The AI Merchandiser: Act 3 - The Fate of Measurement

If you’ve made it this far in my three-act allegory about the changing nature of merchandising in the AI era then you’re in for some raw data insights and concluding thoughts on where we stand today. See, the entire agentic commerce opportunity hinges on retailers ability to prove that this new, emerging channel is worth investing in. I believe it is and I’ll explain why.

Quick Recap

In Act I, I argued that access is the merchandiser's first job: LLMs can't recommend what they can't read.

In Act II, I argued that the work itself has changed shape: Merchandising is now signal design, and confidence is the currency of influence.

Which leaves the question every merchandiser is asked about four weeks after they start: Did it work?

We've spent the summer at Fabric answering that question across multiple enrichment engagements. I’m going to talk about measurement in broad terms with some pointed insights from brands in lifestyle apparel, designer fashion, apparel retail and home furnishings.

We used common measurables from weekly channel data. We’ve observed three patterns and they were consistent enough across very different catalogs.

Let’s Go!

(1) Signals surface first where machines read your catalog directly

Where does enrichment actually show up?

Organic Shopping, consistently and earlier than anywhere else.

Two brands saw significant Organic Shopping volume and enough post-publish history to measure conclusively. The add-to-cart rate on enriched products rose between 35% and 60%. The rest of the catalogs over the same weeks remained flat to modestly up. Both Organic Shopping lifts date to the publish week and not before it.

What we observed wasn’t just a coincidence of channel mix. It makes a lot of sense when you think about it. Organic Shopping is the surface fed most directly by structured product data. Titles, categories, variant relationships and attribute values all go into a feed a machine reads with no human inbetween. It is the shortest path from a signal you designed to a decision made about your product.

The Takeaway: If you want to know whether your signal design is working, look first at the machine-readable surface that already carries volume. When volume begins routing directly through to AI surfaces via protocols (UCP / ACP), your signal design better be ready. We’re predicting the first clear marker of this to be Holiday ‘26.

(2) Conversion moves before revenue does

What moves first, and when?

This one is a bit of good news / bad news but I believe it’s important to look at our findings honestly.

On every engagement the same shape appeared: add-to-cart rate on enriched products went up while revenue per cart went down.

Two explanations fit. Either enrichment is pulling a wider, earlier-stage audience into the cart, which lifts cart rate before it lifts basket value. Or, those products were discounted over the period. We have anecdotal feedback that the latter is true.

But, the sequencing lesson holds either way. Enrichment acts on discovery and consideration. Revenue is downstream of assortment, price and stock — three things the merchandiser is also changing every week for reasons that have nothing to do with AI.

So the practical shape of a measurement plan looks like this:

• First movement appears at one to two weeks as re-indexing not results.
• Confirmed conversion effects took us four to twelve weeks to establish.
• Don't report a number before eight weeks. Re-read at twelve.
• The KPI is add-to-cart rate on enriched products against a comparison group you defined before you published

The Takeaway: If you take one thing from this post it’s the timeline. Most enrichment efforts get judged on revenue at week four, which is roughly like judging a crop by digging it up.

(3) The AI channel is responding to signals, and it's still early

What about AI traffic itself?

This is the channel everyone wants a number for and the number that matters right now is the direction not the size.

On a designer womenswear brand, labeled AI sessions on enriched products rose roughly 3.5x after publish. The step lands one week after the publish date and nowhere before it. Shoppers don't behave on an indexing schedule. The schema is being found, read and acted on. And it responds to what you change, fast.

What AI shopping isn't yet is a mature channel. Labeled AI is a smaller observed slice of traffic than Organic Search or Shopping today for the obvious reason that agentic shopping is a year or two into its life rather than twenty.

But, it is compounding in every dataset we have and everything happening upstream points the same direction. This is a channel in its early growth phase being measured with instrumentation that’s also in its early growth phase.

The share of your catalog carrying rich, structured signals sets the ceiling on how much of that growth you're able to capture. By enriching a subset today you're not chasing AI revenue, rather, you're building the surface area for a channel that hasn't peaked.

The Takeaway: Judge the AI channel on whether it responds to your signals not on what it contributes this quarter. Responses mean you're well positioned. That's the leading indicator and leading indicators are the whole point of measuring early.

Closing the Loop

Over three acts, from access to signal design to measurement, we’ve explored how the role of commerce merchandising is shifting.

Access was a technical problem: Can machines read your catalog at all?

Signal design is a design problem: Is there enough coherent evidence there to reason over?

Measurement is an experimental problem: Can you tell the difference between what you did and what the season did?

Merchandisers have absorbed the first two remarkably fast. The third is the one I'd flag now because it's where good work quietly loses its funding.

If the moment of influence has moved off your site and into a model's reasoning then the proof of influence has to be designed with the same intent as the signals themselves.

Think of it as a channel where the mechanism is visible. Start with a comparison group defined before you optimize and a watch metrics move in weeks rather than quarters after you publish.

Strong signals get you identified.

Rich signals get you reasoned over.

Trustworthy signals get you recommended.

Honest measurement gets you the budget to keep doing it.

Thanks for reading my series.