Short answer
GTM-Bench is a benchmark that tests whether AI agents can do go-to-market research: work out what a seller offers, define its ideal customer, and return a ranked list of buyer accounts backed by evidence. It has 72 tasks built from 59,881 real prospecting queries. Blackpearl Group published it in June 2026; three of its eight authors are on the Levy Street team.
- Weak leads score negatively, so agents cannot win by returning thousands of rows.
- In the June 2026 paper, several general-purpose agents scored below zero overall.
- A purpose-built system with proprietary data led the board by a wide margin.
- Evaluation uses LLM judges for offer, ideal customer profile, match quality and evidence.
- Publisher
- Blackpearl Group Limited (NZX/ASX: BPG)
- Authors
- Maksymilian Polaczuk, Jeremy Baron, Reuben Horne (joint first), Trevor Cavill, Callum Mudgway, Chris Herrmann, Patrick Horne, Sam Daish
- Tasks
- 72 tasks across 11 task types and 15 industries
- Source data
- 59,881 real prospecting queries submitted to Bebop.ai
- Published
- June 2026; leaderboard updated September 2026
Why a go-to-market benchmark matters
Sales prospecting is one of the first jobs businesses hand to AI agents, and one of the easiest to fake. An agent that returns 6,000 companies looks productive. If most of them are wrong, it has created work, not saved it. GTM-Bench was designed to measure what a sales team actually needs: the right buyers, each with evidence a person can check.
How the benchmark was built
The tasks come from real behaviour, not invented prompts. The authors sampled 59,881 opening queries that sellers submitted to Bebop.ai, Blackpearl Group’s AI sales agent, between February 2025 and May 2026. They clustered them into a taxonomy of what sellers actually ask, drafted 140 candidate questions, and reviewed them by hand down to 72 tasks across 11 task types and 15 industries.
How agents are scored
- Offer
- The agent infers what the seller sells. A judge scores it on five criteria.
- Ideal customer profile
- The agent defines who should buy. A judge scores it on six criteria.
- Buyer records
- Every returned row is checked for company fit, contact fit, offer fit, evidence and contactability, then audited for identity, claim support and data consistency.
- Utility
- A-grade records score +1, B-grade 0, anything below B −1. The total is multiplied by offer and profile quality.
The negative score is the important design choice. In the authors’ words, the design “deliberately treats low-quality records as harmful”. It rewards precision, which is what a busy sales team wants from an agent.
What the June 2026 results showed
| System (June 2026 paper) | Net score | A-grade share |
|---|---|---|
| Blackpearl RTSA (purpose-built) | 26,615.6 | 40.9% |
| GPT-5.5 (Codex CLI) | 4,040.9 | 37.7% |
| Claude Sonnet 4.6 (Claude Code) | 400.1 | 27.3% |
| Claude Opus 4.7 (Claude Code) | −2,476.6 | 31.7% |
| DeepSeek V4 Pro | −3,398.0 | 21.8% |
| Gemini 3.5 Flash | −10,671.9 | 13.6% |
| Kimi K2.6 | −15,402.3 | 22.8% |
The paper’s summary: “the purpose-built RTSA system is the clear winner by net score, with many generalist agents even scoring negatively”. One general agent returned 6,342 rows for a single question. The September 2026 leaderboard on gtm-bench.ai adds newer systems; quote the paper or the leaderboard, not a mix of both.
What it teaches about building agents
For anyone deploying agents on commercial work, three lessons generalise:
- Data access beats model size. The leading systems had better data, not just better models. The same pattern shows up in retrieval-augmented generation projects.
- Penalise volume. If your evaluation only counts hits, agents learn to flood. Score false positives as costs, as described in our guide to agent evaluations.
- Judge the reasoning, not just the output. Grading the offer and customer profile catches agents that stumble onto good rows for the wrong reasons.
We apply this thinking in LLM evaluation work for clients, and in our own products. Read more in AI agents for go-to-market research.
Written by the Levy Street team in Wellington. Last updated 19 October 2026.