When you can't choose the audience, choose the outcome

Patient recruitment was already the hardest part of running a trial. Roughly 80% of studies miss their original enrollment timelines, and a large share of sites never enroll a single patient. So when the biggest paid-acquisition channel in the world rewires how health advertisers are allowed to operate, it is worth paying attention. Over the last year Meta did exactly that, and most recruitment teams are feeling the effect without knowing the cause.

Here is what actually changed, and how we have adapted at CliniContact.

Health is now a restricted category, and the targeting inside it grew thinner

Meta removed sensitive interest targeting back in 2022, so you could no longer pick an audience by condition off a menu. This year it went further. Any advertiser "associated with medical conditions, specific health statuses, or provider or patient relationships" now gets sorted into a restricted bucket. In practice that means custom audiences that hint at a condition get disabled, and the platform limits how you optimize. A recruitment campaign for a specific indication is, by definition, exactly the kind of advertiser these rules target.

Automation became the default, and control became opt-out

Meta pushed Advantage+ to the front of campaign setup for leads and sales objectives. Audience expansion and placement automation are now on unless you go turn them off, and for some settings advertisers report the off switch getting harder to find. The pitch is that the model finds your buyer better than you can. For consumer goods, often true. For a trial with hard eligibility criteria, an automated system that widens your audience to hit a volume target is a system that spends your budget on people who will never qualify.

Signal loss made the feedback loop noisier

Since the iOS privacy changes, the click attribution window collapsed from 28 days to 7, view attribution dropped to a single day, and browser-side pixel tracking degraded permanently. Meta backfills the gap with modeled conversions. That is fine when you are counting purchases. It is a real problem when the "conversion" you care about is a genuinely eligible patient, because the platform is now optimizing partly on estimates of an event it can barely see.

The core failure mode

Put those three together and you get the central problem of clinical recruitment on Meta. The algorithm is very good at getting the cheapest possible version of whatever event you tell it to chase. If you tell it to chase form fills, it will find you the people most likely to fill out a form, which is not the same population as the people most likely to qualify. Published recruitment studies have watched this happen at the extreme: one reported generating a large volume of leads and then losing close to 98% of them before enrollment. The clicks were cheap. The patients were not there.

So the real question is not "how do we get more leads." It is "what event are we teaching the algorithm to optimize for."

What we do differently: the screener defines success

A screener is usually treated as a filter between the ad and the coordinator. Someone clicks, answers a few questions, and either passes or does not. Useful for the research site, invisible to the ad platform. We use the screener outcome to define what the campaign counts as success. Instead of optimizing toward a form fill or a click, we set the campaign's target to the qualified outcome. The model stops learning "find me people who click clinical trial ads" and starts learning "find me more people who look like the ones who reached the qualified result."

Two things that make it work

Both are the kind of judgment you only build by running a lot of studies.

First, you have to define "qualified" tightly. If a near miss or a half-finished answer counts as a pass, you are optimizing on the wrong thing again, just one rung up the funnel. The target has to be genuine eligibility and nothing softer, or the whole advantage collapses back into cheap volume.

Second, this only works when there is enough qualified volume to teach the model. Optimizing toward an outcome needs a steady flow of that outcome to stabilize. On a tiny budget with a rare indication, optimizing on qualified passes can trap a campaign in a permanent learning phase, and the honest move there is to optimize higher in the funnel or run a lead form instead. We make that call per study rather than pretending one setting fits every protocol. The point is not that optimizing on eligibility is always right. The point is that the screener gives us a real eligibility signal to work from at all, which most funnels throw away.

The takeaway

The larger shift underneath all of this is that Meta took away the manual levers and asked you to trust the model. You cannot hand-pick a condition audience anymore. What you can still do is control the definition of success you hand the model, and make that definition "an eligible patient" instead of "a curious click." The platform got more automated and less transparent this year. The counter is not to fight the automation. It is to feed it a better target.

When you cannot choose the audience, you choose the outcome, and you let a clean eligibility signal do the targeting the platform no longer lets you do by hand.