A recent test on web extraction services found that adding a simple instruction, "Do not guess," significantly reduced the number of made-up fields returned by AI models. Without this instruction, 71% of fields were fabricated, but with it, that number dropped to 20%. The test, which involved 42 pairs of pages across 7 types, evaluated the honesty of extractors in returning null values for missing fields. The results suggest that explicitly instructing AI models not to guess can improve their accuracy and reliability, allowing buyer agents to trust the information they receive and potentially reducing errors and costs.