AI Merchandising Ranked Your Sold-Out Hero to the Top

AI merchandising means letting software decide which products show first - on your collection pages, your homepage, and your search results - instead of a person dragging items into order. This summer it became a native, on-by-default feature rather than an app you install. The catch is that it ranks on the signals it can see, mostly clicks and sales velocity, and it usually cannot see the three that decide whether a ranking actually makes money: whether the product is in stock, what it really earns after every cost, and how often it comes back as a return.

For years, merchandising was a paid app or a manual chore. A member of your team decided what led each collection, or you rented a tool from the app store to do it with rules and machine learning. That changed in the 2026 platform updates, when AI merchandising moved into the core admin: automatic collection sorting, a featured-products picker, and smarter search ranking, switched on by default. The pitch is obvious and appealing - the store merchandises itself, and you can retire an app or two. This post is about the gap between "the store merchandises itself" and "the store merchandises itself well," and the back-office data you need before those two sentences mean the same thing.

What "AI merchandising going native" actually changed

The mechanics are simple enough. Instead of a fixed manual order or an external app, the platform now watches shopper behaviour and continuously re-ranks your products: the ones getting clicked and bought rise, the ones being ignored sink. It does this across the surfaces that drive most of your discovery - the collection and category pages, the homepage curation, and the on-site search results. For a lot of stores this is a genuine upgrade over a static hand-sorted grid that nobody has touched since launch, and it costs nothing extra to turn on.

Here is the part the release notes skip. A ranker is only as good as the signals it is allowed to optimize on, and the default signals are engagement signals - clicks, add-to-carts, sales velocity. Those tell the model what shoppers want to look at. They say nothing about whether you can actually deliver that product, whether selling more of it helps or hurts your bottom line, or whether it will bounce straight back to your returns desk in two weeks. The AI is not broken when it ranks a popular-but-problematic product to the top. It is doing precisely what you asked. You just asked the wrong question by feeding it only half the data.

Merchandising is not the same as recommendations

It is worth drawing a clean line here, because the two get blurred. Recommendations are one-to-one: the "you might also like" row on a product page, the cross-sell in the cart, personalized to the individual shopper. Merchandising is the store's curation layer: the order products appear in on a collection page, which items lead the homepage, how search results rank - the decisions that shape what most visitors see, whether or not they are logged in or personalized to. We have written before about getting one-to-one AI recommendations right so they lift average order value; this is the different, broader layer sitting underneath it. Both can be automated well, and both fail the same way when the data behind them is thin - but merchandising fails more visibly, because it decides the first thing every shopper sees.

The three signals your ranker probably cannot see

When an AI merchandiser hurts you, it is almost always because one of three facts never reached it. Each one turns a "top product" into a mistake.

• Live stock. The most common own goal. A product sells fast, so the model ranks it up, and it keeps ranking it up right through the moment it goes out of stock, because the engagement signal lags reality. So your best real estate points at a sold-out hero, sending buyers to an out-of-stock page or a back-order they did not expect. The fix is not a merchandising rule; it is a live stock feed the ranker respects. Real-time inventory sync across every channel and location is what lets it demote or hide a product the second it runs dry - and promote it back the moment it is replenished.
• True margin. Velocity and profit are not the same thing, and often they are opposites. Your fastest-moving SKU can be your thinnest earner once you subtract cost of goods, landed cost, shipping, and payment fees - and a ranker chasing clicks will happily push volume through your worst-margin product. It cannot know better, because real margin lives in your ERP, not your storefront analytics. Feeding it true margin per SKU and per channel is what turns "most clicked" into "most clicked among the things worth selling."
• Return rate. A gross sale is not a kept sale. A jacket that sells brilliantly and comes back forty percent of the time is a logistics bill wearing a bestseller badge, and an engagement-only ranker cannot tell it apart from a product people keep. It counts the outbound order and never sees the inbound return. Only when per-SKU return rates are joined onto the catalog can the model rank on kept revenue instead of gross orders - the same hidden-loss lens we applied to which products actually lose money after returns.

Notice the pattern: none of these are settings inside the merchandising tool. They are facts that live in other systems - the warehouse, the accounting ledger, the returns desk - and the ranker will never account for them until they are wired onto the product it is ranking. That is an integration job, not a configuration one.

What margin-aware, stock-aware merchandising needs wired up

The goal is not to switch AI merchandising off - it is a good default. The goal is to feed it the three missing facts so its ranking optimizes for kept, profitable, in-stock demand rather than raw clicks. In practice that means four connections behind the storefront.

