The attribute layer: why good products lose to worse ones

2026-08-02

Ecommerce
Product Data
Merchandising
Strategy

A few years back, I spent most of a fortnight looking at customer service logs. No one ever volunteers to do that, but the same complaint appeared in different words, and after you've read a pattern set out thirty times in thirty different ways, it's difficult to stop seeing it.

Customers could not find a group of products we sold perfectly well. Live, in stock, sensibly priced, decent photography. The problem was that the thing which made them a coherent group in a customer's mind was not something our catalogue recognised. Our attributes had arrived from supplier feeds and had been tidied rather than redesigned. They described the products the way manufacturers describe products, which is not remotely the same as the way people shop for them.

We rebuilt the filtering around the language customers were actually using. Sales of that group nearly doubled. No change to the range, the pricing or the imagery. The only change was that the catalogue could finally answer a question it had previously been unable to hear.

Why a gap hurts more than a flaw

Information foraging theory, developed by Peter Pirolli and Stuart Card at Xerox PARC, models people searching for information the way ecologists model animals searching for food. The central idea is information scent: at every step, a person judges whether the cues in front of them, a link, a filter label, a listing, point towards what they want. Strong scent and they commit further. Weak scent and they abandon the patch, because the cost of persisting in a poor patch is high compared with the cost of moving to another one.

When applied to a product listing page, this means that the absence of a filter is no longer there; it isn't thirty seconds of extra work, it is the elimination of the signal which informed the shopper that they had indeed reached the correct page. The Nielsen Norman Group has observed this behaviour in usability sessions over the course of twenty years: people look for the particular cue that corresponds to their intention, and when that cue is missing they decide that the website does not carry the item they have come for, even though they are often on a page that does list it.

There is another effect, and it is less tolerant. When two products are placed side by side with one giving a specification and the other leaving the attribute blank, the blank becomes apparent. As Huber and McCann showed in their work on inference in consumer choice, people do not regard a missing attribute as being neutral; instead, they assign it a value, and this value generally tends to be low. The failure to state something about the material is interpreted as meaning that the material is of low quality, and the absence of information regarding compatibility is seen as indicating incompatibility.

Kardes and colleagues later described the flip side, omission neglect: away from a direct comparison, shoppers frequently fail to notice what is missing at all and form confident judgements on very little. Put those two findings together, and the implication for merchandising is uncomfortable. Thin data does not damage you evenly across the site. It damages you precisely where a competitor is standing next to you, which means comparison shopping, marketplaces and assisted search results. Which is where a growing share of decisions is now made.

The interface is where data quality becomes visible

Baymard Institute's ongoing benchmark of leading ecommerce sites finds that 38% still do not provide filters for information they already display in their own list items. When they first tracked it in 2015, the figure was 42%. A decade of investment has moved it four points, which tells you something about how the problem is usually diagnosed.

Baymard's own explanation is the useful part. Missing filters are generally a symptom of poor underlying product data rather than a front-end oversight. To offer filtering that behaves predictably, you need harmonised values, and supplier feeds arrive with branded feature names that have to be post-processed into common attributes before a facet can sit on top of them. Nobody enjoys funding that work. It produces no screenshot, no launch, and nothing a board meeting recognises as progress.

The fact that the failure remains unnoticed is precisely why it persists. When a customer navigates themselves into an empty result set, they don't file a complaint; instead, they visit another site, and your analytics show a bounce with no stated reason. The two simplest tools for diagnosis in the building are the internal search terms that yield no results and the filter combinations that produce no results. Whenever I have examined either of them, I have always discovered some demand that the business wasn't aware it had, demand which is typically expressed in words the merchandising team never would have used.

The second reader is not a person

Until recently, poor attribute data mostly cost you at the filter. That is no longer the boundary of the problem. Assisted and agentic shopping tools work by decomposing a natural language request into a set of constraints, then scoring products against those constraints. Waterproof. Under £150. Machine washable. Fits a 60cm gap. The system reads structured fields, not persuasion.

The failure mode here is categorically different from ranking lower. If a required field is empty, the product is not scored badly. It is not scored at all, because it cannot be confirmed to meet the constraint. It never enters the candidate set, and nothing about the experience tells the shopper it exists. NielsenIQ made the same point in its 2026 commerce work: invisibility to the systems consumers delegate to is a harsher outcome than a poor position in a list, because you were never in the running.

The volumes now justify caring. Adobe Analytics reported AI-referred traffic to US retail sites up 138% year on year in May 2026, the highest share of total visits since it began tracking the category in October 2024. Those visitors converted 54% better than non-AI traffic and produced 53% more revenue per visit, a complete reversal from a year earlier when the same traffic converted at roughly half the rate. Adobe's own benchmark also found product pages were less readable to AI systems than several other page types on retail sites. Amazon reported last week that active users of Alexa for Shopping came close to doubling in the quarter, with interactions up more than fivefold.

It is not necessary to believe that agentic commerce takes the place of browsing. All that is required is this narrower point: an ever-growing portion of demand is now reached via a stage which interprets structured fields, and that stage will not overlook your lifestyle photography and grant you the benefit of the doubt.

What it is worth

The commercial effect shows up in revenue per visitor rather than conversion rate, and that distinction matters when you present it. Better attributes do not persuade anybody. They put products in front of people who were already looking for them, which lifts the number of qualified product views per session and pushes sell-through into the parts of the range that were previously unreachable. In my experience, the long tail moves first and hardest, because that is where discovery was doing the least work.

There's also a margin issue, one that arises in a trading meeting. When a stock can't be found, it is subjected to markdowns on a fixed schedule. Each week that a product remains undiscovered is a week brought forward in which it will be sold at the discount you decided under time pressure. Enhancing discoverability is one of the very few methods by which revenue can be increased without having to buy traffic and without having to reduce the price.

Feed quality carries the same logic into paid media. Shopping campaigns match on attributes. Missing or inconsistent values reduce eligibility and mis-match queries, so you pay for clicks against the wrong intent. The catalogue is not just a merchandising asset, it is an input to media efficiency, and most businesses account for it as neither.

Who owns the catalogue

The honest answer in most organisations is that nobody does. Merchandising owns the range. Ecommerce owns the site. IT owns the PIM. The attribute layer sits in the space between them, which means it gets maintained to the standard required for a product to go live and no higher. That standard was set by whoever configured the platform, usually years ago, and it has never been revisited against what customers ask for.

What has worked for me is treating attribute quality as a trading function with a named owner, not a data project with a steering group. The definition of a good attribute set is not completeness against the supplier's specification sheet. It is coverage of the ways customers describe the category, which you get from search logs, service contacts and reviews rather than from the feed. Those two lists overlap far less than people expect.

After that, it is a question of gates and cadence. Nothing goes live without the mandatory attributes for its category, and the mandatory set is defined per category rather than globally, because the fields that decide a camera purchase are meaningless for a sofa. Then you review the failed searches and the empty filter results on a fixed rhythm and feed them back into the taxonomy. It is unglamorous work with an awkward property: when it is done well, the result is that nothing goes wrong, and nothing going wrong is difficult to claim credit for.

Which is probably why it stays broken in so many businesses that are otherwise run well.