Price comparison sounds simple. You take your product, find it in your competitors’ shops and see who is cheaper. In practice you stumble at the very first step. The same product carries a different name in every shop, some have no EAN, photos are shot from different angles and sometimes it is sold as a bundle or in a different pack size. Before you can compare prices at all, you have to be sure you are comparing the same thing.
That work is called product matching. It is the hard part of price monitoring, not reading the prices themselves. In this post we walk honestly through how Comprice finds the same product across shops, what breaks naive matching and why no single method is enough. Product matching runs in layers, they cover each other’s weaknesses and every match gets a confidence score.
Why product matching is the hard part of price comparison
Picture a simple example. You sell “Nivea Men moisturising face cream 50 ml”. A competitor calls the same thing “NIVEA MEN Creme 50ml moisturising” and a third shop just “Nivea men’s cream”. Three spellings, one product. A human sees the same item, but a computer sees three different lines of text.
The naive approach tries to match names word for word. It breaks the moment:
- names are written in a different order or with different capitalisation
- one shop puts the brand first, another last
- the pack size is written in a different place or a different unit
- the product is sold as a bundle or a multipack
- the EAN is missing or wrong
- photos are shot on a different background and from a different angle
When product matching gets it wrong, the whole rest of the price comparison gets it wrong. A false match means you compare a 50 ml cream with a 100 ml cream and draw wrong conclusions. That is why matching accuracy matters more than any pretty chart built on top of it.
The matching ladder: EAN, image, text, meaning, human
Comprice does not rely on one trick. Matching works as a ladder where each rung is more flexible than the last, but also more demanding about verification. Seen from the top it looks like this.
- EAN or GTIN. If both shops carry the same barcode, the match is fast and exact. This is the first and most certain rung.
- Image recognition (CLIP). When the EAN is missing, the AI compares product photos. CLIP recognises the same product even from a different angle or background.
- Text recognition (OCR). If there is text on the packaging in the photo, OCR reads it out. That way brand, product name and variant can be recovered from the image when the shop’s text field is incomplete.
- Semantic analysis. The AI splits the name into three parts: brand, product and variant. Then it compares them by meaning, not by spelling. “Men’s cream” and “Men Creme” come together because the meaning is the same.
- Human verification. The highest confidence tier passes under a human eye. This rung is meant for cases where the machine is confident but an error would be costly.
How the layers cover each other
The point of the ladder is that no single rung has to solve everything alone. The EAN is fast but often missing. The image helps when the text is messy. The text helps when the image is poor. Meaning ties the ends together when spellings drift apart. When one layer stumbles, the next picks up the work. This layered design is what makes matching reliable even on messy data.
Confidence score, not magic
Matching is not a “yes or no”, it is a confidence score. Every pair found gets a rating of how likely it is to be the same product. A high score goes into use automatically. A lower score waits or goes for review. The highest tier is human-confirmed.
That is why we speak honestly about accuracy in the range of 92–99% depending on the verification level, with the top tier human-verified. We are not claiming the machine never makes a mistake. We claim the system knows how confident it is and, when in doubt, does not push a false match on you. That is the difference between market and product intelligence and blind automation.
Comprice is also deliberately advisory only. It monitors public prices and stock and gives you a picture of the market, but it never changes your prices itself. You make the decision. Read more about how product matching works in our system, and how it ties into the whole of online price monitoring.
What breaks naive matching
When someone promises you price comparison, it is worth asking how they link products together. The common traps are these.
- Relying on the EAN alone. Many shops have no barcode or a broken one. A system that stands on the EAN alone misses a large part of the market.
- Relying on the name alone. Text matching breaks the moment a shop writes the name differently. You see an incomplete picture and assume the competitor does not sell that product.
- Mixing up variants. 50 ml and 100 ml, single and multipack, different colour shades. When the variant is lost, the point of the comparison is lost too.
- Automation without checks. A system that shows no confidence score and lets no human review a doubtful match hides its own errors from you.
Comprice built the ladder precisely around these traps. If you want to see how to use it day to day for tracking competitors, see the guide on how to track competitor prices.
What this gives you as a retailer
Good product matching means your price comparison stands on solid ground. You see the same product even when a competitor has listed it under a different name. You do not mix up variants. And you know, for every match, how confident the system is, because the confidence score is visible, not hidden.
That frees you to focus on the right questions. Where you are more expensive on the market, where cheaper, where a competitor is out of stock and where you have room to grow your margin. All of it assumes the matching underneath holds. Once that base layer is in place, the rest of your market and product intelligence becomes trustworthy.
See how much of your assortment is found in other shops: Try it free. If you want to look inside the system first, explore the product matching feature.