Your customers are asking an AI.Make sure it can answer for you.
People increasingly ask an assistant instead of opening ten tabs. Your website was built for the tabs. We make it work for the assistant too: answering from content you approve, and completing the booking rather than suggesting one.
No rebuild. No developer time. Works on WordPress, Shopify, Webflow, Next.js or hand-written HTML.
Search sent people to you. Assistants read you instead.
A search result was an invitation to visit your site and work it out. An assistant does the reading and reports back in a sentence or two. Your site is still the source, but it is no longer the destination.
That is fine when your pages are clear enough to summarize correctly. It goes wrong when the assistant cannot tell what you offer, guesses at your prices, misses that you cover a location, or has no way to book you. You do not get a chance to correct it, because the visitor never arrives.
Being agent-ready means the assistant gets it right, and can act on it.
“I’m not sure whether they treat that.”
“Yes, they do, at their Fitzrovia clinic.”
“Pricing isn’t listed publicly.”
“Consultations start at a published rate. Shall I check availability?”
“You’ll need to call them to book.”
“I can book you in. Thursday at 2pm work?”
Same business. Same website. Different answer.
A chat widget only serves half your audience
There are two very different readers of your website now, and they need different things. Most products sold as AI for your website only handle the first one.
A visitor on your page
A person who has already found you and is looking at your site. A script tag serves this reader completely: it can load an assistant, answer their question and take an action.
An assistant reading your site
ChatGPT, Claude, Perplexity, Gemini and whatever transacts on someone’s behalf next. These never execute your JavaScript, so a script tag is invisible to them. They read your HTML and a few well-known paths, and nothing else.
A script tag cannot make you readable to an assistant that never runs it.
Reaching the second audience means publishing files from your own domain, in the page itself. That is a config entry or a plugin, not another widget, and it is the part that a competitor dropping in a script cannot copy.
Three layers, and you need all three
Answering without acting loses the booking. Acting without being readable means only the people who already found you benefit.
It answers
An assistant on your site, grounded in content you approved.
It reads from a knowledge base built out of your own pages, not from a model’s general impression of your industry. Every answer traces back to a page you signed off. When it does not know, it says so and offers a person.
- Answers from approved content only
- Cites the page it came from
- Hands off to a human on anything it should not answer
It is readable
Machine-readable files published from your own domain.
AI assistants that summarize your business never run your JavaScript. They read your HTML and a handful of well-known paths. We publish those from your domain so what you offer, what it costs and what you can be asked to do is unambiguous.
- llms.txt and llms-full.txt
- Structured data for services, offers and FAQs
- An endpoint assistants can query directly
It acts
The enquiry is submitted. The appointment is booked.
Answering is where most products stop, and it is the half that does not pay. Actions complete the outcome you actually wanted from the visit, with the visitor confirming the exact details before anything is sent.
- Submits enquiries through your existing forms
- Books appointments against your calendar
- Nothing sent until the visitor confirms it
One line, one approval session, live
We are deliberately specific about the approval step. It is the part that stops the assistant inventing services on your site, and it is why we will not claim this is zero-touch.
Under an hour
Verify and install
We confirm you own the domain, then you add one line to your site. On WordPress or Shopify it is a plugin instead. Nothing else about your site changes.
Same day
We crawl and propose
We read your site the way a search engine does and turn it into a proposed knowledge base: services, locations, hours, prices, questions you already answer, and every form we find.
About an hour
You approve
Plain facts like address and opening hours are approved automatically and logged. Anything making a claim about outcomes, treatment or price comes to you first, sorted so the risky items are at the top. We sit with you for this.
Ongoing
Live, and it stays current
When you publish a change we pick it up, re-check only what moved, and queue anything that needs approving again. Your knowledge base does not quietly drift out of date.
The honest version: one line of code for the assistant, one config entry or plugin for the machine-readable layer, and about an hour approving content. Anyone promising zero-touch is either skipping the approval step or has not thought about what happens when it is skipped.
It uses the forms you already have
You do not have to hand over CRM logins, rebuild your enquiry flow or change where your leads land. The assistant fills in and submits the form already sitting on your page.
Your leads arrive where they already arrive
Same form, same routing, same notifications. Nothing about your existing pipeline changes.
Your spam protection and analytics keep working
Because the real form is submitted, everything attached to it still fires. A server-side post around it would break all of that quietly.
The visitor confirms before anything is sent
They see the exact details. If the summary changes after they agree, it refuses and starts again rather than sending something they did not see.
