How a Mid-Market Retailer Grew AI Agent Revenue 63% by Restructuring Its Product Catalog for Agentic Buyers

The Challenge: Invisible to AI Agents Despite Strong Human Traffic

Harrow & Linden Home Co. (name changed for confidentiality) operates a direct-to-consumer e-commerce store selling mid-range furniture, lighting, and home accessories. In Q3 2025, their marketing team noticed a troubling pattern: organic search traffic from traditional Google queries was holding steady, but their conversion rate from AI-powered referral sources — including ChatGPT, Perplexity, and Google's AI Mode — had flatlined at a fraction of what competitors in their segment were reportedly achieving.

The specific problem wasn't visibility. Harrow & Linden appeared in AI-generated shopping responses with reasonable frequency. The issue was completion. When an AI agent recommended one of their products and a user authorized a purchase attempt, the transaction failed or was abandoned at alarming rates — 71% of AI agent-initiated sessions never reached a confirmation screen. By contrast, human-initiated sessions from the same traffic sources converted at 4.2%.

The business stakes were real. Industry data from early 2026 suggested that AI agent-driven purchases would account for an estimated 18–22% of e-commerce transactions for mid-market retailers by the end of 2026. For a $47M business, capturing even a proportional share of that shift represented $8–10M in incremental annual revenue. Failing to capture it meant ceding ground to larger retailers and digitally-native competitors already optimized for agentic buyers.

"We weren't losing on price or product quality. We were losing because our data spoke fluent human and broken robot."
— Head of Digital Commerce, Harrow & Linden Home Co.

An audit revealed three root causes. First, product descriptions were written for emotional resonance rather than machine-parseable specificity — dimensions were buried in paragraphs, materials were described with adjectives rather than standardized terms, and compatibility information was absent entirely. Second, there was no structured checkout API surface that AI agents could interact with programmatically; every purchase attempt required a human to navigate a five-step checkout flow. Third, inventory and pricing data was updated in batch cycles with up to a 6-hour lag, meaning AI agents were citing prices and availability that no longer existed by the time a user acted.

Strategy and Approach: What Was Decided — and What Was Deliberately Skipped

The internal strategy team, working with an external consultant specializing in agentic shopping optimization, identified a narrow, high-leverage intervention set. The core hypothesis was that fixing the data layer would generate more ROI faster than rebuilding any front-end experience. They were correct — but the path was not straightforward.

Three strategic decisions defined the project:

• Prioritize structured data over design. No budget was allocated to redesigning the human-facing product pages. Every dollar went toward the machine-readable layer: schema markup, attribute taxonomy, and API endpoints.
• Adopt UCP as the data standard. Rather than building a proprietary data format or relying on basic JSON-LD alone, the team implemented the Universal Commerce Protocol — a structured framework that AI agents increasingly use to evaluate and interact with product listings. This decision was informed by reviewing the Universal Commerce Protocol guide, which outlines how merchants can align their data architecture with the expectations of major AI shopping agents.
• Do not rebuild checkout — extend it. Rather than replacing the existing five-step checkout flow (a six-month project), the team built a lightweight, agent-accessible checkout API layer that ran parallel to the existing UI. Human buyers saw nothing different. AI agents got a clean programmatic surface.

What they deliberately chose not to do is equally instructive. They did not invest in paid placement within AI recommendation systems — a tempting but expensive option with unclear attribution. They did not attempt to reverse-engineer the ranking logic of specific AI platforms. And they did not pursue a full platform replatform, which a prior proposal had estimated at $1.2M and 14 months. Total project budget for the approach actually taken: $187,000 over 22 weeks.

Implementation: The Four-Phase Catalog Restructure

The implementation ran from October 2025 through March 2026, organized into four sequential phases. Each phase had a defined completion gate before the next began — a discipline that proved critical when Phase 2 uncovered unexpected data quality issues that would have cascaded into Phase 3 had the gates not been enforced.

Phase 1 required the most organizational effort. The existing product catalog had been built by six different teams over nine years and used 14 overlapping attribute naming conventions for something as basic as color. The team created a master taxonomy with 312 standardized attributes, mapped every existing product field to the new taxonomy, and wrote transformation rules to handle legacy data. Tools used: Akeneo PIM for attribute management, a custom Python script for field mapping, and Schema.org Product markup as the baseline schema layer extended with UCP-specific fields.

Phase 2 revealed that 34% of the catalog had missing or non-parseable dimension data — a far worse figure than the initial audit had suggested. A combination of AI-assisted data extraction from manufacturer PDFs (using a fine-tuned document parsing model) and manual review for high-revenue SKUs brought the completeness rate from 66% to 94% in five weeks. The remaining 6% — primarily one-of-a-kind vintage pieces — were flagged with explicit incompleteness markers so AI agents would not cite fabricated specifications.

Phase 3 replaced the batch inventory sync (previously running every 6 hours via a nightly ETL job) with an event-driven feed. Any change to stock level or pricing triggered an immediate push to the product feed. Average data freshness improved from 3.2 hours to 47 seconds. This single change eliminated the majority of the "product no longer available at cited price" failures that had been driving AI agent session abandonment.

Phase 4 built a REST API with four endpoints: product lookup, cart initialization, checkout initiation, and order confirmation. The API used OAuth 2.0 for agent authentication and required no changes to the existing Shopify-based checkout backend. A sandbox environment was made available for AI agent platform partners to test against during a 3-week pre-launch period, which surfaced 11 edge-case bugs before public availability.

