UserApproved surfaced 5 findings, 3 rated critical by independent CRO experts. Claude and GPT surfaced 12 combined, with just 1 rated critical.
60% of UserApproved findings were rated critical, versus 14% for Claude and 0% for GPT.
UserApproved
60%
3 of 5 findings rated critical
GPT Sol 5.6
0%
0 of 5 findings rated critical
Claude Opus 5
14%
1 of 7 findings rated critical
UserApproved scored higher than both leading general agents on all four dimensions, with the widest gaps on tracking real evidence across sources and on defining actionable experiments.
UserApproved Agents
Purpose-built system
85/100
GPT Sol 5.6
High
68/100
Claude Opus 5
X-high
65.2/100
Rated by three external CRO reviewers, blind to which system produced each result.
When a claim isn't tied to a screenshot, a session recording, or a number you can pull yourself, somebody on your team has to go verify it before anyone will act on it. And it only takes one invented claim to make a merchant distrust the entire report. Every finding we publish carries the artifact it came from, and the arithmetic is printed so you can recompute it.
Evidence grounded
A report can tell you mobile bounce is high, and separately that shipping cost appears late in checkout, and leave you to connect them. That connection is the actual analysis. Most systems hand it back to you and call it a finding.
Strategic synthesis
Every finding costs you attention: someone reads it, someone argues about it, someone decides whether it goes in the backlog. Most AI findings are not worth that, and you only find out after you've spent the meeting.
Impact and usefulness
"Improve your mobile experience" is the problem restated as advice. A finding should name what to change, where, and what to watch afterward, so the next step is doing the work rather than another week deciding what the work is.
Action specificity
The store
A live US ecommerce storefront, not a demo site. Audited on desktop and mobile. Named to prospects under NDA.
Same brief
Act as an ecommerce growth expert and identify actionable revenue opportunities.
Same available access
The same store data, APIs, MCP integrations, and browser tools were available.
External review
Three external CRO professionals rated the results from all three systems.
Blinded scoring
Reviewers did not know which system produced each result.
Every finding is rated 1 to 4 on each of the four dimensions, for a maximum of 16 points, then normalized to 100. Every score on this page is the mean across the three reviewers, from the August 4, 2026 run.
| System | Coverage | Elapsed | Quality / $ |
|---|---|---|---|
| UserApproved Agents | Desktop + mobile · 4 persona paths | 46:52 | 3.3 |
| GPT Sol 5.6, High | Desktop · 1 primary path | 40:00 | 3.7 |
| Claude Opus 5, X-high | Desktop · 1 primary path | 39:00 | 2.0 |
GPT had the highest reported quality-per-dollar index. UserApproved had the highest quality score and the longest elapsed time while choosing broader coverage. Provider accounting can differ, so this is a run-specific index.
In this pilot, external blinded reviewers rated UserApproved higher on all four quality dimensions and rated a higher concentration of its findings critical. This is directional evidence for the system design, not proof of universal superiority.
The model is not the product. The product is the system that decides what deserves to ship.
Start with a free baseline scan, then connect your commerce context to see which findings deserve action.
Free baseline audit · No credit card · Evidence stays attached to every finding
