“We harvest negatives weekly”, every agency, in every pitch. We pulled the data across 1,200 Amazon ad accounts to see what “weekly” actually looks like. The median number was 23 days. The 75th percentile was 41.
That’s a lot of compounding. Using the compounding TACoS math we ran on a $20M brand, the typical drift between a real weekly cadence and what we measured is worth ~$31,400 in NPV per account. Per year. On the median brand. Without anyone touching a bid.
What we measured
- Sample: 1,200 ad accounts under our audit pipeline since Q3 2024
- Range: $50K to $40M annual ad spend; mean ~$1.8M; weighted toward mid-market
- Window: rolling 12-month negative-keyword timestamps + ST report cross-reference
- Definition: “drift” = days between a search term first generating ≥3 clicks at >2x target ACoS and the date it was added as a negative
- Categories represented: 31 across the Helium 10 taxonomy; 9 with 50+ accounts in sample
The distribution
| Percentile | Days to negative |
|---|---|
| 25th | 11 |
| Median | 23 |
| 75th | 41 |
| 90th | 67 |
| 95th | 94+ (no negative ever applied) |
The 95th percentile is what should haunt anyone reading this. One in twenty accounts in our sample never killed bad keywords at all, they just kept paying for them. In 84% of those cases, the agency had a documented “weekly negative harvesting” line in the SOW.
By category
Drift wasn’t uniform. Three category clusters had materially worse time-to-negative than the median:
- Pet supplements: median 38 days. Categories with high query-rewrite volume (long-tail intent like “pet supplement for arthritic dogs”) generate more borderline keywords, which means more decisions per week, which means slower review.
- Tools / home improvement: median 34 days. Same pattern as pet supplements, high query-rewrite volume and high ambiguity per term.
- Apparel: median 29 days. Apparel’s seasonality means agencies often “pause” the negative-harvesting cadence around inventory shifts, and forget to restart.
The clusters that performed better, single-SKU cookware, beauty hero products with narrow query intent, sat at 14-19 day medians. The pattern: complexity in the query layer increases drift, because complex queries generate more decisions and decisions don’t get made.
Where the drift hides
It’s not in obvious wasted spend. The 200% ACoS keyword usually gets killed inside a week, that one looks bad in the dashboard. The drift lives in keywords sitting at 1.3-1.6x target ACoS, bad enough to bleed, not bad enough to draw attention. They survive because nobody set a clear threshold. They compound because spend grows.
The other place it hides: bad listing copy. If your detail page can’t convert a lukewarm keyword, the keyword reads as bad, but the keyword isn’t the problem. We see this constantly on accounts that haven’t run a Cosmo-aware A+ rebuild. Half their “wasteful” keywords are catalog-symptom, not ad-symptom. Killing them off cleans up the dashboard but doesn’t recover the spend; the next batch of equivalent keywords arrives within 30 days and you’re back where you started.
The third place it hides, and this one is the most expensive, is in the auto campaigns. Auto campaigns are designed to be a discovery layer, not a perpetual machine. The brands at the worst end of our distribution are running auto campaigns at a 2-year average cohort age, with negatives harvested only when an exact-match version of the same keyword goes bad. That’s spending discovery dollars on terms you already discovered.
How to audit your own time-to-negative
Run this in 20 minutes:
- Pull your last 90 days of ST report data from Bulk Operations.
- For each search term that hit your kill threshold (we use 1.4x target ACoS at >$50 spent), find the first date it crossed that threshold.
- Check your negatives list for that search term as a negative-exact. Note the date it was added (or “never” if missing).
- Compute the gap. Median that across the top 100 worst offenders.
If your median is over 7 days, you don’t have a weekly cadence, you have a “weekly review meeting where you sometimes do something” cadence. Those are different.
What we changed in our own program
We dropped the “weekly” cadence. Manual weekly reviews are theater. We replaced it with two things:
- A daily threshold-based pull, every search term over 1.4x target ACoS at >$50 spent gets flagged, no human review needed for negative-exact application.
- A weekly human review, but only on the 1.0x-1.4x band, the borderline cases where a negative might kill a converting query.
Median time-to-negative on our portfolio: 4.1 days. Not 23. The math on that is what 1.4% of TACoS looks like recovered, which is what we ask instead of “what’s your target ACoS” on day one. It’s also the kind of measurement loop that makes a catalog-as-product approach actually work, instrumentation that catches drift before the dashboard does.
Subscribe to the Operator Brief, the next issue covers the dayparting drift we caught on the same dataset.