How BotPicked works

1. We ask, repeatedly

For every product category, we ask an AI model (currently Claude Haiku 4.5 by Anthropic) for its genuine recommendations. Each category is asked with three different neutral phrasings — "what's the best X", "what would you honestly tell a close friend", and "which X actually delivers value rather than hype" — and each phrasing is repeated multiple times. The model is explicitly instructed to recommend only products it genuinely considers good, and never because something is popular or heavily marketed.

2. The model answers blind

The model answers from its own knowledge with web search and tools disabled, in a structured format: five ranked picks with a reason and rough price. Any run where a tool fired is discarded. Every raw answer is stored.

3. We aggregate

Products are matched across runs and ranked by two numbers: consensus (the share of runs that recommended the product at all) and first-pick rate (the share of runs that made it the #1 recommendation). What you see on every guide is exactly those numbers.

What this is not

Nobody pays to appear here. There are no affiliate links. We don't edit the rankings. We also don't claim AI is an oracle: models can be wrong, out of date, or biased by their training data — that's why we show the consensus strength instead of pretending there's one true answer, and why you should always verify price and availability before buying.

Roadmap

Next: rankings from multiple AI models side by side (Claude vs ChatGPT vs Gemini), month-over-month tracking of how recommendations shift, and category requests.