that a measure stops measuring well once it becomes a target is Goodhart's law.
Two things follow, and both are structural rather than anyone's fault.
Proxies drift from the thing they stood for as soon as anything optimizes against them. The stronger the optimization, the faster the drift, which is why the measure that worked well as a diagnostic often behaves badly as a target.
Proxies also carry side effects. Optimizing for time spent favors material that holds attention, which is not the same as material worth attending to. Optimizing for immediate response favors strong reactions. Optimizing for completion favors short things. None of these are decisions to promote a kind of content; they are consequences of choosing a countable stand in.
This is why a published objective would tell you less than people expect. The objective is a proxy, the weights change, and the behavior that matters is emergent.
When your reach drops, check in this order
Each step is cheaper than the one below it and more often the answer. Work down, and stop when the numbers explain themselves.
Reach drop diagnosis order
- Measurementdid the way you count change?
- Availabilitywas the material reachable?
- Eligibilitydid something make you ineligible?
- Your own behaviordid you change cadence or format?
- Season and newsis the whole category down?
- Competitiondid the ranked pool get bigger?
- A ranking changeonly now, and narrowly
- Measurement.Did the way you count change? A reporting change, a new attribution window, a broken tag, a sampling threshold, a time zone. This is the most common cause of a mystery drop and the easiest to rule out.
- Availability.Was the material actually reachable? A failed publish, a redirect, an expired link, a regional block, an outage during your peak hours.
- Eligibility.Did something make you ineligible rather than unpopular? A policy classification, an age or region gate, a duplicate detection, a formatting requirement you stopped meeting. Eligibility failures look like a cliff, not a slope.
- Your own behavior.Did you change cadence, length, format, posting time, or the thing itself? Compare the month against the month before at the level of what you published, not what happened to it.
- Season and news.Is the whole category down? Holidays, term dates, weather, an unrelated event taking attention. A control series answers this in an afternoon.
- Competition.Did the pool you are being ranked against get bigger or better? Your position can fall while your quality is unchanged, because ranking is relative.
- A ranking change.Only now. And even here the honest form of the hypothesis is narrow: something in the ordering moved in a way that affected material like yours. A change large enough to matter usually shows a distinctive shape, hitting some segments and not others.
The reason the last item gets reached for first is that it explains everything and requires nothing. That is exactly what makes it a poor first hypothesis. It is not testable from where you sit, it suggests no action, and it is right much less often than the six rows above it.
What this page cannot tell you
It cannot tell you what any particular product does, because that is not knowable from outside and would be stale by the time you read it. It cannot give you weights. It cannot predict a change.
What the constraint list does is rule things out. A tactic aimed at scoring cannot rescue material that is failing on eligibility. An explanation involving one worldwide switch is a poor fit for a drop confined to one country. A theory that requires a system to optimize for something it cannot count is not a theory.
That is a smaller claim than most writing on this subject makes, and it has the advantage of still being true next year.
Common questions
Is there one algorithm?
No. There are several systems in sequence, usually plus separate ones per surface, changing on different schedules. The singular noun is convenient and it hides the fact that a symptom can come from any of them.
Should I post more to recover reach?
Only if you can do it without the quality falling, and only after you have ruled out the first three items on the list. Posting more into an eligibility problem multiplies the problem.
Does engagement bait work?
It moves the proxy, which is the point of it, and proxies that get gamed tend to be adjusted. Building on a measure that only works while few people exploit it is a plan with an expiry date on it.
How would I ever confirm a ranking change caused my drop?
From outside, you generally cannot confirm it. You can make it more or less plausible by checking whether the pattern matches a scope a ranking change would plausibly have, and whether a comparable series that the change should not have touched moved as well.
In this guide
- Feed algorithms compared across seven platforms and four structural questionsTikTok, Instagram, YouTube, X, LinkedIn, Facebook and Reddit rank on four structural questions about feed algorithms, and the answers explain creator reach.
- Feed algorithms risks: six to watch, two to plan forFeed algorithms risks: concentration, proxy capture, opacity, volatility, the eligibility cliff and dependence — what a publisher controls and what to plan for.
- Feed algorithms in plain terms: six ordering policies and who each favorsSix feed ordering policies explained in plain terms: who each favors, what readers gain and lose, and what a publisher does under each policy.
