We built an MCP server for goji.agency. It does not improve our AI visibility, and we want to be specific about why, because the gap between those two things is where most of the current advice about AI search goes wrong.
What we actually built
An MCP server is an endpoint an AI assistant can connect to and query directly. Ours sits at mcp.goji.agency. It exposes nine tools over what we have already published: a 104-term plain-English glossary, the fourteen chapters of our AI visibility guide, seven long-form guides including our real Australian pricing for websites and branding, and the sixty-one questions we have already answered publicly about cost, timelines and how we work. One more tool sends a genuine enquiry to our inbox.
Connect it to Claude and you can ask what AEO is, what a website costs in Australia, or which of our service lines fits a problem — and get our answer, with the source URL attached, rather than a paraphrase assembled from whatever the model absorbed during training.
It is read-only against published content. There is no key, no login, and nothing in it that is not already on the website.
What it does not do
It does not make ChatGPT recommend us.
This is the part worth being clear about, because it is easy to sell the opposite. MCP servers are opt-in. A person has to deliberately connect one. Nothing about publishing an MCP server changes what a vanilla assistant says when someone asks it to name a good AEO agency in Melbourne — that answer comes from training data, from live web search, and from what other sites say about you. An MCP server is none of those three.
We looked for evidence to the contrary before building it and did not find any. The MCP servers that surface when you search for AI visibility tools are monitoring products — they measure whether you are cited. They do not make you cited.
So if an agency tells you an MCP server will get you into AI answers, ask them by what mechanism. There is a real answer to that question and it is not the server.
What it is actually for
Three things, and they are all worth having, but none of them is a ranking lever.
It is a distribution channel into a new category. MCP directories are indexed pages that link back to us, in a space with a fraction of the competition of any SEO directory. Off-site mentions do feed AI citation. The listing helps; the endpoint alone does not.
It is a demonstration. We sell AI visibility. Being queryable by an assistant, rather than only readable by one, is the difference between describing the work and having done it. We would rather show a prospect Claude answering questions about us from our own server than explain the concept.
And it is a real route to a conversation. Someone can ask an assistant about answer engines, follow it into what we charge, and send us an enquiry without opening a browser.
The thing that does move AI visibility
Being the source that answers the question. Concretely: publishing specific claims, attributing them, and making them easy to lift. Our glossary entries lead with a two-sentence definition written to be quoted. Our pricing page carries real ranges instead of a contact form. Our service pages answer what things cost and how long they take, in the words people actually search.
None of that is exotic. It is the unglamorous half of the work, and it is the half that determines whether an answer engine has anything of yours to repeat.
One technical finding, since we owe you a specific claim
Building the enquiry tool meant putting a rate limit in front of a public endpoint that can send email. We used Cloudflare's native rate-limiting binding, declared three limiters, and deployed. The bindings attached and appeared in the deploy output.
Then we tested them: twelve calls in a few seconds against a cap of two per minute. None were refused. The binding was being called, the keys were stable, and it returned success every time. The configuration matches the documented API; the platform was not counting.
We mention it because the deploy output looked exactly like proof that it worked, and it was not. If you have declared those bindings and never fired more requests than the limit, you do not know whether your rate limiting works. The protection we actually rely on is structural: the recipient address is pinned in the platform configuration rather than carried in the request, so no bug in our code can widen it.
If you want to try it
Point any MCP client at https://mcp.goji.agency/mcp. In Claude Code: claude mcp add --transport http goji https://mcp.goji.agency/mcp. Then ask it what AEO costs.
The answer it gives you is ours, and it will show you where it came from. That is the whole idea — not for the assistant, for the person asking it.