Results: Before and After Metrics

The primary measurement window ran from April 1 to June 30, 2026 — a full quarter post-implementation — compared against the same quarter in 2025. The headline figure of 63% revenue growth from AI agent-driven sessions obscures several more granular findings that are arguably more valuable for planning purposes.

Several findings stand out. The 26% lift in average order value for AI agent sessions was entirely unexpected — the team had not modeled for any AOV impact. Post-analysis suggested that AI agents, when given complete and reliable product data, were more effectively matching users to higher-specification products that better fit their stated requirements. Human buyers browsing catalog pages were less likely to self-select into premium variants; AI agents recommending based on parsed attributes were not subject to the same anchoring biases.

Importantly, human session conversion rates were unaffected — confirming that the strategy of working on the machine-readable layer without touching the human-facing experience was sound. The project created no cannibalizing effects between channels.

Return on investment as of June 30, 2026: the $187,000 project cost had generated approximately $215,000 in incremental margin over two full quarters of operation (accounting for Q1 2026 ramp-up effects). Full payback was achieved in approximately 7.5 months from project start.

Key Learnings: What Worked, What Failed, and What Was Surprising

What worked: The phase-gated implementation structure prevented cascading data quality problems. Adopting UCP as the underlying data standard, rather than a proprietary schema, meant that new AI agent platforms entering the market during the project timeline could immediately interpret the catalog without custom integration work. The parallel checkout API approach — extending rather than replacing the existing checkout — shaved an estimated four months off the timeline.

What failed: The initial Phase 2 timeline was underestimated by nearly 100%. The assumption that manufacturer data would be clean enough to transform with minimal validation was wrong — 34% of the catalog required manual intervention or AI-assisted extraction. Any future project of this type should budget for a data quality remediation sprint before the main enrichment phase begins.

The team also attempted to create "AI agent-specific" product descriptions — separate content written to be parsed rather than read. This produced no measurable lift and was abandoned in Phase 2 after a four-week test. Structured schema markup attached to existing descriptions outperformed the parallel-description approach on every measured dimension. Writing separate content for machines is unnecessary when the machine-readable layer is properly implemented.

"The biggest surprise wasn't the revenue number. It was that fixing our data for AI agents also made our internal search, our feed-based ad campaigns, and our wholesale EDI exports measurably better. The catalog restructure was infrastructure, not a tactic."
— Project Lead, Harrow & Linden Home Co.

What was surprising: The catalog restructure produced significant second-order benefits that had not been anticipated. Internal site search relevance scores improved by 18% following the attribute taxonomy standardization — because the same structured data that AI agents use to parse products also powers faceted search. Google Shopping feed disapproval rates dropped from 7.3% to 0.9% as a byproduct of the schema improvements. And the wholesale EDI team reported that the new attribute taxonomy made purchase order matching substantially faster, saving an estimated 4 hours per week in manual reconciliation.

How to Replicate It: An Actionable Checklist

The Harrow & Linden project is replicable for most mid-market retailers with catalogs between 500 and 20,000 SKUs. The following checklist is sequenced in order of dependency — each item creates the foundation for the next. Attempting to implement Phase 4 (agent checkout API) before completing Phase 1 (taxonomy standardization) will produce unreliable results regardless of technical execution quality.

• Audit your current AI agent session data. Identify abandonment rate, completion rate, and the specific drop-off points in agent-initiated sessions. If you cannot segment AI agent traffic from human traffic in your analytics, fix that first — you cannot optimize what you cannot measure.
• Standardize your attribute taxonomy. Consolidate all product attribute naming conventions into a single master schema. Prioritize the attributes most commonly used by AI agents to evaluate product fit: dimensions, materials, compatibility, weight, color (standardized), and availability status.
• Audit catalog data completeness. For every SKU, score completeness against the master taxonomy. Any SKU below 70% completeness on high-priority attributes should be flagged for remediation before proceeding. Budget conservatively — expect to find worse data quality than your initial spot-check suggests.
• Implement UCP-aligned structured markup. Deploy Schema.org Product markup extended with UCP-specific fields. Review the Universal Commerce Protocol guide for field-level implementation guidance. Validate with Google's Rich Results Test and at least one AI agent platform's product data validator.
• Eliminate inventory and pricing data lag. Move from batch-based to event-driven inventory feeds. Target a maximum data freshness lag of 5 minutes; sub-2-minute is achievable with most modern e-commerce platforms. AI agents citing stale prices is one of the single largest drivers of failed transactions.
• Build or extend a programmatic checkout surface. If a full API rebuild is out of scope, a lightweight checkout-initiation endpoint that accepts a SKU, quantity, and shipping address and returns a pre-filled cart URL will capture the majority of the value. Ensure it handles error states gracefully — AI agents need machine-readable error responses, not HTML error pages.
• Run a 30-day sandbox period. Before public launch of any agent-facing API, invite at least two AI shopping agent platforms to test against a staging environment. The 11 bugs caught pre-launch in the Harrow & Linden project would each have produced failed transactions at scale.
• Establish a quarterly data quality review cadence. Catalog completeness degrades over time as new SKUs are added and product information changes. Schedule quarterly audits against your completeness scorecard to prevent regression.

For merchants wanting the full strategic and technical framework behind these steps, the agentic shopping optimization complete strategy guide covers channel prioritization, budget allocation models, and platform-specific implementation guidance in detail.

Frequently Asked Questions