The 76% conversion rate that wasn’t.
A patient site reported a 76.21% paid search conversion rate, month after month. Nobody questioned it, because it looked like winning.
The number nobody questioned
We were brought into a gene therapy patient account for routine measurement work and found a paid search program reporting 4,095 clicks and 3,121 conversions for the period. A 76.21% conversion rate. Three of every four people who clicked an ad were, according to the platform, converting.
The reports had been going out monthly. The campaigns were being optimized against the number. And nobody had flagged it, which is the part worth sitting with. A conversion rate of 0.76% gets interrogated in the first week. A conversion rate of 76% gets put in a slide. Data that flatters does not get audited. Data that disappoints does. Every measurement failure we are brought in to fix has survived this asymmetry, usually for quarters.
What it actually was
There was no fraud here, and no single bug. The inflation was assembled from ordinary configuration choices that each looked reasonable alone.
Conversions were being imported into the ads platform from the analytics layer, where engagement events had been promoted to conversion status. Scroll milestones, video starts, interactions with the site’s treatment center locator: each import arrived with the authority of a conversion. Several user actions were counted through more than one path, so a single visit could register multiple times. A separate tracking gap during a platform migration made the historical comparison unreliable in both directions.
The compounding problem is automated bidding. The algorithm was pointed at the inflated signal and did exactly what it was told, spending toward the clicks most likely to scroll and tap, not the ones most likely to matter. The waste was systematic, invisible, and rising with every optimization cycle.
The fix was subtraction
The rebuild defined a small set of native conversion actions tied to actions with business meaning, principally treatment center searches and doctor discussion guide downloads, and removed everything else from conversion status. Each action then received a dollar value under our Proxy ROAS Framework, so spend could be steered by modeled value rather than event volume, without touching HIPAA-restricted data.
Re-baselined and reallocated, the account moved from a $1.10 proxy ROAS to $2.87 within 90 days. Part of that gain is real performance from reallocation. Part of it is simply the truth becoming visible, and it is worth being honest about the split: you cannot improve a number you were never actually measuring.
Proxy ROAS assigns dollar values to high-intent actions, treatment center searches and guide downloads, to create an ROI signal without HIPAA-restricted data. Same engagement as the case study on the Analytics & Measurement page.
Four checks for your own account
- Inventory the conversion sources. List every action currently counted as a conversion and where it originates, platform-native or imported. Most teams have never seen this list in one place.
- Check the counting method. For each action: once per click, or every occurrence? Locator-style interactions counted per occurrence are the single most common inflator we find.
- Sanity-check the denominator. A conversion rate above 20% on cold paid traffic is a measurement claim, not a performance claim. Treat it as a finding to investigate, never a result to report.
- Audit the value weighting. If a scroll event and a treatment center search carry equal weight, your bidding algorithm cannot tell the difference either.
An afternoon of work, and it is the highest-value afternoon available to most biopharma paid search programs. The expensive data quality problems are never the ones that look like problems.