AI agents for go-to-market research.

What GTM-Bench measured

GTM-Bench tests the job sales teams most often hand to AI agents: given a seller, find the buyers. Each agent must state what the seller offers, define an ideal customer profile, and return a ranked list of companies or contacts with evidence for each. The 72 tasks span 11 task types and 15 industries, and were built from 59,881 real opening queries that sellers typed into Bebop.ai between February 2025 and May 2026.

It is a Blackpearl Group publication, released in June 2026. Max Polaczuk, Jeremy Baron and Reuben Horne are joint first authors, and all three are on the Levy Street team. This page is our analysis of what its results mean for teams buying or building prospecting agents.

The scoring rule is the important part. Every returned record is judged on company fit, contact fit, offer fit, evidence and contactability, then audited for identity, claim support and data consistency. A-grade records score +1, B-grade 0 and anything below B −1, and the total is scaled by how well the agent understood the offer and the customer. In the authors’ words, “This design deliberately treats low-quality records as harmful.”

The June 2026 results

These are the paper’s overall results (Table 9), quoted exactly, with 95% bootstrap confidence intervals.

The paper’s summary: “the purpose-built RTSA system is the clear winner by net score, with many generalist agents even scoring negatively: providing more substandard matches than good ones.” Note that Claude Sonnet 4.6’s interval crosses zero, so only GPT-5.5 among the generalists is clearly positive. Blackpearl’s September 2026 leaderboard at gtm-bench.ai adds newer systems; quote one source or the other, not a mix.

Lesson 1: data access matters more than the model

The clearest comparison in the paper holds the model fixed. GPT-5.5 with access to Blackpearl’s structured company and contact data scored +4,040.93. The same model “restricted to web evidence, without structured data access” scored +1,015.40, about a quarter as much.

The leader, RTSA, is a purpose-built system rather than a bigger model; the paper notes such systems can combine “multi-agent configurations, proprietary models and data” among other engineering. And a larger model did not guarantee a better result: Claude Opus 4.7 finished below Claude Sonnet 4.6 on net score. When three models were run through the same Hermes agent, they behaved very differently: DeepSeek controlled and cheap, Kimi expansive and slow, Gemini fast but losing its audit structure.

For buyers, the practical question is not “which model?” but “what data can the agent see, and how fresh is it?” The same lesson drives our RAG development work.

Lesson 2: precision beats volume

Agents that flooded their output were punished. The paper records Kimi returning 6,342 rows on one question, Gemini 3,500 on another and Sonnet 2,617 on a third. By contrast, “Codex’s maximum production trace output is 60 rows, a major reason it wins generalist production utility.”

A higher hit rate does not rescue a flooded list either. “Claude Opus 4.7 has a higher A-grade rate than Sonnet but a lower net score because its broader outputs include enough sub-B rows to offset high-quality leads.”

Spend tells the same story. RTSA cost $470.03 across the benchmark, the most of any system, yet delivered the best net score per dollar (56.63), against 27.44 for GPT-5.5 and −171.56 for Kimi K2.6.

Illustrative example, with assumed numbers: a sales rep at an Auckland software company spends three minutes checking each agent-sourced lead. Agent A returns 1,000 rows, 25% of them A-grade: 250 good leads for 50 hours of checking, or 12 minutes per good lead. Agent B returns 120 rows, 45% A-grade: 54 good leads for 6 hours, or under 7 minutes each. Agent A finds more good leads in total, but at almost twice the review cost per lead, and its 750 weaker rows are a standing temptation to spam.

Lesson 3: judge the reasoning, not just the list

GTM-Bench scores the agent’s understanding of the offer and the ideal customer, not just the rows. That catches agents that stumble onto good leads for the wrong reasons, which will not repeat on the next request.

The trace analysis explains what separates good runs: “successful systems combine broad retrieval with evidence-based pruning, while failures stem from identity errors, unsupported claims, and poor stopping decisions.” The strongest runs pulled a broad candidate pool, then used scripts, website search and row-level filters to cut it down before writing the final file. Sonnet and Opus, in the paper’s words, “often do retrieval well but sometimes skip the hard prune.”

How NZ sales teams should evaluate a prospecting agent

1. Build a test set from real requests. Take 20 to 30 prospecting asks your team actually made, including vague ones, and freeze them.
2. Define the grades before you look. Write down what makes a lead A-grade for you: right company, right person, a verifiable reason to buy now.
3. Count bad leads as a cost. Score +1, 0 and −1 like GTM-Bench, or weight by review minutes. Never count hits alone.
4. Check identity against NZ records. Trading names and legal names often differ; confirm companies against the Companies Register or NZBN before they reach your CRM.
5. Read the offer and customer profile first. If the agent misunderstands what you sell, discard the list.
6. Compare cost per A-grade lead. Include model or subscription cost and human review time.
7. Test the integration. Push a sample into HubSpot or Pipedrive and check fields, duplicates and ownership before scaling.

New Zealand law shapes what happens next. Since 1 May 2026, IPP 3A requires an organisation that collects personal information from a source other than the individual to take reasonable steps to make them aware of specified matters, unless an exception applies. Enriching contacts from the web is indirect collection. Separately, the Unsolicited Electronic Messages Act 2007 says a person “must not send, or cause to be sent, an unsolicited commercial electronic message that has a New Zealand link”, and restricts using address-harvesting software to support such messages. This is general information, not legal advice.

Where prospecting agents go wrong

The paper is candid about its own limits: the benchmark “is not yet able to determine whether a matched prospect would actually buy”, LLM judges “remain imperfect”, and its private data environment “limits immediate reproducibility outside the benchmark operator”. Treat any vendor’s benchmark claim with the same care, and use our AI agent acceptance checklist and guide to agent evaluations when you run your own. If you want help building one, see AI agent development.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.