Agent
Discoverability
ChatGPT methodology

How we measure ChatGPT Plugin discovery.

This methodology explains how we run ChatGPT conversations, measure whether your Plugin is Found, Suggested and Positioned, and calculate the public ChatGPT Discoverability Scores shown in our Category benchmarks. Public benchmarks measure broad Category Intents. Your AgentDiscoverability.com platform score is tailored to the specific Intents your business wants to win.

Looking for Claude? Read the Claude methodology.

Agent DiscoverabilityPublic benchmark protocol
Evidence
Real ChatGPT Runs
Surface
The Plugin picker
Discoverable
Suggested in the last 14 days
Frequency
Recomputed daily
Built in our Research Lab

Grounded in thousands of real agent Intent runs.

Our Research Lab runs real conversations across agent platforms to understand how Integrations are found, selected, and used. The evidence from that work underpins this public benchmark methodology.

A new discovery surface

ChatGPT can now suggest Plugins in response to a Prompt.

At DevDay 2026 on September 29, OpenAI shared that 1.2 billion people use ChatGPT every week, and announced that ChatGPT will recommend relevant Plugins inside the conversation. Until then, a customer had to browse the Plugin directory or already know your name. Now they can start with the job they want to complete, and ChatGPT can suggest your Plugin at that moment. This is organic Plugin discovery.

A customer describes the job. ChatGPT suggests relevant Plugins inside the conversation, each with a one-click Install. The user had none of these installed.
The shift

The recommendation is the new result.

Search gave buyers ten blue links, and the Plugin directory gave them a category page. ChatGPT now goes straight from the Prompt to the Plugins that can complete the task. Your Plugin does not need to rank on a page. It needs to win the recommendation.

01
Plan our launch so owners update their own workstreams

Your buyer asks

The buyer describes the outcome they want. They do not need to know which Plugin can deliver it, or that yours exists.

ATO improves this decision02
ChatGPT found plugins that could be helpful
Booking.comFind all best stays and moreInstall
ExpediaPlan travel, flights & hotelsInstall
TripadvisorBook top-rated hotelsInstall

ChatGPT shortlists

ChatGPT searches a bounded set of candidate Plugins and selects a short list to show in the Plugin picker, with an Install button beside each.

03
ChatGPT found plugins that could be helpful
Booking.comFind all best stays and moreInstalled
ExpediaPlan travel, flights & hotelsInstall
TripadvisorBook top-rated hotelsInstall

One gets installed

The buyer installs one suggested Plugin, several, or none, then connects it to finish the job. Plugins that are not chosen lose that moment.

How ChatGPT decides

One Prompt. Three contests.

Found, Suggested and Positioned describe the path from a buyer’s request to the Plugin picker they see. Your Plugin can drop out at each one, so we measure each stage separately.

Your buyer types

What booking options cover family suites, villas and resort stays with live availability?

Which Plugins get the slot?

Booking.comExpediaTripadvisorHotels.com+ other travel Plugins
1

Found

ChatGPT searches its Plugin catalog, not the public web, for Plugins that could handle the request. The search returns a bounded set of candidates, and it is not exhaustive. Found tells you whether your Plugin was among them.

If you are not Found, you cannot be Suggested.

search_plugins

Candidates for this request

Booking.comFound
ExpediaFound
TripadvisorFound
Hotels.comFound

What the search can match is what your Plugin says about itself: its name, description and capabilities.

2

Suggested

From those candidates, ChatGPT chooses which Plugins to put in the “found plugins that could be helpful” Plugin picker. Suggested tells you whether your Plugin reached the picker the buyer sees.

Suggested is the ChatGPT Discoverability Score.

suggest_plugins
Booking.com
Expedia–
Tripadvisor
Hotels.com–

2 of 4 candidates reached the Plugin picker

The shortlist depends on the Prompt and the user, so it is not guaranteed to be the same every time.

3

Positioned

The suggested Plugins appear in an order. Positioned tells you which slot your Plugin took in the Plugin picker. The first slot is read first.

Position shows prominence after you were Suggested.

ChatGPT found plugins that could be helpful
Booking.comFind all best stays and moreInstall
TripadvisorBook top-rated hotelsInstall

In our runs, position is an observed prominence signal that works like a rank, and it can change with the wording of the Prompt.

Your public score

One number: how often ChatGPT suggests you.

The ChatGPT Discoverability Score is your public benchmark score. We run Prompts within each Subcategory and measure your Suggested rate: the share of contested Runs where ChatGPT puts your Plugin in the Plugin picker. It shows how you perform in that Subcategory, not the specific customer Intents your business may want to win.

ChatGPT Discoverability Score
61/100
Competitive

#2 of 7 in Hotel Search & Booking

Illustrative example

One core score. Three important factors to discoverability.

Suggested

Sets the score

Did ChatGPT put your Plugin in the Plugin picker?

Found

Diagnostic

Was your Plugin among the candidates ChatGPT’s search returned?

Positioned

Diagnostic

Which slot did your Plugin take in the Plugin picker?

Published ChatGPT formula

Score=Suggested÷contested Runs× 100

A contested Run is one where ChatGPT suggested Plugins. Prompts that name a product are excluded. A score is measured in a company’s own category: until its category has enough runs, a Plugin that was suggested shows a green dot and its score is coming.

