AI agents hallucinate prices: LLM grounding with live data

The guide examines why AI agents return convincing but incorrect product prices: they rely on outdated training data, extract a price from the wrong page or offer, or conceal a tool failure. Tests covered 21 product pages across Amazon, Walmart, Best Buy, Target, and Shopify stores; 15 had one clear US price for a new item. Without tools, GPT-6 Luna and Claude Opus 5.5 each matched just 2 of those 15 prices, while often providing a price even when a page showed none. Built-in web search improved accuracy, but still produced incorrect prices, sometimes citing a link to a different product. DIY scrapers also struggled with blocked pages, wrong-country prices, missing data, and variant changes. The guide notes research finding that models may hide bad tool results when a tool reports “ok,” whereas an explicit error status prompted honest answers.

Its proposed solution is a Python agent grounded in live data from Apify’s E-commerce Scraping Tool, which supports many retailers and returns a consistent structure containing price, currency, URL, and scrape time. The key design principle is that code - not the language model - must insert each price into the response, while checking the data and clearly reporting failures. The tool manages browser rendering and proxies, reducing store-specific scraping work. The guide presents this approach as more dependable than memory, search citations, or a basic scraper because it combines current prices with traceable sources and explicit status handling.