Reviews
Part of Why no feed ordering is neutral: the constraints that force a feed
Feed algorithms in plain terms: six ordering policies and who each favors
Six feed ordering policies explained in plain terms: who each favors, what readers gain and lose, and what a publisher does under each policy.
There are about six ways to order a feed, and every product uses one or a blend. Comparing the six is more useful than comparing products, because a product's blend changes and the six do not.
Each policy has a definite answer to who it favors, what it costs the reader, and what a publisher can do under it.
What to take away
- Six ordering policiesarrival time, relationship, predicted response, inferred interest, editorial choice, and reader-set rules. Every real feed is a weighted mix.
- No policy is neutral. Each has a winner it did not choose to have, and the winner is visible from the structure.
- The same item needs a different plan under each policy. Knowing the blend in front of you tells you which plan.
The six policies
Orders by
- Arrival time
- When the item was posted
- Relationship
- How close the reader is to the source
- Predicted response
- How likely the reader is to do the measured thing
- Inferred interest
- Similarity to what the reader and people like them consumed
- Editorial choice
- A person's or a policy's judgment
- Reader-set rules
- Filters, lists, mutes and preferences the reader configured
Who it structurally favors
- Arrival time
- Whoever posts most often, and whoever posts when the reader is awake
- Relationship
- Sources the reader already interacts with
- Predicted response
- Whatever produces fast, strong reactions
- Inferred interest
- Items that resemble popular items in the reader's cluster
- Editorial choice
- Whatever the editor values, and whoever the editor knows
- Reader-set rules
- Sources the reader bothered to organize
What the reader gets
- Arrival time
- Predictability; nothing is hidden
- Relationship
- Familiarity; the people they care about
- Predicted response
- Things they will probably react to
- Inferred interest
- Discovery within their interests
- Editorial choice
- Curation and accountability
- Reader-set rules
- Control
What the reader loses
- Arrival time
- Anything posted while they were away, and any sense of importance
- Relationship
- New sources; anything from the edge of their graph
- Predicted response
- Things they would value but not react to quickly
- Inferred interest
- Anything outside the cluster; pleasant accidents
- Editorial choice
- Scale; personal relevance
- Reader-set rules
- Convenience; most readers never configure anything
Read the third column as facts about structure rather than as complaints. Every feed starts from the same shortage of space and attention, which is where the constraints every feed shares begin.
Six Feed Ordering Policies
Arrival time
- Orders by
- Post time
- Favors
- Frequent posters
- Reader gets
- Predictability
- Reader loses
- Missed posts
Relationship
- Orders by
- Closeness to source
- Favors
- Existing contacts
- Reader gets
- Familiarity
- Reader loses
- New sources
Predicted response
- Orders by
- Likely reaction
- Favors
- Reaction bait
- Reader gets
- Reactable items
- Reader loses
- Slow-value items
Inferred interest
- Orders by
- Similarity to cluster
- Favors
- Popular items
- Reader gets
- In-interest discovery
- Reader loses
- Outside cluster
Arrival time favors volume because volume is the only way to be present more often. Relationship favors incumbents because the graph is built from past interaction. Predicted response favors strong reactions because those are what gets predicted. None of these needed a decision; they are what each policy is.
Where the concentration comes from
Two of the six concentrate attention on their own. Predicted response and inferred interest both use past attention as evidence for future placement, so an item with a head start keeps it, a process known in network science as preferential attachment. Arrival time and relationship do not concentrate in the same way, though relationship entrenches whoever is already close.
How Attention Concentrates
- Past attention used as evidence
- Item with head start keeps it
- Preferential attachment compounds
- Collaborative filtering places reader
- Shown what similar people consumed
- Cold start: nothing to infer from
Inferred interest is the policy most people mean by "the algorithm". Its engine, in the usual case, is some form of collaborative filtering: the reader is placed among people who behaved similarly, and shown what they consumed.
Its strength is finding things no attribute would have connected. Its weakness is that it can say nothing about an item nobody has consumed yet. The cold start and feedback loop that follow are costs of the policy, not defects of an implementation.
Comparing them on what matters to a reader
Predictability. Arrival time and reader-set rules are fully predictable. Editorial choice is predictable in character. The other three are not, and cannot be, because they depend on inference that changes as evidence accumulates.
Manipulation resistance. Arrival time is trivially gamed by volume. Predicted response is gamed by provoking reactions. Inferred interest is gamed by fake behavior, at a cost. Editorial choice is gamed by influencing the editor. Reader-set rules are hardest to game because the reader wrote them. Every policy is gameable; they differ in how expensive the gaming is.
Time sensitivity. Arrival time and relationship handle urgent material well. Predicted response and inferred interest handle it badly, because the evidence that would promote an item takes time to arrive, and an urgent item is stale by then. This is why products that need to carry news lean on the first two.
Fairness to new entrants. Arrival time is fairest: anyone can post now. Predicted response is fair in principle and harsh in practice, because the first hours decide. Inferred interest is hardest, because there is nothing to infer from. Relationship is closed until a relationship exists.
What a publisher does under each
- Arrival time.Post when your readers are present, and often. Quality matters less than presence, and that is uncomfortable but true.
- Relationship.Deepen what exists. Replies, consistency, being recognizable. Growth is slow and durable.
- Predicted response.Front-load. Make the first seconds count and make completion easy. Remember that the item is the only row you hold, and watch for the slow drift toward making whatever scored last month.
- Inferred interest.Be classifiable. Consistency of subject helps the system place you among the right readers; wandering across subjects confuses the inference.
- Editorial choice.Be known to the editor, which means being good and being findable.
- Reader-set rules.Earn your place on a list. Ask for it, once, and make it worth the reader's configuration effort.
Most feeds you meet are a blend, and the blend shifts by surface. The practical move is to estimate the blend from your own numbers: if timing dominates outcomes, arrival time is heavy; if non-followers appear early, inferred interest is; if the first hour predicts the week, predicted response is.
Common questions
Which policy is best for readers?
There is no policy that is best on every column. Reader-set rules give the most control and are used least. The honest answer is a blend with the reader able to shift its weights, which few products offer.
Why did platforms move away from arrival time?
Because it favors volume, buries anything posted at the wrong hour, and gives the platform no way to improve the measured outcome. Whether the replacement was better for readers is a separate question from whether it was better for the measured outcome.
Can I tell which policy a feed is using?
Not exactly, and not stably. You can estimate the weights from how your items behave, and that estimate is more useful than any published description because it is about your material on that surface this month.
Is editorial choice coming back?
It never fully left; policy constraints are editorial choices encoded as rules. Whether human curation at scale returns is a product question, tracked in the three layers of platform change.
