The merchandising teams at some of the most sophisticated ecommerce operations in the world have stopped trying to improve their search tools. It’s not because they don’t want to improve them, but the tools can’t show their reasoning, which means the people closest to the catalog have no way to make them better.
This gets treated as a known limitation — something teams learn to work around. But these workarounds just add to the costs, and the retailers that pull ahead in product discovery over the next few years will be the ones who recognized this early enough to change course.
The Same Frustration Keeps Showing Up
Talk to enough merchandising teams running keyword-first search platforms, and you hear the same story with minor variations. The platform handles head terms fine, conversion ticks along, and for a while nobody questions it. Then, the limitations start stacking in ways that are hard to measure but impossible to ignore.
For example, a tool might make ranking default to global bestsellers regardless of context, so the same hundred SKUs dominate every page. A product manager at a major eyewear retailer called this the “winners keep winning” problem: established products get perpetual exposure, while new inventory sits buried. These products wouldn’t necessarily perform poorly if shoppers found them, but the algorithm simply rewards historical performance and has no mechanism for exploration. This product manager’s merchandising team knew this was happening, but they couldn’t do anything about it because the platform’s reranking feature operated as a binary on-off toggle, with no way to blend strategies or adjust how different signals were weighted against each other.
What made it worse was the opacity. There were three configurable levers, but no visibility into what else influenced the output. Whether the algorithm was weighing color, shape, price, or some combination it had learned from clickstream data, nobody could tell. Without that visibility, the team couldn’t even improve the product data feeding the system, because they had no way of knowing which attributes mattered to the ranking and which were being ignored.
As a result, the merchandising team reverted to manual curation, spreadsheets, and institutional knowledge that never entered the system.
More AI Layered Onto an Opaque System Makes the Opacity Worse
When search underperforms, the brand’s response is often to layer on more intelligence: more machine learning, personalization features, and dynamic capabilities. These get announced, they show up in the dashboard, and they do things. But they arrive as additions to an engine that was never built to be transparent, and each new layer adds another dimension the merchandiser can’t see into or influence.
That’s what “bolted on” AI really means. Even if the AI is quite good, it sits in a separate architectural layer from the core ranking logic, and the two can’t be inspected together. The merchandiser sees a ranked list of products and can’t trace the path from input signals to that output.
For retailers selling commodity products with well-defined attributes, this is tolerable. But consider retailers whose catalogs have nuanced visual and functional characteristics, where internal teams can’t even agree on whether a frame style counts as an aviator based on the bar bridge or the teardrop shape. The algorithm can’t resolve ambiguities that the humans closest to the product haven’t resolved. And when those humans can’t see what the algorithm is doing with ambiguous attributes, the confusion builds up across millions of queries without anyone noticing.
Transparency Is an Architecture Decision, Not a Feature
Being transparent about search ranking isn’t a matter of adding a dashboard or surfacing a confidence score. It means building the system so that every product in every result set carries a visible relevance score and performance score, and a merchandiser can click into any product and see every searchable field, match term, and category association that contributed to its position. The key thing is they have visibility into the data itself, and not just a summary.
It also means the weighting between signals is adjustable at a granular level. A merchandiser should be able to index more on relevance for discovery-oriented queries, shift toward performance for high-intent searches, and layer in additional signals like margin, newness, and ratings — all as testable hypotheses running against live traffic with multivariate testing and measurable deltas across core KPIs. This can’t be a global override that touches everything — the changes need to happen at the query and category levels.
Loomi was built this way. Our search platform processes over 250 million queries per day across 1,400+ brands and exposes the full scoring logic behind every ranking to the merchandiser. The platform combines rules-based AI for governance, machine learning trained on more than 15 years of commerce-specific data for precision, and generative AI for understanding conversational and long-tail queries.
What Changes When Merchandisers Can See and Influence the System
DUSK, a premium home retailer operating in the UK and Australia, generated over £1 million in incremental revenue within six months of moving to a transparent search architecture. These results came from the merchandising team’s ability to deploy price-drop alerts, low-inventory signals, and back-in-stock notifications: campaigns that required visibility into how products surfaced and the ability to influence that surfacing in ways their previous platform couldn’t support. Before using Loomi, DUSK’s team was in the same position many merchandising teams find themselves in — working around a tool rather than with it.
Meanwhile, Unisport, a European sportswear retailer, increased marketing automation revenue by 85% year over year after connecting search behavior with downstream customer engagement. When search data and customer data live in the same system rather than being siloed across a search tool and a separate marketing platform with a fragile integration between them, the platform can trigger a price-drop notification to the specific customers watching a specific product. Better yet, no data engineering ticket was required to connect the two sources. The merchandising team could see which products were being watched, by whom, and act on that information without waiting for someone else to build the pipeline.
What This Means for Retailers Evaluating Search Platforms Right Now
Vendor evaluations for search technology tend to focus on relevance benchmarks, AI sophistication, and feature checklists. Those things matter, because every serious vendor can demonstrate strong relevance on head terms. But they don’t predict whether a platform will deliver value over time or plateau after the initial implementation.
The question that separates platforms is simpler and harder to answer with a demo: How much of your catalog can your merchandising team influence through the tool, without filing an engineering ticket, reverting to manual curation, or accepting that the algorithm’s decisions are final because nobody can see the reasoning behind them?
You need to ensure that the gap between what the merchandiser knows about the catalog and what the shopper sees doesn’t keep widening with every product launch or seasonal shift. Learn more about how Loomi’s search solution can help your team deliver better results more reliably.
