Ghost Bids vs PSA Controls: Which to Use When

What a Control Group Is Actually For

Every incrementality test rests on one question: what would these users have done if the campaign had never run? You cannot observe that directly, so you build a counterfactual, a group of people who are like the exposed group in every way except that they did not see the ad. The IAB and IAB Europe incrementality guidelines frame incrementality measurement around three principles: credible counterfactuals, control of bias, and separation of signal from noise. Incrementality is not attribution. Attribution assigns credit for conversions that happened; incrementality asks whether the ad caused any of them.

The trap is that the control group looks like a detail you can bolt on at the end. It is the entire experiment. If the control group is not comparable to the exposed group, no amount of clean statistics downstream will fix it. Ghost bids and PSA controls are two ways to build that group, and they introduce bias at different points. Knowing where each one leaks is the difference between a lift number you can defend and one that quietly flatters the channel.

Where PSA Controls Break

A public service announcement control is the older design. You run your real campaign to the test group and a neutral placebo, a charity or safety ad, to the control group. Both groups are chosen by the ad platform, both get served an impression, and you compare outcomes. It feels clean, like a placebo arm in a drug trial. The problem is that a modern ad platform does not serve the two creatives the same way.

Platforms optimize delivery in real time toward the users most likely to respond to whatever creative is running. Your product ad and a Red Cross ad have different response profiles, so the platform steers each to a different slice of people. As Eric Seufert's write-up of the ghost ads method puts it, this optimization "distorts the group definitions and makes it impossible to compare behaviors across groups." The users who see your ad and the users who see the placebo are no longer two random halves of one population. They are two populations the algorithm sorted apart.

There is a second, more mundane problem: you pay for it. Every PSA impression is budget spent to learn nothing about your product, which is why PSA tests carry a double cost. For most advertisers the comparability flaw is the fatal one and the cost is merely insult. If the two groups are not exchangeable, the lift you measure is contaminated by selection, and you often cannot tell in which direction.

What Ghost Bids Fix, and the One Thing They Require

Ghost ads, introduced by Johnson, Lewis, and Nubbemeyer, solve the comparability problem by never serving the control group a different ad. Instead the platform records the moment your ad would have won the auction for a control user and logs a ghost impression. The exposed and control groups are split randomly before delivery, and the split is not disturbed by creative-specific optimization because there is no second creative. The authors show the method can, in their words, "reduce experimentation cost, improve measurement precision, and work with ad platforms that optimize ad delivery in real time" (Marketing Science Institute).

A ghost bid is the retargeting-friendly variant of that idea, and it moves the logging point one step earlier. Rather than waiting for an impression to be won, the platform places a real bid for the test group and marks the control group with a ghost bid, a bid that could have been placed but was withheld. Remerge's breakdown of incrementality methods notes the practical payoff: users not found on the real-time bidding exchange are dropped from both groups, "thus significantly reducing noise." You compare like with like, and you never spend a placebo dollar.

The catch is reachability. A ghost bid can only be logged for a user who is found on a real-time bidding exchange and is identifiable enough to bid on. How often that happens, meaning auction density, sets how much data the log collects and how much statistical power the test has. Retargeting a known audience clears that bar easily, because those users are active and biddable. Cold prospecting into thin or non-programmatic inventory does not, and the ghost log ends up sparse. For channels like connected TV, that constraint matters: a ghost design leans on a rich, addressable identifier layer to join exposure and outcomes, which is what a large CTV and ACR feed with roughly 13 to 14 million unique CTV IDs per month is built to supply. Thin identity means thin control.

A Decision Rule for Which to Use When

Start with the outcome you owe a stakeholder and work backward to the design. If you are measuring a retargeting or warm-audience campaign inside programmatic inventory, ghost bids are the default. They give you a randomized, comparable control with no placebo spend, and those users are already found and biddable on the exchange, with enough auction volume to give the test power. This is the case where the older PSA design is hardest to justify, because it costs more and measures worse.

When You Cannot Log at the Device Level

Ghost bids need a biddable, identifiable impression opportunity. When that is missing, because the channel is offline, the inventory is non-programmatic, or privacy constraints block device-level logging, a geo holdout is the honest fallback. You randomize regions instead of users, hold some dark, and read the difference. You lose statistical power and you inherit cross-region confounders like weather and local economy, but the randomization itself is clean and easy to explain to a skeptic. For a full-funnel picture across channels that will not all support the same logging, a cross-channel measurement framework lets you run ghost bids where you can and geo holdouts where you must, then reconcile.

When a PSA Control Still Makes Sense

Rarely. A PSA design can be the only option inside a walled garden that exposes a placebo mechanism but not a ghost log, or in a legacy platform that predates ghost measurement. If you are forced into one, pin the placebo creative's response profile as close to neutral as you can, hold the platform's optimization as flat as the tools allow, and treat the resulting lift as directional rather than precise. Report the comparability risk alongside the number. A lift figure without its control-design caveats is a claim you cannot stand behind.

The discipline underneath all three methods is the same one the standards bodies keep returning to: build a credible counterfactual, control the bias in how it is built, and separate the signal from the noise. Ghost bids do that best where users are reachable on the exchange and auctions are dense. Geo holdouts do it where logging fails. PSA controls did it in an earlier era and now mostly get in the way. Choose by asking where your bias enters, not by asking which method your team ran last quarter.