The Score Is Multiplied by Money
In one line: a ranker's score only has to order things, and a CTR model's score gets multiplied by a bid to decide what someone pays — which makes it a different problem.
The one-line difference
In a recommender or a search system, the score is an ordering device. Apply any monotone transform to it — square it, take its log, add a constant — and the results are identical. The absolute value means nothing.
In an ads system, the predicted click probability is multiplied by a bid:
eCPM = bid x pCTR x quality (expected revenue per thousand impressions)
Now the absolute value is the whole thing. A model that predicts 0.02 when the truth is 0.01 has doubled that advertiser's estimated value and will hand them impressions they should not have won — at a price computed from a fiction.
That is the sentence to open with. Everything distinctive about this system follows from it.
What follows from it
Calibration stops being a nice-to-have. Elsewhere it matters only when a score leaves the slate; here the score always leaves the slate, into an auction. A well-ranked but miscalibrated model is a broken ads system.
Sampling has to be undone. Clicks are rare, so training data is downsampled — and downsampling changes the label distribution the model sees, which changes its output scale. That has to be corrected analytically at serving time, or every prediction is wrong by a known factor.
AUC is not enough. AUC measures ordering. A model can have excellent AUC and be systematically two times high, which AUC cannot see, and which costs real money on every impression.
Model freshness is worth more than model sophistication. New ads, new campaigns and new creatives appear continuously, and a model that has never seen an ad has nothing to say about it. This is one of the few production settings where retraining hourly beats a better architecture retrained weekly.
Three parties, not two
The other structural difference. A recommender balances users against the platform. An ads system has three parties with genuinely opposed interests, and the auction is the mechanism that reconciles them.
| Party | Wants | Failure |
|---|---|---|
| User | Not to be interrupted by irrelevant ads | Blindness, then ad blockers |
| Advertiser | Clicks and conversions at a predictable cost | Overpaying, then leaving |
| Platform | Revenue now, and both parties still here next year | Optimise revenue, lose one of them |
The tension is direct: the platform's short-term revenue is maximised by showing more ads, more prominently, and that is exactly what drives users away and eventually removes the audience advertisers were paying for.
The mechanism that keeps it honest is the quality term in the ranking formula. Ranking by bid alone lets the highest bidder win regardless of whether anyone wants to click. Multiplying by pCTR means an ad nobody clicks cannot buy its way to the top — a low-relevance advertiser must bid enormously to compensate, which prices the harm rather than permitting it.
What is actually being predicted
Worth being precise, because "CTR" hides several different quantities.
pCTR — probability of a click, given the ad is shown in this slot to this user in this context. It is conditional on all of that; the same ad has a different pCTR in position one and position four.
pCVR — probability of a conversion given a click. Needed whenever advertisers pay per action rather than per click, and it is a much harder problem: conversions are rarer, and they arrive late.
Expected value — for a cost-per-action advertiser, the platform must estimate pCTR x pCVR x value to convert their bid into an expected revenue per impression. Every additional factor multiplies the error.
The design consequence: the further the advertiser's payment model is from the impression, the more inference sits between what you observe and what you must estimate — and the more calibration error compounds.
The scale of the imbalance
One number that shapes everything downstream: click-through rates in display advertising are typically well under one percent, and often a fraction of that.
So the training data is overwhelmingly negative. That forces downsampling to make training tractable, which forces a calibration correction. It also makes accuracy a useless metric — a model predicting "no click" for everything is right more than 99% of the time — and it makes each individual positive label precious.
A rare event with money attached, predicted billions of times a day, where the absolute value of the prediction is what matters. That combination is why this is its own problem rather than a variation on ranking.
Key takeaway
A ranker's score only has to order; a CTR model's score is multiplied by a bid, so its absolute value is the product. That single fact makes calibration mandatory rather than optional, forces the downsampling correction, makes AUC insufficient on its own, and makes model freshness worth more than model sophistication. And the bid x pCTR formula is not a revenue trick — it is the mechanism that stops an advertiser buying attention nobody wants to give.
Next: scoping, and the three systems that share this name.