A deep dive into structuring an AI directory business ran by AI agents—my first experiment in creating a fully automated AI-driven venture

Let’s build AI automatons! Wait, we need a plan first. In this post I’ll share how I took the “Online Directory” business model and split it down into AI Agents.

Making a Directory Run by AI Agents

The First Experiment 

Each business model has it’s own scale of complexity. Here we’re focusing on one of the simplest. Web directories collect together resources in a way which makes it easier for the end user to find something that they’re looking for.

Directories range from artful collections all the way down to spammy lists. We will avoid the latter.

Automating directories will be our first experiment for the 100% AI Business Challenge.

Even if you’re not interested in starting a directory, the process of breaking down a business’ core components might be useful to you. Let’s dig in.

My Experience with Web Directories

I’ve launched perhaps 10 different types of online directories over the years. Some have been successful, many have not. 

There can be a lot of cross-over between directories and pSEO (programmatic SEO), so I’ve dabbled with these as what I call ‘microsites’; small standalone websites which attract users but also link out/convert them to a main product elsewhere.

Directories can generate their own revenue through ads/paid submissions, but typically I’ve used them for their SEO benefit for a separate product elsewhere. In this experiment we’ll aim to make them profitable on their own.

The Anatomy of Launching an Online Directory

Here’s how I see the 4 main ‘roles’ needed to successfully launch and run a directory:

Depending on the complexity of your business, this may be a list of 3 or 4 roles, or 300 – 400. These ‘roles’ could compare 1:1 with typical employee roles, or because we’re automating them, they may be split up a bit more (e.g. an accountant role may include data entry + data reconciling, these may be simpler to imagine as 2 separate agent roles). 

Back to directories. As you can see there are 2 ‘one off’ buckets of work:

• Directory Prospector – Generate & Validate ideas for new directory
• Directory Builder – Build the directory (website & data), optimise the SEO

… these only need to be done once to initiate the project.

After that we have 2 cyclical roles which need to periodically check-in on the project and support growth/user requests:

• Directory Marketer – Initially launch the directory, then incrementally apply marketing to enhance it’s value over the long-term
• Directory Manager – Deal with customer service requests (add/change/remove listings, check and report progress, help customers who have pre-sales questions etc.)

Breaking Down Job Roles into Agent Roles

Now that we know what’s needed to launch and maintain a directory, let’s build out a task list and think about how to delineate work to AI agents. 

You’ll see I split each role into Tasks, Inputs, Outputs, Tools, and Tests. This is my own system for building these bots, which is a work-in-progress. If you have a better approach, please do share in the comments.

Role 1: Directory Prospector

This role is all about finding the most advantageous idea for a directory. It needs to be able to look at search/social media trends, and decide which ideas are best to pursue.

Tasks:

• Ideate – Given a remit, come up with ideas for directories
• Explore trends – See if each idea is growing or falling
• Niche research – Learn all about the niche
• Search engine research – Discern how much traffic each idea may capture
• Domain prospecting – Find viable domain name opportunities

Inputs:

Building this as an AI agent, I’d like this agent to be able to take several inputs: 

1. A described niche (e.g. AI tools)
2. Hard-typed list of ideas from a human

… that way the agent has somewhere to start. It might be fun to let it go wild-card for some ideas, but I suspect I’d end up with directories of pet psychologists in Utah or similar. I may test this as we go.

Outputs:

This AI agent needs to pick a single directory idea, (more about scaling & nesting below). It needs to return the most optimum idea it can think of after evaluating many.

I want all of its ideas to be parsed through a matrix such as the following, finally outputting the singular idea.

Idea	Search Traffic Volume	Search Ranking Difficulty	Value Added	Relevancy to Target Niche	Quality of Available Domain	Topic Trending Up or Down
Pet Psychologists Utah	10	1	20%	0%	100%	Stagnant

From this I want the AI to ultimately return something like the following JSON (which is like an object in code):

{
    "idea_name": "AI tools for Designers",
    "keywords": [
        "ai designer tools",
        "best ai tools for designers",
        "using ai in design"
    ],
    "best_available_domain": "aidesigntools.com",
    "second_choice_domain": "aidestools.com",
    "audience_profile": "Designers who...etc.",
    "jargon": [
        {
            "typography": "Text styles"
        },
        {
            "mockup": "A computer generated..."
        }
    ]
}

Tools:

Making this role as an AI Agent I’ll need the agent to have access to the following tools:

• Search engine keyword research – Search volume & difficulty data, keyword generation
• Trend data – Could be social media or search, is this idea growing or falling in the zeitgeist
• Domain prospecting – Generate domain ideas, find available (unregistered, or auctioned) domains

Happily I already have code for a lot of these, and can in-fact wire up NicheReport.io to generate holistic niche reports on the fly.

Tests:

This bot is 20% ideation and 80% tests. The agent will need to come up with ideas, then validate and score them, (see outputs above).

---

Make Sense?

Now that you’ve seen how I break down one of these roles, feel free to skip ahead if you don’t want to read the details of all 4 roles.

An Aside: My First Directory Website: Modified Cars

I first made a directory/collection website in 2007. Back then I was just starting out building websites for myself. My main ‘app’ had been a website called “Wheels-Near-U” (bad name in hindsight), which hoped to replace AutoTrader by being better geolocalised. 

Wheels-Near-U eventually failed, but the directory sites I’d made alongside it prospered. At one point I was travelling the world living off the monthly affiliate revenue they generated. 

At that point I’d built ~250 different sites that could roughly be called directories. But the first one I’d built was the most successful. Check out this ugly throwback:

Back then I used greyhat SEO to get this website to the top of Google for the term “Modified Cars” in the UK & US. 

