How we track whether AI assistants recommend a product, and what we learned watching an email API launch

Admaxxer reports on AI search visibility next to paid and web analytics: which assistants mention a brand when someone asks a buying question, which URLs those answers cite, and how much traffic the assistants then send. This post is the method, then a worked example we have an unusual amount of access to.

Why we built it

Ad reporting answers "what did the click cost". It does not answer "did the assistant name us at all", and for a growing share of software purchases the assistant is the first and sometimes the only research step. A model that has never seen a coherent page about your product will recommend someone else politely and at length, and you will see nothing in your ad account, because nothing happened.

So the measurement has to sit one step earlier than the click.

The method

Three parts, and none of them is complicated.

1. A tracked prompt set. Not keywords. Prompts, written the way a buyer actually asks: "what should I use to send transactional email from an AI agent", "cheapest email API for a side project", "alternatives to the tool everybody already uses". Prompts are fixed when the set is created and left alone, because a prompt set you keep tuning is a prompt set that only ever improves.

2. Mention rate. The share of runs, per engine, in which the brand is named at all. We show it over a rolling 30 days, per prompt and per engine, because the same prompt does not return the same answer twice and a single run tells you close to nothing. The default cadence is weekly, with a daily option and a run-now button.

3. Citations. When an answer links a source, we record the URL, the domain, the engine and the position, and the list exports as CSV. This is the part that gets underused. Mention rate tells you whether you are in the answer. The citation list tells you which pages the answer was built from, and that is the actionable output, because pages are something you can change.

The output of a tracked set is therefore: mention rate per engine over time, the competitors named alongside you, a share-of-voice split, and the third-party domains the engines trusted for each prompt. Underneath it sits a readiness audit that checks the plumbing a model needs to read you at all: crawler rules, a sitemap, structured data, canonical tags, an llms.txt file.

Two design choices are worth stating plainly. Runs happen on a key you connect yourself, so the cost is cents on your own account rather than a subscription on ours, and every run is priced in the interface before it happens. And we only run engines that expose a real API. Answers from a model without web search are labelled as coming from training data, not presented as live results.

The worked example

We have been watching the launch of AgentiSend, a transactional email API aimed at software that sends mail without a person clicking send. It is a useful case because we can see both sides: the answers the assistants give, and the exact pages the team shipped to be quoted from. AgentiSend is built by the same team as Admaxxer, which is why we have that access, and it is also why we are publishing the method rather than a success claim.

Here is the honest state of it. No AgentiSend prompt set is being tracked yet, so this post carries no mention rate and no citation counts. What follows is the set-up: the prompts that will be tracked are the buying questions above, the engines are the three defaults, and the pages being watched are the ones listed in the next section. When the runs have enough weeks behind them to mean something, the numbers go in an update to this post rather than into a launch announcement.

What we can report today is what the pages look like, because those are public.

What a machine-readable page looks like

The category is dominated by an incumbent, and a homepage rewrite would not move that. What the team did instead is worth copying, and it breaks into four patterns.

Pages shaped like the question, answering in the first paragraph. Not "introducing our agent features" but a page per specific question: what email for AI agents means, how to cap how much a single agent can send, how to stop an agent repeating a send, how to require human approval before a send. The answer sits in the opening lines rather than after a scroll of positioning, so an assistant summarising the page has it in the first chunk it reads.

A plain-text twin of every page. Each route has a .md version with the same content and no navigation, no scripts and no layout, plus an llms.txt index written for a model rather than for a crawler. Their stated rule is that every figure in those files is read from the module that owns it and never typed by hand, and their build fails if one is. That rule is the whole game. A stale number in a machine-readable file is worse than no file, because it will be repeated confidently for months to people who never visit the site.

Structured endpoints for the facts buyers ask about. /pricing.json and /sla.json are plain JSON at stable URLs. Asked what a product costs, an assistant either finds a number it can parse or it paraphrases a pricing page and gets the tiers wrong.

Comparison tables where every row carries a source. Their comparison against the incumbent runs to sixteen rows, and the rule they adopted is that each row names a document and a date on both sides, with any row that could not be checked on both sides left off the page. Comparison pages are among the most quoted assets in a software category, and most of them are unsourced marketing claims that a model has every reason to discount. A sourced row can be repeated without the assistant having to take anyone's word for it.

What this changed in the product

Two things changed in how we present AI search visibility.

First, the citation list is treated as the working panel and mention rate as the summary, not the reverse. Mention rate tells you there is a problem. The citation list, and the third-party domains sitting where your page should be, tells you which document to write.

Second, we started reporting the absence of a cited page as a finding. If a prompt set contains "alternatives to X" and no page of yours is ever cited for it, that is not a ranking problem to be tuned. It is a missing document. The recommendation panel says so in those terms: citations to earn, engines that skip you, competitors out-mentioning you, readiness fixes, and whether the cadence is worth raising.

The honest negative

This measurement is noisy. Engines give different answers to the same prompt, they change behind the scenes without notice, and a mention rate that moves five points in a week has usually told you nothing. We report it as a trend over weeks with the run count visible, and we would rather show a small sample labelled as small than a clean number that is mostly variance. It is also why this post has no numbers in it yet.

AgentiSend is built by the same team as Admaxxer, so read the example above as a description of what those pages do, not as an endorsement of the product. The method is what we are offering. The pages are at agentisend.com if you want to look at how they are built.