How to Market an AI Tool

I “It uses AI” is not an offer

Product Hunt counted 1,111 launches in its AI category in the second quarter of 2026, up 46% on the year before. That is twelve a day, on one site, in one category, so telling a buyer your tool uses AI no longer separates it from anything else they were shown this week.

An offer for an AI tool states four things: who it is for, narrowly enough that a person can tell in five seconds whether it means them; the job it takes off their desk; what it costs; and how they will know it worked, in a number they already track. “AI assistant for finance teams” is a category. “Closes your month-end reconciliation in a day, or you don’t pay” is an offer, and the reason it is one is that it can fail. This is the first step of our method: the offer is written as a claim that could turn out false.

If you cannot write that sentence, the problem sits upstream of marketing, and the work is to go back and find the job.

II Fifty people, by hand

Pick the first fifty users yourself, by name. They should be people who have the problem this month and whom you or someone you trust can reach without a cold email.

Fifty is small enough to watch. That matters more than reach right now, because what you need is to see where a real person gets confused, and a dashboard will not show you that. In our method this is the warm step: people we trust see the offer first, then we watch real users try it, and only then does anything go out cold.

There is outside evidence that warm is also where the results are. Growth Unhinged’s 2026 State of B2B GTM report, a vendor-sponsored survey of 300 software go-to-market leaders, ranked network and introductions as the single best channel and intimate events second. Its author’s summary of the alternative was that “AI SDRs aren’t living up to the early hype.” So the robot that emails ten thousand strangers for you is losing to lunch.

III The page is the first experiment

Before you put more work into the product, test the offer on a live landing page with real visitors. The test needs the offer, the page, a funnel, two or three variants, and a pass/fail threshold you wrote down before the first visitor arrived. Agents are good at this part and will draft the variants and build the pages in an afternoon, after which they wait for the data like everyone else.

Kromatic’s Real Startup Book gives benchmarks for this test: 200 unique visitors as a minimum for a directional signal and, for paid search traffic, 5 to 15% email signup as typical. Kromatic offers them as rules of thumb, which is all you need in order to set a bar in advance and hold yourself to it.

The same guide is blunt about what the page proves: “Leaving an email is cheap; an email signup is not a purchase.” For an AI tool we would go further, because curiosity about AI is high and costs the visitor nothing, so signups flatter you. Ask for something that costs a little, such as a booked call, a card, or their own data uploaded.

IV The demo and the launch

Everyone has a demo now, and a clip of a model doing something clever on a clean example proves that models are clever, which the viewer knew. Your demo has to prove the thing a sceptic doubts, which is that it works on their mess. Run it on a customer’s real file with the ugly columns left in, show the failure case and what the tool does about it, and keep the clock on screen. We would take an honest ninety seconds on bad data over a cinematic three minutes on good data.

The launch is a different problem. Product Hunt reported that over 61% of the builders in its AI category that quarter were launching there for the first time. A Hacker News thread titled “Is Show HN dead? No, but it’s drowning” drew more than five hundred points. One commenter put it this way: “The interesting Show HNs still exist, they’re just buried under noise.” Launch platforms are where builders now go to be seen by other builders who are also launching.

So treat launch day as a dated artefact that gives you a link, a deadline and a reason to email your fifty. We would not count it as a channel. In our sequence channels open one at a time, after the warm step, so that when signups arrive you know which door they came through. Opening five on launch morning feels like momentum, but the signups arrive as one blended number that you cannot attribute to any of the five.

V What the press release is for

Answer engines changed who reads coverage, but the press release itself is still something machines rarely quote. PR Daily, reporting on PR software vendor Muck Rack’s latest “What Is AI Reading?” study, said press releases made up 1% of AI citations, against 22% for journalism. Meltwater, which sells press distribution and so had every reason to find otherwise, measured press releases at 0.2% of the citations it tracked across eight models in May 2026, against 37.6% for earned and news media.

The release is for the journalist. Cision’s 2026 State of the Media survey of 1,899 journalists, another vendor study, found that 66% rely on PR-provided content for story ideas, and that 72% say fewer than a quarter of the pitches they get are relevant. Reporters will use your material, in other words, and most of what they are sent is shit.

That is why our method has a single press release: one release with a real claim, a number and a named customer or a named result, sent to the few reporters who cover exactly this. The coverage is what gets read, by people and then by the engines. Keywords come before it, so that when the coverage lands, the page it points to answers the question people are typing.

VI When to stop and change the offer

Set the stopping rule at the start, in writing. If the page misses its threshold with enough traffic to trust, change the offer or drop it. Changing the button colour, the headline font or the ad platform first is a way of not hearing the answer.

Watch whether people come back. Stripe Atlas found that nearly 30% of customers at top-decile solo startups returned the following month, compared with 8% at middle-decile ones. For an AI tool we would count the second use as the conversion, since the first is so often curiosity.

When the offer does clear its bar, that is the point at which to build the rest, launch, and hand it to whoever runs operations with the numbers attached. When it does not, you have lost a few weeks and a landing page, which is cheap next to a year spent marketing a sentence nobody wanted.