Ad platform reporting is the most detailed data in your marketing stack, and it's produced by a party with a direct interest in the conclusion.
That's not an accusation of dishonesty. The numbers are real. But every methodological choice in how they're counted was made by the company selling the advertising, and those choices consistently point the same way.
Here's how to read them as evidence.
1 · The platform grades its own homework
The most fundamental issue. Meta reports on Meta's contribution using Meta's attribution model. So does every other platform.
The practical symptom: add up conversions reported by every channel and the total exceeds the conversions you actually had. Everyone claims the same sales. Nobody is lying; they're each counting any interaction they touched.
The check is your own numbers. Whatever your order system, CRM or analytics says happened is the denominator. Platform-reported conversions should be read as claims about influence, and when they sum to more than reality, they need deflating in proportion.
2 · View-through attribution
Most platforms credit a conversion when someone saw an ad and later converted without clicking.
There's a defensible argument for this — brand advertising works without clicks. There's also an obvious problem: the person may have converted anyway, and an impression is a very low bar for claiming credit.
Check your attribution window settings and know what's included. A campaign that looks efficient on view-through and thin on click-through is telling you something specific, and the two shouldn't be added together without noticing.
3 · Attribution windows
A 7-day-click window credits any conversion within a week of a click. A 1-day window credits only same-day.
Same campaign, same reality, materially different reported numbers. The window is a setting, and comparisons across accounts or across time are meaningless if the windows differ.
Note the window on every report. When someone shows you a historical improvement, check whether the window changed.
4 · Incrementality — the big one
The question ad reporting cannot answer: would these people have converted anyway?
Retargeting is where this bites hardest. You show ads to people who visited your product page. They convert. The platform reports excellent returns.
Some of those people were going to buy regardless. The ad may have taken credit for a sale it didn't cause. This is why retargeting looks efficient by design and why scaling it on apparent return is the most common misallocation in paid social.
The honest way to test incrementality is a holdout — withhold ads from a portion of the audience and compare. Most advertisers never do it because it means deliberately not advertising to people, which is uncomfortable. If you have the scale, it's the single most informative test available.
Absent a holdout, treat retargeting's reported efficiency as an upper bound rather than a measurement.
5 · The metrics that look like outcomes
Some reported numbers describe the platform's activity rather than your result.
Impressions count delivery, not attention. Reach counts people the ad was served to, at any duration. Video views count at a platform-defined threshold that may be a few seconds. Engagements bundle very different actions.
None of these are fake. They're just further from an outcome than they look on a dashboard where they sit next to conversions in the same font. A campaign optimising towards them will faithfully produce them.
6 · Learning-phase noise
New ad sets perform erratically while the delivery system explores. Numbers from the first several days are not a read on your creative.
On Meta the threshold is roughly 50 optimisation events in 7 days before an ad set exits learning — documented in Meta's own Ads Help Center, and worth checking there since specifics change.
Judging a campaign before it clears that is reading noise. Restarting it because of what you read resets learning and guarantees you keep reading noise.
7 · Averages across mixed campaigns
A blended cost per result across prospecting and retargeting is the paid-media version of cross-platform averaging. It moves when your mix moves and gets interpreted as performance change.
Report per campaign job. Never blend.
What the numbers are good for
The tone above is sceptical, so it's worth being clear that this data is genuinely valuable — for comparison rather than for absolute truth.
Comparing two creatives in the same campaign. Same audience, same window, same counting rules. The confounds mostly cancel, which is why hook variant tests work.
Detecting change over time in one account. Consistent methodology means a trend is real even when the level is inflated.
Delivery diagnostics. Frequency, reach, placement breakdowns, where spend actually went. This is first-party operational data and it's accurate.
Spend. What you paid is not in question.
How to report it honestly
Lead with your own numbers. What your business actually recorded, then what the platform claims. The gap is informative rather than embarrassing.
State the window and the model. One line, every time.
Separate measured from claimed. Spend is measured. Platform-attributed revenue is a claim under a model.
Say what you can't know. Naming the incrementality limit protects you. A confident number that later collapses takes the credibility of every other number with it — the same failure described in measuring creator content.
Where Acumin fits
Ad Lab's measurement is deliberately narrow: derivations over real recorded numbers only — nothing projected, nothing forecast. Organic-versus-paid is presented as two real numbers side by side and described as descriptive, never as a causal claim, because the causal claim isn't supported by the data.
Budget comparison across campaigns compares only campaigns that have recorded results, ranks them on actual cost per result, and suggests reallocating only when the gap is meaningful — the leader must beat the laggard by at least 15% before it says anything at all. No invented budgets, no modelled attribution.
Today those results come from Meta's own reporting, entered against the campaign. Automatic read-back from Meta's API is pending Meta's app review — until it lands, you'll read your results in Meta's reporting rather than here. That's labelled in the product, and it's why nothing above promises a closed loop.
What Acumin will not do, then or later, is attribute revenue to a campaign through a model. Any number it produced that way would be an invention wearing the clothes of a measurement.
How to use this today
Take last month's ad report. Find the platform-attributed conversion number and the number your own system recorded for the same period.
Write down the ratio. That's your deflator, and applying it to every future platform number will make you roughly right instead of confidently wrong.
Related: Measured vs reported is the distinction underneath all of this. Budget splits explained is where the incrementality problem does the most damage.