Obility Editorial · · 4 min read
A plugin is found by people who already know your name
Since OpenAI opened ChatGPT to third-party software built on its Apps SDK, now documented under the name plugins, some AEO advice has started presenting a plugin as a way into ChatGPT’s recommendations. The reasoning sounds plausible: if your product lives inside ChatGPT, ChatGPT will surely mention it. OpenAI’s developer documentation describes how plugins are actually found, and it does not support that reasoning.
According to OpenAI’s submission guide, once a plugin is published, people find it in the plugin directory shared by ChatGPT and Codex in two ways: by following a direct link to its listing, or by searching for it by name. Both routes start with someone who already knows you exist. The same guide says a plugin appears on the directory’s main pages only if OpenAI selects it for what it calls enhanced distribution, that few plugins receive it at publication, and that developers cannot request it.
The shortlist answer and the plugin call are separate steps
When a buyer asks ChatGPT for the best payroll software for a small agency, the answer that names three or four companies is written from what the model knows and from the web pages its search step retrieves. OpenAI’s documentation describes no route by which a plugin the buyer has not installed shapes that answer, apart from the proactive suggestions discussed below. A plugin’s tools are called later, inside a conversation, when ChatGPT judges that one fits the request. OpenAI’s metadata guide says ChatGPT decides when to call a tool from the names, descriptions and parameter documentation the developer provides, and tells developers to test with three kinds of prompt: direct prompts that name the product, indirect prompts that describe the outcome without naming it, and negative prompts where another tool should handle the request.
Indirect prompts are where the hope comes from. OpenAI’s planning guide frames them from the point of view of a person who has installed your plugin and has not read its documentation. They are about the plugin being used well by someone who chose it, not about the plugin being suggested to someone comparing vendors. OpenAI does mention proactive suggestions, but as one of the enhanced distribution opportunities it grants to plugins with strong real-world utility and high user satisfaction. That is a reward for being used, not a lever a marketing team can pull.
The rules forbid metadata that steers the model toward you
If the plan is to write tool descriptions that nudge ChatGPT to bring you up more often, OpenAI’s plugin guidelines close that door explicitly. Under a heading called fair play, they say plugins must not include descriptions, titles, annotations or other model-readable fields that manipulate how the model selects plugins, such as instructing it to prefer one over others. Tool descriptions must not favour or disparage other plugins or services, and must not recommend triggering more broadly than the user’s explicit intent. Tool names must avoid promotional or comparative words; the guidelines give pick_me, best and official as examples.
The listing copy has its own limits. Subtitles and descriptions may not compare the plugin with other products or make unverifiable claims, and may not advertise pricing, subscriptions, free trials, discounts or promotions. Plugins must not exist primarily as an advertising vehicle, are expected to provide functionality ChatGPT does not already offer, and for now may sell only physical goods, not subscriptions or other digital products. A plugin built as a brand presence in ChatGPT would breach several of these rules at once, and OpenAI says it may remove approved plugins later found in violation.
Build one for customers, and judge it by use
A plugin makes sense when your existing customers already want to do their work with your product from inside ChatGPT: checking an order, pulling a report, booking an appointment, updating a project board. Then it reduces friction for people who have already picked you, and that is worth something. The test is whether you can list real tasks your customers would ask it to do. OpenAI’s planning guide suggests collecting those requests from support tickets, search queries, feature requests and the workarounds people use to copy data between tools.
Judge the plugin by the measures OpenAI itself suggests: how often the right tool runs for the prompts it should handle, how rarely it fires for prompts it should not, and whether the customers who use it are satisfied, which OpenAI names as a condition for wider distribution. Do not judge it by whether your brand shows up more often when strangers ask ChatGPT for a shortlist, because nothing in the documentation connects the two.
Your shortlist presence is still decided on the open web
The work that changes whether buyers who have never heard of you see your name is the same as before the plugin directory existed: clear pages of your own that answer buying questions in the buyer’s words, and independent pages the engines retrieve that describe you accurately. The two kinds of prompt also belong to different stages of the purchase. In Obility’s Answer Engine Insights, prompts are grouped by audience, buying stage and market, so the question that decides a shortlist is tracked apart from the questions an existing customer asks once they are using you. A plugin can help with the second group. It does nothing for the first, and the first is where most brands are missing.
One honest limit on all of this: OpenAI’s documentation describes the system as it stands today, and the directory is new. Distribution rules may change, and a plugin that becomes popular may earn proactive suggestions. Plan for that as a possible bonus from building something customers use, not as the reason to build it.