AI Shelf Photo Analysis: What It Actually Does for FMCG Merchandising (And What It Doesn't)
Last month I sat in a Carrefour in Deira watching a merchandiser take 14 photos of a single shelf. Fourteen. Then he walked to the next aisle and did it again. His supervisor, sitting in an office in Sharjah, would later scroll through hundreds of these images trying to figure out if the planogram was being followed.
That's the job. Or at least, that was the job.
I've been building Zivni for FMCG teams across the GCC and beyond for a few years now, and honestly, shelf compliance is the one area where I've seen the biggest gap between what brands think is happening in stores and what's actually happening. Sales ops managers show me dashboards with 92% compliance scores. Then we run AI shelf photo analysis on the same stores and the real number is 61%.
So let's talk about what's actually changing. And where the hype is getting ahead of reality.
What AI shelf photo analysis actually does
Strip away the marketing language and here's the mechanic: a field rep takes a photo of a shelf. The image goes to a computer vision model trained on your SKUs. Within a few seconds you get back a structured breakdown — share of shelf, out-of-stock detection, planogram compliance, competitor facings, price tag reading, and sometimes even promotional material verification.
What used to take a supervisor 20 minutes of squinting at photos now takes 4 seconds.
The numbers matter more than the demo though. One of our clients in Riyadh — a mid-sized dairy distributor — was running weekly audits on 340 outlets. Manual photo review. Their compliance team of 3 people could realistically review maybe 40% of submissions before the next week's photos came in. The rest just piled up. Which meant reps figured out pretty quickly that most photos never got looked at, and shelf discipline drifted.
After we turned on AI shelf photo analysis, they reviewed 100% of submissions automatically. Compliance scores dropped from a reported 89% to an actual 64% in the first month. That sounds bad. It wasn't. It was the first honest number they'd ever had.
And here's the thing — you can't fix what you can't see.
Where it works, where it struggles
I want to be straight with you because I've watched brands get burned by overselling on this.
AI shelf photo analysis works really well for:
- Share of shelf calculations against known competitors
- Out-of-stock detection on your own SKUs (assuming you've trained the model on your packaging)
- Planogram adherence when the planogram is reasonably standardized
- Price tag capture in modern trade
- Detecting missing POS materials during promotions
Where it still struggles, and I'm being honest here:
- Traditional trade in Karachi or Lahore where shelves are chaotic and lighting is poor
- Detecting freshness or damage (that's still a human eye job)
- New SKU launches during the first two weeks before the model has enough training data
- Reading Arabic price tags in stores that use handwritten labels — accuracy drops maybe 15-20%
I used to think the tech would solve 95% of merchandising audit work. I was wrong. It solves about 70% and makes the other 30% dramatically faster because your team only reviews exceptions.
That 70/30 split changes the economics though. A supervisor who used to cover 40 stores can now oversee 180. That's the real transformation.
Why the GCC market is moving faster than the US on this
This surprised me at first. You'd assume American CPG brands would be way ahead on AI-driven retail execution software. They're not — at least not the mid-market ones.
A few reasons. Modern trade dominates GCC retail. Lulu, Carrefour, Union Coop, Panda, Danube — these chains have relatively standardized shelves which makes computer vision models much more accurate. Compare that to a US grocery landscape split between Kroger, Publix, HEB, regional chains, and independent bodegas, each with different shelf layouts.
Also, distributor structures in Saudi Arabia and the UAE tend to be consolidated. When one distributor manages 200+ outlets for a brand, rolling out AI shelf photo analysis across all of them is a single decision. In the US, that same brand might work with 15 different broker networks. Getting them all to adopt the same tool is politics, not technology.
Oman and Bahrain are actually moving faster than I expected too. Smaller markets, but the distributors there are hungry for anything that gives them a leg up when pitching multinational brands. I had a conversation with a Manama-based distributor last week who told me his ability to send brands weekly AI-verified compliance reports was the reason he won a new principal.
That's the shift. FMCG merchandising compliance used to be a cost center. Now it's becoming a sales argument.
What to ask before you buy
If you're evaluating any AI shelf photo tool — Zivni, our competitors, whoever — here's what I'd push on:
How many of your specific SKUs does the model already recognize? Not "we support FMCG" — literally, how many. If they can't give you a number, that's a red flag.
What's the accuracy rate on Arabic-language packaging if you're in the GCC? Some models are trained mostly on Western products and their accuracy tanks on regional brands like Almarai, Nadec, or Al Ain.
Can the rep see the AI feedback in real-time in the store? Because if the analysis happens overnight in a data center somewhere, you've lost the moment when a rep could have actually fixed the shelf.
What happens on a bad connection in a store basement? A lot of tools break here. Anything worth using should queue photos locally and sync later.
And honestly, ask for a pilot. Not a demo. A pilot in 10 of your actual stores for 30 days with your actual reps. Any vendor worth their pricing will say yes.
We've done dozens of these. Some go great. Some expose problems in the brand's own planogram documentation that have nothing to do with our software. Both outcomes are useful.
The merchandiser in Deira with his 14 photos? He still takes photos. Just fewer of them now. And his supervisor spends her time actually calling stores where something's wrong, instead of drowning in JPEGs.
That's the whole point, really. The tech isn't replacing anyone. It's just finally letting the humans do the part of the job that actually needed a human all along.