Research summary
This evaluation compared Decisions Lab with five frontier LLMs across buyer personas. Insights generated by these models were then assessed by a human reviewer against actual customer call notes. Decisions Lab achieved the highest supported-insight rate for each persona.
Insight alignment
84.2%
The average share of buyer insights supported by customer call notes across three personas.
Consistency across personas
7.6x
The average variance of five frontier models was 7.6 times higher than Decisions Lab.
Introduction
The evaluation examined whether simulated buyer insights aligned with insights recorded in real customer calls. The simulations used public information to model the buyers. Unsupported assumptions can influence a team’s choice of segment, message, proof point, or objection strategy.
How we ran the study
Buyer personas
Three buyer personas were defined from specific prospect criteria, including company size, IT team structure, role, and operational setup. For each persona, Decisions Lab identified roughly 100 matching real-world prospects by cross-checking public sources such as LinkedIn profiles, company pages, job postings, company structure, and public activity.
Identify buyers
~100
matching prospects per persona
Role · Company size · IT team · Operational setup
Build the simulation
Enrich profiles with public information
Generate insights
Simulated interviews
Pain triggers, beliefs, objections and decision criteria
Generating insights
Those prospects powered simulated customer interviews using Decisions Lab’s persona-building framework and harness. The responses were aggregated into insights covering pain triggers, buying triggers, beliefs, objections, aha moments, and decision criteria. Each of the five general-purpose AI models received a description of the persona and web access, and was asked to generate comparable buyer insights. This compares the Decisions Lab workflow with the models under those conditions.
Reviewing the results
A human reviewer compared each generated insight with the customer call notes. An insight was supported when customers had mentioned it in the notes; it was unsupported when the notes did not support it, including when it was missing from or contradicted by the evidence.
Supported-insight rate
Share of generated insights supported by the call notes.
Variation across personas
How much the supported-insight rates differ across personas.
Variance (rate₁, rate₂, rate₃)
What we found
Each percentage is the share of generated buyer insights supported by real customer call evidence.
Mean supported-insight rate
Decisions Lab’s 84.2% average created a +21.7 percentage-point advantage versus the average proprietary model at 62.5%.
Results across buyer personas
The overall ranking held when results were broken out by persona. Decisions Lab led on every persona run.
Supported-insight rates across buyer personas
| Model | Persona 1 | Persona 2 | Persona 3 | Average |
|---|---|---|---|---|
| Decisions LabProprietary harness | 80.0% (highest) | 82.7% (highest) | 90.0% (highest) | 84.2% (highest) |
| GPT-5.5 ThinkingWeb search enabled | 61.0% | 69.3% | 77.0% | 69.1% |
| Gemini 3.5 ThinkingWeb search enabled | 43.0% | 74.1% | 70.0% | 62.4% |
| Claude Fable 5 HighWeb search enabled | 42.0% | 62.7% | 79.0% | 61.2% |
| DeepSeek Expert ThinkingWeb search enabled | 56.0% | 50.0% | 76.0% | 60.7% |
| Grok 4 FastWeb search enabled | 46.0% | 66.4% | 65.0% | 59.1% |
An average can conceal a weak result for an individual persona. Decisions Lab scored between 80% and 90% across the three personas; GPT-5.5 Thinking scored between 61% and 77%. Using unrounded scores, the average variance of the five other models was 7.6 times Decisions Lab’s variance. This describes differences across personas in this test, not repeatability across runs.
Insight alignment: standard deviation and lowest score
Top-left is better: a higher worst-case score and a lower standard deviation.
Each dot shows a model’s lowest supported-insight rate and its standard deviation across buyer personas. The 7.6x comparison uses variance, the square of standard deviation.
Where the models differed
The models sometimes identified the broad buyer category but added plausible assumptions that were not supported by the call notes. These included adjacent pain points, objections, or decision criteria that customers had not mentioned. Decisions Lab’s insights had a higher rate of support in the notes across all three personas.
What the results mean
In this comparison, Decisions Lab recovered a higher share of insights supported by customer call notes and showed less variation across the three personas than the other systems. The results demonstrate how simulated interviews can surface buyer insights for GTM decisions.