An AI Mission for Retail: Assortment Gap Analysis

Every retailer's analytics stack is a complete record of what customers bought from what was offered. It is a perfect record of nothing at all about the customer who looked, found no version of the thing they wanted, and left without leaving a row anywhere.

A merchant sits down on a Monday with the category review pack in front of her, and the numbers are beautiful. Sell-through by style, margin by vendor, rate of sale by store cluster, week-over-week comps against last year, the whole apparatus rendered in clean tables that took a team of analysts most of a week to assemble. She can tell you which of last season's colorways underperformed and by how much, which price point carried the category, which size curve ran short in the Southeast. What she cannot tell you — what the pack structurally cannot tell her, no matter how many analysts she adds — is what the customer wanted that she never carried. Somewhere in the last quarter, a meaningful number of people came looking for something in her category, found nothing that matched, and went elsewhere. Not one of them appears in the review. They are not a small number in the tables. They are not in the tables.

This is the shape of the problem, and it is worth being precise about it because it gets confused with a data quality issue when it is nothing of the sort. Point-of-sale data is not incomplete or dirty or under-instrumented. It is doing exactly what it was designed to do, which is to record transactions, and a transaction requires that a thing was on offer for someone to buy. The moment you ask sales data a question about demand for something you never stocked, you have asked it a question it has no vocabulary for. It will answer anyway, in the only way it can: with silence that looks like an absence of demand. And a merchandising organization that reads silence as evidence will keep buying more of what already sells, season after season, tightening around a version of its customer that is defined entirely by what that customer was previously willing to settle for.

The record only exists where the offer existed

The trap here is subtler than "we don't know what we're missing," because retailers have always known, in a hand-waving way, that they're missing something. The trap is that the missing part is not merely unmeasured — it is unmeasurable by the instrument in use, and the instrument is the same one that produces every number anyone trusts. This creates a quiet asymmetry in how decisions get made. A proposal to expand a size range, add an occasion, carry a material the category has never carried, or open a price tier below the current entry point arrives at the review with no data behind it, because the data that would support it could only have been generated by making the change. Meanwhile the proposal to deepen an existing bestseller arrives with a beautiful chart. One of them is arguing from evidence and the other is arguing from a hunch, and everyone in the room knows which is which, and so the assortment converges. It converges not because anyone decided narrowness was strategy but because the measurement system quietly taxes every idea that hasn't already been tried.

Notice that the conventional escape hatches don't actually escape. Market-level syndicated data tells you what the category did in aggregate, which is useful for sizing but arrives too coarse and too late to say what your customer, standing in your store or on your site, wanted from you last Tuesday. Customer surveys ask people to reconstruct intentions they mostly don't remember, and reward the articulate over the representative. Competitive assortment scraping tells you what other people chose to stock, which is a record of their guesses, not of demand — copying it just means inheriting someone else's blind spot alongside your own. Each of these is a real input and none of them closes the gap, because all of them are still trying to observe presence. The thing you need to observe is an absence, and absences do not sit in a table waiting to be counted.

Absence has to be inferred from the traces it leaves

Here is the reframing that changes the work: the customer who left without buying did not leave no evidence. They left evidence that is scattered, indirect, and living in systems nobody thinks of as demand systems. They typed something into your site's search box and got a results page with nothing on it, or worse, a results page full of things that weren't what they asked for — and that query is sitting in your search logs right now, in your own first-party data, uncounted. They asked a store associate for something in a wider fit or a smaller pack size, and if that associate wrote anything down at all it went into a note field or a chat transcript. They bought the closest available substitute and then returned it, and the return reason they selected — too tight, wrong length, not what I expected — is a demand signal wearing the costume of a logistics record. They asked a question in a service conversation, they abandoned a cart after filtering to a spec you don't carry, they wrote a review of the thing they settled for that explains what they actually wanted. None of this requires following anyone around or buying data of murky provenance. It is all information those customers deliberately gave the retailer, in the retailer's own systems, in the ordinary course of trying to be served.

The reason it goes unused is not that it is secret. It is that it is unstructured, contradictory, spread across a dozen systems with no common key, and enormously laborious to read. Search-with-no-result logs run to millions of rows of misspellings and junk; the fraction that represents a genuine unmet need has to be separated from noise by someone who understands the category well enough to know that a cluster of queries about a fabric weight and a cluster about a temperature rating are the same customer asking the same question in two vocabularies. Return comments have to be read, not tallied, because the meaning lives in the phrasing. Substitution patterns only become interpretable when you know what was out of stock at that store on that day, which means joining demand evidence to inventory history that lives somewhere else entirely. This is genuine analytical work, and it is the kind of work that never gets staffed, because it produces a maybe rather than a number and it competes for hours against the category review that has a deadline.

Treating the gap as a first-class object

What changes the economics of that work is a system that can do the reading. Not a dashboard that ranks null-result queries by frequency — that is still a place a human has to go and interpret — but something with the capacity to hold a category's structure in mind and go looking for the shape of what isn't there. This is what makes assortment gap analysis a good candidate for an AI Mission rather than a report: the task is not a fixed sequence of queries but an investigation, where specialist agents pull null-result searches, substitution behavior, return reasons, service transcripts, review text and supplier line sheets from wherever they live, and a reasoning layer works out which of the resulting fragments are the same latent need expressed differently, tests each candidate gap against the retailer's own enterprise knowledge of what it has tried and deliberately declined to carry, and hands the merchant a short list of hypotheses with the evidence trail attached. The merchant still decides. What has changed is that the case for the thing you don't stock now arrives at the review with something behind it, instead of arriving as a feeling.

It matters that this be built as an investigation rather than a workflow, and the distinction is where a lot of retail AI has gone wrong. Gartner has predicted that over forty percent of agentic AI projects will be canceled by the end of 2027, naming unclear business value and "agent washing" — older rule engines and dashboards relabeled — among the reasons. A fixed pipeline that surfaces the top hundred zero-result search terms is exactly that kind of relabeling; it finds only the gaps whose evidence happens to live in the one system it was pointed at, which is to say the easy ones. The gaps that matter are the ones assembled from three weak signals in three unrelated places, and finding those requires something that can follow a thread it was not told about in advance. That capacity — reasoning across systems rather than executing a fixed path across them — is the actual dividing line in what the category publication on the autonomous enterprise has been documenting as the difference between automation and autonomy, and it is why platforms like StudioX frame this class of problem as a mission with a question at its center rather than a scheduled job with a query at its center.

The mental model worth carrying out of this is that a retailer's assortment is not a list of products, it is a hypothesis about demand, and every hypothesis has a boundary. Inside the boundary you have transactions, which is to say measurement. Outside it you have customers who came, looked, and left, and the only thing they leave behind is the shape of the hole they walked into. Most organizations have no object in their data model for that hole; there is no table, no owner, no line in the review pack. Creating one — deciding that the gap is a thing you name, track, staff, and argue about with evidence — is a bigger change than any tool, because it converts the most expensive category of commercial mistake from something invisible into something you can be wrong about out loud. The retailers who make that move will stop asking their sales data what customers want, which it was never able to answer, and start asking it only what customers bought, which was always the only question it knew.