How to choose anti-counterfeit software
This category is unusually easy to sell badly, because the demo is always a genuine pack and an obvious fake, and every system on the market gets that right. The differences appear on marginal cases, poor photographs and categories the vendor has not seen before — none of which show up in a scripted demonstration.
Evaluate anti-counterfeit software on six things the demo will not show: how many genuine reference samples it needs, whether it has a distinct INCONCLUSIVE verdict, whether its evidence is itemised rather than a score, whether the vendor will name their failure modes, what capture tooling they provide, and what you get back if you leave. Then run a blind pilot on your own packs.
The six questions that separate vendors
How many genuine samples do you need per SKU, and why that number?
A vendor who says one is comparing against a golden image and will reject genuine product on ordinary print variation. A vendor who says 20–50 across several batches understands that genuine is a distribution. Ask what happens when a new batch prints slightly differently.
Is there an INCONCLUSIVE verdict?
There must be. Field photographs are often unusable, and a system that outputs only genuine or fake will call a glare-blown photograph a counterfeit. Ask what proportion of real verifications come back inconclusive — an honest answer is not zero.
Show me the evidence output for a real counterfeit.
Not the dashboard, the evidence. You want itemised, human-reviewable differences — 'header band olive-gold versus cobalt blue', 'guilloche coarser' — because that is what survives an argument with a distributor. A score and a red banner does not.
Where does your system fail?
Every honest vendor can list: curved surfaces, transparent packaging, flat-colour artwork, re-encoded images, phones without macro capability. A vendor who claims no weak categories has not deployed enough to have found them, which is itself the answer.
What capture tooling do you provide?
Live sharpness, glare and steadiness gates; a fixed shot list; manual capture rather than auto-shutter; untouched image storage. Capture quality is the dominant failure mode in deployment, and a vendor who has not built for it has not run a real programme.
What do we get back if we leave?
Reference images, verification history, verdicts with their evidence, and the geographic data. These accumulate into your brand-protection case file and should be unambiguously yours.
Design details that indicate a serious system
- Tolerance bands learned from your samples, with the spread across genuine references treated as legitimate variation rather than error.
- Layers that degrade rather than guess. When print structure is not resolvable, that layer should lose weight and say so.
- Cross-checking between layers. A model conclusion that contradicts the measurements should produce SUSPECT, not a confident verdict either way.
- Every image retained, because today's verification is tomorrow's training data and, if it comes to it, legal evidence.
- Location and device recorded with every verdict, so findings aggregate into a map.
- Explicit calibration phase, with the vendor saying out loud that early verifications are tuning rather than production.
The pilot that actually settles it
Scripted demos prove nothing. Run this instead:
- Supply your own packs. Twenty genuine units across at least three batches, plus any confirmed counterfeits you hold.
- Include hard cases deliberately — a genuine unit from an unusual batch, a genuine pack photographed badly, a genuine pack from a different production site.
- Run it blind. Do not tell the vendor which is which.
- Measure false positives, not just detection. Rejecting genuine product is the failure that destroys field trust, and it is the number vendors quote least.
- Have a non-expert do the capture. Your field officers are not photographers; the pilot should reflect that.
- Review the evidence for every verdict, including the correct ones. Right answer for the wrong reason will not generalise.
Scoring the result
| Outcome | What it tells you |
|---|---|
| Catches obvious fakes | Table stakes — every vendor does this |
| Catches your best counterfeit | The actual capability question |
| Passes all genuine units, including odd batches | The trust question — more important than detection |
| Says INCONCLUSIVE on the bad photograph | Verdict design is sound |
| Evidence is specific and reviewable | Usable operationally and legally |
| Vendor volunteers what went wrong | The strongest signal of all |
That last row is not sentimental. In a category where every system looks identical in a demo, the vendor who tells you where their system struggles is giving you the one piece of information you cannot get anywhere else. The category framing is in anti-counterfeit platform versus loyalty platform, and the general vendor process in vendor selection and RFP.
Frequently asked questions
What should I ask an anti-counterfeit software vendor?
How many genuine reference samples they need per SKU and why, whether there is a distinct INCONCLUSIVE verdict, to see the evidence output for a real counterfeit rather than the dashboard, where their system fails, what capture tooling they provide, and what data you get back if you leave.
How do I run a fair anti-counterfeit pilot?
Supply your own packs — twenty genuine units across at least three batches plus any confirmed counterfeits — include deliberately hard cases, run it blind, have a non-expert do the capture, and measure false positives rather than only detection rate.
Why do false positives matter more than detection rate?
Because rejecting genuine product is what destroys field trust. A team that sees the system flag good stock stops using it, and no detection capability survives that. It is also the number vendors quote least often, which is why you have to measure it yourself.
Is a vendor claiming 99% accuracy credible?
Only with context. Ask on what test set, against which tier of counterfeiter, and at what capture quality. Accuracy on clean studio images tells you nothing about accuracy on a photograph taken by a field officer in a badly lit godown, which is the actual operating condition.
What is a warning sign in an anti-counterfeit demo?
A vendor who cannot name any weak categories. Curved surfaces, transparent packaging, flat-colour artwork, re-encoded images and phones without macro capability are genuinely difficult for every system, so claiming none of them are problems means the vendor has not deployed enough to have found them.
Should the vendor guarantee results from day one?
No, and a vendor who does is overselling. Thresholds tuned on anything other than your own packs need recalibrating, so the first several dozen real verifications are legitimately a calibration phase. Vendors who say this openly are describing how the technology actually behaves.