Merchandising used to be a hands-on craft: a buyer or category lead deciding which products lead a page, which sit below the fold, and which campaign gets the hero slot. AI can now do parts of that job at a scale no team can match — ranking thousands of products per shopper, in real time. The fear, understandably, is that automation buries the products you need to sell and turns your storefront into a black box. It doesn’t have to. The goal is a system where AI does the heavy lifting and merchandisers steer.
What “AI merchandising” actually means
The term covers several distinct capabilities, and conflating them causes most of the confusion:
- Automated ranking orders products within a category or search result using signals like conversion rate, margin, stock, and recency.
- Personalised ranking adjusts that order per shopper based on behaviour and context.
- Recommendations populate “you may also like”, cart, and post-purchase modules.
- Dynamic curation assembles or reorders collections without manual sorting.
Each automates a decision a person used to make by hand. The question is never “AI or human” but which decisions to delegate and how much override to keep.
Why control matters commercially
Pure performance-optimised ranking sounds ideal until you account for the things the algorithm can’t see. A model maximising click-through has no idea that:
- You’ve over-ordered a line and need to clear it before season end
- A supplier funds a co-marketing placement that must appear top of category
- A product is discontinued and shouldn’t gain visibility
- Margin on the best-converting item is thin, and a slightly lower-converting alternative is far more profitable
- Brand strategy requires a new range to get airtime it hasn’t yet earned on data alone
These are legitimate business goals an engagement model will quietly work against. Control is how you encode commercial reality the algorithm has no way of knowing.
The layered model we recommend
The reliable pattern is to treat AI as the default and merchandising rules as the override layer that sits on top. In practice that means three tiers.
1. The automation layer
Let the model handle the long tail: the thousands of ranking and recommendation decisions no one has time to make manually. This is where AI earns its place — surfacing relevant products, adapting to demand, and keeping out-of-stock items from cluttering the top of the page.
2. The rules layer
Give merchandisers explicit, durable controls that constrain the model:
- Boost and bury specific products, brands, or attributes
- Pin a product to a fixed position
- Promote in-stock, demote low-stock, hide out-of-stock
- Margin weighting so profitability influences ranking, not just conversion
- Hard exclusions for discontinued or compliance-restricted items
These rules should be readable and reversible — a person should be able to look at a category and understand why it’s ordered the way it is.
3. The campaign layer
Time-boxed overrides for promotions, launches, and seasonal pushes that take precedence over both automation and standing rules, then expire automatically. This is what prevents stale “summer sale” pins from haunting your store in November.
The principle behind all three: automate the default, govern the exceptions. The same philosophy underpins our work on search relevance tuning, where business signals are blended into algorithmic relevance rather than replacing it.
Keeping merchandisers in the loop, not out of a job
The teams that get the most from automation reframe the merchandiser’s role from sorting products to setting strategy and supervising outcomes. That requires the tooling to make automation legible:
- Explainability. When a product ranks first, a merchandiser should be able to see why — high conversion, a boost rule, a campaign, or personalisation.
- Preview and simulation. Test a rule against live data before it goes out, and see which products move.
- Guardrails. Limits that stop personalisation from doing something absurd, like showing only one brand to a shopper or hiding an entire category.
- Override without code. Changes should be a merchandiser’s job, not a ticket to engineering.
When people can see and shape what the system does, trust follows — and trust is what stops teams from ripping out automation the first time it makes an odd-looking choice.
Measuring whether it works
Automation should be held to the same standard as any merchandising decision. Track:
- Category and search conversion rate before and after
- Revenue and margin per visit, not just conversion — a model can lift conversion while eroding margin
- Product discovery breadth, so a handful of winners don’t crowd out the catalogue
- Rule coverage, how often manual overrides fire, which tells you where the model needs work
Where possible, validate changes with controlled experiments rather than before/after guesses; our guide to A/B testing with AI covers how to run those cleanly. And tie it back to the channel value — site search and category browse are often your highest-converting surfaces, so small ranking gains compound.
Pitfalls to avoid
- Optimising for clicks alone. Engagement metrics drift away from profit. Always weight ranking with margin and stock.
- Set-and-forget rules. Boost rules accumulate and contradict each other. Review them quarterly and prune.
- Personalising too aggressively. A storefront that looks wildly different on every visit erodes the merchandiser’s mental model and can trap shoppers in filter bubbles.
- No off switch. You should always be able to revert to a stable, rules-based ordering if a model misbehaves.
- Hiding the logic. If only the data team understands the ranking, merchandising loses control by default.
A pragmatic starting point
Begin where the stakes are contained: automate ranking on a handful of high-traffic categories, give merchandisers boost/bury and pin controls, and add margin weighting. Run it alongside your current setup, measure conversion and margin per visit, and expand once the team trusts the override tools. The aim is not to remove human judgement but to free it from manual sorting so it can focus on the decisions that genuinely need a person.
If you’re weighing how much to automate and how to keep your team in control, our AI search and recommendations work is designed around that balance. Talk to us about your catalogue and we’ll map out a layered approach that fits how your team merchandises today.