How G2 evaluates AI Outbound SDR agents

Our first attempt at evaluating Outbound SDR agents starts with the prospecting foundation: finding prospects, filling in missing details and deciding who to contact first. We give products the same tasks and score what they return. This measures preparation for outreach, not performance across a complete sales cycle.

The benchmark defines 30 tasks across Discover, Enrich and Prioritize. Each scored task combines Accuracy (50%), Breadth (25%) and Depth (25%) on a 0–100 scale. This methodology covers prospect research and selection; outreach generation, message delivery and follow-up are outside this scope.

Discover10 tasks

A prospecting brief. Find people or companies that meet its eligibility rules.

Example task: Find up to 15 current Director-level-or-above cybersecurity leaders at companies headquartered in California, Oregon or Washington with 200–5,000 employees, with at most one person per company.

Enrich10 tasks

Partial records. Preserve the supplied identities, verify current details and fill required fields.

Example task: For 50 supplied name-and-company rows, verify current employment and return the current title and a credible professional-profile URL. Preserve each supplied person and use null when a detail cannot be verified.

Prioritize10 tasks

A supplied candidate list. Select eligible prospects and rank them when the task requires an order.

Example task: From 40 supplied companies, select the top 15 whose specified product description explicitly names both Salesforce and Marketo. Rank qualifying companies by supplied review count, then company name.

Each task specifies the goal, eligibility or exclusion rules, requested count or range, and required output fields. Ranked tasks also state the priority criteria and tie-breakers. Products receive the same task and inputs and complete the work independently.

These examples are shortened from the task definitions. Execution uses the full prompts, input files and output contracts.

Repeated runs

We run each task four times with the same vendor and setup. Scores are calculated per task attempt, then combined across runs and tasks.

Discover and Enrich are open-world tasks with multiple valid answers. Once the agent has returned its list of people, we deduplicate people by profile URL or name-and-company. We then attempt to validate the returned answers through multiple sources to ensure these are real people and that their returned title and company are valid.

To reward agents that return more prospects, we apply a coverage penalty for scenarios where an agent returned fewer than 50% of the requested entries, such as 10 sales leaders instead of the 50 requested.

Final score = (Accuracy × 0.50) + (Breadth × 0.25) + (Depth × 0.25) − coverage penalty

Accuracy (50%)

Are the details we can check correct? Validate the returned facts and eligibility using Lix and Exa, with evidence-based semantic review and email validity using NeverBounce.

100 × correct resolved checks ÷ all resolved checks

Breadth (25%)

Did the agent return enough distinct prospects? Remove duplicates and prospects confirmed not to meet the task’s rules. Compare the remaining count with the requested range.

100 × total returned eligible prospects ÷ total requested prospects

Depth (25%)

How much required information is filled in? Count the filled required fields across the unique returned records.

100 × filled required fields ÷ total required fields across unique returned records

Coverage penalty

Can we check enough of the requested list? Deduct up to 40 points if we can check fewer than 50% of the requested prospects. This is a points deduction, not a multiplier.

40 × max(0, 1 − coverage ÷ 50%)

Prioritize tasks ask the agent to select prospects from a supplied list and rank them using the task’s criteria. Before scoring, we assign eligibility labels and priority grades to the supplied candidates. We then compare the agent’s selections and order with this reference answer, which is not shown to the agent.

We score Accuracy, Breadth and Depth using the same weights as Discover and Enrich. Prioritize has no coverage penalty.

Final score = (Accuracy × 0.50) + (Breadth × 0.25) + (Depth × 0.25)

Accuracy (50%)

Are selected prospects eligible and correctly ordered? For ranked tasks, compare the points earned by the returned order with the best possible eligible order. For qualification-only tasks, each distinct qualifying selection earns equal credit.

100 × earned ranking points ÷ best possible ranking points

Breadth (25%)

Did the agent return enough distinct prospects? Count distinct returned prospects and compare with the requested count. Ineligible or unmatched selections affect Accuracy, not this count.

100 × distinct returned prospects ÷ requested prospects

Depth (25%)

How much required information is filled in? Count the filled required fields across the unique returned records.

100 × filled required fields ÷ total required fields across unique returned records

Sample Prioritize task

The following is a breakdown of our scoring methodology for the task below.

Select the top 3 companies from the supplied list based on intent (priority 1) and fit (priority 2). Exclude existing customers and rank the remaining companies by intent and fit.

Required fields: Company name, intent, fit.

Suppose the supplied list contains four companies: A has high intent and strong fit (grade 3); B has medium intent and strong fit (grade 2); C has low intent and moderate fit (grade 1); and D is an excluded existing customer (grade 0). The best possible order is A, B, C.

The agent returns B, A, D, with all 3 required fields filled for each company. This gives us 3 distinct returned companies and 9 filled fields out of 9 possible fields.

The agent loses points for placing B above A and selecting the excluded company D instead of C. Missing or repeated entries add no ranking credit; candidates without a reference label are excluded from scoring.

Final score = (78.96 × 0.50) + (100 × 0.25) + (100 × 0.25) = 89.48