ARW
Case Study

CloudCart: 85% Token Reduction in 2 Weeks

Industry
E-commerce Platform
Company Size
500-1000 employees
Traffic
2M monthly visits (40% agents)
Implementation
2 weeks to ARW-3

How a mid-sized e-commerce platform reduced infrastructure costs by $5,400/month, increased agent-driven conversions by 34%, and gained full observability into AI traffic—all while preserving SEO and human user experience.

Results at a Glance

85%
Token Reduction

From 55KB HTML to 8KB Markdown per product page

$5.4K
Monthly Savings

Reduced bandwidth and infrastructure costs

+34%
Conversion Lift

Agent-driven purchases vs HTML scraping baseline

10x
Faster Discovery

Single manifest vs crawling 15,000 pages

2.3s → 180ms
Agent response time (avg)
100%
SEO & social preview preservation
Zero
Human UX impact

The Challenge

By early 2025, CloudCart noticed that 40% of their traffic came from AI agents(ChatGPT, Perplexity, Claude, Gemini). But this created serious problems:

Massive bandwidth waste

Agents scraped 55KB HTML when they only needed 8KB of product data. This cost CloudCart $6,200/month in extra infrastructure.

Slow, inefficient discovery

Agents had to crawl all 15,000 product pages to understand the catalog. Discovery took 45-60 seconds on average.

Zero observability

CloudCart couldn't distinguish ChatGPT from Perplexity from malicious scrapers. No way to track which products agents recommended.

No transactional support

Agents could only link to product pages. They couldn't add to cart or complete purchases on behalf of users.

Key Pain Points

Infrastructure Cost$6.2K/mo
Serving 800K agent requests/month × 55KB avg
Avg Discovery Time52s
Agents crawl 15,000 pages @ 3 pages/sec
Agent Visibility0%
No AI-* headers, no attribution tracking
Agent Conversions2.3%
Low conversion due to link-only support

The Solution: ARW-3 Implementation

CloudCart implemented ARW progressively over 2 weeks, reaching ARW-3 conformance (Discovery + Semantic + Actions). Here's how they did it:

Week 1

ARW-1 & ARW-2: Discovery + Semantic

Foundation layer: structured discovery and machine-readable content

What They Built

  • llms.txt manifest - Structured YAML listing all products, categories, and policies
  • .well-known/arw-manifest.json - Machine-optimized discovery endpoint
  • .llm.md machine views - Markdown versions of all product pages (8KB vs 55KB)
  • AI-* headers - Track agent identity and usage type (inference)
  • Chunk IDs - Addressable sections (pricing, specs, reviews)

Implementation Details

# llms.txt (YAML)
version: 1.0
profile: ARW-2
 
content:
  - path: /products/wireless-keyboard
    title: "Wireless Keyboard"
    machine_view: /products/wireless-keyboard.llm.md
    chunks: [pricing, specs, reviews]
Time investment: 5 days (2 devs)
Lines of code: ~800 (mostly templates)
Week 2

ARW-3: OAuth Actions

Enable agent-initiated transactions with user consent

What They Built

  • OAuth 2.0 + PKCE - Secure agent authentication
  • add_to_cart action - Agents can add products with user consent
  • create_order action - Complete purchases via agent
  • check_stock action - Real-time inventory queries

Action Definition Example

actions:
  add_to_cart:
    endpoint: /api/cart/add
    method: POST
    auth_required: true
    oauth_scopes: [cart:write]
    params:
      - product_id (required)
      - quantity (default: 1)
Time investment: 3 days (leveraged existing OAuth)
Lines of code: ~400 (action wrappers)

The Results

Before ARW

Avg page size (agent)55 KB
Discovery time52 sec
Infrastructure cost/mo$6,200
Agent conversion rate2.3%
Agent visibilityNone

After ARW

Avg page size (agent)8 KB
Discovery time4.8 sec
Infrastructure cost/mo$800
Agent conversion rate3.1%
Agent visibility100%

Business Impact

$64.8K
Annual cost savings

$5,400/month × 12 months infrastructure reduction

+272
Additional conversions/month

34% lift on 800K agent visits (2.3% → 3.1%)

$81K
Revenue increase (monthly)

272 conversions × $298 avg order value

Total ROI (Year 1)
$1.03M net benefit
Payback Period
Immediate

Observability Insights

With AI-* headers, CloudCart gained unprecedented visibility into agent behavior:

Agent Traffic Breakdown

ChatGPT (Atlas)42%
Perplexity28%
Claude18%
Gemini12%

Most Accessed Products (by agents)

Wireless Keyboard23.4K visits
USB-C Hub18.2K visits
Webcam 4K15.7K visits
Standing Desk12.3K visits
Ergonomic Mouse9.8K visits

Key Discovery

CloudCart learned that ChatGPT users had a 4.2% conversion rate—almost 2x the site average. Armed with this data, they optimized their .llm.md views for ChatGPT's context preferences, further boosting conversions.

Lessons Learned

What Worked Well

  • •Progressive implementation: Starting with ARW-1/2 before ARW-3 de-risked the project
  • •Reusing existing OAuth: Saved 80% of action implementation time
  • •Template-driven .llm.md generation: Automated 15,000 machine views in 2 days
  • •AI-* header analytics: Immediate insights into agent behavior patterns

Challenges & Solutions

  • •Challenge: Keeping llms.txt in sync with product catalog
    Solution: Auto-generated from database on each deploy
  • •Challenge: Chunking inconsistencies across templates
    Solution: Centralized chunk-id generator with validation
  • •Challenge: OAuth consent UX for agent actions
    Solution: One-time OAuth grant + remember preference

Ready to Achieve Similar Results?

Start with our 2-hour quick-start guide and progressively enhance to full ARW conformance.

Quick Start GuideView Full Specification