Author(s): Anubhav

Originally published on Towards AI.

My First $5,000 Month Writing About AI Engineering on Medium

In May, my Medium earnings crossed $5,000 from writing about AI engineering.

A month earlier, the same account had done 10.9K views. Two months earlier, it was at 3.9K. The jump looks like an overnight success if you only look at the final revenue screenshot.

The tempting explanation is that two posts exploded and carried the entire month. That’s true, but it misses the point. If I had published the exact same viral posts on a brand new account, they would have spiked and died. Instead, they spiked and woke up fifteen other articles. Readers clicked on a post about local LLMs, finished it, looked at my profile, and clicked through three more posts on retrieval-augmented generation and agent architectures.

The Month That Changed The Account

March was quiet. I was publishing consistently and seeing small signals of life — a stray comment here, a highlighted code snippet there. The traffic was random.

By mid-April, the compounding started to become visible. I wasn’t getting viral hits, but my baseline daily views were rising. I would publish a new article and older pieces would get a secondary wave of traffic. The catalog was waking up as readers who found one piece of my work started clicking through to the others.

Entering the second week of May, the breakout happened. The recommendations stopped being random. The same kind of builder-focused reader kept showing up. When a post took off, the traffic spilled over into the rest of the catalog. I remember refreshing the dashboard on a Tuesday morning and seeing an old RAG post from March suddenly back on the daily charts, sitting right next to an article I had published twelve hours prior.

The Wrong Lesson Would Be “Write Viral Posts”

When people see a big month, they immediately try to reverse-engineer the biggest hits. Looking at my May dashboard, a few specific articles did most of the numbers.

My guide on Claude Code setup pulled in 39,000 views and nearly 20,000 reads, earning about $1,360. A breakdown of the best local LLMs for coding did 39,000 views and 18,900 reads, bringing in just over $2,000 — my single highest-earning post. A curated list of 12 AI books brought in another 28,000 views, 12,800 reads, and $714.

The obvious conclusion is that I should just write more listicles and setup guides. Outliers gave me the spike. But the spike only mattered because there were 20 other posts for those readers to land on next.

When a developer clicked on the Claude Code setup guide, they got a useful tutorial. At the bottom of that page, however, they saw links to my other work. They found deep technical dives like “RAG Chunking That Works,” “Multi-Agent Systems” and a detailed comparison of LangGraph vs Temporal.

If my profile had only contained generic AI news, that developer would have closed the tab. Because it contained dense engineering articles, they realized I was a builder. They hit the follow button, joined the email list, and clicked through to older posts.

The Niche Was Narrower Than “AI”

I drew a hard boundary around my topics. The niche was not artificial intelligence broadly, but AI engineering specifically.

If I wrote about AI in general, I would be competing with news sites and generic content farms. I’d end up writing about how AI is going to change the future of work, which no engineer wants to read. Instead, I broke my niche down into very specific lanes.

This specific mapping worked. It had enough search volume to matter, but the code snippets naturally filtered out people looking for ChatGPT prompt hacks. It was also deeply connected. If someone reads an article about RAG chunking, they are a natural fit for a piece about reranking or hybrid search.

I had to stay disciplined here. Writing a great article about Python agents today and a generic productivity piece tomorrow would just dilute the exact reader base I was trying to build.

Compounding Looked Boring Until It Was Obvious

By late April, I realized that compounding in content creation looks like absolutely nothing for weeks. It looks like small, boring movements. I would publish an article and get 50 views. I’d publish another and get 70.

Then, older posts started firing together.

The pattern repeated every time. One new post would catch attention. Readers clicked on it, finished reading, and the recommendation engine pulled up an article from three weeks ago at the bottom of the page.

The Claude Code spike directly helped my older AI coding workflow posts. The AI Books article drove traffic back to my roadmap pieces. The RAG Chunking article kept my technical RAG authority alive and fed traffic to my reranking guide. You cannot judge a technical article by its first week of traffic, because an article about Python agents might sit dead for a month until a related LangGraph piece suddenly revives it.