That site just listed ebay auctions of modified cars, but made me a salary each month on autopilot, which went on for almost 2 years untouched.

This is a little example of how directories can work, though now the web is a far busier place.

But now, back to the role-splitting:

---

Role 2: Directory Builder

This AI agent will build the actual directory. 

There are many platforms for building directories online, but here I’ll stick to a simplistic setup which provides me flexible outcomes without loads of boilerplate. I’ll either Vibe code this initial logic myself or use an existing template.

Tasks:

• Retrieve starting data
• Register domain – point the domain at hosting, set up default email accounts, and SSL
• Branding – design logo & look and feel
• Build the directory files
• Deploy to web hosting
• Bonus: Connect to beneficial platforms – Google Search Console, Google Analytics, Stripe, URLmonitor

Inputs:

The inputs for this agent will be the outputs from the previous (Directory Prospector), i.e. ‘what to build’.

Outputs:

This AI agent will make the actual directory, so after it’s finished we should have a sparkly new website (hopefully not PetPsychologistsInUtah.com).

Tools:

• Data retrieval – Hunt down viable listings to pre-populate the directory
• Registrar/Hosting – Register the domain name, setup email, hosting, database, and SSL
• Development – AI Vibe Coder. Directory Builder will need to be able to generate the website files
• Deployment – Upload the website files to hosting
• Bonus: Access Google products, Stripe, URLMonitor

Tests:

As it goes this agent will need to check that the website loads properly, has no errors, is performant, and has sensible initial data.

Role 3: Directory Marketer

This role is all about getting the directory out into the wild internet.

It will run initially (launch), but also periodically (growth maintenance). This is perhaps the hardest role to build out as AI agents as much of what is produced needs to be high level human writing.

Tasks:

• (Initially) Submit to other directories – e.g. Product Hunt
• (Initially) Social share – X, reddit, indie hackers
• Guestpost outreach – find places to write about directory
• Setup search engine ranking tracking
• Social replies – the ability to reply to ongoing social media conversations where this directory will be useful
• Bonus: SEO writer – this may be useful ongoing marketing (never spam)
• Bonus: Cold email

Inputs:

This agent will take the inputs from the two previous agents, as well as:

• Social network searches
• Search engine ranks

Outputs:

Social posts, directory listings, guest posts, and search engine ranking results.

Tools:

• Social network posting & searching
• Submissions to directories
• Search engine results – to find places to guest post
• Legit content creation

Tests:

Tracked via search engine rankings and traffic volume (Google analytics).

Role 4: Directory Manager

The Directory Manager role will need to be checking in everyday to make sure things are operating as intended.

Tasks:

• Check website is up & fast
• Reply to any customer messages
• Report back to me

Inputs:

• Email inbox
• Social network searches

Outputs:

Replies to emails, messages, and social posts. Periodic summary messages to me.

Tools:

• WWW – website retrieval and snapshot comparison
• Email retrieval
• Social network searches

Tests:

Directory Manager will need to test the website against some predefined goals (performance, uptime). It will need to sustain/improve a conversion rate for replying to pre-purchase questions via email.

An Overview of an AI Directory Team

Putting that all out into words makes me aware of the scale of this thing! No matter, I’m going to attempt it anyway.

Here’s an overview of what’s needed:

This week I’ll be starting building this thing, so subscribe now if you haven’t already and learn along with me.

The Side Benefits of Choosing this Experiment (AI Directory) First

In last week’s post I talked about how I chose the first few experiment ideas for this AI business challenge. The key reason for picking Directories first is that a lot of the agent development needed for this simplistic business model will be reusable for nearly every other web business I might make.

By experimenting here first, I should walk away with:

1. Good overall understanding of which Workflow tools (e.g. n8n) work best.
2. Collection of AI tasks as executable code (e.g. I already have a good domain prospecting tool I made, but this’ll give me a reason to build it out to be even more AI)
3. An idea of where the limitations and blockers might be for this challenge.

The Potential Pitfalls of Splitting Up Directory Building into AI Roles

It’s hard to predict exactly where this experiment might fall over, but I am aware that an amount of online directories are purely spam. I will endeavour to make each output directory legitimately add value to end users.

That might mean that the % of AI automation is reduced. Let’s see.

Another way of looking at that is that even if these directories end up only trivially useful to others, the AI agents I build will likely be reusable for more meaningful projects later.

Scaling and Nesting an AI Directory team

This post works on splitting down the act of making one web directory into separate AI agent roles. A beautiful multiplier to all this is that it’s relatively simple to scale once, (or if),  I get something working.

So while I can have the above work through the process to create and launch one directory, I will probably aim for a collection of directories to fully test the viability here.

That may mean that I end up with another layer of AI agents, let’s call them `middle management`. Who oversee the agents in this post in their operations.

It’d be cool to see what happens if I charged these AI agents to make:

1. 10 directories based on me specifying a niche/niches
2. 10 directories based on me specifying exact ideas
3. 10 directories based purely on AI ideas

… I’ll need to see what each directory ends up costing to build/maintain, but I think that it’d be better science to do it at this scale (if my AI employees can sustain genuine quality).

Let’s Build a Team of Agents to Make Directories

So there we have it – the rough scope of the first experiment here on Profit Swarm.

By next week I’ll make an experiment page where you can check in on specific experiments, and in the meantime I’m going to go and dig in, making new AI agents for this.

I hope this has been useful or at least interesting to you, and I look forward to sharing some of the highs and lows as I build this out. Please do comment below and let me know if you like this post format/what I could improve.