Read the commerce platform announcements from the first half of this year, and you'd conclude the merchandising role is being dismantled. Agents that reorder category pages from a plain-language goal. Agents that write product copy. Agents that enrich attributes, generate descriptions, optimize promotions, and recommend the next assortment. One vendor's material describes freeing merchandising teams from spreadsheets; another describes automating merchandising decisions outright.

We build a commerce platform, so we're not neutral, and we've shipped plenty of automation ourselves. But the gap between how this work is described and how the merchandisers we actually talk to spend their weeks is now wide enough that it's producing bad buying decisions.

What Changed

The catalog acquired a second audience, and it doesn't browse. This is the real shift, and the coverage of it has been about discovery. The work it creates is data work. A meaningful share of product discovery now happens through systems that never render your page. They read structured attributes. Which means catalog completeness, historically an internal hygiene metric that nobody outside the team cared about, became a revenue variable. The long tail of SKUs with a one-line supplier description and three populated fields isn't just underperforming now. It's absent from consideration entirely, and the absence leaves no trace in your analytics.

Merchandisers inherited governance responsibilities they didn't ask for. When ranking was a set of rules a person wrote, explaining a product's position was trivial. Now that some portion of ordering comes from a learned system, someone has to be able to answer for it to a supplier whose product dropped, to a brand partner with a placement agreement, to a compliance team in a regulated category. That someone is the merchandiser. The tooling to answer has lagged well behind the tooling to automate.

Content production stopped being the bottleneck. Writing 4,000 product descriptions used to be a quarter-long project with a budget. It's now fast, and good enough for most of the catalog. This is a real gain. What it exposed is that description writing was never the hard part of merchandising. It was just the visible, expensive, easily-quantified part.

The feedback loop got slower and noisier. You could A/B test a PDP layout and know within two weeks. You cannot A/B test your way into visibility on a discovery surface whose ranking criteria aren't published, change without notice, and vary by query phrasing. Merchandisers who built their instincts on fast experimental feedback are now operating with considerably less of it, on decisions that matter more.

What Didn't Change

Judgment about assortment. No system currently in market decides what to carry. It can tell you what sold and model what might. Deciding to bring in a line because a category is shifting, or to drop a supplier whose quality is slipping, or to hold inventory through a soft quarter because the brand position matters more than the quarter. That's still a person with context that isn't in the data.

The politics. Much of a merchandiser's week is negotiation. Vendor placement expectations. The buying team's inventory position versus the marketing team's campaign calendar. Regional teams who want a different assortment than corporate approved. None of this is a data problem, and none of it automates. An agent that reorders a category page has, if anything, added a new party to negotiate with.

Ownership of being wrong. When an automated boost cannibalizes full-price sales or an inventory-driven suppression buries a partner's product for a week, nobody escalates to the model. The accountability sits exactly where it did before. That is the strongest practical argument for insisting on systems whose decisions you can inspect.

Compatibility, fit, and configuration. In categories where the right answer depends on what the buyer already owns (parts, industrial supply, telecom, anything configurable), merchandising is largely a data modeling problem that generative tools have not solved. Fitment either resolves deterministically or it doesn't, and no amount of fluent description covers a missing relationship.

Where the Work Went

The work moved from making things to setting conditions.

Less of the week goes to producing merchandising artifacts (copy, curated collections, manual sort orders) because that production got cheap. More of it goes to defining the conditions the automation operates inside: what the attribute model needs to contain, which constraints are inviolable, where a human has to approve, which categories the automation has earned and which it hasn't.

That's a more senior job than the one it replaced, and it's one most teams have not been resourced or trained for. The merchandiser who was excellent at hand-tuning a category page is now being asked to reason about data models, governance boundaries, and the failure modes of systems whose behavior they can't directly observe. Some will find that a promotion. Others will find it a job they didn't apply for, done with tools that weren't built for it.

Three Questions for Your Platform

Given all of that, the evaluation questions have changed. Three that separate the platforms in practice:

Can I add an attribute without waiting for you? Discovery surfaces will keep changing what they expect from product data, faster than any vendor roadmap moves. If extending your catalog schema requires a vendor release, you're structurally one cycle behind the surfaces that mediate your discovery. The question isn't whether the platform ships enrichment features. It's whether your team can model something the platform's designers never anticipated.

Can I see why a product ranks where it does? The decision path itself, not a performance summary. Which rules were active, who set them, when, and what the learned layer contributed. If the answer is a support ticket, you cannot answer for your own category pages, and every expansion of automated scope becomes a bet you can't hedge.

Can I constrain the system with rules it cannot learn around? Contractual placements, regulatory exclusions, brand safety. These belong in a deterministic layer the optimization runs inside, not as a training signal it treats as a preference. A platform that can only express these as soft weights will eventually violate one at the worst possible moment.

Those three questions are more useful than any feature comparison, because they're about what happens when the system does something you didn't expect. That moment decides whether you can keep using it.

The Mismatch

Merchandising is being restructured rather than automated away. The skills it is restructuring around (data modeling, governance design, constraint specification) are the ones vendor marketing has spent this year describing as things you will no longer need.

That mismatch has a cost. Teams buy automation on the promise of freeing merchandisers from tedium, then discover the merchandisers now need to reason about systems the platform won't let them see inside. Tedium was the cheap part. Judgment is what the job costs, and judgment requires visibility.

Automate aggressively, and insist on being able to see what you automated.