Bol.com

Finding differentiation through reviews (with AI analysis)

Lars Hurkmans·Co-founder·11 February 2026·Reading time 10 minutes

You find differentiation by analysing the reviews of existing products, supplemented with the filterable specifications. Collect the reviews in bulk via the review feed in the panorama or niche explorer, separate positive and negative, and have an AI tool analyse them along a SWOT structure. Always validate the outcome against the real reviews.

This is part 6 of our product research series. In part 4 we found two candidates through the niche explorer: computer glasses (low ticket) and outdoor wildlife cameras (mid to high ticket). In part 5 we looked at whether there's room for a new player. In this article we really go deep for the first time with specific products, their specifications and their reviews.

What is differentiation and what does the research consist of?

Differentiation is the way your product is better or different from the existing supply. You research it along two lines: the reviews of the current products and the filterable specifications on the category page. At this stage we go deep on the reviews and keep the specifications light for now.

That distinction between the two lines is deliberate. Fully digging through the specifications takes a lot of time, and we do that in the deep-dive specification research in the next video. Right now we mainly want to know whether there are even enough different specification options at all to differentiate on. We do tackle the reviews thoroughly, because that's where the direct feedback from consumers lives.

Why are reviews the richest source for differentiation?

Reviews give direct feedback from consumers about what does and doesn't go well with the current supply. From that you get two things at once: the baseline level you need to at least match yourself, and the weak points you can differentiate on. Reviews also weigh heavily in the purchase decision: according to Thuiswinkel.org, over 90% of consumers read online reviews before buying something.

That feedback works on two levels. At the technical level of the product itself, you see which specifications don't work well or are entirely missing. At the listing level, you see which points you want to bring out strongly in your images and text, in other words where you do better than the rest.

There's one more downside for a new seller. Once consumers compare products with each other, your product can be just as strong as the rest, but you don't yet have a name and haven't built up reviews yet. Buyers therefore can't validate their gut feeling against your ratings, while competitors often already have those reviews. That's why you need an edge at the start, and you get that edge from the weak points of the existing supply. How you then get reviews yourself is covered in the importance of reviews on bol and how you get them.

How do you collect all the reviews without going through each product separately?

Via the review feed. Instead of opening each product separately on bol and scrolling down, you pull all the reviews into one feed via the panorama or the niche explorer. That saves a lot of time compared to the way many bol sellers currently do it, namely reading through reviews product by product and summarising them in an Excel or Word document.

We'd already saved the products from part 4 in the product radar, with a group of relevant products per niche. From the panorama or the niche explorer you then click on analyse reviews and get all the reviews listed one below the other in a feed. There's a small difference between the two entry points:

• Via the panorama you choose a category, for example computer glasses, and see the review feed of all products that belong to that category.
• Via the niche explorer you type in a keyword and get the reviews of the top thousand products found for that keyword.

In this article we continue from the category in the panorama, because that's clearest for this research.

How do you separate positive and negative reviews and which period do you choose?

Set the chart above the feed to display per score and choose a current period. This way you see the split between positive and negative reviews per time period, and you then filter on the scores you want to read. On bol, most reviews are predominantly positive, so the darker periods with one, two or three stars are exactly where the feedback is.

You choose the period based on relevance. Reviews from two years ago can be about products that have since changed and are then less current for the present catalogue. In this article we choose the last six months, which in the example yielded 223 reviews. If a category yields too few reviews, you go further back; if you already have enough, you stay close to the present to stay current.

Then you make two groups. Negative reviews are the one to three star scores, positive reviews the four and five stars. From both groups you take the full list, including the review text itself, because the AI tool needs that text.

How do you write a good AI prompt following a SWOT structure?

Give the model the goal first, then the research questions, then the desired output. The prompt determines how good the analysis turns out, so you set it up carefully. The core is a SWOT-like structure: the strengths of the current supply, the weaknesses, the improvement opportunities, the USPs for your listing, and the question of whether there's market room.

Concretely you build the prompt like this:

• Goal. You want to analyse the reviews to determine whether there's market room for you as a seller on bol.
• Research question. The first question is whether there's room for new players. Specify that with objectives: have the model analyse the shortcomings of the current products, name the real improvement opportunities, and check whether the complaints are representative. You want to know whether multiple products get the same negative points, not that one product with many complaints skews the picture. Close with the conclusion: is there market room, yes or no?
• Strengths. Which characteristics of the existing products are found good? That's the baseline level you'll probably also need to offer.
• Weaknesses. Which problems come out of the negative reviews, and are there already products that solve them? This is where the difference you can make lies.
• Improvement opportunities. Which unsolved points can you implement yourself? Those are the differentiating specifications.
• USPs for the listing. Which advantages do you reflect in your listing, on the technical side, on solving problems and on findability, in other words what people actually search for?
• Report format. Indicate what the output should look like, so you get back a usable report.

