Search Page Optimizer methodology

This is for:

System Administrator

Search Page Optimizer (SPO) models use an approach known as learning to rank. This article explains the methodology at a high level: how the model learns from visitor behavior, what information it considers, and how it applies what it has learned at query time.

For an overview of the feature, see About Search Page Optimizer (SPO).

What is learning to rank?

Coveo traditional search ranking assigns each product a score by combining many relevance factors, such as keyword matches, popularity, and Coveo Machine Learning (Coveo ML) outputs, using a set of weights. Choosing those weights is difficult: they interact with each other, they differ from one catalog to another, and the best combination changes over time.

Learning to rank takes a different approach. Rather than relying on manually chosen weights, the model learns the best way to combine ranking signals directly from examples of what visitors preferred. For each search query, past visitor behavior indicates which products should have ranked higher. The model studies many of these examples and learns how to order results so that the products visitors actually want appear first.

How SPO learns from visitor signals

Not all interactions carry the same meaning. A click shows interest, an add-to-cart shows purchase intent, and a purchase confirms value. The model treats these signals accordingly: for each query, every product shown is graded by the strongest interaction visitors gave it, and stronger signals weigh more heavily in what the model learns.

Diagram showing how clicks, add-to-carts, and purchases are converted into signal grades, then used in daily learning to rerank search results | Coveo for Commerce

Training is grounded in conversion signals, meaning add-to-cart actions and purchases, rather than clicks alone. This keeps the model focused on ranking products that visitors buy, not only products that attract attention.

Note

The model uses aggregate observations to learn which ordering best matches visitor choices. It doesn’t personalize results for an individual visitor in real time.

What the model considers

When learning how to order results, the model considers the signals that already contribute to your ranking:

Explicit merchandising rules, such as pin rules and ranking expressions, are deliberately excluded. The model never learns to override them; they’re reapplied after reranking and always take precedence.

How daily learning works

The model retrains every day on a rolling window of recent behavioral data. This keeps the ranking aligned with current visitor behavior as catalogs, seasons, and preferences change. The model compares the ranking shown with the products visitors chose and learns which product ordering better matches observed behavior. Each reranked result page generates new interactions, which feed the next learning cycle.

How SPO applies what it learned at query time

At query time, the model doesn’t replace the index ranking; it refines it:

  1. A visitor submits a query.

  2. The Coveo index retrieves and initially ranks candidate products.

  3. The model rescores the top results and reorders them based on what it learned. Results beyond the top set keep their original order.

  4. Merchandising rules are reapplied.

  5. The final search result page is returned to the visitor.

If the model is unavailable for any reason, the search result page falls back transparently to the standard index ranking, so search availability is never affected.

How to interpret SPO behavior

Keep the following in mind when interpreting SPO behavior:

  • SPO optimizes relative ordering. It determines the order of the top results of each result set; it doesn’t determine whether a product is indexed, and it doesn’t expand or replace the candidate result set.

  • SPO learns from aggregate behavior over time. A single interaction doesn’t necessarily cause an immediate ranking change. The model learns from many observations across visitors.

  • Learning quality depends on your signals. The model learns effectively only when it has sufficient, correctly tracked behavioral signals. Evaluate SPO using appropriate search engagement and business metrics.

Each SPO deployment is validated through a test that compares the optimized ranking against the standard ranking on live traffic before the model is fully activated. This ensures the impact is measured on your own visitors and catalog rather than assumed.

What’s next