The article
One page that answers one question you are currently losing — written against your brand knowledge graph, not against the category in general.
Sophyx MCP connector · ChatGPT · Claude · Copilot · Gemini
From “we should write about that” to a live page — one MCP connector, one conversation
Switch on the Sophyx MCP connector in the assistant you already pay for. It knows which question you are losing, writes the page against your real brand data, attaches the schema everyone else skips — and ships it. WordPress, Webflow, a Git repo, or one HTML file you upload yourself.
No credit card · the connection is included in every plan · about four minutes to set up
Gap to published page, without leaving the chat
Walkthrough video landing shortly
Can ChatGPT publish content directly to my website?
Yes — when you connect a publishing destination alongside Sophyx. With a WordPress or Webflow connector the assistant writes the post, attaches the schema and pushes it as a draft. With a Git connector it opens a pull request against your repo. With neither, it hands you one complete self-contained HTML file that you upload yourself. Sophyx supplies the visibility data and the optimised assets — the article, the FAQ block, the JSON-LD, llms.txt and robots.txt — and the destination connector puts them live. One address connects it to every assistant: mcp.sophyx.app.
Nobody in this category has a generation problem any more
They have a folder of finished drafts nobody uploaded. The FAQ that never got its schema. The llms.txt still sitting in a chat window three weeks later. The last hundred metres is where the whole thing quietly dies, and it is invisible in every dashboard — the work looks done, and the AI answers keep naming somebody else.
A page that exists in a chat window is worth exactly nothing to an AI answer. MCP is the first thing that closes that distance without asking you to learn a new interface: your assistant reads your visibility data on one side and holds a connection to your site on the other.
What comes out of one conversation
Six assets, generated together, so the page ships complete instead of shipping and then waiting on the parts nobody enjoys doing.
The FAQ block, with FAQPage schema
Paired question and answer is the format assistants lift from most readily, because the boundaries of the answer are unambiguous.
Per-page JSON-LD
Your entities made machine-legible instead of implied — and merged with the schema already on the page rather than stacked on top of it.
llms.txt
The file at your site root that tells AI systems what your brand is, what it does and which pages matter.
robots.txt
Explicit rules for the retrieval crawlers. This one decides whether anything else on the list can be read at all.
The social cut
The same answer sized for the feed, so the page has somewhere to be distributed the day it goes live.
Three routes to live. Pick the one that matches what you built on.
All three end the same way: the page is on your site and an assistant can read it.
CMS
WordPress or Webflow
Add your CMS connector alongside Sophyx and the assistant writes the post, attaches the schema and pushes it as a draft. You approve in the CMS you already use.
One extra connector, then it is one sentence: “publish this as a draft.”
Code
A repo, or a vibe-coded site
If your site lives in Git, the assistant writes the page component and the schema into a branch and opens the pull request. Review the diff like any other change.
Built something in v0, Lovable or Cursor? This is the path that fits it.
Hand-built
One HTML file you upload
No CMS, no build step, no framework. You get a complete self-contained page — inline CSS, metadata, FAQ, JSON-LD — that you save and upload however you already do it.
Often the fastest route of the three, and the one nobody else supports properly.
Add the MCP connector in about four minutes
One step stops almost everybody, and it is not the one you would guess.
Server address
https://mcp.sophyx.app/
The same address works in ChatGPT, Claude, Microsoft Copilot and Gemini. The steps below are ChatGPT’s, because it is the only one that hides the feature behind a toggle.
- 01
Open Settings → Connectors
Your name, bottom-left in the desktop app, or your avatar top-right on the web. Depending on your build the panel reads Connectors, Apps & Connectors, or Plugins — it is the same panel.
- 02
Scroll to the bottom and switch on Developer mode
It is the very last item. Do this before anything else.
Until this is on there is no add button anywhere, and searching Settings for “MCP” returns nothing. It is not that you are missing it — it genuinely is not rendered yet.
- 03
Add the connector
Browse connectors → scroll to the top → the + button to the right of the search bar. Name it Sophyx, paste the server address, choose OAuth, tick the trust confirmation, Create.
- 04
Approve with your Sophyx login
Sign in, scroll to the bottom of the permission screen, Allow access. That window often opens behind the app — check your taskbar if nothing seems to happen. Then tick Sophyx in the composer’s + menu and you are live.
Developer mode is a paid-plan feature. On ChatGPT Free there is no Connectors panel at all. Claude, Copilot and Gemini need no equivalent toggle — paste the address and approve.
The three prompts that do the work
Take them. They are the whole system, and they work the moment a workspace is connected.
Prompt 1 — write the page
paste into ChatGPT or Claude
Using the Sophyx connector, take my highest-value visibility gap and write the
page that closes it.
Deliver one standalone HTML file:
- <title> and meta description filled in
- one <h1>, then <h2> sections, no skipped levels
- the gap question answered directly in the first paragraph, in full sentences
- a 6-question FAQ section in plain Q/A text
- FAQPage + WebPage JSON-LD in a <script type="application/ld+json"> block
- inline CSS only, no build step
- a call to action at the top and the bottom
Before you write: pull my brand knowledge graph and my existing content so the page sounds
like my business and does not repeat a page I already have.Comes back as one file you can save and upload, or push straight to your CMS if that connector is on.