For the analysis itself you can use ChatGPT, Gemini or Claude. In this article we use Claude, because the model can handle enough context to read in and process all the reviews at once. So choose a model that can handle that amount of text. If you don't have an AI tool, you follow the same steps manually; that often gives you a better feel for the reviews, but takes more time.

Why do you always validate the AI analysis against the real reviews?

Because an AI model can make mistakes or make up points. You go through the report, but then scan the real reviews to check whether the mentioned points really appear in them. With the knowledge from the report, that scanning goes a lot faster.

For the computer glasses, the model reported 92% positive and 8% negative across 204 analysed reviews. Despite that positive sentiment, real improvement opportunities did come out. The clearest point, mentioned seven times, was that the glasses don't block blue light, while that's exactly the function people buy them for. On top of that there was the quality of the frame and the lens, and frames that are too one-size-fits-all. The strengths that came up often were use before going to sleep and an acceptable price-quality ratio.

The advice came down to two things: a better frame (in more durable material) and better lenses with an anti-reflection and anti-scratch coating. According to the model, that solves about 90% of the negative points. The model also made things up, such as a cleaning kit, and gave price segments that weren't right because it didn't have all the data. That's exactly why you validate.

Back in the review feed, the first review immediately confirmed that the blue light wasn't filtered out, with a plastic frame that had quality issues. A longer review mentioned the same thing about the blue light and the frame, and there was even a review that used the word carnival for the cheap look. The improvement points from the report thus matched well with what consumers actually wrote.

What did the review analysis deliver for the wildlife cameras?

For the wildlife cameras, five critical improvement points came up, with enough structural complaints to see market room. Here the model analysed 100 positive and 29 negative reviews and reached a conclusion faster than with the computer glasses. The five points:

• Wifi range. With a number of cameras the range is limited, and according to the analysis no competitor solved that.
• Genuine 4K. Some cameras claim 4K but deliver jerky footage, forcing you to switch back to 1080p. For that kind of camera you specifically don't buy a 4K model.
• Detecting small animals. Certain models struggle to register mice and birds.
• Battery. The battery runs out fast on some cameras.
• Weather resistance. Water or moisture sometimes gets in, so the camera isn't fully weather-resistant. That's what the IP rating is for; you then do need to genuinely comply with it.

On the positive side were points you'd better keep: the complete package with batteries, SD cards, cables and a tripod, ease of use, good image quality on the cameras that genuinely deliver 4K, a fast-responding sensor and a Dutch-language interface. That interface was notably strongly appreciated, so there's differentiation to be made there.

Based on that, the USPs for the listing are obvious: emphasise that you're genuinely 4K and don't interpolate, that you have a large wifi range, that the battery lasts, that you detect all animals and that everything is in Dutch. The validation in the review feed confirmed most points, from the wifi range and the detection of small animals to the fake 4K. One complaint came up that wasn't in the analysis: the build quality and a lacking manual. That too argues for keeping checking yourself.

How do you read the filterable specifications on the category page?

The filters on the left of the category page are sorted by what consumers search for most, from top to bottom. This way you see at a glance which specifications you can differentiate on, and which ones you can be found by. That findability matters, because the filters that get clicked on a lot are the ways people need to be able to find your product.

For the computer glasses, lens strength is at the top, followed by colour, price, reviews, frame shape, target group and brand. So the three base specifications to include are strength, colour and frame shape. Per specification you can check in the data whether products with that setting also sell: if it shows zero or unknown, that specification isn't relevant enough to steer on.

For the wildlife cameras, the order is different. Night vision is an important type of feature here, and price comes right after it, which differs from the computer glasses. On top of that you differentiate on connectivity (the wifi capabilities we already heard about from the reviews) and on resolution, where Full HD and 4K are at the top and 4K stands out above the rest in the data. Not every specification is filterable: things like an included tripod or an SD card are present in the products but not in the filters, so you research those separately later.

Both product groups therefore deliver both specifications from the reviews and filterable specifications, and you save that combination in the product radar. If there are enough improvement points and enough filterable specifications, you move on to the deep-dive specification research in the next step.

Summary: differentiation from reviews

• Research along two lines. Reviews (thoroughly) and the filterable specifications (lightly at this stage); together they show where you can differentiate.
• Collect reviews in bulk. Use the review feed in the panorama or niche explorer instead of going through each product separately; then separate positive and negative reviews and choose a current period.
• Analyse with a SWOT prompt. Provide the goal, research questions and output format: strengths, weaknesses, improvement opportunities, USPs for the listing and the question of whether there's market room.
• Always validate. An AI model can make mistakes or make something up, so scan the real reviews to check whether the points hold up.
• Read the filter order. The filters on the category page are sorted by most searched; that points you to both differentiation and findability.