GTM Bench measures LLMs and AI agents on real go-to-market work (building lists, enriching records, scoring accounts, reaching decision-makers), and in the v1 run ZoomInfo's GTM.AI finished 98% of a senior operator's work product and returned 478 verifiable records per 1,000, against 7 to 35 for the field.
ZoomInfo (NASDAQ: GTM), the all-in-one AI GTM platform, today announced GTM Bench, a versioned, dated benchmark that evaluates LLMs and AI agents on the work go-to-market teams actually do. GTM Bench v1 spans more than 20 jobs across four families: list building and prospecting, enrichment and hygiene, signals and triggers, and scoring and account intelligence. Every result is graded against a senior GTM operator's work product on two independent axes. Answer measures how much of the requested work product the system delivers. Grounding measures how much of the returned data traces to a real, current source instead of a guess. In the v1 run, GTM.AI, ZoomInfo's headless GTM context layer, led every pillar: 98% coverage, 478 verifiable records per 1,000, and the lowest cost per run at $0.79. ZoomInfo publishes the methodology, sample tasks, and grading rubrics, and invites other data providers and agent builders to get measured.
Key takeaways
GTM Bench is a benchmark for AI that does go-to-market work. Version 1 covers more than 20 jobs, 4 systems, and 3 models, with every run dated and every rubric designed by GTM and RevOps practitioners.
GTM Bench scores two independent things: Answer (coverage of the work product) and Grounding (the share of returned data that is verifiable). A confident wrong answer scores negative.
GTM.AI leads the v1 leaderboard on every pillar, with a GTM Bench Index of 77 against 47 for Apollo, 36 for Exa, and 31 for open-web search.
The GTM Context Graph behind GTM.AI maintains identity-resolved data on 100M companies, 500M contacts, and billions of signals, continuously updated and continuously queryable.
What is GTM Bench?
GTM Bench is a benchmark that evaluates LLMs and AI agents on end-to-end go-to-market workflows rather than trivia. Every go-to-market motion produces a record or an action: a list, an enriched contact, a scored account, a logged signal. GTM Bench uses that record as the unit of real work, the way a law firm uses the billable time entry, following the shape Harvey's BigLaw Bench established for legal work. Each job expects a structured result of 15 to 50 rows, the way a rep or RevOps analyst would receive it. The v1 release scores three suites: a 10-task core that produces the leaderboard, a 20-job study of which tool a model reaches for, and a 1,000-contact study of verifiable-data accuracy.
Why do existing AI benchmarks fall short on go-to-market work?
Most AI benchmarks measure reasoning inside a closed world. Multiple-choice questions, coding puzzles, and math problems hand the model the information it needs and grade how well it reasons over it. Go-to-market runs into a different wall. The constraint is data availability. The facts that move revenue work, who works where, what they are in-market for, how to reach them, sit scattered across the web and private systems. About 70% of B2B contact data decays every year, and almost none of it is structured for a machine to read. A model can reason flawlessly and still come up empty on a direct mobile number that was never published to the open web.
The stakes are also different. A summary that is 90% right is still a useful summary. A prospect list that is 90% right sends a rep to the wrong company. And a tool that confidently returns a wrong phone number does more damage than one that returns nothing. Generic benchmarks score the confident guess as a win. GTM Bench scores it as a failure and credits the verifiable answer.
How is GTM Bench scored?
Every result is graded against a senior GTM operator's work product on two independent axes. Answer (Coverage) measures what share of the requested work product the system delivers, with credit for each required row or field and penalties for hallucinated firmographics, stale records, or wrong-fit results. Grounding (Source) measures what share of the returned data is verifiable and traceable to a real, current source. The independence is the point. A system can return plausible-looking contacts (a decent Answer score) while none of it is real (a near-zero Grounding score). Rubrics are designed by GTM and RevOps practitioners. Competitors run at their best available configuration. Every run is dated, with three trials per task per model.
