Conversion Rate
A click is cheap. It costs the user nothing and tells you only that something looked interesting. Conversion rate asks the harder question: of the people shown a recommendation, how many went on to do the thing that was actually wanted?
The definition, and the choices hidden inside it
Conversion rate is the share of impressions that result in a defined completed action — a purchase, a subscription, a booking, an episode watched to the end. Three decisions determine what the number means: the denominator (per impression, per click or per session), the attribution window inside which the action still counts, and what you are willing to call a conversion at all. Sitting downstream of click-through rate, it is far sparser — conversions per impression land orders of magnitude below clicks — so an experiment needs considerably more traffic to move it detectably.
Why it is the metric that decides
- Optimising clicks drifts towards sensational thumbnails and misleading titles. Conversion is where that strategy is caught, because the click arrived and nothing followed it.
- It is the usual online counterpart to offline NDCG, and the pair diverges routinely — a model that wins on the offline ranking metric can ship a worse conversion rate.
- In a multi-task CTR and CVR model the conversion head trains only on clicked impressions, so it inherits a selection bias that has to be corrected rather than ignored.
Read it with care
Price, stock, checkout friction and seasonality all move conversion without the ranking changing at all. On its own it is a noisy judge of recommendation quality, which is why it belongs beside engagement and coverage measures rather than replacing them.
Conversion rate is the share of recommendation impressions — or clicks, or sessions; the denominator must be stated — that result in a defined completed action such as a purchase, subscription or finished episode, counted within an attribution window. It is the online metric closest to business value and the standard check on click-through rate: a model that wins clicks with sensational thumbnails is caught here, where the click led nowhere.
Shops count two quite different things. How many people stopped at the window, and how many actually walked out with a bag. The first number is flattering; the second one pays the rent.
Conversion rate is the second number, applied to recommendations: out of everyone who was shown a suggestion, how many went ahead and bought the thing, took out the subscription, or watched the whole film. It matters because chasing clicks alone has an obvious cheat — a sensational picture and a misleading title will get the click every time. Conversion is where that cheat is caught, because the person arrived, looked, and left again.