Technical Deep Dives Worked, But Not Alone

I used to think you had to pick: deep tutorials or broad overviews. May taught me you need both, for different reasons.

Deep technical dives proved my competence. My article on RAG Chunking was dense, filled with regex patterns for markdown parsing and detailed explanations of semantic boundaries. It didn’t go viral. It did 2,700 views and 1,100 reads. But the people who clicked it read it, holding a read ratio around 40%.

Meanwhile, my deep-dive architectural comparison of LangGraph vs Temporal did only 959 views, but it earned $119.91. That is a high earnings-per-view ratio. It monetized well because it answered a specific architectural question that senior engineers were searching for. They were trying to decide between state machines and durable execution for their agent orchestration, and they spent ten minutes reading every word of that post.

The biggest discovery came from posts with broader entry points. The Claude Code setup guide and the Local LLMs for coding post were accessible to any developer, not just AI specialists. They brought thousands of new people into the catalog who then trickled down into those dense architectural pieces.

Read Ratio Was The Metric I Started Trusting

By May 17, Claude Code was still the main engine. It had reached 27,000 views, 14,900 reads, and about $1,074 in earnings. But the Local LLM post was starting to move fast behind it. It was only eight days old and had already reached 7,200 views, 4,000 reads, and $473.

That was the moment I stopped thinking about May as a two-post month. Claude Code had created the first wave. The AI Books post had widened the audience. Now the Local LLM post was becoming the third engine.

By the end of the month, the Local LLM post had caught up in traffic. Both posts were sitting around 39,000 views. But Local LLM earned just over $2,000, while Claude Code earned about $1,360.

Same rough traffic. Very different money.

That gap changed how I read the dashboard. Views told me how many people clicked. Earnings told me whether paying Medium members stayed long enough for the article to matter.

The funnel is simple, but easy to misread. Medium first shows your title in the feed. That is a presentation. Some people click. That is a view. Some stay past thirty seconds and scroll. That becomes a read. The part that actually pays is narrower still: reading time from paying Medium members.

I started monitoring my read ratio obsessively. If an article dipped below 40 percent, I immediately tried to figure out why. Did I bury the code snippet too far down? Was the opening confusing?

I also started watching the presentation-to-view drop-off. Across the whole catalog, May did 863K presentations but only 132K views — the platform showed the work to a lot of people who didn’t click. The Reranking post was that same gap in miniature: 43K presentations, 454 views. The title was far too academic for the general feed it landed in.

The Titles Mattered More Than I Wanted To Admit

Engineers generally hate marketing. I used to believe that if the code was good and the technical explanation was accurate, the audience would magically find it. I had to drop this mindset.

I started studying which title patterns worked for technical audiences. They are very different from standard clickbait because technical readers want to know exactly what they’ll get and whether it’ll save them time.

These patterns worked consistently:

• “I spent 6 months…” (Establishes skin in the game)
• “What is the best…” (High search intent for tooling)
• “12 X worth reading…” (Curation and taste)
• “What actually happens when…” (Mechanism deep dive)
• “X vs Y…” (Architectural comparison)
• “Here’s how…” (Direct procedural instruction)

A title can get the click, but it cannot save the read. If I promised a deep architectural breakdown of LangGraph and delivered a generic summary copied from their documentation, my read ratio would have collapsed instantly.

Publications Helped More Than I Expected

I couldn’t build a $5,000 month entirely alone since I didn’t have an external following on Twitter or LinkedIn. I needed distribution.

Publishing my work in the Data Science Collective mattered. It put the article in front of the right readers on day one. When I published independently on my small account, the feed had to guess who might like the work. When I published in DSC, the editors provided initial placement and formatting that pushed my pieces directly into the feeds of people who cared about data engineering and AI architecture.

A publication won’t rescue a boring topic or a broken read ratio — but it gives a genuinely good article a fair first test instead of leaving it to the cold-start lottery.

Writing To Learn Kept Me From Burning Out

If I only wrote articles based on what I thought would get traffic, I would have burned out by April. Chasing trends is exhausting, and I would have found myself writing about tools I didn’t care about.