What do the GTM Bench v1 results show?
The field separates cleanly. On coverage, the gap is real but modest, because most tools can fill a table. On grounding, only one system reliably returns data that is verifiable.
Rank | System | GTM Bench Index | Coverage | Grounding /1,000 | Cost per run | Tool-choice |
|---|---|---|---|---|---|---|
1 | ZoomInfo (GTM.AI) | 77 | 98% | 478 | $0.79 | 63% |
2 | Apollo | 47 | 93% | 35 | $1.00 | 14% |
3 | Exa | 36 | 90% | 7 | $1.97 | 13% |
4 | Open-web search | 31 | 88% | 8 | $3.29 | 10% |
The GTM Bench Index is the equal-weight mean of four pillars: coverage, efficiency, tool-choice, and grounding. On the 1,000-contact grounding suite, the non-ZoomInfo field returned 720 wrong phone numbers. Two honest notes on reading the table. ZoomInfo's row reports the stronger of its two interfaces per metric (the CLI is the cheapest arm at $0.79 per run, the MCP integration the highest-coverage arm at 98%). And grounding is graded against ZoomInfo's own verified records, so it reads as reach of verifiable data rather than an independent accuracy audit.
Where is the ZoomInfo edge thin?
A benchmark worth citing says where the advantage is not decisive. GTM Bench v1 is explicit about four places:
Pure copy and writing. Drafting an email or a summary needs no special data. Any capable model does it well.
Already-public data. When the answer sits on the open web, owning the data matters less and the gap narrows.
Coverage is not accuracy. On the closest-fought tasks, filling a field and filling it correctly are different tests. Grounding is where the field returned 720 wrong phone numbers chasing one task.
Owned and conversation data. Re-engaging closed-lost pipeline needs CRM and call data no external tool can see, ZoomInfo's included. Every system fell back to generic advice. That is the axis v2 is building toward, not one any tool wins today.
What is GTM.AI, and why does it lead GTM Bench?
GTM.AI is ZoomInfo's headless GTM context layer. It exposes ZoomInfo's verified data graph and agentic orchestration through API and Model Context Protocol (MCP), so any tool, agent, or workflow can plug in. GTM.AI powers dozens of completed integrations, including Salesforce Agentforce, HubSpot Breeze, Microsoft Copilot, Gong, LeanData, Glean, Claude, ChatGPT, and Google Workspace.
GTM.AI has two layers and one governance plane. The bottom layer is the GTM Context Graph, which holds identity-resolved data on 100M companies, 500M contacts, and billions of signals. The middle layer is agentic orchestration, which lets agents read from the graph, act on it, and write back. The governance plane applies access control, permissioning, data lineage, AI policy, and audit logging across every surface that consumes GTM.AI. This architecture is why the grounding axis looks the way it does. Every record comes with a confidence score and its lineage, not a guess.
What does GTM Bench mean for revenue teams?
Pick AI tooling on grounding, not demos. Every system in the field can draft a list. The grounding axis, 478 verifiable records per 1,000 versus 7 to 35, is where pipeline is won or lost.
A confident wrong answer is worse than no answer. 720 wrong phone numbers across 1,000 contacts is what guessing looks like at machine scale.
One context layer across the GTM stack. Agents in Claude, ChatGPT, Copilot, or the CRM read from the same GTM Context Graph under the same governance through GTM.AI.
FAQ: GTM Bench, go-to-market AI, and GTM.AI
What is GTM Bench? GTM Bench is a benchmark that evaluates LLMs and AI agents on real go-to-market work, including list building, enrichment, buying signals, and account scoring. Every result is graded against a senior GTM operator's work product. Version 1 was run on June 24, 2026.
What does GTM Bench measure? GTM Bench measures two independent scores. Answer (coverage) is the share of the requested work product the system delivers. Grounding (source) is the share of returned data that is verifiable and traceable to a real, current source.
