How to Market an AI Startup Where Technical Buyers Ask

Why an AI product is harder to sell in public

Nobody evaluating your product can tell how the model behaves on their own data until they try it. "AI-powered" is on nearly every homepage in your category, so the phrase tells a buyer nothing. Technical buyers ask other people instead, in threads, with specific questions.

• Which transcription tool copes with heavy accents and crosstalk without inventing names?
• Has anyone compared this tool against just prompting the base model well?
• Does this product train on customer data? The privacy page is vague.
• What does it cost per seat once you pass the usage cap?
• Is there a local-model alternative that runs on a 16GB laptop?

None of these is answered by a feature list. They are answered by a test someone can repeat, a plain statement of what the product does badly, and a fact about where data goes.

There is a legal side too. A 2023 FTC business blog post asked advertisers whether they exaggerate what an AI product can do, claim it beats non-AI products without proof, have weighed its risks, and whether the product uses AI at all (summary by Kelley Drye). In September 2024 the FTC announced Operation AI Comply, five enforcement actions against companies that used AI hype or sold AI tools that could be used to deceive, and the FTC chair at the time said there is no AI exemption from existing law (FTC). A claim in a reply under your name deserves the same care as one on your pricing page.

The awkward part: many technical communities ban AI-written text

AI startups are the companies most likely to draft marketing with a model, and their buyers' communities are among the strictest about it. Know where AI-drafted text is not welcome before you plan anything.

Where a community bans AI-written or AI-edited text, write every word yourself or stay out. Saying "drafted with AI" does not make a post allowed where the rule bans it. Our longer piece on whether AI agents may post on Reddit covers the platform-wide side.

Build every answer from proof

A September 2026 playbook for AI startups, written by an agency that sells this kind of work, makes the same point: reproducible evals against a baseline, openly stated limits and a free tier or sandbox are what move technical buyers. Get these four things ready before you answer a single thread.

1. Publish an eval people can rerun
   Name the dataset, the baseline (include a well-prompted base model, because someone will ask), the metric and the cases where you lose. Link to it in replies instead of pasting a number that has no context.
2. Write the limits down
   Languages, file sizes, latency, the inputs where it gets things wrong. A founder who states a limit before being asked is believed on everything else.
3. Put data handling on one page
   Whether you train on customer data, where it is hosted, which sub-processors see what. Say only what is true: if a sub-processor stores data, do not write that you never store it.
4. Make it easy to try
   A free tier, a sandbox or a demo with sample data. Show HN expects this, and a thread asking for a tool usually ends with people trying two or three.

Base models change, so a comparison that held in spring can be wrong by autumn. Rerun the eval when the base model changes and update every reply and page that cites it.

Say who you are, every time

Reddit's site-wide rules ask people to take part authentically in communities they have a personal interest in, not to spam or manipulate content, and not to intentionally mislead others or impersonate anyone. The FTC's endorsement guidance says that employees who mention their company's products should disclose the relationship, that listing the employer on a profile is not enough, and that an employee or co-owner cannot cover it with a vague word such as "paid".

Whether and how a reply discloses is a decision the person or company running the account makes, but the platform's rules and the law expect material connections to be disclosed. In practice that means one plain sentence in the reply itself. Our disclosure templates have wording for different cases.

The example answers before it mentions anything, names the weakness and points elsewhere when something else fits better. The guide to answering tool recommendation threads walks through that pattern.

Where AI buyers ask, and how to take part

Where an agent fits, and where it should stay out

We know how this looks: an AI company selling an AI agent to AI companies, for communities that are tired of AI text. So here is the honest division of labour.

The slow part of the work is reading. Agent Agency reads up to five public pages from the link you give it, plus your llms.txt if you have one, and writes a Product DNA from them: who the product is for, how it sounds, what it never says. You approve it. Keep your eval, limits and data pages on your own site so the facts it works from are the ones you published. Then the agent reads each community's rules and finds the threads where your buyers ask.

• Approve mode. Every post, comment and DM waits for your OK. For an AI startup we would start here and rewrite drafts in your own words.
• Mentions only when they answer the question. No bare links, no copy-paste replies, and product mentions under a daily cap.
• A daily report. What the agents did and what they skipped, and why.
• Not everywhere. Reddit and X are live and YouTube is in testing. Agent Agency does not post on Hacker News, so HN launches stay with you.

In communities that ban AI-written or AI-edited text, the right setting is no agent at all: decline the draft, or write the reply yourself. Measure the work by technical buyers who sign up and try the product, not by how many people saw a comment.

Questions

Can I use AI to write Reddit posts about my AI product?

It depends on the community. Some subreddits remove AI-written posts or ban them outright, and Hacker News asks for no generated or AI-edited text at all. Where it is allowed, use drafts as a starting point and rewrite them in your own words. Where it is banned, write the post yourself or skip that community.

Should I post benchmark numbers in a reply?

Only numbers from an eval you have published, with the dataset, baseline and method, so anyone can check them. A figure without a link invites a hostile thread, and the FTC expects comparative claims about AI products to have adequate proof. Link the eval and name the cases where you lose.

How do I answer "isn't this just a wrapper?"

Directly. Say which model you build on, what you add on top, and how that compares with prompting the base model well, ideally with your eval showing both. If the honest answer is that a well-written prompt gets most people most of the way, say so and explain who needs more.

Do I have to say I built the product?

Reddit's rules ask you not to intentionally mislead people, and the FTC's guidance says employees should disclose their connection when they mention their company's products, and that a co-owner's disclosure has to name the relationship. One sentence in the reply is enough, for example: I'm one of the founders, so weigh this accordingly.

Add your product URL and a few answers about the product while you wait for an invite. Join the waitlist