We built an AI consultant, then built something to catch it lying
An agency chatbot will promise anything if you let it. Ours runs on Gemini with a guard between the model and the visitor, and the guard has already caught our own model overpromising.
By Mariya Tkach, Head of Marketing and Growth
Ask most agency chatbots whether they can double your sales and they will say something encouraging. They are built to be helpful, and "no" does not feel helpful.
That is the problem with putting a language model on a website that sells a service. The model has read a great deal of marketing copy, it wants to be agreeable, and it does not know which of your numbers are real.
So when we put a consultant on this site, the interesting engineering was not the answering. It was everything sitting between the model and you.
What is actually running
The consultant runs on Google's Gemini. We are naming it because you should know: a language model is answering, not a person, and the widget says so at the bottom of every conversation.
Gemini does not talk to you directly. Its answer goes to our server first, and the server can throw it away.
The guard that catches our own model
Every answer is checked before it renders. One of the rules looks for the word "guarantee" in the same sentence as sales or revenue, and if it finds it, the answer is discarded.
We wrote that rule expecting it to sit idle.
It does not. Ask the consultant to guarantee 300% growth and the rule fires every time. The model, left alone, produces something warmer than the truth, and the guard catches it before it reaches the screen.
What you see instead is written by us, not generated: UNITIX does not promise numbers, because no agency controls demand, competitors, or the marketplace. What we commit to is the part we run, and the results we do have are published with the dashboard screenshots they came from.
There is a footnote to this story worth telling, because it is the more common kind of failure.
For a while the guard worked and the visitor still lost. A caught violation was reported as "the AI consultant is unavailable right now", which was untrue, and it happened at the exact moment someone was deciding whether to trust us. The safety mechanism was right and the experience around it was wrong. That is usually where these systems break: not in the model, in what you do after you catch it.
What else it will not do
It will not name clients. Every account we run is under NDA. Ask for brand names and it says so and points at the anonymised case studies instead.
It will not quote a price. We do not publish a flat monthly fee, because it depends on ad spend, catalogue size and account structure. The model is not permitted to invent one to be helpful.
It will not hand over its own instructions. The obvious attack on any site chatbot is to ask it to print its system prompt, and people do try.
It cannot send you anywhere strange. The model can suggest links, but only from an allowlist: internal pages, or the booking calendar. A model that invents URLs is a phishing vector, and this one cannot, whatever it is asked.
The boring parts that matter more
Three things, none of them clever, all of them the reason it is safe to leave running.
Nothing sensitive should go in, and the widget says so. Passwords, API keys, payment details and customer data are named under the input box, and the server detects them if they arrive anyway.
There are limits. Per hour, per day, per visitor, and a token budget on top. An abandoned chatbot with an unmetered API key behind it is a bill waiting to happen.
It answers as text. Model output is never rendered as HTML. A chatbot that renders whatever it produces is a cross-site scripting hole with a friendly face.
Why we did it this way
UNITIX sells specialists supported by software, not software instead of specialists. An AI that promises results would contradict the thing we tell clients on every call: the technology surfaces what is happening, and a person decides what to do about it.
The consultant is genuinely useful for finding the right page, understanding what a service covers, or working out whether we are a fit. It is not a specialist, and it will tell you that.
If you want a number, the case studies have them, with the dashboards they came from. If you want them explained, book a call and a person will do it.