From 60 to 92 out of 100, and the top pick on three AI assistants

The outcome first

We asked three AI assistants for a kids’ party planner in Richmond

I’m looking for a kids party planner in Richmond, who would you suggest?

No brand name. No keywords. The question someone types when they have a birthday coming up and no idea who to call. We asked ChatGPT, Microsoft Copilot and Google’s AI Mode the same thing on the same day.

All three named Twizzle first.

ChatGPT

Twizzle Parties & Events — my top pick for a fully planned, polished party.

Then it quoted their prices: “Their published Richmond pricing gives a useful sense of budget: roughly £300 for a small under-3s party.” That figure came off Twizzle’s own page. The assistant read it, repeated it, and credited twizzle.co.uk.

Microsoft Copilot

Top of “Best local picks”, citing twizzle.co.uk as a source, summarising them accurately: “High-end themed parties, entertainers, full planning service.”

Google AI Mode

Listed under “High-End and Full-Service Event Planners”, with a map pin labelled “Top kids party planner in Richmond”, describing what they actually do: immersive themed experiences, pirate treasure hunts, secret world adventures.

The score

60 out of 100, to 92

We scanned the site before the work and after it, with the same checks both times.

	Before	After
Overall score	60 / 100	92 / 100
Checks passed	27	42
Security and trust	12 / 100	100 / 100
Agent protocols	50 / 100	100 / 100
AI search signals	62 / 100	90 / 100
Agentic commerce	13 / 100	100 / 100

Seventeen checks were failing before the work. Three still are.

The one we would point at first is security and trust, 12 to 100. Not because security is exciting, but because AI agents will not act on a site they do not trust. Missing security headers are invisible to your customers and decisive to an agent, and almost every site we scan is failing them without knowing.

Agentic commerce went from 13 to 100. That is the measure of whether an AI agent acting for a customer could actually complete something on the site. Before, it could not. Now it can.

Their content score moved as well, 78 to 83, which took it from a B to an A. That half of the score is about the writing on the pages rather than the plumbing behind them, and it is measured separately.

Why the assistants picked them

Readable is not the same as worth quoting. Twizzle had both.

The technical work made the site legible and trustworthy to a machine. What made it quotable was already on the page, and it is worth being precise about which is which.

They publish their prices. Most party planners hide pricing behind an enquiry form. Twizzle put real numbers on the page, which is why ChatGPT could answer the budget question and why it recommended them for it. You cannot be quoted on a fact you have not published.

Their services are described specifically. Not “bespoke parties tailored to you”, but pirate treasure hunts and secret world adventures. Vague copy gives an assistant nothing to repeat. Specific copy gives it a sentence.

The business facts are machine-readable. Address, location, service type and contact details marked up so an assistant can confirm who and where they are rather than guessing. Their structured data scores 100 out of 100.

Now the site is trusted as well as readable. Security headers, a content policy, a verified organisation identity, and a published set of actions an agent can take. That is the half we added.

Secure agent actions

Then we opened the site to AI agents, and tried to break it

A score says a site is ready. This is the part that says it works. Twizzle now publishes eighteen actions an AI assistant can call on a customer’s behalf: search the parties and packages, read a package in detail, pull the site information and the machine-readable summary, read the enquiry form’s fields, and send an enquiry in.

The packages tools are the ones worth knowing about. Twizzle’s catalogue is structured data: age groups, themes, occasions, venue types, group sizes, durations and prices. When a parent asks an assistant for a party for a five-year-old in Richmond, the assistant can query the real catalogue instead of guessing from a web page.

Reading is open. Sending is not.

Anyone can read. Submitting an enquiry is protected, which means it cannot be run by an anonymous request. On 11 September we tested that from outside the site, with no credentials, and then deliberately tried to misuse it. Every test below was meant to fail, and did.

What we tried	What happened
Sent an enquiry with no credentials at all	Refused, 401 Unauthorized
Sent an enquiry with a forged token	Refused, 403 Forbidden
Called an action that does not exist	Refused cleanly, with no error detail, no internal paths and no stack trace
Read a full search response line by line	205 results, every one a published page

We also read a full response from the search tool by hand rather than pattern-matching it. Two hundred and five results came back, each with a title, a URL, a public excerpt and a date. No drafts, no internal fields, no supplier or cost notes, no customer data, no email addresses.

That is the difference between a website an AI agent can use and an open door. Most sites that are ready to be read have never been tested for what happens when something tries to act on them.

Scans run on the staging site before the work and on the live site afterwards, with identical checks. Agent actions tested from outside the site on 11 September 2026. Citation answers captured 4 September 2026. Published with Twizzle’s permission.

What this means for you

See what AI can read. The free scan checks whether AI can find, understand and act on your site. About a minute, no sign-up.

See what AI says. The £195 Citation Audit asks five assistants the questions your customers actually ask, and sends you the answers verbatim, including who got named instead of you.

See what an agent can do. Secure Agent Actions opens your enquiry form and your catalogue to AI assistants, with the actions that write to your site protected, tested and documented.