A benchmark test evaluated 16 AI models and web extraction services on their ability to correctly return null for missing fields rather than inventing data. Adding the instruction 'Do not guess' reduced hallucinated fields from 70.7% to 20.2%, with Firecrawl performing worst and plain fetch plus GPT-6 Luna performing best at low cost. A cheap verification method using GPT-6 Luna caught 20 of 24 hallucinated values for under a cent.