During a recent review of our Amazon Canada advertising account, we found an apparent discrepancy. For one automatic-targeting recommendation ("substitutes"), our platform and Amazon Ads Console showed quite different pictures.
At first glance, this looked like a data error. After tracing the recommendation back to Amazon's source reports, we found the platform's numbers were correct. The discrepancy came from different metric definitions rather than bad data.
Orders attributed to this target
The same target, counted two ways.
| Window | Our platform30-day attribution, includes halo sales | Amazon Ads ConsoleIn this case: 7-day, advertised SKU only |
|---|---|---|
| 7 days | 1 | 0 |
| 14 days | 3 | 0 |
| 30 days | 5 | 0 |
In this case the Console "Orders" figure reflected a 7-day attribution window counting only the advertised SKU. Console attribution behaviour varies by report and view.
Why the numbers differ
1. Halo sales and SKU scope
In this case, none of the attributed purchases were for the advertised product itself. Customers clicked the ad and later bought different products from the same catalog, and Amazon still attributed those purchases to the original ad interaction. Our platform counted them; the Console's advertised-SKU view did not — which is why its count stays at zero no matter which window you look at.
2. Different attribution windows
Our platform uses Amazon's 30-day attributed purchase data. A customer can click today and purchase several days or even weeks later, and Amazon may still credit the original click — which is why our own count rises from 1 to 3 to 5 as the window lengthens. Shorter windows discard those late conversions.
Different systems, different definitions
Four definitional differences explain most of the gaps we see:
| Metric definition | Ads Console | Our platform |
|---|---|---|
| Attribution window | Typically shorter | Up to 30 days |
| Late conversions | Limited | Included |
| Halo sales | May differ by report | Included |
| Data freshness | Recent dates | Settled reporting window |
What we learned
The interesting finding is that the platform was being more conservative than the Console data suggested.
Using our platform's longer attribution data
- ACOS was approximately 43%
- The recommendation was a modest 11% bid reduction
Using the Ads Console interpretation
- 92 clicks
- $57.67 spend
- 0 orders
On the Console numbers alone, the same target could have qualified for a much larger bid reduction. The additional attributed sales softened the recommendation rather than making it more aggressive.
The bigger lesson
When Amazon advertising numbers don't match, it does not necessarily mean one source is wrong.
Before concluding that a recommendation is incorrect, it is important to understand exactly what each metric represents. In this case, the discrepancy was not caused by a data bug. It was caused by different attribution and reporting definitions.
When evaluating advertising recommendations, understanding the metric definition is just as important as understanding the metric itself.