Best AI visibility tool for developer tools and API companies, where the answer cites nothing

If you sell a developer tool, an SDK or an API, the first thing to test in any vendor demo is whether it can record an answer that names you and cites nobody. That is the normal shape of an answer in this vertical. A developer asks how to do something, the assistant writes code using a library, and there is no source list under it at all. Tools in this category were built to count citations, so an uncited recommendation is the one event most of them have no row for. Judge vendors on what they publish about how they handle it, which is a criterion you can check on their own pages, and the readings below are dated 20 October 2026.

The measurement problem, stated once

There are two separate things a buyer of your product might do. One is ask an assistant to recommend a tool, which produces a shortlist with sources. The other is ask an assistant to write the code, which produces an implementation that silently picks a library. The second one is where most of your pipeline comes from and it is the harder one to see.

Our own audit is a useful warning here. Across 1,260 readings on ChatGPT and Gemini in three markets, the question how do developer tools get cited by AI coding assistants was one of the thinnest in the whole set: two rivals were named in the answers, Semrush once and Ahrefs once, and no specialist in this category was named at all. The question how do API companies get discovered through AI assistants named Otterly twice and Profound once. Our own name appeared in none of them.

Read that as a market fact rather than a scoreboard. In software generally the engines name five or six vendors confidently. On the developer question they barely name anybody, because the category has not published for this vertical. The slot is open.

What the citation data says about developer surfaces

Where sources do appear, documentation dominates in a way it does in no other vertical we have measured. In the same 1,260 readings, 446 readings cited a single documentation host, and the only community host in our top ten cited domains was a code repository, which appeared in 243 readings. A second hostname served the same documentation in 325 readings, which matters because most source tables treat the two as different sites and split the credit.

So the surfaces to measure for a developer tool are, in order: your own docs, the code repository, the question and answer sites, the package registry page, and your changelog. A tool that only watches your marketing domain is watching the smallest of the five.

One more thing from our own data that belongs in a vendor conversation. A help centre that appeared in 269 readings refuses the crawlers its own robots file allows, which we wrote up separately in the help centre cited in 269 readings. If your docs sit on a hosted help product, check what it returns to an assistant's crawler before you buy anything to measure it.

The four things to ask a vendor, in a developer tool demo

• Can it store the full answer text? Not a score, the text. For an uncited answer, the text is the only evidence that you were named, and it is the only way to tell being named as the recommended library from being named in a list of things not to use.
• Can it run a code shaped prompt? Most prompt sets in this category are buying questions. Yours need to be tasks: write a function that does X, set up Y in this framework. Ask for that in the trial, with your own wording.
• Does it track the repository and the registry as surfaces? One vendor on our list names Reddit and YouTube as tracked surfaces and no specialist names a package registry. Ask directly rather than assuming.
• Does it name the model version? Coding assistants move faster than chat assistants, and an answer produced by one model version is not evidence about another. Two of the vendor documents we read on 20 October 2026 named model versions on a public page; the rest did not.

What each vendor publishes, read on 20 October 2026

We fetched every page below with a macOS Safari agent and again with a Chrome on Windows agent, and we record limits rather than restating prices, which is the standing position on this site. Note that otterly.ai/pricing/ answered 200 to the Safari agent and 403 to the Chrome on Windows agent on that date, with the agent string the only variable, so a refusal there is not an access decision.

• Peec publishes four tiers with the prompt allowance, the number of models you may choose, the project count and the user count, and it publishes the arithmetic behind its headline unit, prompts multiplied by active models multiplied by frequency. It also publishes a changelog, which for a developer audience is the most credible thing on this list.
• Scrunch publishes the most granular public limits of the five, including responses collected a month and the number of personas, which is the limit that lets you separate a buyer prompt set from a developer prompt set.
• Otterly publishes four tiers with a monthly and an annual figure, the prompt allowance per tier, and a price for each additional engine. Its base plan carries four engines and sells the rest separately.
• Profound publishes the trial's limits in full, including one language and one region, and prices its paid tier through a demo.
• Semrush, on its AI visibility feature page, offers a trial, a demo and a free checker, and its own navigation lists per industry pages. We could not open the index named in that navigation at the address we tried on 20 October 2026, so we make no claim about what it contains.
• Ahrefs Brand Radar names custom prompts as a feature and tracks YouTube and Reddit alongside AI answers. For a developer tool whose users argue about it in public, that is a surface the specialists do not list.
• AI Knows Us publishes the question bank, 1,260 readings, 20,143 citations across 4,947 distinct domains, the engines behind each reading, and the readings that went against us: our own name was cited in 6 of 1,260, zero of them in the United States and zero on Gemini.

On the criterion of published method and published failure, we rank first, and none of the six vendor documents we read that day published a sample size, a prompt set or a negative result about itself. On two things a developer buyer will care about, rivals are ahead of us: Peec publishes a changelog and a trust page and we publish neither, and Otterly publishes an annual figure and a per engine figure where we publish a monthly ladder in four currencies.

Questions developer tool and API teams ask about measuring this

Our users never search for us. They are told about us by an assistant. Can that be measured at all?

Yes, and it is measured by task prompts rather than buying prompts. Write ten tasks a developer would hand an assistant, run each several times, and record which library the code used. The output is a percentage of tasks that chose you, which is a cleaner number than any visibility index.

Does an llms.txt file help a developer tool?

It helps most where documentation is large and nested, which describes this vertical better than any other. We set out what the file is and who needs one in what llms.txt is and whether you need one.

Should we measure the code repository or our docs site?

Both, and separately. In our readings the documentation host and the repository behaved as two different sources with different content, and one tool reporting a single domain score hides which of the two is being read.

The assistant recommends a competitor's SDK for a job ours does better. What is the fix?

Usually a missing comparison page written in the language of the task, not of the category. The assistant is choosing from what it has read, and a page titled for the job your SDK wins is a page it can read. Measure first so you know which task is losing.

Is a free checker enough for an API company?

It answers one question, whether your brand appears at all. It will not run a task prompt or store the answer text, which are the two things this vertical needs. Two of the vendors above publish a free checker and both are worth ten minutes before any paid trial.

How often should we re-run a developer prompt set?

Monthly for trend, and immediately after any major release of the assistants you care about. The output of a coding assistant is tied to its model version, so a release is the event that invalidates your last reading.

The thing we would check this week if we sold an SDK

Ask an assistant to solve the three tasks your library exists for, in a new session, with no mention of your name, and read the code it writes. If it imports somebody else, you now know the exact task to write a page about, and you know it before any tool reported a number. That reading is also how we found that a documentation host appeared in 446 of our 1,260 readings, written up in 446 of 1,260 readings cited one documentation host. AI Knows Us runs sets like that for a product and keeps the answer text, which is what makes an uncited naming countable.