What can attribution honestly not tell you?
It cannot tell you what would have happened without the ad. Attribution observes paths; it never runs the counterfactual. It also cannot see untracked touches such as a neighbor's recommendation, a truck on the street or a mailer on the refrigerator, so it silently reassigns their credit to whatever it can see. Read every attribution report as an upper bound on channel credit and a lower bound on offline influence.
Three distinct blind spots, often confused with each other
- Unobserved touches. Word of mouth, signage, a truck in a driveway, a mention at a job site. These influence real decisions and leave no record, so their credit lands on whichever tracked touch happened to be last.
- Harvesting versus creating. Attribution sees the touch closest to the sale, which is usually the one that captured intent rather than the one that produced it.
- Survivorship. You only have paths for people who converted. The far larger set of people who saw the same ads and did nothing has no path, so the model never learns what failed to work.
Why more tracking does not close the gap
It is tempting to believe the blind spots are a tooling problem. They are not. Even with perfect identifiers, attribution would still be describing an observed sequence rather than testing a cause. A better pipeline improves coverage; it does not change what kind of claim the output supports.
The practical implication is that the return on additional tracking investment flattens quickly. Once you can match most revenue to a source and you know your match rate, the next meaningful gain comes from testing, not from instrumenting one more surface.
Sentences that overclaim, and their honest versions
Reports go wrong in the wording more often than in the math. Replacing a few phrases changes what people do with the number.
Instead of stating that a campaign generated a sum of revenue, say that jobs worth that sum were last attributed to the campaign. Instead of stating that a channel drove growth, say that revenue rose in a period when that channel was the largest attributed source. Instead of stating that pausing a channel will cost you a projected amount, say that the attributed amount is at risk and that the incremental amount is unknown until tested.
What to do with the honesty
Being straight about limits is not the same as giving up. It changes the shape of the work: publish match rates, keep an explicit unknown bucket, label correlational findings as correlational, and reserve controlled tests for the spend lines big enough to justify them.
That framing is why operational AI work is useful here at all. The value is not a cleverer model of credit; it is reliably joining what your systems already know so the honest numbers arrive on time, which is what our revenue intelligence work is built to do.
Topics: limits · counterfactual · measurement · honesty
Have a version of this question about your own business?
The useful answer usually depends on which systems you run and how they're connected. That's a conversation, not a blog post.