Your brand protection queue is the product
Detection recall is the metric vendors compete on and the one that matters least. What decides whether a programme works is whether a human can finish the queue it produces.
Brand protection vendors compete on recall. More platforms, more listings processed, more links per day. Those numbers are real, and they are close to irrelevant to whether your programme works.
The number that decides it is much less impressive: how many findings can one person actually close in a day, and is that more than the number arriving?
If arrival rate exceeds closure rate, the queue grows without bound. Everything downstream — takedown volume, follow-up, reporting — degrades quietly. Nobody declares failure, because detection is still running and the dashboard still moves.
Why more detection can make things worse
Doubling recall on a queue you already cannot clear does not double removals. It adds latency to every finding, because the important ones are now buried deeper. In the worst case it inverts your priorities entirely: the counterfeit storefront doing real revenue damage is on page nine, behind two hundred fan accounts.
This is the central reason we think filtering deserves more engineering attention than crawling.
Four filters, cheapest first
The order matters, because each filter should be cheaper than the one after it.
Allowlist your own estate. Your domains, your properties, your authorised sellers. Costs nothing, and removes a surprising share of findings. A programme that surfaces your own storefront has an obvious credibility problem.
Exclude what you have already decided about. Names that are similar but unrelated, known partners, businesses in different classes. This is a list your team maintains once.
Remember every verdict permanently. This is the filter most often done badly. If a URL was judged last month, it should not be judged again — regardless of whether the verdict was a match. Time-windowed deduplication means the same cleared listing comes back around every few weeks and consumes attention again.
Then score what survives. Model inference is the expensive step, so it should run on the smallest possible set. A confidence threshold on the output decides what a human ever sees.
Recall you should measure instead
If you are evaluating tooling, the questions worth asking are not about platform counts.
- Of the findings that reach a human, what proportion are actioned? Below roughly half, the filtering is too loose.
- How many findings sit in the queue older than thirty days? That number is the honest measure of whether closure rate is keeping up.
- When a finding is dismissed, does the same URL ever appear again?
- What proportion of sent notices got a response, and who chased the rest?
Those four tell you whether a programme is working. Total links processed tells you how much the vendor spends on crawling.
The design consequence
Once you accept that human attention is the constraint, most product decisions follow from it. You judge candidates before a person sees them. You remember every clearance forever. You capture evidence at detection so nobody reassembles it later. You draft the notice automatically and ask a human only for approval.
That is the shape of how TracBrand works, and the reason the queue is the thing we optimise rather than the crawl.