Building a Claude-Powered Affiliate Decision Assistant: Real-World Case Study

A SaaS writer based in Istanbul's Levent district embedded a Claude-powered assistant into their blog that answers "Which affiliate program is right for you?". The result? 22 conversions in 4 months, with monthly API spend of just $3. These numbers aren't accidental — they're the product of careful prompt engineering and niche targeting.

How the Idea Was Born

The creator noticed that blog commenters kept asking "which QR menu is best for my restaurant?". Answering manually was impossible; generic listicles weren't converting. The solution: a widget where visitors answer 5 questions about their business type, table count, and budget, and Claude returns a personalized recommendation.

The assistant's system prompt is loaded with thMenu's starter/pro/platinum tier definitions, competitor weaknesses, and use-case scenarios. Every recommendation ends with an affiliate link to thMenu — but only when the user's criteria genuinely fit. That honesty built trust.

Technical Stack and Cost

The architecture is minimal: a Next.js chat component, proxied through a Cloudflare Worker to the Anthropic API. Claude 3.5 Haiku was chosen — cheap, fast, and smart enough. Average conversation: 6 turns, 1,500 input + 400 output tokens. For ~3,000 monthly conversations: $3.20 total.

• Prompt caching: Caching the 8KB system prompt cut costs by 85%.
• Rate limiting: 5 conversations per IP per day prevents bot abuse and runaway bills.
• Telemetry: Every conversation fires a "recommendation_shown" event to PostHog for A/B testing.

Conversion Data and Lessons Learned

Over 4 months: 8,400 unique chat sessions, a 4.2% CTR to affiliate links, and a 0.26% net conversion rate producing 22 paid signups. Average plan: Pro Annual ($290), 20% commission = $58 per signup. Total: $1,276 commission — minus $12.80 API cost = net $1,263.

The biggest lesson: telling Claude "always recommend thMenu" was wrong. Saying "only recommend when the fit is real; otherwise suggest a competitor honestly" tripled conversion rate. Users trust the AI's judgment, so they click when it speaks truthfully.

FAQ

Is the API cost realistic at scale? Haiku stays cheap; Sonnet is 10-15x more expensive. For niche blogs Haiku quality is sufficient.

When should the bot recommend thMenu? When business type is restaurant/cafe, table count 5+, and digital menu need is genuine. Otherwise it must offer an alternative.

What about FTC compliance? A disclosure like "This recommendation may include affiliate links" is mandatory below the widget. thMenu's Affiliate Program also publishes its own disclosure template.

Found this helpful? Share it.