Skip to main content

LLM Based Ranking and Relevance

Marqo's ranking system uses large language models trained specifically on your customer interaction data to rank products for conversion, revenue, and margin — not just semantic relevance. The Marqo pixel automatically collects per-search interaction data (clicks, add-to-carts, purchases) with no manual submission required, and the ranking layer updates continuously as new data arrives.

What Makes This Different​

Traditional learning-to-rank systems are built around historical data per product. They can only rank products they've seen before, struggle with new launches, and optimise for a relevance proxy rather than business outcomes. They also share a single model across all customers.

Marqo's approach is different across every dimension:

  • Business-objective alignment: the model is trained to optimize the metrics you care about — conversion rate, revenue per session, gross margin — not to return the most "semantically similar" result
  • Per-customer specificity: a separate embedding model is finetuned for each retailer on their own data, so the system understands your catalogue and your shoppers specifically
  • Continuous optimization: the ranking layer learns in near-real-time from incoming interaction data, improving automatically as your business runs
  • Per-search signal: interactions are attributed to the specific search that produced them, giving higher-quality training data than session-level tracking

Per-Customer Embedding Model Finetuning​

Most search platforms use a single shared embedding model across all their customers. Marqo does not.

Marqo finetunes its embedding models on a per-customer basis using each retailer's own clickstream and purchase data. Critically, the training objective includes your specific business goals — if you want to optimise for gross margin over raw volume, the model learns that. If certain categories are strategically important, the model encodes that.

This means the core semantic layer — the model that determines which products are relevant to a given query — is trained to reflect both your customers' behaviour and your commercial priorities.

What the model learns:

  • Which product attributes your customers actually respond to when searching
  • Which synonyms and phrasings your audience uses (e.g. "trainers" vs "sneakers" vs "athletic shoes")
  • Which products convert for which queries, even when the relationship isn't obvious from text alone
  • Which products drive the margins and revenue you've defined as priorities

Periodic retraining:

The platform automatically schedules periodic full model retraining as your catalogue grows and seasonal patterns shift. As new products launch and buying behaviour evolves, the embedding models are updated without requiring any action from your team. Between retraining cycles, the ranking layer continues to update continuously from incoming interaction data.

A unique capability:

Per-customer embedding model finetuning — aligned to business objectives — is not offered by any other ecommerce search provider. Competitors use shared or generic models, sometimes with shallow personalization layered on top. Marqo trains the embedding model itself on your data against your goals, which produces improvements at a more fundamental level than post-hoc re-ranking alone.

Continuous Ranking Optimization​

The LLM ranking layer does not require a scheduled retraining cycle to improve. As the Marqo pixel collects interaction data from your store, the ranking model updates in near-real-time.

This means:

  • Rankings improve automatically as your store runs, without any manual intervention
  • New products receive better rankings faster as interaction data accumulates for them
  • Seasonal shifts and changing customer preferences are reflected in rankings without a lag
  • There is no "batch update" window during which rankings stagnate

Merchandising teams can layer manual controls (boosts, buries, pins) on top of the AI ranking at any time — these are applied as a priority layer over the continuously-optimized base rankings.

Advantages Over Traditional Approaches​

Immediate Performance for New Products​

Traditional learning-to-rank requires historical data per product. New launches, seasonal items, and niche products struggle because there's nothing to learn from.

Marqo understands patterns across your catalogue: what price points convert, what styles resonate, what attributes drive purchase decisions for your customers. When a new product launches, these patterns apply immediately — no waiting period, no poor initial rankings.

Long-Tail Coverage​

Traditional systems need exact matches. If customers haven't searched for a specific variation before, the system cannot rank it well.

Because Marqo operates on learned semantic patterns rather than frequency counts, it handles novel queries across your full catalogue, delivering relevant results even for queries with no prior search history.

Built Specifically for Your Business​

Every retailer is different. Your customers have specific preferences; your products have unique characteristics; your margins vary by category and supplier. The LLM ranking system learns all of this from your data — it's not a generic algorithm applied uniformly.

Business Outcomes​

The business impact of LLM-based ranking is measurable and consistent across retailers:

  • Higher conversion rates — products that match customer intent and business objectives rank higher
  • Improved revenue per session — the ranking layer optimises for revenue, not just clicks
  • Better new product launches — new items benefit from catalogue-wide pattern learning from day one
  • Reduced manual merchandising — the AI handles routine ranking optimization; merchandisers focus on strategy
  • Long-tail performance — queries that previously returned poor results now convert

Monitoring Performance​

Track ranking performance in the Analytics section:

  1. Go to Analytics → Ranking Performance
  2. Review key metrics:
    • Conversion rate vs baseline
    • Revenue per search trend
    • New product performance — rankings for recently launched products
    • Query coverage — proportion of queries returning high-confidence results

Best Practices​

  1. Deploy the pixel fully — per-search clickstream is the training signal. The more coverage, the faster and better the model learns.
  2. Define your business objectives — work with your Marqo representative to configure the training objective (conversion, revenue, margin, or a weighted combination).
  3. Layer manual controls strategically — use boost/bury and pinning for promotions, clearance, and brand commitments; let AI handle everything else.
  4. Review monthly — check analytics for drift or unexpected patterns as your catalogue and customer base evolves.