How estimated action rate works
Before showing your ad to someone, Meta estimates the chance that this particular person will do the thing your ad set is optimising for. Not click, not engage in general — the specific event you chose, whether that is a lead form submission, an add to cart, a purchase or a message. The estimate is built from what the system knows about that person’s past behaviour, what it has learned from everyone who has already responded to your ads, and signals about the ad itself.
That estimate is then multiplied into the auction, forming part of the total value score alongside your bid and ad quality. A high prediction lets a modest bid win. A low prediction means you pay more for the same placement, or lose it.
Why estimated action rate matters
It is the mechanism behind the single most consequential setting in a Meta campaign: the optimisation event. Change the event and you change what is being predicted, which changes who the ads are shown to. Optimise for link clicks and the system finds practised clickers. Optimise for leads and it finds people who fill in forms. Those are different populations, and no amount of audience targeting outweighs that difference.
It also explains why event data quality has a price attached. If your pixel or conversions setup reports events late, incompletely or not at all, the model is learning from a blurred picture, its predictions get weaker, and the auction treats you accordingly. Clean measurement is not only a reporting concern — it is a cost concern. Getting that plumbing right is what a Meta pixel and Conversions API setup is for.
Common mistakes with estimated action rate
The first is optimising for a shallow event because it is plentiful. Choosing page views or clicks feels safer when conversions are thin, but the system will faithfully deliver exactly what you asked for: traffic that goes no further.
The second is the opposite error — choosing an event so rare that the model never gathers enough examples to predict anything, which leaves the ad set stuck and unstable. The third is looking for the figure in reporting. It is not published, and any tool claiming to show yours is guessing.
How to act on it
Pick the deepest event that still happens regularly enough for the system to learn from, and be honest about which one that is. For a considered purchase in a small market, optimising for a qualified lead often works better than optimising for the sale itself, because the sale is too rare to teach anything.
Then protect the signal. Check events fire once and only once, send server-side events as well as browser ones, pass good customer information where you are allowed to, and keep the event names consistent when the site changes. Finally, give the model something worth predicting on: an ad that states clearly what happens next, and a page that delivers it quickly, produce a better prediction than any targeting tweak.