If you change the form, it stops
We fingerprint the form when you approve it. If the structure moves, the action fails closed and alerts us, instead of posting into fields that shifted.
What the visitor sees
Confirm before booking
- Appointment
- Consultation, 45 minutes
- When
- Thursday, 2:00pm
- Where
- Fitzrovia clinic
Illustration of the confirmation step. Nothing is sent until it is accepted.
The objection is always the same
What happens when it says something wrong about my business? Four answers, and none of them is asking the model nicely.
Grounded in your content
It cannot answer from anything except records approved for your site. There is no general-knowledge fallback to wander into.
A human approves first
Nothing extracted is answerable until someone signs it off. This step is not removable, and we will not remove it to make onboarding look faster.
Refusals route to a person
Questions it must not answer are caught before the model, not discouraged in a prompt. They go to your team with the context attached.
Confirmed before it acts
Anything that books, sends or notifies shows the visitor an exact summary first. If the details change after they confirm, it refuses and starts again.
Built for businesses that cannot afford a wrong answer
Clinics, pharmacies and healthcare providers carry rules a general chatbot has no concept of. We treat refusal as a feature to sell, not a limitation to apologize for.
These are enforced before the model is asked, so a cleverly worded question does not get around them. Each one routes to your team with the conversation attached, which is usually a better outcome than an answer anyway.
Also carried per client
Advertising rules for your sector, publication-approved claims only, and the AI disclosure the EU AI Act has required since August 2026. The disclosure is not a setting we will switch off on request.
It will not answer
- Whether symptoms are urgent
- What a condition might be
- Whether a treatment suits someone
- Anything about prescriptions
- Outcomes framed as a promise
- Prices presented as a final quote
Every one of these goes to a person on your team instead, with what the visitor asked and everything they said before it.
What it does to your site
Your developer or IT reviewer will ask these before it goes anywhere near production. Here are the answers up front.
Payload
Under 15KB gzipped, deferred, never render-blocking
Isolation
Renders in a shadow root. Cannot read or affect your CSS
Failure mode
Removes itself and shows nothing. Your page is never broken
What it sends
A content hash. Never page text, never form values
Crawl boundary
robots.txt and noindex respected. Logged-in pages never ingested
Layout impact
Zero cumulative layout shift, verified against your Core Web Vitals
Accessibility
Keyboard operable, screen-reader announced, WCAG 2.2 AA
Disclosure
Visitors are told it is AI before they interact. Not a setting
We will provide a content security policy snippet and a written statement of exactly what is read and transmitted, before you install anything.
Pick the layer you need now
Every package includes the crawl, the approval session and the guardrails. You can move up a layer later without redoing any of it.
Answer
Sites that mainly need to stop losing enquiries to unanswered questions.
- Crawl, extraction and approval session
- Grounded assistant on your site
- Guardrails and refusal routing
- Handoff to your team
- Monthly freshness re-crawl
Priced as a one-off setup plus a monthly fee banded by the size of your site. We do not sell unlimited flat plans, because one busy site would end up subsidized by everyone else.
What clients ask first
No. The assistant arrives as a single script tag. The machine-readable layer needs one config entry or a plugin, depending on your platform. Your site, your CMS and your hosting stay exactly as they are.
A chatbot answers the visitor looking at your page. It cannot help the AI assistants that read your site without running JavaScript, and most cannot complete an action. Agent readiness covers both audiences and finishes the job: the enquiry is submitted, the appointment is booked.
It can only answer from content a human has approved. We crawl your site, propose a knowledge base, and you approve it before anything is answerable. Every answer traces back to a page on your site.
Refusal classes are enforced outside the model, not asked for in a prompt. No diagnosis, no urgency assessment, no treatment suitability, no prescription guidance. Each one routes to a human instead of being answered.
Yes. It can drive the forms already on your site, so leads arrive wherever your leads already arrive and you grant no credentials at all. Direct calendar and CRM connections are available when you want them.
The loader is under 15KB, deferred, and never blocks rendering. It draws inside a shadow root so it cannot touch your CSS. If anything fails it removes itself and shows nothing rather than breaking your page.
The crawler fetches your pages anonymously, the way a search engine does, and respects robots.txt and noindex as hard boundaries. Anything behind a login is never ingested. The script on your page transmits a content hash, never page text and never form values.
The snippet takes minutes. Expect around an hour of reviewing extracted content for a typical site, and a working day for a large catalogue. We do that review with you rather than leaving you to it.