Completed —

Semantic search with Vertex AI for e-commerce.

Project owner @ Funiglobal

20 semantic search engines on Vertex AI Search for Commerce for Funidelia, an e-commerce company from Zaragoza, lifting post-search conversion across 20 markets.

Replacing Funidelia's in-house search engine and those of its sibling brands with a fleet of 20 specialized search engines built on Vertex AI Search for Commerce. Goal: lift conversion on internal searches — the traffic flow with the highest purchase intent and the worst close rate in fashion and costume e-commerce.

Starting point

Funidelia and its 19 vertical markets (by country and language) used a custom search engine based on exact token matching plus manual category boosting. Post-search conversion was ~40% below the site average: a user would type "red carnival dress" and get an empty page or literal results that failed to grasp intent. Search was a bottleneck, not a lever.

Approach

Vertex AI Search for Commerce offers three key advantages for this case:

  • Free semantic understanding. Synonyms, misspellings and mixed-language queries resolved by the model without manual dictionaries.
  • Catalog as a first-class citizen. The index is fed from the structured product catalog (not scraped HTML), with typed attributes — category, size, color, brand — that the model uses as ranking features.
  • Personalization signals in the pipeline. Browsing, addToCart and purchase events feed the ranking with no extra code; the A/B test runs against the previous search engine behind a flag.

Execution

5 months to reach production across all 20 markets:

  • M1-M2: ingest pipeline from the product catalog (Python + GCP Dataflow) and product-to-Vertex AI Catalog schema mapping. Index quality validation against a historical dataset of real searches.
  • M3: A/B test with 5% of traffic in the pilot market (ES). Measuring CTR on results, post-search conversion and revenue per search.
  • M4-M5: progressive rollout to the remaining 19 markets with per-market tuning of boosting rules (in some countries brand weight matters more than category).

Outcome

A measurable improvement in post-search conversion in every market where it was deployed. The project shows that GenAI applied to e-commerce isn't chatbots — it's semantic ranking over existing structured data, with far clearer ROI and less regulatory risk.