Why Your Recommended Rail Keeps Showing the Same Five Products

The rail is supposed to surface the long tail. Instead it shows the bestsellers, every time, on every page, to every shopper. That is not a bug in the algorithm. It is the algorithm doing exactly what it was trained to do.

The problem is called popularity bias. Recommendation algorithms learn from past behavior, and past behavior is dominated by the products that were already popular, so the algorithm recommends the popular products again, which makes them more popular, which makes the algorithm recommend them more. It is a feedback loop, and it is why the rail shows the same five products. The fix is to measure what share of recommendations go to the long tail, and to treat a rail that only shows bestsellers as a broken rail, not a working one.

The feedback loop, in plain terms

A recommendation engine learns from what shoppers clicked and bought. The products with the most clicks and purchases are, by definition, the products that were already visible and already popular. The engine sees that signal, concludes those products are what shoppers want, and shows them more often. More exposure brings more clicks, which brings more exposure, and the loop tightens. Researchers call it popularity bias, and the mechanism is simple: the algorithm focuses on already-popular items because the training data is full of them, and the long tail starves because it never gets the exposure that would generate the data.

The promise of recommendations was the opposite. The long tail was supposed to get its moment, because an algorithm could surface a niche product to exactly the shopper who would love it, without the cost of a shelf. That promise has not held up in practice. Studies of deployed recommenders find that the long tail stays invisible, and the same handful of products dominate every rail on the site. A store that measures its recommendations will find the same thing: the top five products take a share of impressions that no honest reading of shopper taste can explain.

Related reading on this site: PersonalizeIQ and Personalization on the PDP: Showing the Right Variant Before the Shopper Scrolls cover the neighbouring problems in more depth.

Content-based filtering makes it worse in a specific way

The other common approach, content-based filtering, recommends products similar to the ones a shopper already looked at. That sounds sensible, and it is, until you notice what it does to discovery. A shopper who looked at black boots gets more black boots, and then more black boots, and the rail becomes a hall of mirrors. Researchers describe this as a homogeneity problem, or a filter bubble: the recommendations converge on a narrow set of similar items, and the shopper never sees anything different enough to be interesting.

Content-based filtering also fails new users, because it needs a history to work from. A first-time visitor has no history, so the system falls back to the most popular items, which are the same five products again. The new shopper, who is the person most in need of discovery, gets the least discovery. That is the tell that the rail is not a discovery tool at all. It is a popularity amplifier with a discovery label on it, and the label is doing more work than the algorithm.

What a healthy rail actually looks like

A healthy rail shows a mix. The bestsellers have a place, because some shoppers do want the popular product, but they should not take every slot. Measure the share of recommendations that go to products outside the top 20 by sales, and set a floor for it. If the long tail takes less than a fifth of the impressions, the rail is a bestseller list, and the store should say so rather than calling it personalization. The metric is easy to compute and hard to argue with, which is why so few stores publish it.

The second check is the new visitor test. Open the store in a fresh browser, with no history, and look at the rail. If it shows the same five products as the rail on a loyal customer's page, the system is not personalizing. It is broadcasting. A real recommendation system should look different for a stranger than for a regular, and a system that cannot tell the difference has nothing to personalize with. Fix the rail by feeding it signals beyond sales rank, and by measuring the long tail share every week, because a rail that is not measured will drift back to the bestsellers within a month.

Frequently asked questions

What is popularity bias in recommendations?

It is the tendency of recommendation algorithms to focus on already-popular items, because the training data is dominated by them. The algorithm recommends popular products, which get more exposure and more clicks, which makes the algorithm recommend them more. Researchers call it a rich-get-richer feedback loop.

Why does my recommended rail show the same products to everyone?

Because the algorithm is learning from past sales, and past sales are concentrated on the bestsellers. The long tail never gets the exposure that would generate the data to recommend it, so the rail converges on the same handful of products.

Does content-based filtering help with discovery?

Not much. It recommends products similar to what a shopper already viewed, which narrows the rail into a hall of mirrors, and it falls back to the most popular items for new users who have no history. That is the opposite of discovery.

How do I know if my recommendation rail is working?

Measure the share of recommendations going to products outside the top 20 by sales, and open the store in a fresh browser to see what a new visitor gets. If the long tail takes less than a fifth of impressions, or a stranger sees the same five products as a loyal customer, the rail is a bestseller list, not personalization.