How to get your product recommended by Cursor

Cursor is an editor first. When it makes a decision the project is already open and already indexed. That changes what it reaches for.

We ran 3,373 sessions with Cursor, on Grok 4.6 and Grok 4.7, across 29 categories. Here is what moved them.

Where Cursor landed

The pattern: Cursor reaches for what drops into the project

Look at the maps row. Cursor picked Leaflet, a library rather than a service, where Claude Code and Codex picked hosted map platforms. Look at deploy, where Cursor picked Vercel while Codex picked Render. The search row cuts the other way: Cursor picked Meilisearch, a separate engine to run, while Claude Code and Codex both picked Postgres full text search.

One reading is that Cursor leans towards the option that fits into an existing project with the least new infrastructure, as a tendency rather than a rule. That is what you would expect from an agent working in an editor with the project open. Adding a dependency is cheap. Standing up a new service is not.

If your product needs an account, a dashboard action and a deployed service before the first call succeeds, that is a real cost in a Cursor session. If it can be a package plus an API key, that is cheap.

What to do

1. Make the first call work from inside the project

Count the steps in your quickstart that happen somewhere other than the code. Create an account. Open a dashboard. Create a project. Copy a key. Configure a webhook.

Each one is a place a session stalls. You cannot remove all of them. You can do this:

• Offer a test or sandbox key that works without an account, even if it is rate limited.
• Make the first call work with a single environment variable.
• Put the dashboard steps inside the code block as numbered comments, so nothing is implied.
• Fail loudly when a key is missing, with the fix in the error text.

2. Ship a template for the stacks your users run

In our data, what is already in the repository beat an explicit written instruction to use something else, two times out of three. A maintained official template is the cheapest way to be already in the repository.

3. Keep the quickstart pasteable

One complete file. The install command in the same block. Every environment variable named. An agent that cannot make your example run does not open an issue. It picks something else.

4. State the version on the page

Current version: 4.2.1 (September 2026)

Models write code for interfaces you removed. One line prevents it.

The self-preference question

In agent frameworks, Cursor put the Cursor SDK first while Claude Code and Codex both put the Vercel AI SDK first. It did the same in two more categories: Cursor Bugbot first in code review, and Cursor Automations first in AI SRE.

This is worth stating carefully. The Cursor SDK is a real product in that category and a reasonable choice inside the Cursor environment. We also wrote the agent framework prompts to name the makers in alternating order, with a linter that checks it, precisely because we expected the question to come up.

It is also not only Cursor. In code review, Claude Code and Cursor each put their own maker's reviewer first: Claude Code review for Claude Code, Cursor Bugbot for Cursor. Codex put Codex code review second.

So: in three categories, Cursor put its own maker's product first, and in code review Claude Code did the same. It is worth watching rather than a proven pattern, and we will keep measuring it.

For a vendor the practical consequence is real. In a category where an agent's maker also sells a product, expect that agent's numbers to be harder, and measure it rather than assume it.

Where content fits

We measured web search behaviour for every agent. Cursor ran at least one web search in 72% of its sessions, between Codex at 97% and Claude Code at 28%, so published pages can reach it in most decisions.

The safe conclusion is the one that holds everywhere. Content is a lever and the repository is a bigger one. With Cursor, spend on the template and the quickstart before the blog.

How to check whether it worked

Take three repositories in the stacks your users run. Write ten requests the way your users describe the problem, not the way you name the category. Run each five times in Cursor. Record what got installed.

Then read the losses. In our data the most common losing pattern is not that the agent preferred a rival on the merits. It is that the agent started with the product and then hit a step it could not finish from inside the editor.