I run growth at a dev tools startup. How can I tell whether Codex recommends us or the category leader?

Developer tools companies are the clients where AI recommendations show up in revenue conversations earliest. A developer asks a coding agent how to add a feature and the agent picks a library. If it picks your competitor, you will not see the lost deal in any funnel report. So the first job is to find out how often that happens, and the answer takes more care than asking Codex once.

Why developer tools are a special case

Three things differ from consumer or B2B software categories.

The agent chooses while working. A buyer in most categories asks "what are the best options?" A developer often asks an agent to do a task, and the agent selects a tool as part of the implementation. Your visibility depends on being the tool it picks when writing code.

The sources are technical. Documentation, repositories, package pages, community threads, and tutorials matter more than review sites.

The surface is different. Coding agents and chat assistants are separate products with their own behavior. Developers use both, and they may recommend different things.

How we build the prompt set

For a dev tools client, we start with the real work developers do and the questions they ask around it. We group prompts into four types:

Category questions: "What is the best way to add observability to a Python service?"
Task prompts: an instruction to build something, where the agent picks a library.
Comparison prompts: "X versus Y for queues."
Brand prompts: "What is [client product]?" to see what the agent knows and whether it is accurate.

We weight task prompts heavily. We also set personas, such as a solo developer versus a platform engineer, since their questions differ.

Why we sample many times

A single agent response is a draw from a spread. We report the share of responses that recommend the client, the share that recommend the category leader, and the share where neither appears, aggregated over many runs and several weeks. A quick calibration we share with clients: if a product truly shows up in 30% of responses and you sample 10, the standard error is about 14 percentage points. At 100 responses it is about 5. That stops teams from reacting to noise.

What we run it on

We run client programs on Promptwatch. The relevant feature is its coding agent visibility: it tracks what Codex, Claude Code, and OpenCode recommend. Codex and Claude Code are monitored through real-UI monitoring, and OpenCode is tracked via API, which we note in client reports. All active models are included on every paid plan, so we can compare coding agents with ChatGPT, Gemini, Perplexity, and the others in the same project.

What we use around it:

Prompt tracking by type and topic with daily panels per model, country, and language, and personas for the developer segments.
Competitor visibility and heatmaps to see where the category leader wins, for example on Codex but not on ChatGPT.
Citation analytics at page and domain level, with offsite mentions, to see which sources are cited around the recommendation.
Agent Analytics for IP-verified AI crawler logs, to check that the client's docs are reachable by GPTBot, OAI-SearchBot, ChatGPT-User, ClaudeBot, and the rest.
Visitor analytics to connect AI-referred visits to signups.
Content tools to draft and refresh pages, with review before publishing.

Crawler logs start on the Professional plan ($245/mo). Essential ($95/mo) has none, and the free Explore tier is ChatGPT only.

Diagnosing why the leader wins

When the category leader is recommended more, we work through a short list:

Do the agents read the client's docs? Check crawler logs and look for blocks on the docs subdomain, which often sits behind different bot rules than the marketing site.
Does the documentation answer the task prompt? A page that explains concepts but never shows a working example will lose to one that does.
Who writes about the leader? Look at cited third-party pages: tutorials, community threads, comparison posts.
Is the client's brand clear? If a name is ambiguous or shared with other things, agents may skip it.
Is the information current? Old install steps or deprecated API examples hurt when an agent reads them.

We are careful not to claim we know exactly how Codex chooses tools. We do not. We measure what it recommends, look at what it cites, and fix what we can influence.

What to tell your leadership

The data is a sample of what agents say to the prompts you chose.
It will show you who gets named more often, and the trend.
It cannot guarantee a recommendation.
Improvements come from clearer docs, working examples, reachable crawlers, and being represented accurately on third-party pages.

The short version

If you run growth at a dev tools startup and want to know whether Codex recommends you or the category leader, build task-shaped prompts, sample them repeatedly, track coding agents as their own columns, and read the sources. That is the setup we use, and it is why our developer-tool programs run on Promptwatch.