Choosing the Right Campaign Optimization Event
The deepest event that stays reliable, timely, and frequent enough to train on. Alignment sets a ceiling you cannot buy your way past, precision decides how close you get, and only one of them improves with volume.
Every delivery algorithm you buy media through asks you to nominate an event. It then spends the budget finding people who look like the people who did that event.
The instinct is to nominate the deepest thing you care about, usually a purchase or a qualified opportunity. The instinct after that, once delivery stalls and someone points out the account is stuck in learning, is to nominate something shallower with more volume. Both instincts are right in different accounts, and the thing that decides which is right for yours is not a matter of philosophy.
Two forces, and only one of them improves with spend
Alignment is how closely the ranking implied by an event matches the ranking implied by value. Plenty of people browse heavily and buy nothing, so a page view identifies a browser rather than a buyer. Deeper events align better because they sit closer to the money.
Precision is how well the rate can be estimated from what happened. A rare event produces few observations, and a ranking built on few observations is partly noise.
The asymmetry between them is the whole lesson. Precision improves as volume grows. Alignment does not improve at all, because it is a property of what the event measures rather than of how much of it you have.
- Optimizing on page view
- Optimizing on high-value purchase
Show the data
| Weekly reach | Optimizing on page view | Optimizing on high-value purchase |
|---|---|---|
| 20,000 | 64.8% | 79.7% |
| 60,000 | 64.8% | 91.0% |
| 200,000 | 65.2% | 96.7% |
| 600,000 | 65.7% | 98.7% |
| 2,000,000 | 66.1% | 99.8% |
The flat line is the point. Optimizing on page views leaves roughly a third of the achievable value on the table at every account size, and no amount of extra budget recovers any of it, because the algorithm is ranking people accurately on a thing that is not what you want.
The rising line starts as the worse option and ends as the better one. At the smallest account it is guessing, because the event happens 19 times a week across the whole audience. At the largest it happens 1944 times a week and it is close to optimal.
Where the crossover falls
| Weekly reach | Page view | Add to cart | Checkout started | Purchase | High-value purchase |
|---|---|---|---|---|---|
| 20,000 | 65% | 82% | 89% | 87% | 80% |
| 60,000 | 65% | 84% | 94% | 94% | 91% |
| 200,000 | 65% | 84% | 97% | 98% | 97% |
| 600,000 | 66% | 84% | 98% | 99% | 99% |
| 2,000,000 | 66% | 84% | 98% | 100% | 100% |
The best event moves one step deeper as the account grows, and the two shallowest are never the answer at any size. Their ceilings are too low, a different problem from being underpowered.
Read the columns rather than the rows and the practical rule falls out. The question is not “how deep can I go”, it is “how deep can I go while still observing the event often enough to rank on it.” Every account has an answer and the answer changes about once a year in a growing one.
The two shallowest events never win at any volume here. That is worth separating from the sparsity problem, because the fixes are opposite: a low ceiling is a reason to change the event, and low precision is a reason to wait, consolidate ad sets, or widen the window.
The third constraint, which no simulation here captures
Timeliness. An event that arrives six weeks after the click cannot train a system that adjusts daily, however well aligned and however frequent it is.
This is why the practical answer in long-cycle businesses is often a shallower event than the volume arithmetic alone would suggest. A sales-qualified opportunity is beautifully aligned and lands a month after the algorithm needed it, so the model spends most of its life optimizing against a window that has already closed. A proxy that arrives in a day and correlates decently often beats a perfect signal that arrives too late to act on.
The way to make that trade is to measure the correlation between the fast proxy and the slow outcome on your own data, rather than assuming it. If a demo request predicts a closed deal at all, the proxy is defensible. If it predicts mainly that someone likes filling in forms, you have bought volume and given up the ceiling.
What this does not fix
An optimization event is a targeting instruction, and every event on this list is observed among people the platform already served. Getting it right means the algorithm finds people more likely to do the valuable thing. It does not mean the advertising caused them to do it.
