How to Build an AI Product Recommendation Chatbot

An ai product recommendation chatbot for ecommerce acts like your best salesperson, available 24/7. I build these for stores where customers struggle to pick from a large catalog. The bot asks a few questions, understands the need, and suggests the right products with reasons why.

This is not a basic filter. The AI understands context. "Something for my dad who golfs and hates the cold" returns heated golf gloves, not just random golf gear.

What You Need Before Starting

• Your product catalog in a structured format (CSV, database, or API)
• Python 3.8+ with the Anthropic SDK
• A chat interface (web widget or messaging platform)
• Product images and URLs for links in recommendations

Step 1: Structure Your Product Catalog

Load your catalog into a format the AI can reference:

import json

def load_catalog():
    with open("products.json") as f:
        return json.load(f)

# Example product structure
# {
#   "id": "SKU-001",
#   "name": "Heated Golf Gloves",
#   "category": "golf-accessories",
#   "price": 89.99,
#   "description": "Battery-powered heated gloves for cold weather golf.",
#   "tags": ["golf", "winter", "heated", "accessories"],
#   "rating": 4.7,
#   "url": "/products/heated-golf-gloves"
# }

Step 2: Build the Recommendation Engine

Use the AI model with your catalog as context and tool use for lookups:

import anthropic

client = anthropic.Anthropic()

SYSTEM_PROMPT = """You are a product recommendation assistant for [Store Name].
Your job is to understand what the customer needs and recommend 2-3 products that fit.

For each recommendation, include:
- Product name and price
- One sentence explaining WHY it fits their need
- The product URL

Ask clarifying questions if the request is vague. Keep it conversational.
Never recommend products not in the catalog.

PRODUCT CATALOG:
{catalog}"""

def get_recommendations(messages, catalog):
    response = client.messages.create(
        model="claude-sonnet-4-20250514",
        max_tokens=600,
        system=SYSTEM_PROMPT.format(catalog=json.dumps(catalog[:100])),
        messages=messages
    )
    return response.content[0].text

Step 3: Add Guided Discovery Questions

For vague requests, the bot should ask smart narrowing questions:

DISCOVERY_PROMPT = """When a customer gives a vague request, ask ONE clarifying question.
Good examples:
- "Is this for yourself or a gift?"
- "What is your budget range?"
- "Do they prefer practical gifts or fun ones?"

Never ask more than one question at a time. Never list all options.
Pick the one question that narrows it down the most."""

Step 4: Track Recommendation Performance

Log what gets recommended and what gets clicked:

import sqlite3
from datetime import datetime

def log_recommendation(session_id, product_ids, clicked_id=None):
    conn = sqlite3.connect("recommendations.db")
    conn.execute("""
        INSERT INTO recommendation_log (session_id, products_shown, product_clicked, created_at)
        VALUES (?, ?, ?, ?)
    """, (session_id, json.dumps(product_ids), clicked_id, datetime.now().isoformat()))
    conn.commit()

Step 5: Handle Large Catalogs

If your catalog has thousands of products, pre-filter before passing to the AI:

def filter_catalog(query, catalog, max_results=20):
    """Pre-filter catalog by keyword matching before AI selection."""
    query_words = set(query.lower().split())
    scored = []
    for product in catalog:
        product_text = f"{product['name']} {product['description']} {' '.join(product['tags'])}".lower()
        score = sum(1 for word in query_words if word in product_text)
        if score > 0:
            scored.append((score, product))
    scored.sort(key=lambda x: x[0], reverse=True)
    return [p for _, p in scored[:max_results]]

What to Build Next

Connect recommendation data back to your sales analytics. Track which bot-recommended products convert to purchases. Use that data to refine the system prompt and improve recommendation quality over time.

Related Reading

• AI in Customer Service - how recommendation bots fit into the support ecosystem
• Lead Scoring with AI - scoring buyer intent from chatbot conversations
• The Data Flywheel Explained - how recommendation data compounds into better suggestions