ChatGPT Plugin Metadata Is a New Kind of AEO
Optimize plugin names, descriptions and parameters for correct tool selection with direct, indirect and negative golden prompts.
Direct answer: ChatGPT plugin metadata creates an optimization surface, but it is not conventional SEO. Names, descriptions and parameter documentation help the model decide whether a tool fits a request. The measurable outcomes are invocation precision and recall—not a web ranking, citation share or guaranteed installation.
OpenAI’s current plugin guidance recommends a labelled “golden prompt” set with direct, indirect and negative cases, followed by repeatable evaluation in developer mode.
Optimize for correct selection
A broad description can increase accidental calls. A narrow description can miss relevant indirect requests. Treat metadata as a classification contract:
- Name the domain and action clearly.
- Start descriptions with the situation in which the tool should be used.
- State important disallowed cases.
- Describe each parameter with examples and allowed values.
- Declare read-only, destructive and open-world behavior accurately.
These fields help the assistant distinguish neighboring tools and decide whether it should act. They do not make an unrelated tool relevant, and they should not hide side effects in persuasive copy.
Build a golden prompt set
| Prompt class | Expected behavior | Metric |
|---|---|---|
| Direct | User names the product or data source | Recall and correct arguments |
| Indirect | User describes the outcome without naming the tool | Recall without over-broad matching |
| Negative | Built-in or another tool is a better fit | Precision and correct abstention |
Record whether the right tool ran, whether arguments were correct and whether the component rendered. When selection fails, change one metadata field and replay the complete set. Preserve timestamps and diffs so improvement can be attributed.
A practical first benchmark is 60 prompts: 20 direct, 20 indirect and 20 negative. Direct and indirect prompts create 40 expected calls; negatives create 20 expected abstentions. One mistaken call on a negative prompt produces 95% negative-prompt precision in this simplified gate (19 correct abstentions out of 20), while four missed relevant calls produce 90% recall (36 of 40). Report the confusion matrix, not only an average score.
Separate plugin AEO from web AEO
Plugin invocation happens inside an assistant’s tool-selection system. Web AEO concerns discovery, retrieval, citation, answer use and referrals across web sources. Both require clear scope and testable language, but the systems, denominators and business outcomes differ.
A plugin team should report precision, recall, wrong-tool calls, argument failures and task completion. A publisher should report crawl access, retrieval, citations, absorption, referrals and outcomes. Combining those metrics into one “AI visibility” score hides the mechanism.
OpenAI recommends prioritizing precision on negative prompts before chasing marginal recall. That is a healthy optimization rule: a tool that acts in the wrong situation can be worse than one that occasionally needs clearer user intent.
| Field | Documented purpose | Failure to count |
|---|---|---|
| Name and title | Stable action identifier and concise human-readable action. | Wrong neighboring tool selected. |
| Description | User intent, trigger conditions, limits and distinctions. | Missed relevant call or accidental activation. |
| Input schema | Required fields, types, allowed values and limits. | Invalid, guessed or incomplete argument. |
| Output schema | Stable fields the model can inspect and reuse. | Result cannot support the next step. |
| Safety annotations | Accurate read-only, destructive and open-world behavior. | Incorrect permission or confirmation path. |
Primary documentation
For a web-publisher measurement model, see the AEO guide and AI visibility crosswalk.
Ask a question or join the discussion