Alem data methodology and benchmark limits
Updated 1 October 2026 · Alem product methodology
The Data page reports aggregate product telemetry from the public JSON endpoint. It is not a market survey, independently audited revenue report or leaderboard of paying customers.
How to read the counters
| Metric | Definition | Limitation |
|---|---|---|
| Agents registered | Agent records in the store | Includes internal and sandbox agents; not unique customers |
| Economic actions evaluated | Recorded spend-request events | Repeated attempts can count separately |
| Human approval required | Approval-created events | Not the number of completed purchases |
| Policy denied | Requests whose stored decision is deny | Not independently verified money saved |
| Sandbox spend settled | USD amounts in sandbox settlement records | No claim that funds actually moved |
| ROI cost and attributed value | Recorded cost and benefit events | Can be entered by an operator |
Aggregation and formulas
The endpoint sums the records currently present in the store. It is an all-time snapshot, not a monthly cohort. The attributed return multiple is total attributed benefit divided by total recorded ROI cost. It is not the arithmetic average of each agent’s ROI. When recorded cost is zero the return multiple is unavailable.
Sandbox settlement totals and ROI cost totals come from different event types. They may differ. Do not rename either number “payment volume” or “revenue”. An internal example with $500 recorded cost and $1,000 attributed value produces 2× return and 100% ROI; it does not demonstrate a customer earning $1,000.
What a public result proves
A share page shows an owner-created snapshot with self-reported costs and outcomes. Its random link is public to anyone who receives it, even though we mark it noindex. Owners can revoke the original link. Revocation cannot erase screenshots or third-party copies that have already been made.
Requirements for a meaningful benchmark
- Define the task and evaluation window before observing results.
- Use a comparable baseline and include failures, retries and human review.
- State sample size, selection criteria and missing data.
- Separate independently verified amounts from operator estimates.
- Report both absolute value and cost, not only a percentage.
FAQ
Are visitor counts identified people?
No. Daily pseudonymous visitor counts and user-agent classifications are directional. A crawler claiming to be an AI agent is not an authenticated agent.
When were these figures measured?
Use the endpoint’s updated_at value and record your retrieval time. The figures change as new events are recorded.
Can this establish typical customer ROI?
No. The current aggregate includes internal tests and sandbox activity, without a representative customer cohort or independent audit.