Do plugin and MCP tool descriptions affect discovery?
For installed plugins, yes: OpenAI says ChatGPT and Codex decide when to call a tool based on the metadata the developer provides, which includes MCP tool names and descriptions. For directory search, our small pilot on 9 Oct 2026 saw results that mostly matched the words in a listing name or short description. For suggestions to people who have not installed a plugin, no source shows that descriptions matter.
Three surfaces, three answers
Each point carries the page it came from. Platform behaviour is described in OpenAI's own words where we can.
OpenAI advises names that pair the domain with the action, descriptions that start with "Use this when" and state the cases the tool should not handle, and replaying a prompt set after each change.
OpenAI, Developer docs, optimize metadata
Logged out on 9 Oct 2026, a search for CRM returned 47 listings, 45 with CRM in the name or short description. HubSpot, which sat on two directory shelves that day, was not among them. A search for "customer relationship management" returned none. Nine queries, one day: exploratory, not a finding.
KailxLabs, KailxLabs directory pilot, 9 queries and one shelf snapshot · OpenAI, ChatGPT plugin directory, read logged out
The plugin manifest has an optional keywords field described as discovery terms. How it is used is not documented.
OpenAI, Developer docs, optimize metadata
OpenAI ties suggestions to utility and user satisfaction, and says nothing about descriptions.
OpenAI, Developer docs, publish and review
What we do not know
- The pilot is nine queries on one day, logged out, from one location. It shows the directory can be measured; it does not show how search works.
- A logged in search may behave differently. We have not tested it.
- Correlation between listing words and search results does not show that changing the words will change the results.
Nobody outside OpenAI knows the full recommendation method. We report what is documented, what we counted, and what is still unknown, and we keep the three apart.
Sources and dates
Source date is the date on the source itself. Checked is the day we last read it.
| Source | Publisher | Source date | Checked |
|---|---|---|---|
| Developer docs, optimize metadata | OpenAI | Not dated | 9 Oct 2026 |
| KailxLabs directory pilot, 9 queries and one shelf snapshot | KailxLabs | 9 Oct 2026 | 9 Oct 2026 |
| ChatGPT plugin directory, read logged out | OpenAI | Live page | 9 Oct 2026 |
| Developer docs, publish and review | OpenAI | Not dated | 9 Oct 2026 |
Related: Can you optimize a plugin for discovery? · How to measure plugin visibility · ChatGPT plugins research hub
- 9 Oct 2026: answer published, from OpenAI pages read at the source that day.
Talk to the founder.
20 minutes with Kailesk Khumar. You leave knowing where AI names you today, and what it would take to become the answer.
- A live look. We ask five of your buyer questions on ChatGPT, Google and Perplexity while you watch.
- The reasons. The three biggest reasons AI names a rival instead of you.
- The plan. A clear next step and its price, whether or not you hire us.
Pick a time that suits you. The calendar invite arrives straight away. Prefer a link? Book on Cal.com.