Every time an ad slot opens, the same question gets asked: what is this single view worth to me? A person could never answer fast enough, or often enough. A machine-learning model can — millions of times a second, for a fresh decision every time. Here's roughly how it does it, minus the maths.
The question behind every bid
Two impressions are almost never worth the same amount. A view from someone who's about to buy is worth far more than a view from someone who'll never convert — even if the ad looks identical. The whole game is telling those two apart in the few milliseconds you have, and bidding accordingly.
What the model looks at
When a bid request arrives, the model reads the signals that come with it and asks how this opportunity has played out in the past. Roughly, it weighs things like:
- the context — what app or page, what kind of content;
- the format and placement — is the ad likely to actually be seen;
- device, time and rough location;
- how similar past impressions performed against your goal.
Out of all that comes a single, useful number: an estimate of how likely this view is to lead to the result you care about.
Predict first, then shade the bid
Once the model has predicted value, it works backwards to a price. It doesn't just bid high to be safe — that wastes budget. It figures out the lowest price that should still win this particular auction and bids that. This is called bid shading, and across millions of auctions it's the difference between your budget buying a lot or a little.
Predicting value tells you which impressions to chase. Bid shading tells you how much to pay for them. You need both — value without price discipline overpays; price discipline without value buys cheap junk.
Auto-optimization: it tunes itself
Here's the part that saves you the most time. The model doesn't make one decision and stop — it keeps learning from results and quietly adjusts. If a certain kind of placement is converting, it leans in; if another is wasting money, it backs off. Bids, budget pacing and targeting all drift toward what's working, 24/7, without anyone touching a dial. You set the goal once; the system chases it continuously.
Why "explainable" matters
Automation only earns trust if you can see what it did. A good system doesn't just act — it tells you, in plain language, why it bid the way it did and what changed. That's the difference between an AI you supervise and a black box you hope is behaving. We're firmly on the "show your work" side.
This is exactly what adZoic's machine-learning engine does on every bid. Want the auction context around it? Read RTB explained.