Making something and getting it into the hands of the right people are different jobs. The second can feel strangely harder at the beginning. You have a product, perhaps even a good one, but no established way for the people who would care to find it. Every message is a guess about who needs it, where they spend time, and which part of the work will make sense to them.

I don't think that first stretch can be skipped. The useful version of distribution begins with contact: showing the thing to someone specific, listening to what they understand, and noticing where their interest ends. A launch can start that conversation. It cannot stand in for it.

The first audience is found one person at a time

Early distribution is often described as a channel problem. Which platform? What posting schedule? Which headline? Those questions matter later. The first question is smaller: who has the problem strongly enough to try this now?

Paul Graham's “Do Things That Don't Scale” argues that founders commonly have to recruit their first users manually. That essay is advice drawn from startup experience, not a controlled study of every product. Its practical value is that it puts the work in the right order. A conversation with one well-matched person can tell you more than a broad announcement to people who were never likely to care.

Manual does not have to mean a cold pitch copied into a hundred inboxes. It can mean making a useful example for a particular community, asking a designer to try a creative tool with their own image, or watching someone attempt the first task without explaining the interface for them. The goal is to learn enough to improve both the product and how it is introduced.

That is why the beginning feels slow. You are doing product research and distribution at the same time. A reply may show that the wording is wrong. Silence may mean the audience was wrong, the channel was wrong, or the message simply did not reach anyone. One attempt rarely tells you which.

Repetition can build an asset

The work should become less blind over time. Not because an algorithm rewards persistence, but because each honest attempt can leave something behind:

• A clearer account of who uses the product and what they call their problem.
• Examples that show the result without requiring a long explanation.
• A list of communities and conversations where the work is relevant.
• Relationships with people who gave useful feedback, even if they did not become customers.
• A record of which messages led to meaningful use, rather than which ones merely collected views.

Some of these are reusable. A strong example can be shown again. A question heard three times can change a help page. A useful article can keep answering the same question after the day it was published. But none of this guarantees compounding growth. A channel can change, an audience can move, and repeating a weak message can make you more efficient at being ignored.

The thing that compounds is understanding. Volume only helps when it creates enough real contact to improve that understanding.

Measure the step after attention

Views and likes can tell you that something was seen. They say much less about whether the right person understood the product, tried it, or came back. For a small company, those later steps matter more. If a post brings attention but nobody reaches a useful first result, the next job may be to fix the page or the product, not write another post.

There is no universal dashboard for this. A useful record can be modest: where someone came from, what example they saw, whether they tried the first task, what confused them, and whether they returned. Keep counts only where the data is actually collected; keep observations as observations. A handful of conversations should not be dressed up as a conversion rate.

This is also why a free product changes the distribution question. Grainform currently lets someone bring in an image and make grain artwork in the browser, then save PNG or SVG without an account or watermark. That removes some friction from trying it, but it does not solve discovery. The work becomes showing a specific result to people who might use it and seeing whether the tool helps them get there. “Free” is useful information. It is not an audience.

What I would give an agent

Agents can make the repeated parts of this work cheaper. They can collect relevant questions from approved sources, compare how people describe a problem, prepare several honest explanations of a feature, check that a claim matches the current product, and turn a set of notes into a draft. They can also look for contradictions: a page saying “free” while an old help article still says “watermarked preview,” for example.

That is different from sending an agent out to manufacture activity. I would not judge a system by the number of posts it can produce. I would judge it by whether it helps us make a better choice about what to publish, where it belongs, and whether the product actually supports the promise. An agent can prepare options. Someone still has to own the message, check the context, and see what happened afterward.

Anthropic's guide to building effective agents recommends starting with simple workflows and adding complexity where it improves the task. Its evaluator-and-revision pattern is a useful fit here: draft a piece, check it against the product and sources, then revise it before publication. That is a workflow design principle, not evidence that agents will automatically make a distribution channel work.

There is a quality boundary too. Google's guidance on AI-assisted content distinguishes useful, accurate material from large amounts of low-value generated content. The same distinction matters even when search is not the channel. More pages and posts can fill a calendar while making the company harder to trust.

A loop worth repeating

The process I want is small enough to run and honest enough to learn from:

1. Pick one group of people and one problem they already recognize.
2. Show a real result, with a clear path to trying it.
3. Make the first contact appropriate to the place and person. Do not disguise a pitch as participation.
4. Record the response, including confusion and silence, without pretending either has one certain explanation.
5. Improve the product, example, or wording based on what the evidence can actually support.
6. Reuse what still helps; retire what does not.

An agent can help with research, assembly, fact checking, and the record. The decision to speak and the responsibility for what is said stay with us.

At the start, distribution is hard because so much is unknown. Over time, the work can get easier in a specific sense: fewer guesses have to be made from zero. We know more about the people, the examples, the objections, and the route from first glance to first use. That knowledge has to be earned, and it can be lost if we let automation turn the conversation into noise.

Sources and scope

This is an operating view, not a measured account of Einhar's acquisition results. The sources below inform the approach; none proves that every company's distribution becomes easier with time or that agents cause growth.

• Paul Graham, “Do Things That Don't Scale”, on manual recruitment and early user contact.
• Anthropic, “Building effective agents”, on simple agent workflows and evaluation.
• Google Search Central, “Guidance on generative AI content”, on accuracy, value, and scaled low-quality content.