Answers
How do I get my API recommended by AI assistants?
Short answer
How do I get my API recommended by AI assistants?
In three moves, in this order. First, find the pages an assistant actually retrieves when a developer asks for an API like yours, because those pages and not your own site are where the answer comes from. Second, get accurately represented on them, which means the roundups, the comparison posts and the community threads that already rank for those questions. Third, measure with enough repeats that you can tell a real change from noise, because single readings flip: in our own step-zero run, 7 of 12 queries returned different outcomes across five identical repeats. Work done in that order is buyable. Work done in the reverse order produces a redesigned docs site and no movement.
Three moves, in this order
Reversed, this produces a redesigned docs site and no movement, which is the common way the quarter gets spent.
01
Start from the query, not from your product
The instinct is to begin with your own surfaces, because they are the ones you control. It is the wrong end. Begin instead with the exact sentences a developer types when they have your problem and no idea you exist, which are rarely your product category and almost always a job: pulling data from somewhere, avoiding a rate limit, replacing something that shut down. Those sentences are the queries that decide whether you get recommended.
Write them down before doing anything else, because everything downstream is scoped by that list. A programme built against ten questions your buyers genuinely ask is a different thing from one built against the keywords your category uses about itself.
Where to start the query set
Buyers describe a job, not a category. The list of sentences they type is what scopes everything downstream.
02
Go where the answer is actually assembled
Once you know the queries, look at what gets retrieved for them. In our own measurement across 901 citations and 82 hosts, 91.5 percent of cited pages were not owned by any vendor named in the answer. That is the working surface. It means the asset that moves most for an API company is usually not another page on its own docs site, it is accurate presence on the comparison post that already answers the question.
This is slower than publishing and it is the part that actually moves. It also has a floor of honesty: getting listed somewhere you do not belong produces a citation that damages you, because the model will happily quote a page saying your API lacks the one feature the buyer came for.
82 distinct hosts, our own run
The first three are where the answer is assembled. The fourth is the one most teams spend their quarter on.
03
Instrument it before you believe it
The measurement problem here is real and most of the category waves it away. Asking an assistant once and writing down what it said produces a number that feels like data and behaves like a coin flip. Our permutation test over 20,000 splits with no intervention at all put the null difference at a mean of roughly plus or minus 0.0016 with a standard deviation of 0.215, which is the useful finding: the noise is unbiased, so it cancels, and a difference-in-differences design works.
What that costs is samples. The 12 queries by 5 repeats design has a minimum detectable lift of 51.1 percentage points, which is unusable. Getting to a 9.9 point minimum detectable lift takes roughly 40 queries by 40 samples, about 3,200 calls and six hours per timepoint. That is the honest price of knowing whether anything worked.
Minimum detectable lift, our own design
The second design is about 3,200 calls and six hours per timepoint. That is the price of being able to tell whether anything moved.
Asked next
The questions that follow this one
Do I need to publish more content to get recommended?
Usually less than you think. Our own measurement puts 91.5 percent of cited pages off-site, so the constraint is presence on other people's pages rather than volume on yours. Publishing helps once you are being retrieved and it is a slow way to become retrieved.
Does llms.txt make my API citable?
It is not where the evidence points. The pages carrying recommendation answers are overwhelmingly third-party, and a file on your own domain does not change which third-party pages get retrieved. Treat it as cheap housekeeping rather than a lever.
How many queries do I need to track?
Enough that a real change clears the noise. On our numbers, 12 queries by 5 repeats detects nothing under 51.1 percentage points, while roughly 40 queries by 40 samples brings that to 9.9 points. Below that you are reading noise and calling it a result.
Can I do this without an agency?
Yes, and the measurement design is published for that reason. What it costs you is the sampling run and the discipline to hold a control group, which is the part teams skip when they are running it themselves.
The order that pays
Free gap report
When the model answers,
be the one it names
Tell us your category and the questions your buyers ask. We run them against live AI answers and walk you through what came back. If you are already winning, we will tell you that too.
12 queries · 5 repeats each · 5 working days
Free · Walked through live · Five working days
Queries we run
Cited instead of you
reddit.com38%
g2.com26%
news.ycombinator.com21%91.5% of citations point off-site