First, a live inventory signal from the system that actually knows your stock. For most brands that is the ERP or warehouse system, not the storefront, and the two drift apart constantly. Keeping Shopify and Odoo in sync so the merchandiser reads true, per-location availability is what stops it from promoting what you cannot ship. Second, real cost data - COGS and landed cost held in your accounting system - surfaced as a margin figure per SKU, so "rank by demand" can become "rank by profitable demand." Third, a return rate per product from your returns and fulfilment records, so the model can discount the SKUs that come back. Fourth, clean, complete product attributes from a product information layer, because a merchandiser also groups and filters, and it can only do that on structured data - size, material, category, use case - not on a title string. Get those four onto the catalog and the native AI has something worth optimizing; leave them out and it optimizes on vibes. Wiring exactly this kind of scattered operational data into one place a system can act on is the core of the AI in ecommerce and AI workflow work we do.

The India angle: RTO, high-return categories, and COD

For Indian D2C brands the return-rate blind spot is the expensive one, because a return here is often not a return at all - it is a cash-on-delivery order that failed and became a return-to-origin, where the product travels both ways and you collect nothing. An engagement-only merchandiser has no idea that the fast-selling category it keeps promoting is also your worst RTO offender, or that a chunk of its "sales" in certain regions never convert to cash. Wire in RTO and failed-delivery data and the ranker can quietly favour the products and regions that actually stick, instead of inflating gross orders that dissolve on the doorstep. It is the same principle as everywhere else - rank on kept, paid revenue, not on the order that looked good at checkout - and it is exactly the join between storefront, ERP, and courier data that decides whether the number means anything.

How to tell if your AI merchandising is helping or hurting

Click-through on a merchandised row is the vanity metric, because a ranker can drive clicks straight into products you should not be pushing and still report a healthy number. Three readings tell you the truth. The first is out-of-stock exposure: what share of the products in your top positions are actually unavailable or nearly so right now - anything above a trace means the ranker is spending your best real estate on things you cannot ship. The second is margin per impression rather than clicks per impression: are the products getting the most exposure the ones that earn the most kept profit, or just the most attention? The third is return-adjusted ranking: if you re-scored your top products on kept revenue instead of gross orders, how much would the order change - a big shift means high-return items are riding too high. Watch those three and you are managing merchandising as a profit engine, not admiring a click-through chart. A simple operations dashboard that puts stock, margin, and return rate next to the ranking is usually all it takes to see the gap.

Frequently asked questions

Is native AI merchandising enough on its own?

For a small catalog where everything is in stock, high-margin, and rarely returned, it is probably fine - the engagement signal and the profit signal point the same way. The bigger and more varied your catalog, the more they diverge, and the more the ranker needs stock, margin, and return data to avoid confidently promoting the wrong things. The feature is a strong starting point; it is not a finished system unless it can see what a sale actually costs you.

How is this different from AI product recommendations?

Recommendations are personalized, one-to-one suggestions - the "you might also like" row tuned to an individual shopper. Merchandising is the store-wide curation layer: the default order of a collection, the homepage lineup, the search ranking that most visitors see. They are related and often powered by similar models, but merchandising decides the first impression for everyone, which is why a data gap in it is more visible and more costly.

Do I need Odoo specifically for this?

No - the principle holds with any ERP or warehouse system. What matters is that live stock, real cost, and return data live somewhere authoritative and can be fed to the storefront in near real time. We work in Odoo a lot because that join is clean there, but the requirement is the join, not the brand. If your systems already agree on stock, margin, and returns, you mostly need to connect them to the ranker; if they do not, that disagreement is the first fix.

Where do we start if our product and stock data are a mess?

Start with one collection page and one product. Take the item ranked first and try to answer three questions without switching tools: is it in stock everywhere it claims to be, what does it truly earn after all costs, and how often does it come back? If you can answer all three cleanly, you have the foundation and just need to pipe those facts into the merchandiser. If you cannot - and most brands cannot for at least one of them - that gap is exactly where the ranker is flying blind, and closing it is the highest-return thing you can do before you trust the store to merchandise itself.

The short version: native AI merchandising is a real convenience, but the store merchandising itself well is not the same as the store merchandising itself. Left to engagement signals alone, it will rank the sold-out, the low-margin, and the high-return to the top, and do it with total confidence. Join up your live stock, your real margin, and your return rates - the same clean spine that a machine-readable catalog needs everywhere else - and the ranking finally optimizes for what you actually want: in-stock, profitable, kept demand. That is when letting the store merchandise itself starts making you money instead of quietly costing it.