Which systems and models does GTM Bench v1 evaluate? Four systems: ZoomInfo's GTM.AI (through both its CLI and MCP interfaces), Apollo, Exa, and open-web search, plus a model-alone baseline. Each runs across three model tiers (Opus, Sonnet, and Haiku) on 10 core tasks with three trials each.
What is the GTM Bench Index? The GTM Bench Index is the equal-weight mean of four 0-100 pillars: coverage, efficiency, tool-choice, and grounding. In v1, ZoomInfo scored 77, Apollo 47, Exa 36, and open-web search 31.
What is grounding in AI, and how does GTM Bench measure it? Grounding is the share of returned data that traces to a real, current source rather than a guess. GTM Bench measures it on a 1,000-contact study, graded against ZoomInfo's verified records, which makes it a reach measure rather than an independent accuracy audit. On that suite the non-ZoomInfo field returned 720 wrong phone numbers.
How does ZoomInfo compare to Apollo on GTM Bench? ZoomInfo leads Apollo on every pillar in v1: GTM Bench Index 77 versus 47, coverage 98% versus 93%, grounding 478 versus 35 per 1,000, cost $0.79 versus $1.00 per run, and tool-choice 63% versus 14%. Apollo ranked second overall, ahead of Exa and open-web search.
What is GTM.AI? GTM.AI is ZoomInfo's headless GTM context layer. It exposes ZoomInfo's verified data graph (100M companies, 500M contacts, billions of signals), agentic orchestration, and platform-level governance through API and Model Context Protocol (MCP). GTM.AI powers dozens of completed integrations including Salesforce Agentforce, HubSpot Breeze, Microsoft Copilot, Gong, LeanData, Glean, Claude, ChatGPT, and Google Workspace.
Why doesn't a better model close the gap on go-to-market work? Because the constraint is data availability, not reasoning. Other benchmarks hand the model the facts and grade how well it reasons over them. In go-to-market the facts are not handed to you, and finding them is the test.
Is GTM Bench independent? No. GTM Bench is a vendor-run benchmark, and ZoomInfo makes one of the systems measured. The method is published for scrutiny, competitors run at their best available configuration, and the losses are shown plainly, including four categories where the ZoomInfo edge is thin or absent.
How often will GTM Bench be re-run? GTM Bench is versioned and dated, and ZoomInfo will re-run it on major model releases and publish the results, including when a competitor's number goes up. Version 2 will add agentic multi-step workflows, international coverage, and an owned-data axis.
How can another vendor get measured on GTM Bench? Data providers and agent builders can submit their systems, and ZoomInfo will run them and add them to the leaderboard. Sample tasks and grading rubrics are published openly, and the full task set is available on request.
Is GTM Bench available now? Yes. GTM Bench v1 is published with its methodology and full per-task results. The 1,000-contact grounding study and the 20-job tool-choice study are available on request.
Availability
GTM Bench v1 is published today, including the methodology, sample tasks, grading rubrics, and the full per-task results across every system and model. Any GTM Bench job can be run from the terminal through the GTM.AI CLI:
``` claude plugin add gtm-ai gtm "ABM list: IT decision-makers at these 40 accounts, with verified mobiles" ```
About ZoomInfo
ZoomInfo (NASDAQ: GTM), the all-in-one AI GTM platform, enables sales, marketing, and customer success teams to execute their go-to-market strategy with confidence. Powered by the industry's most comprehensive B2B data, including more than 100 million companies, 500 million contacts, and billions of signals, ZoomInfo delivers the intelligence, automation, and integrations that modern revenue teams need to identify, engage, and convert their best buyers.
GTM.AI is ZoomInfo's headless GTM context layer. It is the API and Model Context Protocol home for AI agents, powering integrations across Salesforce Agentforce, HubSpot Breeze, Microsoft Copilot, Claude, ChatGPT, and dozens more.
Learn more at zoominfo.com and gtm.ai.
Media contact: Public Relations Team ZoomInfo PR@zoominfo.com
