← Guides
benchmarks

GTM Bench v1: Specialist Retrieval vs. Web Research

Measuring how much a structured data tool helps an AI agent on B2B research tasks.

·4 min read

Sales and marketing teams now point AI agents at research questions that used to be manual analyst work. Find 50 New York companies that just raised funding, including headcount and whether they are hiring sales. Find 20 AI CTOs in the Bay Area with their email and LinkedIn profiles.

Two things determine whether an agent answers these well. The first is the model's reasoning. The second is the data tools it can reach. We built GTM Bench v1 to isolate the second factor. Given the same agent and the same prompt, we wanted to see how much the retrieval tool matters.

We compared web research against four retrieval tools: the ZoomInfo CLI, the ZoomInfo MCP, the Apollo.io CLI, and Exa.ai neural search. We evaluated a single agent (Claude Code running Opus, Sonnet, and Haiku) across these five settings.

The results show a massive divide in contact data. Specialist data tools return verified work emails roughly four times as often as web or neural search. They also operate at a fraction of the cost and token usage.

Download the full GTM Bench v1 report (PDF)

Complete methodology, per-task scores, and the raw cost and token data behind this guide.

The email gap

The defining difference between specialist tools and open-web research is contact data. On the five contact-discovery tasks, the structured data tools filled 84 to 94 percent of the verified work email cells. Web and neural search filled 22 to 27 percent.

Retrieval ToolVerified Work-Email Coverage
ZoomInfo MCP94%
ZoomInfo CLI91%
Apollo84%
Exa.ai27%
Web Search22%

Open web research generally cannot return a verified work email. The agent guesses an email pattern like first.last at domain, or it leaves the cell blank. On a task to find AI CTOs in the Bay Area, the data arms reach near-total coverage. The web baseline struggles to reach 80 percent. The shortfall is concentrated entirely in the email column.

Cost, tokens, and time

Structured tools are also more efficient. Running the ten benchmark tasks on Claude Opus 4.8 with the web baseline costs $43.88 and consumes 11.1 million work tokens. The web agent issues multiple searches. It fetches large pages and reads them.

A structured query returns compact records. The ZoomInfo CLI arm completes the same ten tasks for $10.12 and 460,000 work tokens. The ZoomInfo MCP records the lowest total time at 32 minutes. The web baseline takes 102 minutes. Apollo costs $14.78 and takes 101 minutes, though it struggles on Haiku due to rate limits on email reveals.

Agent ArmTotal Cost (USD)Work TokensTotal Time
ZoomInfo CLI$10.12460K53 min
Apollo$14.78670K101 min
ZoomInfo MCP$15.801.48M32 min
Exa.ai$36.753.38M55 min
Web Search$43.8811.10M102 min

Coverage vs. cost per run

10 GTM tasks, averaged across 3 models · run 2026-06-24

What coverage means for companies

On company-discovery tasks, such as finding fast-growing US fintechs by headcount, every arm achieves high coverage. This result requires careful reading.

GTM Bench v1 measures coverage. It checks whether the requested cells were filled. It does not verify the values themselves. The scored columns on company tasks are firmographic basics like company name, website, and headcount. A web researcher can find these for almost any company. The actual task constraint (fast-growing, recently funded, or running Salesforce) is not a scored column.

The web arm infers Salesforce use from a job posting for a Salesforce Administrator. The ZoomInfo CLI cites technographic install records. Fill-rate treats the two as equal. We plan to measure correctness and constraint satisfaction in v2.

The risk of coverage metrics

A completeness metric on its own can rank a less accurate configuration above a more accurate one. We saw this on an enrichment task where the agent had to find executives at acquired companies.

For this task, Opus recognizes the acquired-company ambiguity and leaves the cells empty, scoring 73 percent coverage with the ZoomInfo CLI. Haiku takes the top fuzzy match and enriches the wrong company, scoring 95 percent. It returns an email at an unrelated firm also named Segment. The more careful model scores lower.

This is a general caution for agent benchmarks. GTM Bench v2 will pair agent coverage with correctness using ZoomInfo's data-quality systems, which rely on a 300-person data research team. For now, the clearest signal is in contact data. The ability to supply verified emails dictates performance.