There is a number that shows up in almost every retail computer vision programme, and once you have seen it a few times you start expecting it.

The model scores 94% in the store where it was built. By store forty, it is somewhere in the low seventies.
Nothing broke. No code changed. The model that was excellent is now mediocre, and the team's instinct is to go improve the model — which is almost always the wrong response, and reliably wastes a quarter.
Why the estate breaks your model
The pilot store is not a sample. It is the training environment, and every subsequent store differs from it along axes nobody enumerated:
-
Lighting. Skylights, different fixture lighting, west-facing windows at 5pm in summer. Colour temperature and shadow behaviour vary enormously.
-
Fixture geometry. Shelf heights, bay depths, the distance from camera to product.
-
Camera angle. Nominally identical mounting specs produce meaningfully different views. Then someone knocks one six degrees out of alignment during a shelf reset and nobody records it.
-
Product mix. Regional ranging means store forty stocks items the pilot store never carried.
-
Seasonality. Christmas displays change the visual environment more than most teams anticipate, and they do it estate-wide at once.
Each of these is a distribution shift. Together they mean your model is being evaluated on data that does not resemble what it learned from, which is exactly the condition under which supervised models degrade.
Aggregate metrics will hide this from you
Here is the monitoring mistake that turns a manageable problem into a crisis.
If you track accuracy across the estate as a single number, a store that has effectively gone blind gets averaged away by two hundred stores performing fine. The dashboard reads 91%. Someone in store 214 has been ignoring every alert for six weeks because they are all wrong, and nothing in your telemetry says so.
Track accuracy per store. Always. The distribution matters far more than the mean, and the tail is where your operational failures live.
The second monitoring habit worth building: watch confidence distributions, not just accuracy. Confidence distributions shift measurably before accuracy visibly degrades, which gives you warning while you can still act. A store whose mean prediction confidence has drifted down over three weeks is telling you something even if its accuracy has not moved yet — and confidence monitoring has the significant advantage of requiring no ground-truth labels, so it runs everywhere continuously rather than only where you have review capacity.

The correction loop
The fix is not a better model. It is a loop that runs permanently.
Sample production frames continuously, weighted toward low-confidence predictions and toward stores whose confidence distribution has shifted. Random sampling wastes review capacity on cases the model already handles well.
Route uncertain cases to human review. This is a real labelling function with real headcount, running forever — not a phase that completes before launch. This is the line item that gets cut from business cases and it is the one that decides whether the system improves or decays.
Feed corrections into retraining, then roll out new versions in stages with per-store accuracy gating progression rather than deploying estate-wide and hoping.
Attribute regressions to versions. During a staged rollout different stores run different model versions, so every event needs to carry the version that produced it. Without that field, a false-positive spike is unattributable and you are debugging blind.
Budget somewhere around fifteen to twenty-five percent of initial build cost annually for this loop. That number consistently surprises people, and it is the difference between a system that gets better and one that is accurate on launch day and quietly ignored a year later.
What the two-tier pattern changes
One recent development genuinely helps here, though not in the way it is usually pitched.
Vision-language models can be asked in plain language whether a shelf image shows a gap, with no task-specific training. They are too slow and expensive for continuous frame-rate inference, so the working architecture is two tiers: a cheap always-on detector triggers, and the VLM adjudicates ambiguous or novel cases.
The drift-relevant benefit is that tier two degrades far more gracefully. A purpose-trained detector that has never seen a Christmas display fails in ways that are hard to predict. A general model handles novelty better, because novelty is closer to what it was built for.
The catch: without task-specific training you have no training metrics, so your evaluation set becomes your entire quality control. Build it from real footage across genuinely different stores, keep the awkward cases in it, and run it in CI. An eval set assembled from clean pilot-store footage will pass everything and predict nothing.
What to ask a vendor
If you are buying rather than building, one question separates products that survive year two from those that do not:
How do you detect that the model has degraded in store 214 specifically, and how does a fix reach that store?
Most vendors demo capture, inference and dashboards competently. Model operations is where they differ, and it is where buyers rarely look. Ask to see the per-store accuracy view from a live deployment — not a slide, the actual dashboard.
The full guide covering architecture, cost structure, accuracy targets by use case and pilot design is on the TechCirkle blog: Computer Vision in Retail: What Actually Ships in 2026.
Frequently Asked Questions
Why does a vision model lose accuracy as it rolls out to more stores?
Because each store differs from the training environment in lighting, fixture geometry, camera angle and product mix. Together these constitute a distribution shift, which is precisely the condition under which supervised models degrade — no code or model change is involved.
Why is aggregate accuracy monitoring inadequate?
A single estate-wide figure averages away individual site failures. A store whose model has effectively stopped working disappears into two hundred healthy stores, so the dashboard looks fine while staff at that location have been ignoring every alert for weeks.
What is the advantage of monitoring confidence distributions?
Confidence shifts before accuracy visibly degrades, giving earlier warning. It also requires no ground-truth labels, so it can run continuously across every store rather than only where you have human review capacity available.
How should production frames be sampled for review?
Weighted toward low-confidence predictions and toward stores whose confidence distribution has drifted, rather than randomly. Random sampling spends scarce review capacity on cases the model already handles correctly and surfaces few useful corrections.
Why must events carry a model version?
Because staged rollouts mean different stores run different versions simultaneously. Without version attribution on each event, a spike in false positives cannot be traced to a specific release and you are left debugging without the one piece of information that would resolve it.
Do vision-language models eliminate the drift problem?
No, but they degrade more gracefully on novel situations because generality is what they are built for. The tradeoff is that without task-specific training you have no training metrics, so the evaluation set becomes your only quality gate and must be built from genuinely varied real footage.