By Ryan Richardson · Published 8 October 2026
Knowing the three distortions, because they stack, and knowing the one that runs the other way, are steps 54 and 55 (Part IX), both marked hard difficulty, almost nobody does this. They're treated together here because the Field Manual itself groups all four distortions under one heading; splitting them into two pages would duplicate the same source material rather than add a second fact.
The reporting window excludes today, so anything that converts after the dashboard was last read is invisible until the next pull. The value passed back is usually front-end price alone, so order bumps and upsells vanish from the reported figure even though they landed in the bank. Some conversions get missed outright, through blocked pixels, expired attribution windows, or a dedup failure.
Broad targeting over-claims credit from the narrower testing that found the buyer. A campaign widened after a narrow test succeeds can end up reported as the source of sales the earlier, narrower test produced, which is the one mechanism in this set that runs in the opposite direction from the other three and is easy to miss precisely because it looks like good news.
One platform's default attribution window moved from a 28-day click to a 7-day click, effective 12 January 2026; a 25-experiment comparison found observational estimates running roughly nine times the randomised ones behind them. Across fifteen separate paired experiments, about half were off by a factor of three; one study found an observational lift of 316% against an experimental 73%, another 4,074% against an experimental 2.4%, and a 663-experiment run found a median error of 62 to 115 percentage points.
A weekly page that demotes every platform number to a signal, reads settled cash as the grade, and names which of the four distortions is live in any given week rather than averaging them away.
Reading a platform improvement as pure upside without checking whether it's the broad-targeting distortion borrowing credit from a test that already did the work. Treating the gap between platform and ledger as a constant, when it moves with attribution-window policy changes outside the business's control.
Vendors selling a fix for the platform-versus-ledger gap routinely quote a 'healthy' range of fifteen to twenty-five percent overstatement, or claim their own server-side tracking brings it under ten percent. These figures carry no sample, no method and no date behind them, and they come from the party with something to sell. Any claim phrased as 'X% lift' with no control group or randomised holdout is observational, not experimental, whatever the platform calls it in its own dashboard; treat an unsourced vendor benchmark the same way, as a reason to check your own number, not a substitute for it.
| Claim | Value | Source |
|---|---|---|
| Attribution window change | one platform's default window moved from 28-day click to 7-day click, effective 12 January 2026 | Measured in Real Money, Field Manual |
| Vendor-quoted 'healthy' attribution overstatement range, with no disclosed sample or method | fifteen to twenty-five percent, or under ten percent claimed by vendors selling server-side tracking | Measured in Real Money, Field Manual |
| Observational vs randomised estimate gap | 25-experiment comparison found observational running ~9x the randomised estimates behind them | Measured in Real Money, Field Manual |
| Paired-experiment error rate | across 15 paired experiments, about half were off by a factor of 3; one study found 316% observational vs 73% experimental, another 4,074% vs 2.4% | Measured in Real Money, Field Manual |
| Largest cited study on reporting error | 663-experiment run found a median error of 62 to 115 percentage points | Measured in Real Money, Field Manual |
| Matrix status for these steps | Both hard difficulty, almost nobody does this (step 54 automated tooling, step 55 partial) | The Sixty Steps matrix |