0

Invisible

0–9

10

Buried

10–39

40

Emerging

40–59

60

Competitive

60–79

80100

Winning

80–100

Discoverable now, or no longer visible

A Plugin is discoverable now if ChatGPT’s own plugin search returns it, or if ChatGPT suggested it unprompted in the last 14 days.

No longer visible means we have seen ChatGPT suggest a Plugin before, but it is not in plugin search and has not surfaced in the last 14 days. That is “not observed recently”, not “removed from ChatGPT”: we have not confirmed why Plugins drop out.

Where the data comes from

Every day we run realistic customer prompts across product categories on ChatGPT and record which Plugins appear and in what order. We also check which Plugins ChatGPT’s plugin search returns.

Plugin search is read as a snapshot, so it shows what is in search today rather than a history of additions and removals. The day-to-day movement in the discoverable count comes from suggestions entering and leaving the 14-day window.

Your turn

How discoverable is your Plugin?

Enter your domain below to see whether your product can be discovered and used in ChatGPT and Claude.

Public registry data only. No account required.

A new optimisation discipline

You do not optimize with AEO. You optimize with ATO.

There are two discovery areas, and they need different disciplines. AEO is about being cited when an AI answers a question. ATO is about being found, suggested and installed when an agent completes a job.

The Plugin picker creates a new discovery layer: a high-intent surface where relevant companies are found, compared and chosen at the moment a user is ready to act.

For each Prompt, ChatGPT searches a bounded set of candidate Plugins, selects which make the Plugin picker, and shows them in an order. That search runs over ChatGPT’s Plugin catalog, not the public web, so Answer Engine Optimization (AEO) is not the direct mechanism for this surface.

AEO improves how accurately a company is represented, cited and recommended when an AI system answers a question. It works on the content and sources answer engines retrieve and trust. It does not change whether an agent finds, ranks or picks a callable tool.

Agent Tool Optimization (ATO), as coined by AgentDiscoverability.com, improves whether a product’s callable tools are found, suggested and positioned when an AI agent decides how to complete a user’s request. It works on the names, descriptions, schemas, metadata and documentation agents use to choose between tools.

Answer Engine Optimisation

AEO wins the citation.

AEO helps information get found, trusted, and cited in an answer.

Your content is cited.

Agent Tool Optimisation

ATO wins the install.

ATO helps a callable Plugin get discovered, suggested, installed and used.

Your Plugin is installed.
  1. 1

    Understood

    ChatGPT can tell what your Plugin does and which tasks it completes.

  2. 2

    Matched

    Your capabilities fit the Prompt better than nearby alternatives.

  3. 3

    Surfaced

    You earn a slot in the Plugin picker shown inside the conversation.

  4. 4

    Installed

    The user picks you over the other slots.

AEO vs ATO at a glance

Both matter. They win different moments.

AEOATO
The user’s momentAsking a questionTrying to get a job done
What you winA citation or mention in the answerA slot in the Plugin picker, then the install and the call
What you optimiseWeb content, sources, reviews, entity dataPlugin name and description, tool names, tool descriptions, schemas, metadata
Who decidesThe answer engine, choosing sources to quoteThe agent, choosing tools to use
How it’s measuredMentions, citations, share of voiceShare of suggestion and slot position, measured through repeated runs (ChatGPT does not report them)
How to optimise

Measure. Diagnose. Improve. Verify.

Run the customer Prompts your Plugin should win, identify where visibility is lost, make focused improvements, and measure the same Prompts again.

Agent Tool Optimisation cycle: Measure, Diagnose, Improve, and Verify
01

Measure

Run priority Prompts repeatedly and track share of suggestion, slot position and which competitors appear.

02

Diagnose

Find where visibility is lost: not Found, Found but never Suggested, or Suggested but outranked.

03

Improve

Make focused changes to the signals ChatGPT reads: Plugin name and description, tool names, descriptions, schemas and metadata.

04

Verify

Run the same Prompts again and confirm that outcomes improved without losing Prompts you already won.

Continuous, not one-off.

Scores refresh daily. Suggestions shift as ChatGPT, your competitors and your own Plugin change, so each new window shows whether your Plugin is getting easier to find and choose.

Unverified by OpenAI

Signals that may matter over time

OpenAI has not documented these as ranking factors. We expect them to influence whether your Plugin is returned and suggested as the ecosystem matures, so treat them as hypotheses, not as part of the score.

Ratings

Customer ratings, agent ratings

Reliability

Error rates, tool calls, latency

Reviews

Customer reviews, agent reviews

Evidence and research

Real Runs. Evidence you can open.

We run natural Prompts under consistent conditions, record the visible outcome, and keep the evidence needed to review every result.

  1. 01

    Represent real jobs

    Prompts represent the jobs buyers ask ChatGPT to complete.

  2. 02

    Run live conversations

    We repeat them under consistent conditions so results are comparable.

  3. 03

    Record the outcome

    Each Run keeps the Plugin picker, the slot order and supporting evidence.

  4. 04

    Review the evidence

    Each result can be checked against the captured conversation and evidence.

Evidence from a measured Run
A captured ChatGPT Plugin picker showing Booking.com, Expedia and Tripadvisor suggested for a travel booking request
Let’s talk

Got questions?

Want to improve your discoverability? Schedule a call with our team.

Schedule a call