Transactional agents must act on underspecified requests. "Book me a flight from Chicago to San Francisco" omits budget, timing, stop tolerance, and airline loyalty, any of which can change the best option by hundreds of dollars. The challenge is identifying, without a user-specific prior, which unstated attribute preference is most likely to change the recommendation.

We introduce ELICITBENCH, a benchmark of 1,387 travel and housing scenarios pairing real search-result pages (SERPs) with transaction-grounded preference profiles. Each scenario partitions attributes into stated and hidden preferences. An ex-post analytical oracle assigns question-value labels based on the normalized reduction in market-conditioned recommendation regret. Agents are evaluated by task success, question count, and net utility under a super-linear patience cost.

ELICITBENCH reveals that plausible questions are often not decision-relevant: on Gemini 2.5 Flash and Claude Sonnet 4.5, unguided agents ask natural follow-up questions with little downstream value. SERP-aware prompting raises net utility from 0.073 to 0.306 and task success from 21.7% to 57.6% on Gemini, and net utility from 0.221 to 0.346 on Sonnet while asking fewer questions. The selection of items is a second bottleneck: with every preference stated, agents still pick the wrong item in about 11% of the scenarios.