Chatbot Accuracy Evaluation
A written, per-question report
on how accurately your chatbot answers.
We test your chatbot against an agreed test set and report, per question: whether the answer is supported by your approved sources, whether it declined when it should have, and where exactly it failed.
What we do
One question at a time, against your own sources.
What the report tells you
The test questions are written by us from your FAQ and help pages, and you approve them before the run. We then test your chatbot question by question and deliver a written report.
- Whether each answer is supported by your approved sources
- Whether the bot declined when it should have
- Where exactly it failed
- Scope
- This is an evaluation of the agreed test set — not a certification of every possible answer.
- Language
- English (Japanese available on request)
- Scope limit
- One chatbot, one knowledge base per order
Sample reportSee exactly what you receive: a complete sample report from an actual evaluation run on our own in-house demo bot with synthetic data. In that run (20 questions, 13 November 2026), the bot answered 13 of the 16 questions its sources could answer correctly and fabricated 0 answers — and it answered 4 questions it should have declined, which is exactly the kind of failure this report exists to catch. Read the sample report
Pricing
Three fixed prices, in USD.
All delivery times count from your approval of the test set.
Starter
$95fixed
20 test questions / delivery in 3 business days from your approval of the test set
- 20 test questions, written by us from your FAQ/help pages and approved by you before the run
- Per-question pass/fail
- Written report (PDF)
Standard
$245fixed
50 test questions / delivery in 5 business days from your approval of the test set
- 50 test questions
- Failure-pattern analysis
- Concrete fix list
White-label available from this tier (client-facing report carries your branding)
Premium
$495fixed
100 test questions / delivery in 7 business days from your approval of the test set
- 100 test questions
- Fix list
- One re-test of the same test set: you may request it within 30 days after your fixes, and the re-test report is delivered within 5 business days from the day we receive access to the updated bot
White-label included
Process
Four steps, fully asynchronous.
Share your materials
You share your FAQ and help pages, and access to the chatbot, by secure link.
Test set: written and approved
We write the test questions from your FAQ/help pages. You approve the test set before we run anything.
Run
We test your chatbot against the approved test set, question by question.
Report delivered
You receive the written report within the delivery time of your tier, counted from your approval of the test set.
Data handling and terms
Agreed in writing before payment.
| Data handling | Your materials — source documents, chat logs, files, and any credentials — are transferred by secure link, used only for your job, never used to train models, and deleted 30 days after delivery. |
|---|---|
| Personal data | If the work involves processing personal data on your behalf, a data processing agreement is signed before work starts. |
| Price and acceptance | Fixed price and acceptance criteria agreed in writing before payment. Changes after work starts, delays on your side, and cancellation terms are confirmed in the same written agreement. |
| Payment | First orders under $500 are 100% prepaid (Stripe invoice or Wise); larger first orders are 50% to start, 50% on delivery. |
| Contracting party | GlobeZenith OÜ (Estonia) |
| Working style | Work is fully asynchronous. |
| Privacy | If we contacted you about this service, please see our Privacy notice for business outreach. |
Who you are contracting with
A registered Estonian company that runs its own AI products.
GlobeZenith OÜ
This evaluation is run by GlobeZenith OÜ, the Estonian company that builds and operates its own AI products — Cuevane (real-time sales-meeting support), Mensetsu Kanpe AI (interview support, with paying subscribers in Japan) and Pentra (AI content operations, in production on our own media). It applies the principle Cuevane is built on: an answer is only as good as the source behind it. Every answer in your report is graded against your approved sources, and facts that appear in none of them are counted separately as fabrications.
- Annual reports for 2024 and 2025 filed with the Estonian Commercial Register
- Fixed prices agreed in writing before payment — no hourly billing
- Company overview and CEO profile
- Legal form
- Estonian private limited company (osaühing)
- Registry code
- 16979948 (View in the Estonian Commercial Register)
- Founded
- 3 May 2024
- Registered office
- Sakala tn 7-2, 10141 Tallinn, Estonia
Tell us about your chatbot.
Send us the bot's URL, its FAQ or help pages, and the tier you are interested in.
Email: info@globezenith.net