Ecommerce teams have long focused their energy on optimizing the search box. After all, a customer typing something into the search box is a clear declaration of intent. But what about the moments when a shopper lands on a category or listing page without entering a query?
These moments, i.e. category pages, brand listings, and curated collections, are a major part of the customer journey, yet they’ve traditionally lacked the same level of optimization as search. These pages drive visibility, discovery, and ultimately revenue and yet they remain under-optimized. Many ecommerce teams still rely on static sort orders or basic merchandising rules to determine which products appear and in what order. But for teams managing thousands of listing pages, manual curation simply isn’t scalable — so they focus their efforts on a handful of top-performing pages while the rest are left untouched and unoptimized.
At Coveo, we saw this gap, and built a solution to fill it. Our new Listing Page Optimizeruses machine learning techniques to dynamically reorder listing pages based on what’s most likely to drive Revenue per Visitor (RPV). Unlike traditional rule-based ranking or limited relevance tuners, the Listing Page Optimizer AI model balances business goals with user behavior, giving merchandisers the edge they need to stay competitive, and profitable.
In this blog, we’ll break down why listing pages require a different optimization approach than search, how Coveo’s new model works behind the scenes, and what makes our Listing Page Optimizer a smarter, more scalable solution for driving revenue and relevance — without relying on static rules.
Why Listing Pages Matter
Ecommerce is increasingly discovery-driven. A growing portion of shoppers don’t come to your site with a specific product in mind — instead, they’re exploring. And in these cases, it’s not necessarily search results pages doing the heavy lifting; it’s the listing page.
Whether it’s “New Arrivals,” “Men’s Jackets,” or “Best Sellers in Tech,” these pages shape which products are seen, clicked, and added to cart. And yet, because they don’t include a query, traditional search optimization models often struggle to handle them effectively. It’s a bit like a customer walking into a physical store, standing silently in a department or aisle, and waiting to be helped. No context, no request. How do you assist them?
The reason is simple: listing pages are fundamentally different from search — and they deserve their own optimization model. That’s the challenge our new model addresses head-on. It’s designed to rank all listing pages, even those with no prior interactions.
A New Model for Listing Pages
The Listing Page Optimizer is designed to address these gaps. It ranks the entire product list for each listing page based on precomputed performance weights and real-time page context. This allows merchandisers to automatically promote products that are more likely to convert or drive higher revenue — without relying on manual pinning, boosting, or burying.
Among the advantages:
- Optimize ranking across the entire category, not just the top results
- Deliver better relevance without relying solely on prior interactions
- Incorporate business goals like profitability, inventory, or strategic product priorities
- Empower merchandisers with more control and visibility over how products appear
The model learns from behavior at scale while making each product’s performance on each page the basis for smarter ranking. Whether it’s a category page for a new product line or an evergreen listing with thousands of SKUs, the Listing Page Optimizer helps ensure that what shoppers see first is both relevant to them and strategically valuable to the business.
What Powers the New Model? Data, Context, and Flexibility
So how does this model rank products without a search query? It relies on three core types of data; catalog attributes, measured KPIs, and computed variables. Let’s break down what each of these mean.
Catalog Attributes (Static Metadata)
These are the product attributes available at indexing time. They include brand, category, price, color, margin, and review score and they provide a foundational layer of understanding across all products.
Measured Variables (Behavioral Signals)
These are real-world signals from how users interact with products including metrics such as revenue generated, conversion rate, add-to-cart actions, and bounce/exit rates. Crucially, we don’t just track these globally — we measure attributed impact per listing page. For example: Product A might generate $10,000 in revenue overall, but only $500 of that comes from the “Running Shoes” listing page. That distinction matters for optimization.
Computed Variables (Generalized Insights)
These are derived features, helping us generalize behavior across products such as brand and category affinity, and price influence. These computed signals enrich the model’s understanding of what drives performance — not just what’s visible in raw data.
Under the Hood: Precomputed Weights + Real-Time Inference
At the heart of the Listing Page Optimizer is the use of measured, computed, and indexed product data to build a foundation for ranking. These inputs include both business KPIs, such as total revenue, conversion rates, and basket impact, and relevance-based signals derived from user behavior.
Importantly, these metrics aren’t assessed in aggregate alone. The model is designed to understand how a product performs within the context of each individual listing page. For example, if a product generates $10,000 in revenue across the site, we want to know how much of that revenue comes specifically from when the product appears on a particular listing page. How often does it get clicked there? How often does it convert there? That contextualized performance matters when ranking products within listing pages.
These contextualized weights, covering both relevance and business impact, are computed for every product across every listing page. The results of this training process are then encoded, not as readable values, but as numerical vectors. Each vector contains scores specific to a given listing page and is highly meaningful within the model’s framework.
Once computed, these vectors are stored as external fields in Coveo’s index, where they are available at inference time. The model is retrained on a daily basis to reflect updated behavioral trends and business priorities.
Real-Time Ranking in Context
When a visitor accesses a listing page, such as “Canoes & Kayaks” illustrated below, a query is initiated and then routed through the relevant query pipeline, where it is recognized as a listing context.
At this point, the Listing Page Optimizer is invoked. It fetches the precomputed product weights corresponding to that page and uses a Query Ranking Function (QRF) to retrieve the appropriate score from each product’s vector. This score is added to the product’s baseline index score to produce a final, context-aware ranking.
Because each product carries a unique score for each page, the system can rank all products returned, not just a subset. There are no performance bottlenecks between the search API and the index, and no limitations on how deep the ranking can go. The entire framework is designed for full-scale, real-time reranking, driven entirely by the data.
A Flexible Framework for the Future
The architecture of the Listing Page Optimizer is inherently flexible, allowing for expansion as new business goals and variables come into focus. While today the model accounts for metrics like attributed revenue, conversion rate, and click-through rate, its structure offers the future inclusion of additional variables that impact profitability and customer satisfaction such as margin, fulfillment cost, return rate, or cost-to-serve.
In addition, the framework is well-suited to evolve toward more dynamic, real-time applications. For example, it could be extended to adapt rankings based on dynamic visitor context, such as location or audience segment, enhancing both the user experience and business outcomes.
A New Era of Ecommerce Ranking
Listing pages are powerful drivers of product discovery and conversion, but they’ve traditionally relied on static rules or simplistic models. Coveo’s Listing Page Optimizer introduces a dynamic, data-informed approach that adapts to real shopper behavior, aligns with business goals, and improves continuously over time.
As the model is designed to learn from both relevance signals and business KPIs, it is expected to help brands:
- Improve revenue performance across different product categories
- Drive higher conversions, especially on curated, seasonal, and top-of-funnel pages
- Reduce manual merchandising work by aligning product ranking with performance goals
Early results from select Coveo customers are already demonstrating this potential:
- Revenue uplift of 2% to 6%, depending on product category and traffic volume
- Noticeable conversion gains on curated, seasonal, and early-funnel landing pages
- Operational efficiency, with merchandisers spending less time managing manual rules and achieving greater alignment with business performance goals
These outcomes represent a meaningful shift in how ecommerce teams can approach product discovery— moving from rigid rule sets to flexible, intelligent systems that scale. And as the model continues to evolve to incorporate additional KPIs and dynamic visitor context, it sets the stage for a smarter, more adaptive approach to merchandising. A win for the shopper and for the business.