Trust
Sources for every claim we make
Each promise on this site, where it comes from and what kind of evidence it is. A figure we haven’t measured stays off the site until we have.
- Claims listed
- 7
- Corrections
- 0
Claims
7 claims
| Claim, and where we make it | Evidence |
|---|---|
| “Demo scorecard: 14 of 14 answered from the manual with the right page, 3 of 3 out-of-manual questions declined, 4 of 4 safety warnings shown first.”Home · Live demoLive demo | 17 questions about the GF-3200 operator’s manual, Oct 17, 2026, listed on the demo page. Each answer scored against what we recorded before the test: the facts it must contain, the page it must cite, and the safety warnings it must show. Re-run after every model or retrieval change.Our measurementMethod |
| “73% of customers use self-service, but only 14% of issues are fully resolved there.”Demos and pilot conversations | Gartner surveyed 5,728 customer-service journeys in December 2023 and published the findings in August 2024. 43% of customers who couldn’t resolve their issue in self-service said they couldn’t find content relevant to it. A broad customer-service benchmark, not a measured Glassleaf or hardware-specific outcome.Third-party evidenceRead the source (Gartner) |
| “Your manuals and your customers’ questions never train models.”Home · Questions to askHome · Trust row | We don’t train models on your content. Our AI providers (Anthropic and OpenAI) process it under their commercial terms; under their current policies, API content isn’t used to train their models.Our commitmentPrivacy policy, section 6 |
| “Safety warnings, word for word, first.”Home · ProtectSafety | How DANGER, WARNING and CAUTION notices are taken from your manual word for word, approved before they appear (DANGER by two people) and shown before the answer.Our commitmentHow safety works |
| “Every answer shows its page.”Home · HeroLive demo | Check it yourself: ask the live demo a question, open the page it cites and compare it with the manual (PDF).Our commitmentTry the live demo |
| “Contact details removed before anyone sees them.”Pricing · On every plan | Questions are redacted when they’re stored, before Insights shows them: email addresses, phone numbers, serial numbers, names and addresses become tokens. Redaction is best effort and catches common formats.Our commitmentPrivacy policy, section 5 |
| “WCAG 2.2 AA widget.”Home · Trust row | Our accessibility statement: the standard we build the widget to, and its known limitations.Our commitmentAccessibility statement |
- “Demo scorecard: 14 of 14 answered from the manual with the right page, 3 of 3 out-of-manual questions declined, 4 of 4 safety warnings shown first.”Home · Live demoLive demo17 questions about the GF-3200 operator’s manual, Oct 17, 2026, listed on the demo page. Each answer scored against what we recorded before the test: the facts it must contain, the page it must cite, and the safety warnings it must show. Re-run after every model or retrieval change.Our measurementMethod
- “73% of customers use self-service, but only 14% of issues are fully resolved there.”Demos and pilot conversationsGartner surveyed 5,728 customer-service journeys in December 2023 and published the findings in August 2024. 43% of customers who couldn’t resolve their issue in self-service said they couldn’t find content relevant to it. A broad customer-service benchmark, not a measured Glassleaf or hardware-specific outcome.Third-party evidenceRead the source (Gartner)
- “Your manuals and your customers’ questions never train models.”Home · Questions to askHome · Trust rowWe don’t train models on your content. Our AI providers (Anthropic and OpenAI) process it under their commercial terms; under their current policies, API content isn’t used to train their models.Our commitmentPrivacy policy, section 6
- “Safety warnings, word for word, first.”Home · ProtectSafetyHow DANGER, WARNING and CAUTION notices are taken from your manual word for word, approved before they appear (DANGER by two people) and shown before the answer.Our commitmentHow safety works
- “Every answer shows its page.”Home · HeroLive demoCheck it yourself: ask the live demo a question, open the page it cites and compare it with the manual (PDF).Our commitmentTry the live demo
- “Contact details removed before anyone sees them.”Pricing · On every planQuestions are redacted when they’re stored, before Insights shows them: email addresses, phone numbers, serial numbers, names and addresses become tokens. Redaction is best effort and catches common formats.Our commitmentPrivacy policy, section 5
- “WCAG 2.2 AA widget.”Home · Trust rowOur accessibility statement: the standard we build the widget to, and its known limitations.Our commitmentAccessibility statement
Time from upload to a live widget
- What we measure
- For each new account, the time from uploading its first manual to the first visit to its widget from one of the account’s allowed domains. That visit is how we know the widget is installed and working.
- How we report it
- The median across new accounts in a stated period, with the date we measured it: half of new accounts were live sooner, half later.
- Status
- Not measured yet. Until it is, the site makes no claim about how fast you go live.
How we score the live demo
- The questions
- A fixed set of questions about the GF-3200 operator’s manual, written before the test. For each one we record what a correct answer must contain, which pages it may cite and which safety notices it must show. Some questions are deliberately outside the manual.
- How it runs
- Each question goes to the live demo through the same API the widget uses, in a new session. The answer is scored automatically against what we recorded.
- What counts
- Answers correct: the answer contains the required facts and cites an accepted page. Correct refusals: when the manual doesn’t cover a question, the widget says so instead of answering. Safety notices: every notice the question needs is shown with the answer.
- Status
- Not published yet. The scorecard appears on the home and demo pages once it is measured, with the number of questions and the date.
Numbers we don’t show
We don’t turn answered questions into tickets deflected, publish a return on investment without your own baseline, or present vendor marketing as independent research. An estimate of support savings uses your own ticket volume and cost per ticket, as the calculator on the pricing page does, and is labelled as a projection. A pilot publishes its sample size, question set and handoff rules with its result.