Prompt 2 — the three files at your site root
paste into ChatGPT or Claude
Using the Sophyx connector, get my generated llms.txt, robots.txt and the
per-page JSON-LD for my site.
Show me llms.txt and robots.txt as complete files I can save directly.
Then tell me, for each JSON-LD document, exactly which page it belongs on, and give me the
<script type="application/ld+json"> block to paste into that page's head.
Before you give me those blocks: check whether the page already has JSON-LD on it, and if it
does, merge rather than adding a second one.The merge instruction on the last line matters. Two JSON-LD blocks fighting on one page is worse than none.
Prompt 3 — upgrade a page you already have
paste into ChatGPT or Claude
Here is the HTML of one of my existing pages: [paste the file]
Using the Sophyx connector, check my visibility gaps and tell me which of them this page is
closest to already answering.
Then rewrite the page so it actually answers that question — keep my layout and my voice, add
an FAQ section targeting the gap, and add or merge the FAQPage JSON-LD.
Show me a short summary of what you changed, then the complete updated file.Usually cheaper and faster to move than a new page, because the URL already has history.
The prompts are free and always will be. They only return your pages once there is a tracked workspace behind them.
Create your workspace freeTwo things worth knowing before you publish
Q: Will it sound like AI?
A: It will sound like whatever you accept. The connection hands the assistant your brand knowledge graph and your existing content, which fixes accuracy — it does not fix voice. Read the draft. Cut the padding. The customers who get the best results run one more pass in their own words before publishing, and they are right to.
Q: Can two JSON-LD blocks conflict?
A: Yes, and it is the most common self-inflicted injury in this whole workflow. If a page already carries global schema, adding a second block gives crawlers two contradictory descriptions of the same entity. Always ask the assistant to check for existing JSON-LD and merge rather than append — it is one line in the prompt above, and it is there on purpose.
How Sophyx builds the schemaWhat you get, end to end
Every one of these is included. None of them is a separate line item.
The question, chosen with evidence
Your visibility gaps ranked against real search demand, so the page you write is the one that is actually costing you.
The draft, written against your brand
Your knowledge graph, your existing content and your tracked prompts go in first — so it does not repeat a page you already have or describe a business that is not yours.
The whole AI-readable layer
FAQPage schema, per-page JSON-LD, llms.txt and robots.txt. The parts everyone skips, generated in the same pass.
A route to live for whatever you built on
CMS connector, Git branch, or a single file you upload. All three end with the page live, which is the only state that counts.
The re-measure
Four to six weeks later, the same prompt runs again and you find out whether the AI answer changed. Publishing is the work; the changed answer is the result.
From USD 39 per month, and the first audit is free. One published page that an assistant starts quoting is worth more than the year of drafts that never left the folder. See pricing.
Publishing from ChatGPT — your questions answered
- Yes, when you connect a publishing destination alongside Sophyx. With a WordPress or Webflow connector the assistant writes the post, attaches the schema and pushes it as a draft. With a Git connector it opens a pull request against your repo. With neither, it hands you one complete self-contained HTML file that you upload yourself. Sophyx supplies the visibility data and the optimised assets; the destination connector puts them live.
- llms.txt is a plain-text file at the root of your site that tells AI systems what your brand is, what it does and which pages matter. It does not guarantee a citation — nothing does — but it removes the ambiguity that stops one, which is why it is worth the ten minutes. Sophyx generates yours from your brand knowledge graph, along with a robots.txt that grants the retrieval crawlers explicit access.
- Answer one question per block, in short complete sentences, and mark the block up as FAQPage so the question and its answer are explicitly paired. Assistants lift paired Q&A more readily than prose because the boundaries of the answer are unambiguous. Keep the visible text and the JSON-LD identical, and never put a citable claim only inside an image.
- Answer the question in the first paragraph rather than building to it, keep answers to two to four sentences, define your terms in full sentences a model can quote, name your brand inside the answers so attribution is possible, and mark the page up with FAQPage and WebPage schema. Then measure it properly: re-run the prompt four to six weeks later and check whether the AI answer changed.
- No. A single self-contained HTML file works fine, and for hand-built sites it is usually faster than a CMS. Sophyx supports three routes — a CMS connector, a Git pull request, or a file you save and upload — and the assistant will produce whichever one you ask for.
- The Sophyx MCP connector is a Model Context Protocol server at mcp.sophyx.app that lets an AI assistant read your Sophyx workspace inside a conversation — your prompt visibility and gaps, your brand knowledge graph, your existing content, and the generated llms.txt, robots.txt and per-page JSON-LD. It is read-only apart from moving a piece of content between draft, published and archived, so an assistant can compose and ship a page but cannot silently change your settings. One address works in ChatGPT, Claude, Microsoft Copilot and Gemini.
- Any MCP-capable client. ChatGPT, Claude, Microsoft Copilot and Gemini all connect to the same address, mcp.sophyx.app, and the connection is included in your Sophyx plan at no extra cost. Sophyx is an AI visibility platform that tracks whether ChatGPT, Gemini, Claude, Perplexity and Copilot mention, recommend or exclude your brand inside generated answers. The first audit is free; plans start at USD 39 per month.
Publish the first one this week
Create a workspace, connect your assistant, and the first page can be live before the end of the day. The prompts above are the whole method — the workspace is the part that makes them return your data instead of generic advice.
Doing this for clients? Report it back to them under your own name.