AI Recommendations Tab: Which Suggestions Customers Actually Click (6-Month Data)

A 14-table cafe in Istanbul's Kadikoy district turned on the AI recommendations tab and watched its average basket climb from roughly $4.20 to $5.20 over six months. It's not an outlier — six months of thMenu Pro+ AI recommendation data across 52 restaurants tell us exactly which suggestion types diners actually click and which they ignore.

Click-Through Rates: Past Orders Lead by a Mile

Across 186,000 recommendation impressions, the hierarchy was crystal clear. "Based on your past order" headers pulled 38% CTR — when diners see something they already love, the yes-tap is almost instant.

"Trending this week" landed at 29%, "personalized for you" at 17%, "chef's pick" at 11%, "goes well with dessert" at 5%. Social proof and personalization both work — but personalization only when it's powered by real behavior, not made-up "for you" labels.

Effect on Basket Size

The 52-restaurant average showed a 14% basket lift in the first 30 days, 22% at day 90, and 24% stabilized at day 180. Kadikoy cafe slightly outperformed because table-service plus hot-drink mixes lend themselves to natural pairings.

• Main dish selected → drink suggestion: 31% add-to-cart
• Hot drink selected → dessert suggestion: 24% add-to-cart
• Salad selected → soup/starter suggestion: 12% add-to-cart

Setting Up the Engine Right

thMenu's Pro+ AI engine runs on Cloudflare Workers AI (LLaMA 3.1 8B) with a 7-day KV cache that keeps costs negligible. Suggestions are built from real order history plus allergen/diet tags plus time-of-day signals — never hallucinated.

For the cold-start week, "chef's pick" suggestions are manually flagged in the admin panel; once 200 orders accumulate, the engine auto-enables history-based recommendations. Boosting low-volume high-margin items instead of bestsellers grew gross margin by 8% at day 90.

FAQ

Which plans include AI recommendations? Pro and Platinum tiers ship it on by default. Starter is excluded. Stripe webhook normalizes the tier so upgrades activate within five minutes.

Is the diner data GDPR-compliant? Yes — the engine uses anonymous device fingerprint plus order history; no PII is stored. Encrypted pgcrypto columns are reserved for KYC, not menu analytics.

Do bad suggestions annoy diners? In A/B tests, only 0.4% flagged a "wrong category" issue. Suggestions are framed as "want to add?" — never auto-added to cart.

Found this helpful? Share it.