How We Built an AI Content Writing Agent (Starting From 16 Impressions)

Most agencies will never show you their own Search Console. We will: at the moment this article was written, digitalbees.ai had 16 impressions. That is effectively zero — and it is exactly why it makes a good starting point for the story of how we handed that problem to a swarm of agents instead of a content team.

A site with no history has no organic visibility

A new site starts from zero no matter how good it looks. No history, no links, no signals that would give a search engine any reason to show it. For a company that claims to automate business systems, that raises an uncomfortable question: if the automation is really that good, why not apply it to yourselves?

The answer is that we did. A blog is the fastest channel to organic visibility because it is the only one where you can answer the specific question someone is typing into a search box right now. But that means publishing focused articles regularly, each with a single primary keyword and no overlap between them — precisely the kind of disciplined, repeatable work people hate and systems handle well.

Why an agent instead of a content calendar

The conventional approach is a content calendar in a spreadsheet and someone trying to catch up with it every week. We built an agent in the swarm that treats the job as a process instead. Here is what it consists of.

Written rules instead of a style guide

The agent is not told to “write in our tone”. It gets an editorial rulebook — a document defining identity, voice, forbidden patterns and article structure, and it applies unconditionally. It also spells out what is off limits: empty hype, a chatbot-centric narrative, half-translated hybrid terms, clickbait headlines that do not deliver. A rule you can read is a rule you can check against; a “feel for the brand” is not.

Vetted facts as the only source of truth

The biggest risk in automated content is not weak style, it is invention. So along with the assignment, the agent receives a file of vetted facts — services, positioning, legal details, our stance on references. Anything not in that file or in the brief for a specific project simply does not exist and cannot be claimed. We have no public client references, so we do not invent any; when we have them with the client’s permission, they go into that same file, and only then into articles.

Topic selection that knows what has already been published

Before writing, the agent receives a list of everything already published. If a topic is covered, it has to find a fresh angle or propose something else. That solves the classic failure mode of an automated blog: ten variations of the same article cannibalising each other in the search results and diluting every primary keyword.

A validator allowed to say no

A draft does not go straight out. It is checked by a separate AI validator looking for unsupported claims, deviations from the rulebook and structural gaps. The crucial part is that the validator has the authority to send the article back. An agent grading its own work is not quality control — it is just a longer prompt.

A swarm, not a tool

None of these steps is remarkable on its own. The value appears once they are connected: one agent picks the topic, another writes, a third checks, a fourth publishes and records the work in a public activity log. A single agent is a tool; a swarm of connected agents is a system that runs without someone pushing it every morning.

This is the same architecture we build for clients, applied to a problem of our own. Swap “blog article” for “incoming email”, “quote”, “meeting minutes” or “compliance form” and you get the same pattern: clear rules, a verified data source, an agent that executes, a validator that checks, and a trail you can audit.

So what about those 16 impressions

We are not going to pretend we know what that number will look like in three months — that would be exactly the kind of promise we avoid. What we can say is that the problem no longer depends on whether anyone finds time to write. The system is in place, it runs, and its effect is measurable with the same tool that showed us those 16 in the first place.

This article is not an exception to the rule — it is a product of that same process. It was written by an agent in the swarm, working from written editorial rules and vetted facts, about a system that same swarm built. If you are curious what this would look like applied to a process that genuinely costs you time, you do not have to trust our promises. Inspect our systems.

Frequently asked questions

How does an AI content writing agent decide what to write about?

It starts from the site’s actual visibility data and a list of everything already published. If a topic is covered, the agent drops it or looks for a fresh angle. The goal is one primary keyword per article, with no internal overlap between pieces.

How do you stop the agent from inventing clients, numbers or references?

Two things: an editorial rulebook and a file of vetted facts. Anything not in that file or in the project brief does not exist and cannot be claimed. A separate AI validator checks the draft before publication and sends it back for revision if it finds a claim without backing.

Does that mean no human is involved at all?

A human does not write or edit sentences, but does set the rules and hold the final word on what goes out. The agent does the work, the rulebook sets the boundaries, and the system’s activity is visible in a public activity log.

Why publish the fact that you only had 16 impressions?

Because it is the only honest way to demonstrate capability while we have no public client references. Instead of promises, we show the starting point and the system running on it — that can be verified, unlike marketing claims.

Can the same approach work outside marketing?

Yes. The same architecture — rules, a vetted data source, an agent that executes and a validator that checks — works just as well for email handling, documentation or compliance processes. Content is simply the domain where the result is easiest to see.