It cites, or it declines.
Your site already answers most questions, and visitors cannot find the answers, so your team repeats them by email and phone. The standard fix is an AI assistant. But the standard fix invents answers, and an invented answer about your organisation is a liability you will be asked to explain. AnswerHatch answers only from the pages you publish, attaches the page to every answer, and declines when it has no source. One line of code, 14 days, no card.
<script async src="https://cdn.answerhatch.com/api/widget.js" data-agency="your-site" data-key="your-key" data-env="prod"></script>
Every answer carries the page it came from, so a reader can check it before acting on it.
I do not have a source for that. My sources are AnswerHatch's own published pages, and that question sits outside them.
no registered source The reason goes to the reader. Nothing plausible is composed to fill the gap.
Your site already answers most questions. Visitors cannot find the answers.
People ask things a website does not say directly: which form, which deadline, which address, which fee. They give up, email, or call, and your team repeats what the site already publishes. A general-purpose AI assistant makes that worse, because it invents an answer, and an invented answer about your organisation is a liability nobody can audit.
The problem
The answer is on your site, but not where the visitor is looking. Support time goes on repeating published material.
What AnswerHatch does
It reads the pages you publish, answers questions from them in plain language, and shows the page behind every answer.
Why that matters
Anything it says can be checked in one click. Nothing is invented. With no source it declines, and points to your help page instead.
This page runs the widget. Try to break it.
The Ask button on this page is the real product, answering from AnswerHatch's own published pages. Ask something we cover and it answers with links you can open. Ask something we do not cover and it declines, and says why. Courts in Canada and Germany have held operators responsible for what their chatbot told a customer. Declining is how we avoid that.
What happens when AnswerHatch has no source for a question?
What is the best pizza in Chicago?
The widget opens in the corner and answers, or refuses
Three steps, and the third one is allowed to say no.
Governance sits outside the model, in code that runs before and after it. A prompt can be talked out of a rule. A gate in the request path cannot.
Refused before the model
A deterministic scope screen and a prompt-injection screen run before the model is called. Off-topic and adversarial requests are turned away without reaching it.
Only what we crawled
The question is matched against the pages crawled from your site. Every document in the registry carries the URL it came from, the date it was fetched and its hash.
No source, no answer
An answer ships only with registered sources attached. With no source, the widget declines and gives the reason. This is a gate in the platform, not an instruction in a prompt.
The category sells AI that guesses. AnswerHatch answers with its source, or it declines. Call it SI if you want a name for the difference. The gate is the difference.
The test run.
Acceptance is evidence, not a feeling. Every trial includes a test run on your own pages: the questions it must answer, with the sources attached, and the questions it must refuse. The results are yours to read in the portal, page by page, and to hand to procurement.
Pass list
The questions it must answer, each with the page it used.
- What are the fees for a development application?
- Where do I report a missed bin?
- What are the council's opening hours?
Refusal list
The questions it must refuse, with the reason shown.
- Jailbreak attempts
- Off-topic steering
- Requests for figures your site does not publish
One line of code, and a key bound to your domain.
One script tag
The widget loads from our CDN and renders on your pages. Nothing to install on your server, and no rebuild of the site around it.
Keys bound to your origin
Each key is issued against the domain you nominate. A key copied out of your page source does not serve answers somewhere else.
Preprod and prod
Every subscription runs a preprod and a prod environment, each with its own corpus, keys and bindings. Test against preprod, promote when you accept it.
A registry, not footnotes
Crawled pages go into a provenance registry of sources, derivations and decisions. The citations in the widget are a view of that registry, and a compliance officer can ask for the registry itself.
Ask out loud. Hear the answer, sources included.
A visitor speaks the question and the answer is read back through their speakers. For a visitor who cannot read the screen, the whole exchange runs by ear.
Speech to text in the browser
The Ask by voice button transcribes the question in the visitor's browser. The same pipeline answers it as a typed question: screen, retrieve, cite or refuse.
Who does the transcribing
Chrome and Edge send the audio to their own vendor's speech service. Safari transcribes on the device. AnswerHatch receives text, never audio.
Answers read aloud
Every answer carries a Listen button. Setting data-voice="on" on the script tag makes the widget open listening, read the answer aloud, and listen again for the next question.
It is not a screen reader
It reads its own answers, not the rest of your page. Native screen readers work beside it through ARIA live regions, so nothing is read twice. Adding it does not make your site conformant, and we will never say it does.
Where the data goes, stated literally.
The category standard is "your data is not used to train models", which says nothing about where the data is processed, how long it is kept, or whose model API it passed through on the way. Here is the specific version.
Processed on our own machines
Visitor questions are processed on AnswerHatch-operated infrastructure, on machines we run.
No third-party model API
A question is never handed to an outside model provider in order to be answered. Models are not fine-tuned on your content either.
Thirty days, then deleted
Visitor questions are retained for 30 days so a fault can be diagnosed and a refusal reviewed, and are deleted after that.
One sub-processor
Cloudflare carries the network in front of the platform. That is the whole list. We hold no third-party security certification today and will not imply otherwise.
Start with a free trial. Pricing is quoted for your site.
Run the widget free for 14 days on up to 250 pages, no card. After that, pricing is on request: send us what you run and we come back with a yearly figure in writing. Answers are fair use: no per-answer meter, no credit balance to burn down, and no charge that arrives as a surprise.
The four objections, named and resolved.
It will say something wrong, and I will be the one explaining it.
Every answer carries the page it came from, so a reader checks it before acting on it. With no registered source the widget declines and gives the reason. Courts in Canada and Germany have held operators responsible for what their chatbot told a customer. Refusal is what keeps you out of that case. The demo on this page is the real product: ask it something in scope and it answers with links; ask it something out of scope and it declines.
Our data will end up in someone else's model.
Visitor questions are processed on machines we run. No third-party model API is in the path, and your content is never used to train a model. Questions are retained for 30 days so a fault can be diagnosed and a refusal reviewed, then deleted. Cloudflare carries the network in front of the platform and is the only sub-processor. That is the whole list, stated literally.
IT and procurement will block an AI tool.
Give them evidence instead of an argument. The governance sits outside the model, in code that runs before and after it, so a prompt cannot be talked around. Every crawled page carries its URL, its fetch date and its hash in a provenance registry, so a reviewer can audit exactly what the widget is allowed to say. A compliance officer asked us for that registry before sign-off. Run the free trial on your own domain and hand procurement the test results.
Our content is not plain enough for a machine.
We answer at your site's reading level, in your site's words. The crawl measures how each page reads, plain English for a council, statutory language for an agency, and the answers are generated to match. Nothing is rewritten into a voice your readers would not recognise as yours. You run the widget against a preprod copy of your site and accept it only when the tests pass on your own pages.
Start with a trial: 14 days from signup, 250 pages, no card. Give us an email address and the domain you will answer from, add your site, paste one line of code. When the trial ends, subscribe to keep answering. Start your trial. For an estate larger than Scale, tell us your estate.