AskMyBusiness avoids hallucinating by answering only from a business's own published content, checked by a second model, with a deterministic safety gate underneath.
AskMyBusiness avoids hallucinating by answering strictly from the business's own published content and by checking every answer with a second model before it ships. If it can't verify something, it says it doesn't know and offers a person instead of guessing.
The assistant doesn't draw on general knowledge about what businesses in a given field usually charge or usually offer. It's restricted to a specific business's own published pages: services, pricing, hours, policies, and booking pages. When it answers, it shows the source page behind that answer, so the business owner and the visitor can both see exactly where it came from. That source line is part of every answer, not an optional detail.
Grounding alone doesn't catch everything. A second model reviews each answer against the published content before it reaches the visitor. If the first answer drifts from what's actually published, this check is what catches it before it goes out. When neither the grounding nor the check can confirm an answer, the assistant says it doesn't know rather than filling the gap with something plausible-sounding.
Before any model even sees a message, a deterministic crisis-and-hazard check runs in code. On top of that, safety guardrails are tuned per field. A med spa's assistant carries a no-medical-advice line and offers to connect the visitor with a clinician. A gym's assistant declines to advise on an injury and offers to pass the report to staff, with permission. This isn't a model deciding what's safe in the moment. It's a standing rule the assistant is built with, and every answer is separately checked against the business's published pages before it ships.
A visitor writes: "I hurt my knee during my workout, what should I do?"
The assistant doesn't attempt medical guidance. It answers: "I can't give medical advice, but I want to make sure you're taken care of. If it's urgent, please call 911. Otherwise, would you like me to pass your name and number to our staff so they can follow up?" That routing comes from the field's standing safety rules, standing rules the assistant is built with — not from the model improvising how to handle an injury report.
None of this runs unsupervised without oversight. The dashboard shows what visitors have asked and where the assistant needed help instead of answering confidently on its own. The owner sets the sources it draws from and the boundaries around what it answers, and can retract or correct content at any point, which matters if a published page changes and the assistant needs to catch up.
No. If it can't verify an answer against the business's published content, it says it doesn't know and offers to connect the visitor to a person.
It shows exactly which published page the answer came from, so anyone can check the assistant's work against the business's actual content.
Crisis and hazard messages are caught by a deterministic check in code before any model responds, and the conversation stops with instructions like calling 911. Medical or injury questions get a standing rule instead: the assistant declines and offers to pass the visitor's name and number to a person, only with their permission.
Give us your website and we build your private assistant — free, usually within a day, no card. Nothing goes live without your approval.
Build my free previewUpdated 2026-09-01 · AskMyBusiness by Vision Genesis · Knoxville, TN