thought leadership

The Anatomy of a Commerce Intelligence Stack

A technical breakdown of what a modern commerce intelligence system actually looks like — from raw data to automated decisions.

Ivy ·
Architecture Commerce Intelligence Engineering

“Commerce intelligence” is a vague term that vendor decks have stretched beyond meaning. Let’s define what an actual system looks like — no buzzwords, no fluff.

The Five Layers

Every real commerce intelligence stack has five layers. You can’t skip any of them.

Layer 1: Data Ingestion

Raw material for decisions.

Sources:

  • Marketplace APIs — competitor pricing, inventory, reviews, rankings
  • Web scraping — for channels without APIs (still surprisingly common)
  • First-party sales data — orders, returns, refunds, by SKU
  • Customer sentiment — review text, support tickets, social mentions
  • Operational data — inventory turns, restock lead times, fulfillment costs

The ingestion layer’s job is reliability and completeness. Data missing from this layer cannot be recovered later.

Layer 2: Normalization and Enrichment

Raw data is messy. The normalization layer makes it queryable.

  • Convert all prices to a consistent currency and time zone
  • Map competitor SKUs to your own catalog (fuzzy matching)
  • Deduplicate signals from multiple sources
  • Tag products by category, subcategory, and competitive tier

This is unglamorous work. It’s also the layer that determines whether your intelligence is trustworthy.

Layer 3: Analysis and Reasoning

The “brain” layer. Takes normalized data and produces insights.

This is where LLMs or rule engines:

  • Identify anomalies (price spikes, review bombs, sudden rank changes)
  • Calculate competitive positioning across the catalog
  • Surface opportunities (underserved niches, demand spikes)
  • Generate hypotheses for testing

Modern systems use LLM-based reasoning to handle ambiguity that rigid rule systems can’t. “This looks like a new competitor entering at aggressive pricing” is something a rule can catch. “This competitor is changing their strategy” requires inference.

Layer 4: Decision Routing

Not every insight needs the same treatment.

Signal TypeResponse
Competitor price drop on commodity SKUAuto-reprice (no human needed)
New entrant in premium categoryAlert + analysis brief to category manager
Review velocity spike on your top productUrgent alert to ops + CS lead
Demand trend emerging in adjacent categoryResearch task for growth team

This layer is business logic — it decides what happens to each insight.

Layer 5: Action and Verification

The closed loop.

Action systems:

  • Execute reprices in marketplace platforms
  • Create tasks in project management tools
  • Draft communications for human review
  • Update internal dashboards and reports

Verification:

  • Confirm the action was applied correctly
  • Measure outcome vs. expected
  • Feed outcome data back into the analysis layer

The Common Failure Modes

Layer 1 failures: Incomplete data, no ownership, gaps go unnoticed Layer 2 failures: Inconsistent normalization, SKU mapping errors, trust issues Layer 3 failures: Analysis paralysis — too many insights, none prioritized Layer 4 failures: Decision routing doesn’t exist — everything goes to humans Layer 5 failures: Actions are manual, no verification loop, no learning

Most companies have Layer 1 and occasionally Layer 5. Layers 2-4 are where the value actually is.

The Architecture Decision

You can build this yourself with commodity tools:

  • Airbyte for ingestion
  • dbt for normalization
  • A chat-with-data layer over your warehouse
  • Zapier or similar for action routing

Or you can use an integrated platform. Neither is wrong. The build-vs-buy tradeoff here depends heavily on your team’s engineering capacity and how specific your requirements are.

The minimum viable stack for meaningful commerce intelligence has all five layers. You can’t shortcut it.

Running an integration right now?

The research is clear: discovery-first integrations deliver on time. ACQI has the modules to get you there in weeks, not months.