Search Is Now the Storefront — Not the Homepage
The homepage is a brand statement. The search bar is where commerce happens. According to Forrester Research, 43% of retail website visitors navigate directly to search upon arrival — and those visitors convert at 2–3× the rate of browsers. Yet the average enterprise retailer treats search as an afterthought: a bolt-on ISV running on top of keyword matching and a catalog with 40% attribute gaps.
This is the gap that creates the AI search opportunity — and the gap that Iksula’s catalog-first approach closes before any ISV is deployed.
“ISVs give you the search index. Iksula gives you the catalog that makes the index worth something.”
The revenue cost of bad search
Zero-results pages are revenue sinkholes. When a user searches and gets no results, 68% abandon the session immediately (Baymard Institute). A 2% zero-results rate on a $500M retailer with 40M site visits per year represents approximately $8–12M in avoidable lost revenue annually — a number that a CDO can take to a CFO in 10 minutes.
Estimates based on Baymard Institute, Forrester, and Iksula deployment benchmarks. Individual retailer results will vary with traffic volume and average order value.
From Keyword to Intent: How AI Search Actually Works
Legacy e-commerce search is BM25: a probabilistic ranking function that counts keyword frequency in product titles and descriptions. It works when users type exact product names. It fails when users type intent — “something warm for Sunday hiking,” “birthday gift under $50 for a 10-year-old,” “office chair for bad backs.”
AI-powered search closes this gap in three layers
Layer 1: Semantic search (dense retrieval)
Semantic models (BERT, Sentence-BERT, E5, or domain-fine-tuned variants) convert both queries and product descriptions into dense vector embeddings — multi-dimensional numerical representations of meaning. At query time, the user’s intent vector is matched against product vectors by cosine similarity, surfacing semantically relevant products regardless of keyword overlap.
Layer 2: Hybrid retrieval (keyword + dense)
Pure dense retrieval excels at intent matching but can miss exact product codes, SKU numbers, and branded terms. Hybrid retrieval combines BM25 (keyword precision) and dense vectors (semantic recall) using reciprocal rank fusion — giving the search engine precision on exact matches and recall on intent-based queries simultaneously. This is the architecture Iksula deploys in production.
Layer 3: Query rewriting with LLMs
When a query is ambiguous, malformed, or outside the retailer’s catalog vocabulary, a lightweight LLM rewriting layer expands, clarifies, or translates the query before it hits the index. This reduces zero-results rates by 40–60% in production deployments without changing the underlying index or product data.
The Catalog Prerequisite No ISV Will Tell You About
Algolia, Klevu, and Constructor are excellent search engines. They are not catalog enrichment tools. Every AI search ISV requires clean, complete, attribute-rich product data to perform at its advertised benchmarks. What they will not tell you during the sales process is that most enterprise retailers’ catalogs are missing 30–60% of the attributes required to make semantic search accurate.
What “attribute-complete” actually means
For a power-tool retailer, attribute completeness means every SKU has: product type, voltage, battery type, compatible battery brands, weight, dimensions, included accessories, application type (professional vs. DIY), and material compatibility. Without this, a query for “18V compatible drill bits for Milwaukee” returns nothing — or worse, irrelevant results that erode user trust permanently.
Iksula’s catalog-first search approach
Before deploying any AI search ISV, Iksula runs a catalog quality audit (Athena’s 300+ rule engine) to measure attribute coverage, taxonomy depth, and marketplace-native requirements. Gaps are closed via PC² (attribute generation), WordsworthAI (description enrichment), and PictureAI (visual attribute extraction) before the ISV index is populated. The result: an AI search engine that works in production, not just in demos.
In one engagement with a $1B+ hardware retailer across 45,000 SKUs, this catalog-first approach delivered a +30% on-site search success rate within 90 days of deployment — with zero ISV changes, purely through catalog enrichment.
Agentic Shopping: When Search Becomes a Conversation
The next frontier of AI search is not better ranking. It is conversational product discovery — where a customer describes a need in natural language and an AI agent surfaces, narrows, compares, and configures the right product without the customer ever clicking through ten filter facets.
Iksula’s Product Selector Agent (part of the Deep Agent AI platform) handles this for complex B2B and B2C discovery scenarios — guided selling workflows for industrial equipment, configurators for furniture and custom products, and compatibility-aware search for spare parts and accessories. The Universal Commerce Protocol (UCP) makes this agent available across Adobe Commerce, voice interfaces, and API commerce channels.
The 60-Day AI Search Upgrade Roadmap
Frequently Asked Questions: AI Search & Discovery
AI-powered search combines keyword retrieval with dense vector embeddings to understand user intent, not just literal query terms. Semantic models (BERT, CLIP) match queries to products by meaning, enabling searches like “something cozy for winter evenings” to surface relevant products without exact keyword matches.
Vector search converts products and queries into numerical embeddings — multi-dimensional representations of meaning. When a user searches, their query is converted to a vector and matched against product vectors by similarity (cosine distance), enabling semantic and visual matching beyond keyword overlap.
Buy the search index and API layer (Algolia, Klevu, Constructor). Build proprietary ranking signals and attribute layers tuned to your catalog. Use Iksula’s Athena as the quality gate ensuring the ISV’s index receives clean, complete, attribute-rich data — this is what the ISV will not tell you is required.
Iksula production benchmarks: +30% on-site search success rates, 2–3× conversion rates versus browse traffic, and zero-results rate reduction of 40–60%. For a $500M retailer, a 2% zero-results rate reduction delivers $8–12M in measurable revenue uplift within 90 days.
Your Search Is Losing Revenue Every Hour
Book a 30-minute AI search audit with Iksula. We'll baseline your zero-results rate, attribute coverage, and search success rate — and show you the exact enrichment gap between your current catalog and production-ready AI search.
Book AI Readiness Review
