OpenAI’s documentation now describes a way for a ChatGPT Site to become a source of callable tools inside an assistant conversation. For publishers and SEO teams, that opens a different question from whether a page gets cited: can a useful information product become something the assistant can actually use?
According to OpenAI’s Site-hosted plugin guide, ChatGPT or Codex can add an MCP server to a new or existing Site, with tools that read information or make specified changes. The owner publishes the Site to create its associated plugin. Users then install, connect and invoke it in supported ChatGPT or Codex chats.
From a page to a usable capability
A cited page contributes evidence to an answer. A callable tool can supply a structured result or perform a defined operation. Consider a hypothetical publisher with an original database of industry announcements: an article explains what changed, while a tool could return matching records for a date range. Both are useful, but they serve different jobs.
This distinction makes the news relevant beyond website generation. A publication’s value may reside partly in its maintained data, comparisons, archives or specialist workflows. Packaging a narrow capability around those assets could give a user another reason to return to the publication’s information product through an assistant.
That is an editorial implication of the capability, not an OpenAI promise of greater discoverability. Creating a plugin does not establish that an assistant will recommend it broadly, retrieve a publisher more often or increase referral traffic.
Availability comes with important boundaries
The guide says hosting is available on all plans, with gradual rollout. Direct sharing is currently supported in Business and Enterprise workspaces; personal and Pro users cannot directly share Site-hosted plugins through invitations or share links. Workspace recipients need access to both the Site and plugin, plus their own connection.
Those limits matter to the business case. An internal research tool used by a team is a different proposition from a broadly distributed consumer service. Publishers should assess the audience they can actually reach before treating this route as a new acquisition channel.
Sharing a website and sharing its tool capability also represent different outcomes. A team evaluating distribution should record which users can open the information product, which can use the plugin, and which successfully complete a useful task.
Installation is not unrestricted access
OpenAI’s plugin documentation explains that plugins can combine skills, connected apps and other capabilities. Availability depends on the surface and account. Installation does not bypass provider authorization or workspace permissions.
For an information business, the practical design question is what a particular user should be able to do. A tool returning public archive records has a different purpose from one changing a shared dataset. The value proposition should make that distinction clear in ordinary language: what information the tool returns, what action it performs, and how the user can check the result.
What SEO teams can test
A useful first experiment would choose one repeatable task where structured access adds value. Examples might include looking up a documented product change, filtering an original research table, or checking a maintained reference record. These are proposed applications, not announced publisher features.
Measure task completion, result accuracy and repeat usage. Separately measure citations and website visits. Tool usage can deliver value without a browser visit, while an assistant’s citation can expose a brand without anyone installing a plugin. Combining these measures into one visibility score would conceal how the audience actually uses the product.
Maintain a clear connection between returned information and its provenance. A tool result should make it possible to understand which record or source supports an answer. That helps a user assess the information and gives the publication a concrete role in the workflow.
A documented rollout, not a verified launch timestamp
The official guide confirms the capability and describes a gradual rollout. It does not establish the precise launch timestamp supplied with this news tip. The September 30, 2026, 21:22 CEST timestamp is therefore treated as the reported observation time rather than a verified release time.
The broader signal is that participation in AI workflows can include both evidence and functionality. For publishers, becoming a useful tool is a potential extension of their information products. Whether it becomes a meaningful distribution or revenue channel will depend on access, adoption and demonstrable usefulness.