AI search agents stumble when queries are unclear, not when searching

AI search agents handle multi-step research tasks well, but their real challenge lies in ambiguity—not in the search itself. According to a new benchmark, models that repeatedly guess instead of asking clarifying questions perform worse, with accuracy as low as 51.9 percent. Even top models only achieve 43 percent overall accuracy when faced with vague queries.

The issue isn’t the search process but the inability to refine unclear inputs. When ambiguity is removed, accuracy improves dramatically, sometimes by up to 40 points. This suggests that AI search agents need better mechanisms to handle uncertainty in user requests.

The cost of guessing over asking

Researchers behind the DiscoBench benchmark found that models often default to searching repeatedly rather than seeking clarification. This approach backfires, as vague queries lead to inconsistent or misleading results. The data shows that even the most advanced models fail to bridge the gap between unclear inputs and precise outputs.

A benchmark for clarity

DiscoBench evaluates how well AI search agents manage ambiguous queries. The results highlight a critical gap: models perform poorly when users don’t specify their intent. The benchmark’s findings underscore the need for systems that can detect and resolve ambiguity proactively, rather than making assumptions.

For now, users may need to refine their queries manually to get better results. Until AI search agents improve their ability to ask follow-up questions, clarity in input remains key to accuracy.

---

Source: The Decoder. AI-assisted editorial synthesis — TechnoExpress.