Why ads perform better on a lower budget

You drop your daily budget from €120 to €40 for a week because cash is tight, and your CPA quietly slides from €38 to €22. So you draw the obvious conclusion: the cheaper setting is the better setting. A month later you push spend back up to chase volume, the CPA climbs again, and now you have a theory. My ads just work better when I spend less.

I have believed that theory. It cost me real money before I learned to take it apart.

Short answer: Ads look more efficient at lower spend for two unrelated reasons. One is real: each extra euro buys a worse impression, so the marginal cost of a conversion rises as budget grows. The other is an illusion: small-spend weeks swing wildly and memory keeps the good ones. Telling them apart needs a deliberate budget step, not a comparison of two random weeks.

The takeaways

• Diminishing returns are real, and they are marginal. Your average CPA can look fine while the last euro you spend converts at twice the cost of the first. A bigger budget reaches lower-intent people, so the marginal customer always costs more than the average one.
• Most "low spend wins" stories are sample-size noise. A week at €40 a day might be 15 conversions. That is a coin flipped 15 times: a cheap streak proves nothing, and the number drifts back toward your real average the next time you look.
• Two different weeks at two different budgets is not a test. To know whether spend level caused the change, you change one budget on purpose and watch the marginal cost, ideally beside an untouched control.

Why do my ads perform better when I spend less?

Two forces produce that pattern, and only one of them is a genuine property of the auction. The first is diminishing returns: as you raise budget, you exhaust the cheapest, highest-intent impressions and start paying for people who were always going to cost more. The second is statistical noise: at low spend you have few conversions, so your CPA bounces around a lot, and the weeks you remember as "cheap" are the lucky low end of that bounce.

Both can be true in the same account at the same time. The danger is treating a noisy good week as proof of a permanent rule. Before you cap your spend forever, it helps to know which force you are actually looking at, because the two call for opposite responses.

How much of the effect is real diminishing returns?

The real part is the auction. Spend buys attention, and attention is sold cheapest to the people most likely to convert. Raise the budget and the platform has to find you more impressions, so it dips into audiences with lower intent and higher cost. Your next conversion is genuinely more expensive than your last.

The trap is judging this with average CPA. Average CPA blends your cheap early conversions with your expensive late ones, so it stays calm while the marginal cost climbs out of sight. Marketers call the cost of the last block of spend the marginal CPA, and it is almost always worse than the average. A campaign can be profitable on average and bleeding money on the margin at the same time. If you only ever read the blended number, you cannot see the point where the next euro stops paying for itself.

How much of it is a measurement illusion?

A large share of it. At €40 a day you might be buying a dozen or two conversions a week, and a dozen conversions is a tiny sample. One extra cheap sale swings the CPA hard, and you notice the swings that flatter you. When you look again the next week, the number has crept back toward your true average. That is regression to the mean, and it fools buyers constantly.

There is a second illusion underneath it: you are comparing two different time windows, with different competition, seasonality, and creative age, and crediting the whole gap to budget. The honest position is that a single cheap week tells you very little. I dug into how to separate a real winner from a lucky one in reading an ad test without kidding yourself, and the same caution applies to a budget level.

How do you tell diminishing returns from noise?

You run a deliberate step instead of mining history. Pick one campaign, raise its budget by a fixed amount, hold everything else steady, and let it gather enough conversions to mean something before you judge. Watch the marginal cost of the new spend, not the blended average, and keep an untouched ad set at the old budget as a control. If the control drifts too, the market moved and budget was never the story. If only the scaled set gets worse, you have found a real ceiling.

This is the same holdout logic that keeps you from breaking winners when you scale them, which I covered in scaling winning ads without breaking them. The point is to make the budget the only thing that changed, so the result means something.

Where this lives in the work

The annoying part is doing this for every campaign, every week, by hand. The scoring half of it is what I eventually built into Adscalr's ad intelligence: a composite of six metrics so one lucky CPA day cannot crown a campaign, and Bayesian shrinkage with format-specific priors that pulls a wild early score back toward what the format normally does, so a cheap streak never reads as a permanent win. Fatigue checks read a multi-day slope against the whole market first, so a bad market week does not get blamed on your budget. If you want spend decisions that start from a number you can trust, that is the part of the product built for it.