ChatGPT Plugins for Brands Are Not Organic Visibility

The ChatGPT release notes of 11 September 2026 carried a short entry headed "Planned custom GPT retirement and migration to plugins". Two weeks later, on 25 September, Anthropic published "Build plugins for Claude" and opened a directory where approved plugins are listed. On 29 September, in its DevDay recap, OpenAI added a line that matters more to anyone selling than the rest of the announcements: it is improving how it ranks and recommends plugins in the directory and in conversations. Within three weeks, ChatGPT plugins and their Claude equivalent had become the official way for a brand to sit inside the assistant.

Sitting inside the assistant is not the same as being named by it. This piece explains how a brand plugin works, and why its numbers and your organic visibility have to live in separate columns.

What a brand plugin is, in ChatGPT and in Claude

A brand plugin is a package a company publishes so that an assistant can use the company's instructions, documents and systems on request. It is software you own, running inside someone else's interface.

In ChatGPT, OpenAI's help centre describes plugins as a way to bring reusable instructions, reference files and connected apps together. The migration path from custom GPTs shows the parts clearly. The GPT's instructions become a skill inside the plugin. Its knowledge files become the plugin's reference files. Its connected apps are carried over as apps. Custom actions, the old way of calling an external API from a GPT, do not transfer, and replacing one may need a custom MCP server.

In Claude the definition is almost the same. Anthropic's post of 25 September describes plugins as bundles of MCP connectors, Agent Skills, or both. Developers on paid Claude plans can submit them, and the approved ones appear in the Claude directory.

So in both assistants a plugin is three things you control: what the assistant should do (the skill), what it should read (your files) and what it can call (your systems, through MCP). None of the three is the model's own view of your company.

How a plugin reaches a buyer

A plugin does nothing until a person has it. That is the core of the mechanism, and it is where plugins differ from everything else that decides whether an assistant mentions you.

There are three ways in. The person installs it from the directory. The person is given it by a colleague or by you; a GPT migrated to a plugin starts private and does not inherit the old GPT's sharing settings, so even former users have to be invited again. Or the assistant suggests it during a conversation, which is what OpenAI's line about recommending plugins "in conversations" refers to.

Even then, the user decides. OpenAI's recap says users stay in control of which plugins they use and approve access for each one. Every connected app inside a plugin still needs its own authorisation.

Once installed, a plugin can run without being called by name. OpenAI's migration FAQ says ChatGPT may use a relevant installed skill automatically when its description matches the request, and adds that this will not happen on every request. Keep that sentence in mind: it is the one that will contaminate your tests.

On the Claude side, Anthropic has said that a single discovery experience will roll out across Claude and Claude Code over the coming weeks. Neither company has published the criteria behind its ranking. Anyone who claims to know them is guessing.

Why ChatGPT plugins are not organic visibility

Organic visibility is what an assistant says about your market when nobody has installed anything. A buyer types "manufacturers of stainless steel valves for food processing in Europe", the model draws on what it learned in training and what it retrieves from the web, and your company is named, or it is not. You did not write those words. You can work on the sources the model reads, and you can measure the result.

A plugin answer is the opposite case. The words come from your skill and your files, and they appear in a conversation where the user, or the assistant on the user's behalf, has already picked you. It answers a different question: once someone is working with our tools, do the tools help?

Between the two sits a third thing that is easy to confuse with both: your plugin being recommended in the directory or in a chat. That is a placement in a catalogue the platform ranks, closer to an app store listing than to an answer. It is not the model describing your company, and it is not your content being used.

OpenAI's customer story of 7 October, about a hotel group, shows the separation from the other side. The group runs its brand plugin and sponsored ads in ChatGPT as one strategy. Those are two owned or paid channels side by side. Neither tells you whether ChatGPT names the group when a traveller asks an open question. We drew the same line for advertising in An Ad Beneath the Answer Is Not in the Answer. A plugin moves that line inside the conversation. It does not erase it.

Keeping plugin and organic answers apart when you measure

The practical risk is a contaminated test.

Your marketing team builds the plugin, installs it in their own accounts to try it, and then runs the usual check: half a dozen buyer questions typed by hand. ChatGPT may now pick up the installed skill, because its description matches the question. The brand appears in every answer, and the team concludes that visibility has improved. Nothing has changed for a buyer who has never heard of you.

Four rules keep the columns separate.

• Test organic answers from a clean account. No plugins installed, memory off, a new conversation for every question. If a work account cannot guarantee that, use a separate one.
• Measure the plugin from your side. Its usage arrives at systems you run: the MCP server behind it logs every call. Claude's directory also shows how often your listing is viewed and which searches lead to it. Those are plugin metrics, and they belong in the plugin report.
• Record a recommendation of the plugin as its own event. If you see the assistant suggest your plugin in a chat, write it down as a placement, not as a mention of the brand.
• Never add the two into one number. A score that blends answers written by your own skill with answers written by the model is counting your plugin twice.

PSentry sits entirely in the first column. It sends its prompts to ChatGPT, Claude, Gemini and Perplexity through their APIs, with no plugin installed, twice a month, in each language you track. It records what the models say about your market when nobody has steered them, and it does not see your plugin at all: not its usage, not its place in a directory. That is a limit, and it is the right limit for this question.

What changes for a B2B supplier before 11 December

OpenAI's FAQ gives 11 December 2026 as the retirement date for custom GPTs. Enterprise workspaces with an approved deferral have until 11 February 2027. On the retirement date, custom GPTs and their pages become inaccessible.

If your company built a GPT with datasheets and a product finder, the migration is mostly mechanical, with two catches. Only the latest published version migrates, so drafts are lost unless you publish them first. And any custom action that queried your stock or price list has to be rebuilt, probably as an MCP server.

The harder question is whether a B2B supplier needs a public plugin at all. A plugin serves people who already know your name: existing customers, distributors, your own sales staff. It is a sensible home for a product selector or a documentation lookup. It does very little for the buyer who is still drawing up a shortlist, the person we followed in The Supplier Search That Happens Behind the Login. That buyer meets you, or does not, in the organic answer.

Build the plugin if your customers will use it. Measure organic visibility because the customers you do not have yet never will.