Feed algorithms platforms: what to know and why

Ask what "the algorithm" does on a given platform and you get a company name and a rumour. Ask instead what kind of surface you are looking at, and you get an answer that holds across companies and survives their updates. A feed is not one thing. It is a family of surfaces, each solving a different version of the same problem, and each surface's behavior follows from its version, not from who built it.

This page names no platform. Every surface below exists, in some form, on several.

What to take away

• Seven surfaces cover almost everything you will publish into. They differ on four structural questions, and the answers predict behavior better than any leaked detail.
• "Reach" means something different on each surface. A number that is good on one is meaningless on another.
• Advice ported from one surface to another fails at a predictable rate, because it was tuned to a structure the new surface does not have.

Four structural questions

Before anything else, ask these about the surface in front of you.

1. Is there a query? A search box knows what the person wants. A feed has to guess. Everything about how the surface behaves flows from this: with a query, matching dominates; without one, prediction from behavior dominates.
2. Is there a graph? Some surfaces only show material from sources the person chose. Others draw from everything. A graph bounds the candidate pool and changes what "new" means.
3. Does time matter? Some surfaces are worthless if stale. Others happily surface something from years ago. Where time matters, recency is a hard input; where it does not, it is barely one.
4. How many slots? A single slot is a decision. A short tray is a selection. An endless scroll is a ranking. The fewer the slots, the more each choice costs the system and the less room there is for exploration.

The seven surfaces

Surface	Query	Graph	Time-sensitive	Slots	What it has to infer	What "reach" means here
Following feed	No	Yes	Usually	Many	Which of your chosen sources to show first	The share of your followers who saw it
Discovery feed	No	No	Sometimes	Many	What a stranger might respond to	How far outside your graph it traveled
Search results	Yes	No	Rarely	Many, ranked	Which items answer the words typed	Position for the queries that matter to you
Notifications	No	Yes	Very	Few	Which of the things that happened deserve an interruption	Whether it was worth interrupting for
Ephemeral tray	No	Yes	Very	Few, ordered	Which chosen sources the person is most likely to open now	Open rate among followers
Listing or marketplace	Partly	No	Sometimes	Many, filtered	Which items match a partly specified need	Impressions against relevant filters
Inbox or messaging	No	Yes, tightly	Very	Few	Almost nothing; ordering is mostly by time and thread	Not applicable; there is no audience

Look down the "infer" column. The surfaces that must infer most, discovery and listing, are where behavior evidence dominates and where the two families of evidence a recommender uses matter most. The surfaces with a graph and strong time pressure, following and ephemeral, infer least; they are mostly ordering a bounded set by a few inputs.

The search column is its own world. With a query, the system's job is closer to classic retrieval, trading off precision and recall over a set of items that can be matched to words. Advice about search is advice about matching. It does not transfer to a feed, where nothing was typed.

Why the same material behaves differently

Publish one item and it enters several of these surfaces at once, and does differently on each. That is not inconsistency. Each surface is asking a different question of it.

• On the following feed, it competes only with other things your followers chose. The question is whether it is worth their limited time compared with those.
• On discovery, it competes with everything. The question is whether early evidence suggests strangers will respond, which is why the constraints that force a feed into stages bite hardest here: a cheap first pass has to find it before anything else can.
• In search, it competes with everything that matches the words. Whether it is any good matters less than whether it is findable.
• In the tray, it competes with a handful of sources the person is close to. Relationship dominates.

A following feed's value depends on the graph existing, which is a network effect: it is worth more to each person the more of their chosen sources are there. Discovery has no such dependence, which is why platforms can bootstrap it faster and why it is the surface they tend to push.

Measuring per surface

The mistake is a single reach number. Instead, one number per surface, defined by that surface's column in the table.

• Following: share of followers reached. The denominator is yours to know.
• Discovery: fraction of views from non-followers, and how that fraction changed over the item's life.
• Search: position for a short list of queries you care about, checked on a schedule.
• Tray: open rate, which is a relationship measure and moves slowly.
• Listing: impressions on the filters your item genuinely matches, not on everything.

When any one of these moves and the others hold, you have located the change to a surface, and the number of possible explanations drops sharply. The place to go next is which producer of evidence could have moved, asked about that surface alone.

Common questions

Is a discovery feed just a following feed with the graph removed?

Structurally, close to it, and that removal changes everything. Without the graph the candidate pool is unbounded, so the cheap first pass matters far more, and early behavior evidence becomes the main way anything gets found.

Why does search advice keep failing on feeds?

Because search has a query and feeds do not. Matching words to items is a different task from predicting response, and tactics for the first have nothing to act on in the second.

Which surface should I prioritize?

The one whose "reach" definition matches what you want. Followers who see you, strangers who find you, or people who search for what you do are three different goals with three different surfaces.

Do the seven surfaces change?

Products add, merge and remove surfaces all the time; the four structural questions do not change. When a new surface appears, answer the four questions and it will behave like whichever row it resembles. Track the appearances in your own change register.