That is a separate question with a separate instrument, and the two get confused constantly because both produce a number labeled conversions. Why Agentic Marketing Still Needs Holdouts works through what happens when an optimizer is pointed at observed conversions and left to run.
Context
Best suited for
Accounts choosing what to optimize delivery toward, especially where the valuable event is rare, slow, or both, and where somebody is about to swap it for a proxy.
Data requirements
- Event counts per week for each candidate, at the ad set level rather than the account level, since that is where learning happens
- The correlation between each candidate proxy and the outcome that pays, measured rather than assumed
- Time from event to the outcome it stands in for, because a well-aligned event that arrives late is not usable
- A holdout, if the question is ever going to move from targeting quality to whether the spend works at all
What changes by context
- B2B SaaS
- The valuable event is a closed deal months later, and there is almost never enough of it. The practical choice is which mid-funnel event correlates with pipeline rather than with form-filling, and that correlation is measurable on your own history.
- Sales-led
- Lead volume is the classic trap, because it is abundant, fast, and easy to hit while being nearly uncorrelated with revenue. Optimizing on it produces what it asks for, which is leads.
- Product-led
- Activation events are the sweet spot: frequent enough to train on, fast enough to be timely, and predictive of retention. This is the setting where the three constraints most often have an overlapping answer.
- B2C ecommerce
- Purchase volume is usually sufficient, so the interesting question moves up a level: whether to optimize on purchases or on purchase value, since a policy that maximizes orders will find cheap ones.
When it breaks down
- Accounts where even the shallowest usable event is too rare to leave the learning phase, which is a budget problem rather than an event choice.
- Categories where browsing and buying are the same behavior, so every candidate implies the same ranking.
- Platforms that do not expose the choice, or that silently substitute a modeled event of their own.
- Any use where the output is read as evidence that the advertising worked rather than that the targeting improved.
What this does not establish
- That checkout events beat purchase events below some specific reach. The crossover here sits between two of the volumes tested and depends on this event mix, this audience, and how sharply value concentrates.
- That a page view always caps at two thirds of the achievable value. That ceiling is set by how much of the shallow event is driven by something unrelated to buying, which varies enormously by category.
- That platform algorithms work this way. They are far more sophisticated than ranking estimated rates over cells. The mechanism modeled is the one they cannot escape: an estimate built from a finite number of observed events.
- That more volume is always available. Consolidating ad sets to concentrate events is usually cheaper than growing the account, and this says nothing about which to try first.
What to test next
Before changing the event, measure the ceiling you are currently under. Take whatever you optimize on today, rank a recent cohort by it, and look at value within each band. If value is flat across the ranking, the event is telling the algorithm almost nothing about who is worth reaching, and no budget increase will change that.
Then count. Weekly events per ad set, not per account, because learning happens where delivery is optimized. If the deepest well-aligned event clears the platform’s threshold with room to spare, use it. If it clears by a little, the next question is whether consolidating ad sets buys enough concentration to make it safe, which is usually cheaper than the alternative of moving a step shallower and accepting a lower ceiling permanently.
How the comparison was run Code
An audience split into 144 cells on two independent traits: purchase intent, which value depends on, and browsing propensity, which it does not. Shallow events load heavily on browsing and weakly on intent, deep events depend on intent alone and are rare. That structure is the assumption doing the work, and it is stated above rather than buried here.
For each candidate event and each account size, the observed count per cell is drawn binomially over 4 weeks, the cells are ranked on the observed rate, and the top 25% are served. Value delivered is scored against the best any ranking could achieve on the same audience, so the numbers are shares of what was available rather than absolute returns. Ties at zero events break at random, as an algorithm with no information does. Every figure is the mean of 200 replications.
Nothing here models a real platform’s algorithm, far more capable than ranking cells. It models the constraint no algorithm escapes: an estimate is built from events that happened, and rare events do not happen often. These are not measurements from a real account.