Features
Part of A ranking signal is not the same thing as a ranking factor
Why there is no honest list of ranking signals, and the four stages instead
Why nobody can give you a list of ranking signals, plus the four jobs every ranking system has and how to read a symptom back to the stage it belongs to.
Nobody outside a platform's engineering team can tell you how its ranking works, and the people inside are maintaining something that changes faster than any description of it survives. So the usual question, "what are the ranking signals," has no honest answer.
A more useful question is structural. Whatever the implementation, what jobs does a ranking system have to do? Those jobs follow from the problem rather than from a company's taste, which is why they look the same wherever you find them.
What to take away
- Anything that puts items in an order has to gather candidates, score them, apply hard constraints, and spread the result out. The internal names differ. The four jobs do not.
- Each stage fails in its own way, so a symptom points at a stage.
- This map tells you which explanations of a change are even possible. It will never give you a formula, and anyone selling you one is guessing.
Why there are stages at all
A single careful pass over everything is not buildable. A large catalog holds hundreds of millions of items, and a request has to return in a fraction of a second. So the system cannot ask an expensive model about every item, and it asks a cheap process for a few hundred plausible ones first.
That is candidate generation, the same move classical information retrieval has relied on for decades.
The same shortage of time and slots is what makes a feed necessary at all, and how a feed gets assembled covers the assembly step.
Scoring runs the expensive model over the survivors. This is the stage people mean when they say "the algorithm." A modern learning to rank setup lives here: a model trained to predict some measurable outcome, ordering the shortlist by that prediction.
Constraints come next, because some rules cannot be a score. Age gates, regional availability, legal removals and safety policy are not preferences to be traded against popularity. They are yes or no, so they sit outside the model and override it.
Diversification is last because a pure score maximiser is unusable. Order strictly by a predicted outcome and the top ten become ten near-copies of one item. Something has to break the run up, and that something is not the score.
The stage map
What it must do
- Candidate generation
- Cut a huge pool to a small shortlist, cheaply
- Scoring
- Predict an outcome for each shortlisted item
- Constraints
- Apply yes or no rules that beat any score
- Diversification
- Stop a page of near-duplicates
What it cannot do
- Candidate generation
- Judge quality with any care
- Scoring
- See anything that was not retrieved
- Constraints
- Express a degree of preference
- Diversification
- Improve any individual item
Characteristic failure
- Candidate generation
- Good items never reach scoring, so no scoring change can rescue them
- Scoring
- Predicts the measured proxy well and the thing you actually wanted badly
- Constraints
- One blunt rule catches a whole category in order to stop a few cases
- Diversification
- Something good is held back because a similar item already holds the slot
Read the table as a set of tradeoffs rather than a design. Cheapness at the first stage buys speed and pays for it in recall. Precision at the second stage is bounded by whatever the first stage handed over. The overview of ranking signals sets out the terms these stages use.
Four ranking stages and their tradeoffs
Candidate generation
- Must do
- Cut pool cheaply
- Cannot do
- Judge quality
- Failure
- Good items never scored
Scoring
- Must do
- Predict outcome
- Cannot do
- See unretrieved items
- Failure
- Proxy beats real goal
Constraints
- Must do
- Apply yes/no rules
- Cannot do
- Express preference
- Failure
- Blunt rule over-catches
Diversification
- Must do
- Break near-duplicates
- Cannot do
- Improve an item
- Failure
- Good item held back
The third stage is deliberately crude because the cost of a rare miss is judged higher than the cost of a common false positive. The fourth stage exists to make the whole thing bearable to a human, and it necessarily demotes items the scorer liked.
Reading a symptom back to a stage
These are hypotheses to test on your own numbers, not diagnoses. They are still worth more than a guess about a signal.
- A page loses reach across every item you publish, evenly, all at once. That pattern is more consistent with a retrieval or eligibility change than with a scoring change, because scoring changes usually reshuffle rather than flatten.
- One item does badly while similar items do fine. That looks like an item-level problem, not a system-level one.
- Everything holds up except items that resemble each other. That is what a diversification change looks like from outside.
- Reach falls in one country or one age band only. Constraints are regional and demographic; scoring rarely is.
None of these is proof. Each one narrows the question from "what changed in the algorithm" to something you can actually check, which is the whole point of having a map. Where the same four jobs run with no query at all, recommendation systems is the place to follow the thread.
What the map does not give you
It does not give you weights, tell you which stage a given platform has invested in this quarter, or predict a change before it happens. It rules things out.
A tactic claiming to influence scoring cannot help an item failing at retrieval. A theory that one policy change explains a shift across every region conflicts with how constraints are usually scoped.
The stages also explain why the same content behaves differently in different products. A search box and an endless feed share all four jobs but weight them differently, because one has an explicit query and the other has to infer intent.
That difference is structural, and it is why advice ported from one surface to another so often fails. What the inference itself gets wrong is a separate register, kept in the risks that come with recommendation.
Common questions
Does knowing the stages help me rank better?
Indirectly. It stops you spending effort on the wrong stage and stops you believing explanations that the structure rules out. It does not substitute for making something people want.
Why will a platform not just publish the stages and the weights?
Publishing the stage names costs nothing and several platforms describe them in general terms. Publishing weights would be both useless and harmful: useless because they change constantly and vary by surface, harmful because a published objective is immediately gamed. The gap is not evasiveness alone; it is that a precise answer would be wrong within weeks.
Is a signal the same thing as a ranking factor?
Not quite. A signal is an input the system can observe. A factor is a claim that the input changes the output. Most public lists of ranking factors are lists of observable signals with the second half asserted rather than shown.
How often does the structure itself change?
Slowly. Specific models, thresholds and features change constantly, but the four jobs are forced by cost, policy and human tolerance for repetition, so they persist. That is exactly why the structure is worth learning and this month's signal list is not. A side by side look at feed designs shows how stable the shape is across very different products.