This strategy worked because the topics were problems I was genuinely trying to understand better for my own work. The RAG posts helped me clarify my own mental model. Writing the agent pieces was the first time I had to pin down where a ReAct loop stops and a state machine begins — I’d been hand-waving that distinction for months. The Claude Code and local LLM comparisons came directly from me trying to tune my own local development workflow.

I got two things out of one effort: I learned the tool, and I had a post. It also prevented fake authority. Readers can feel when a writer has spent four hours debugging a broken integration. When I wrote to learn, my articles naturally included the edge cases, the weird error messages, and the undocumented behaviors that proved I built the thing.

Trial And Error Was The Actual Strategy

Looking back at a successful month makes the strategy look perfectly planned. In reality, the content system was not obvious from day one. Throughout March and April, I published a lot of different formats to see what the audience wanted.

I tested different post angles, moving from deep tutorials to aggressive tool comparisons. Eventually, I started curating reading lists. Some things failed. Generic overviews of AI news got zero traction. Posts that were too short felt incomplete and had bad read ratios.

I treated my publishing history as a dataset. I looked at which posts produced clicks, which produced reads, and which converted readers into the ~1,500 new subscribers I added that month. I looked for catalog spillover. If an article got low views but ten subscribers, I knew I had found a vein of high-trust readers.

Having Drafts Ready Changed Everything

I couldn’t maintain consistency if my publishing schedule relied on daily inspiration. Waking up on a Tuesday and thinking “what should I write about today?” guarantees failure. I would have rushed a weak idea, written a bad draft, and published it just to maintain a streak.

The operational reality behind the May revenue was an inventory of drafts. I usually kept around 20 drafts in various stages of completion.

By late May, I was publishing off the queue twice a week. This meant I always had something ready to publish. If I had a busy week at my job, I just pulled a finished draft from the queue and hit publish. It also let me respond to what was working. When the Claude Code post took off, I looked at my draft queue, found a related outline outline about the eight skills every Claude Code setup needs, finished it quickly, and published it while the audience was still warm.

I treated my writing like an engineering workflow with a strict pipeline: an idea backlog, an outlining phase, an ugly drafting phase, a polish phase, and a strict SEO pass before submission.

What I Would Repeat

If I had to start over with a new account, I would follow a very specific playbook based on what worked.

• I’d publish a broad post only when I already had two deep ones behind it for the new readers to fall into.
• I would track the member read ratio obsessively and largely ignore daily view counts.
• I would keep a deep inventory of drafts so I never have to write in a panic.
• I would change titles aggressively after publishing if the initial promise proved unclear to the audience.

What I Would Not Repeat

The process was messy and I made several distinct mistakes that cost me time and energy.

I would not overvalue presentations. In March, I used to refresh my stats and get excited when the feed showed my article to ten thousand people. But if only fifty people clicked, those presentations were useless.

I would not worry when a good post starts slowly. Assuming a post failed before the platform finished testing it is a mistake.

I would not treat every technical deep dive as a broad traffic play. Some articles are meant to be read by five hundred targeted engineers. Expecting them to do 50k views just leads to disappointment.

What The $5K Month Actually Means

By the end of May, hitting $5,000 in a single month proved that technical writing in a specific niche can generate meaningful revenue. But it doesn’t mean every subsequent month will hit that exact number.

The real asset is the catalog and the audience. The account now has a backlog of technical pieces that reference each other, and a clear sense of what developers want to read. June’s non-outlier baseline is still roughly ten times what March was, simply because the catalog now exists.

If this helped you, consider clapping 👏 so others can find it too.

Continue Reading

1. I Spent 6 Months Tuning Claude Code. Here’s the Exact Setup That Finally Worked.
2. If You Had To Read Only 5 AI Papers, This Should Be It.
3. Why 90% Of Engineers Fall Behind In This AI Era? (And How To Fix It)
4. How I Would Become an AI Engineer in 2026 If I Had to Start Over
5. Building an AI Agent That Runs My Twitter Account

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI