Independent methodology

Evidence before verdicts.

Product facts trace to first-party research records. Recommendations are deterministic. Unknown earns no credit. Benchmark results stay empty until real, reproducible testing exists.

Finder score

Frozen at 100 points.

Affiliate commission is absent from the data passed into the scoring engine and is never a tie-breaker.

Must-have capability match30

Pricing / budget fit20

Industry fit15

Integration compatibility15

Call-volume economics10

Language fit5

Multi-location / scale3

Setup simplicity / trial2

Hard gates

Verified false on a required feature makes a product ineligible for Best Overall. Unknown stays visible, earns zero, and creates a warning.

Tie-breakers

More verified must-haves, better integration match, lower verified cost, newer verification, then alphabetical name.

Cheapest rule

Only exact or estimated results from verified public pricing can be called cheapest. Base price alone is excluded.

AI Receptionist Benchmark 2026

20 frozen calls. Zero published results.

Same fictional business: Northstar Heating & Air. Same knowledge base, facts, booking availability, escalation number, language expectations, and scripts.

Benchmark data is intentionally empty.

0 product results exist. No winner, score, or performance claim will appear before real testing and review.

01Opening-hours factual question
02Service-area factual question
03Supported-service factual question
04Appointment booking — simple
05Appointment booking — unavailable requested time
06Reschedule existing appointment
07Cancel appointment
08Human transfer request
09Urgent no-cooling HVAC call
10Ambiguous problem description
11Pricing question where exact answer exists
12Pricing question where answer is absent
13Unknown / out-of-scope factual question
14Interruption mid-response
15Caller changes mind mid-call
16Full Spanish call
17English to Spanish language switch
18Controlled background noise
19Long caller story requiring summary
20Question + booking + SMS request

Benchmark score

100 points after real testing.

Each call stores pass, partial, fail, or not applicable; notes; audio/reference ID; latency where measurable; expected and actual behavior; and hallucination flag.