The product your filters can't see
2026-09-30
When a shopper ticks "Long sleeve", the site keeps every dress marked long and drops everything else, including the dresses where nobody filled in the field. She never sees them, and she has no reason to think anything is missing. The filtered list looks like the answer, so a short list reads as a thin range.
- Ecommerce
- Product Data
- UX
- Google Shopping
In my previous post, I spoke about the person who is responsible for product data; today I'd like to trace one blank field from the website all the way through to Google, since the cost involved is almost never taken into account and accumulates silently.
A filter reads an empty field as a no
Faceted filters work by exclusion. When someone ticks "Long sleeve", the site keeps every product whose sleeve attribute says long and drops the rest, including products where the field is blank. The system can't tell that a blank means nobody filled it in, so it treats the product as not matching.
In the case of 400 dresses, 70 of them have no given sleeve length and those 70 are not a random selection. According to my experience, they appear in new arrivals which have been introduced under time pressure, in partner and concession brands whose data was obtained directly from a supplier's spreadsheet, and in ranges that have been moved from an old platform. It is therefore the products you most want to see that are hidden by these gaps: the new season ranges and those in which you have made a commitment to a partner.
Someone who has applied two or three filters has usually done most of their deciding already. Baymard's usability testing found it wasn't unusual for participants to apply five or six filters in a session, and some applied as many as ten, to get a list down to a handful of real candidates. When the list comes back shorter than it should be, the shopper has no reason to suspect the data. She takes the list as the full answer. Daniel Kahneman described the tendency in 2011 as "what you see is all there is": people build a judgement from the information in front of them and rarely ask what's missing. A thin filtered list reads as a thin range.
The commercial impact operates in two ways: the session has a lower conversion rate since the product she originally intended to buy never showed up, and the invisible lines sell more slowly, which means they reach their markdown with more stock remaining. You incur the cost of the gap once in terms of conversion and once again in terms of margin, and this doesn't appear in any report that has product data in the title.
Three ways of saying navy
Blank fields are the obvious problem. Inconsistent ones are harder to spot. If one supplier sends "Navy", another "Navy Blue" and a third "Dark Blue", the colour filter shows three options with small counts where there should be one with a large count. The shopper picks one and misses two-thirds of what you have.
Facet counts carry more meaning than most teams realise. People read the number next to a filter value as a quick estimate of how much choice lies behind it, and they use it to decide which route is worth taking. Splitting one value into three makes each of them look thin. It also lengthens the list of options, adding effort at the exact moment the shopper is trying to narrow things down.
Baymard has written about this for years. Its 2018 benchmark found that 52% of sites weren't post-processing vendor-supplied product data into consistent, comparable values, and a 2019 article noted that missing filters are often a symptom of poor product data underneath. Both are more than four years old, and I couldn't find a newer public figure for the data-processing point specifically, so I'd treat them as directional. The more recent work points the same way. Baymard's 2025 product list benchmark rates 58% of desktop sites and 78% of mobile sites as mediocre or worse, and 51% don't offer all five of the filter types users rely on most: price, rating, colour, size and brand. One example in that article is Tesco's bedding category, where shoppers couldn't filter a list of 279 items by price, rating, size or colour.
You can't offer a size filter on bedding unless size is a clean, structured field across the whole range. Most of the filter problems I've been asked to fix turned out to be data problems wearing an interface costume.
The same gap, outside your site
Google is unusually direct about this. Its Merchant Center specification says product data is used to match products to the right queries and as a foundational input to its AI-powered ad formats. It lists missing or incorrect variant attributes, colour and size among them, as common problems that can stop ads and free listings from showing. For clothing targeted at UK shoppers, colour, size, gender and age group are required fields.
So the empty field that hides a dress from your own filter also weakens it in Shopping. A search like "long sleeve navy midi dress" can only be matched with confidence to a product that states those things in structured form. The listing that says exactly what the shopper typed is also easier to take in at a glance in a crowded grid of near-identical results. That's processing fluency: the easier something is to read and match against what you were looking for, the more relevant and trustworthy it feels. Better matching earns more relevant impressions, and more relevant impressions are where ROAS on Shopping comes from.
The same logic now applies to AI agents. Amazon's new Seller Assistant reads listings, stock and performance to decide what to recommend. ASOS said last week that recommendations inside its search convert materially better than its usual journeys. Any system like that can only reason about what's in the catalogue. An attribute that doesn't exist can't be matched, compared or recommended against, however capable the model reading it.
Filling the gaps with AI, carefully
This is where Tuesday's article comes back in. Generative models are good at reading a product image and description and proposing attribute values, and for many catalogues that's the only realistic way to close thousands of gaps before peak.
The risk is not equally distributed among the different attributes. Those that are visible, such as colour, neckline or sleeve length, involve a low level of risk since the model is looking at what the photograph shows and any incorrect value is easy to spot. The situation is different with regard to composition, fit, dimensions and care instructions. If a model states that a jumper is 100% merino when in fact it is a blend, or claims that a relaxed cut is slim, this results in a listing that looks good and creates an expectation which the product cannot meet. The shopper who has selected items based on the merino filter gets a blend instead, and that breach of expectation is all that is needed to lead to a return, a negative review, and the customer trusting your filters a little less in future.
Currently, Google is requiring merchants to use a distinct structured attribute to mark titles and descriptions that have been generated, which shows that the company considers such content to be a separate category. I would apply the same division within our own organisation, dividing them into three groups: attributes which AI can fill in and publish, attributes which it can only suggest for a person to approve, and attributes which it should not touch at all. Making a decision about where each of these categories belongs requires someone who has a good understanding of the different types of attributes so that they can identify where errors could become costly. That is why the Product Data Manager has this responsibility, and that is the reason I would never give this task to anyone who simply had a bit of spare time.
Measuring it so finance listens
Completeness is usually reported as one figure for the whole catalogue, which hides nearly everything that matters. A catalogue can be 90% complete overall and still be missing the one attribute most shoppers in a category filter by.
The method I have found to be useful assigns a weight to each attribute according to how frequently shoppers actually use it to filter or search, and it assigns a weight to each product according to the revenue it is expected to generate. This means that the gaps are ranked according to their commercial cost, which is why the team addresses sleeve length on new-season dresses before looking after the care instructions on last year's basics. It also provides the discussion with a figure that the finance director can keep track of, since this has usually been the case with product data.
The solution is seldom impressive. It might involve an improved supplier template, a validation rule when importing data, or having a few hundred attributes checked by someone who knows what a midi length really is. The benefits appear in areas that people already keep an eye on—conversion rates on filtered sessions, sell-through on new product lines, and ROAS on Shopping. Yet it is rarely given credit to the individual who carried it out.