Most advice on marketing an AI startup assumes people already understand the product. Usually they don't, and the people who do are not where reach gets built. So the work is translation: find positioning ground the incumbents left open, make the product explainable without a demo, and carry that explanation to channels native to the people you want.
The product where I learned this is Mimir, an AI research copilot for R&D and engineering teams. When I joined in March 2026 the product was real and the audience was zero: no followers, no content system, no launch plan. I ran marketing, communications, and social through July, covering individual acquisition, the content operating system, the launch plan, and the GTM playbook. Underneath all of it sat the problem this guide is about. The people who understood Mimir instantly were researchers. Researchers are not where reach gets built. The channels where reach gets built are full of people who will never read a paper.
Crypto has its own version of this, and I've written that guide: How to Market a Crypto Project With No Users. The AI version differs in a way that changes the work. In crypto, almost nobody understands the product yet, and your audience is still forming the question. With an AI product, a small expert group understands it completely and immediately, and everyone else does not. The temptation is to market only to the group that already gets it. Their channels are small, and every competitor you have is already standing in them. The reach lives with the people who need every sentence translated.
Find the ground the incumbents left open
The reflex positioning for an AI product is a borrowed comparison: the AI for X, or a famous tool's name with your industry attached. It feels efficient because it rides on understanding people already have. It also spends your introduction on someone else's product, because a reader who files you as a version of a thing they know finishes the thought with the thing they know.
Before settling Mimir's positioning I ran competitive intelligence on nine tools in the space and used it to pick ground the incumbents had left open. Read how every product in your space describes itself, and lay the claims side by side until you can see which are crowded and which nobody has made. Position on an unclaimed one that is true of you; it is the one place a newcomer can meet the product without an incumbent already standing in the frame.
Make it explainable without a demo
AI products demo well and describe badly. In a live demo the product carries itself: a question goes in, the answer comes back, the room leans in. Almost none of your marketing happens in that room. It happens in a caption, a search result, a thumbnail, an answer assembled by someone else's AI, and nobody there can watch your product think.
Founders usually fail this test in a specific way: the explanation is accurate and useless. It describes machinery, agents, retrieval, a model tuned on your data, when the person scrolling needs a scene: who was stuck, on what, and what changed. That was the bar I set myself at Mimir: put the product in front of students, engineers, and researchers in words that don't require a demo to follow. This is People Don't Commit to What They Don't Understand applied to a product whose best argument is a live experience. If understanding only transfers in a demo, growth is capped at the number of demos you can give.
Pick channels for the people, not the product
Standard advice for a technical product says LinkedIn, webinars, a whitepaper. That fits when a department is the buyer. When growth runs through individuals deciding on their own time, the question flips to where those individuals already are, and the answer is usually consumer channels. At Mimir I owned individual acquisition across TikTok, Instagram, X, and Reddit, because that is where students and engineers spend attention, off duty, between everything else.
Make each platform's native thing; the crypto guide linked above treats that discipline in full. The harder part is accepting who is in the room. Most people a reach channel touches will never buy, and that is the deal. The ones you want are in the same crowd, and you earn them by being worth a stop in the feed.
What honest numbers look like at two months
Here is what this produced, with the readings that belong to it. From late April, in two months of posting, Mimir's Instagram went from zero to 1,281 followers, and over 350,000 accounts were reached in that period.
The distance between those numbers is the useful part. Reach of that size says the translation traveled: content about an R&D copilot found its way in front of over 350,000 accounts on a consumer platform. The follower count says what the durable asset looked like at two months, which is small, because a following compounds slow before fast and two months is early on that curve. If a marketer reports reach, ask what is being kept. If a marketer reports followers at month two, ask what curve they are on. The numbers answer different questions, and neither substitutes for the other.
Then the reading specific to AI. The role ended when Mimir pivoted to a different strategy, and the channels I built stayed built for the direction the company moved away from. The ground under an AI company moves: a model release or an incumbent's feature launch can redraw the space while a campaign is still running, and strategy moves with it. So when you plan a pivot, count the distribution in its cost. The audience, the formats, the positioning ground: some of it transfers, some of it doesn't, and an honest plan writes the loss down instead of assuming the marketing will stretch. You can execute well against a direction you do not control, and the plan should price that in.
Making a product explainable to someone who will never see a demo, and carrying that explanation to the channels where they already are, is the kind of early-stage work I do with founders.