In March 2023, I talked with Ben Fitzpatrick about how we were using AI for content marketing and data journalism at Fractl. ChatGPT’s API had just launched, and most of our work ran on GPT-3. A lot has changed since then, but most of the core ideas held up better than the tools did.

This recap covers what we did with AI in 2023, why it still works, and how AI agents now handle work I didn’t expect them to manage for years. Originally presented on the Strategy Led Podcast, hosted by Ben Fitzpatrick.

How We Used AI for Content Marketing in 2023

Fractl creates data journalism for brands, then pitches it to major publishers through digital PR to earn media coverage and links. By 2023, I had been experimenting with generative AI for a few years, starting around GPT-2. We were testing it at almost every step of our process:

Ideation Generating topics, subtopics, data sources, and campaign ideas to start each brainstorm.	Data work Cleaning, analyzing, and augmenting the datasets behind each story.
Technical SEO Automating research tasks like keyword grouping and SERP analysis.	Outreach Building pitch lists and writing and revising pitches.

None of it replaced our team. AI was a research and brainstorming assistant inside a process that people ran.

Why One Prompt Gives You Generic Content

Most people use ChatGPT one prompt at a time. The problem is that a single prompt returns the most common, most generic answers first. To get past them, you’ve got to widen the range of possible ideas, then narrow it down in stages.

That’s prompt chaining: each prompt’s output becomes the input for the next. For data journalism ideation, a chain might look like this:

1. Find sources. List every dataset or data source related to the topic.
2. Filter. Pick the sources most relevant to the readers you want to reach.
3. Generate. Use those sources to propose story ideas.
4. Refine. Extend, critique, and improve the strongest ideas.

In 2023, I ran these chains in Python so many prompts could run at once. Cost mattered: each stage multiplied the number of API calls, and a three-stage chain could easily pass a thousand requests. I go deeper on modern versions of this in AI in Content Marketing: How To Build Pipelines That Move Past Single-Prompt Thinking.

Feed AI Your Own Data

The second big shift was injecting outside data into prompts. That can be a third-party API, your own data table, or the output of an earlier prompt. For example, you can pull a thousand Reddit headlines and run a series of questions against all of them.

Outside data also protects you from made-up answers. When a model suggests datasets, you can check them against Google Scholar or search results to confirm they exist. Or you can reverse the process: give the model a verified list and ask it to choose the best fits for a topic.

Use Embeddings To Map Topics at Scale

One of the scripts I shared on LinkedIn used embeddings, a natural language processing technique that turns text into numbers that capture how similar items are. Machine learning clustering then groups the most similar items together.

Clustering has worked for years. What it couldn’t do was tell you what each cluster was about, which took a person reading every group. GPT-3 could label each cluster automatically, at scale.

Two ways to use it for content strategy:

Keyword research Export every keyword in your space from a tool like Ahrefs. Then embed the keywords, cluster them, and label each group to map topics and content gaps against competitors.

Community research Pull the top 5,000 posts from a subreddit, cluster and label them, and keep the upvotes and comments to see which topics people engage with most.

Treat AI as an Assistant, Not an Author

We didn’t use pure AI content generation for any client deliverable. AI filled out our daily ideation briefs with topics, data sources, and sample ideas, and our creative team built from there. Sometimes an idea was great out of the box. More often, it sparked a better one.

We also used it to draft survey questions, then had it evaluate which questions were most likely to produce a newsworthy result for a given group of readers.

You need to be an expert in your field to use AI well. Expertise is what lets you spot the few great outputs among the many average ones.

How Original Data Became the Content Moat

Generative AI changes the economics of information. Anything a model can produce on its own becomes cheap, so content built from aggregated information loses value fast.

What gains value is content AI can’t make on its own: video and other human content, person-to-person communication, and complex data journalism. A data story requires collecting the data, cleaning it, analyzing it, finding what matters, and telling it well with strong visualizations. In 2023, that was far beyond any model, and it’s why we leaned harder into data, not away from it.

How a Police Data Project Became a Nonprofit

The project that best showed me how far original data can travel didn’t involve AI at all. My county in Florida, like most in the U.S., has a public portal for police data: arrests, traffic stops, and details about each incident. But it only lets you pull one date at a time, which makes trend analysis nearly impossible.

We wrote a scraper to collect about 10 years of data, and I wrote up what I found. In May 2020, I posted about the project on Reddit, where it reached the front page of the Privacy subreddit. A follow-up post invited people to help build an open-source portal for local police data, and about 2,500 people joined the Slack channel I set up.

A core group of volunteers took it from there and organized themselves. With help from a pro bono law firm, the group became the Police Data Accessibility Project, a nonprofit that later received a quarter-million-dollar grant and hired staff. Today it’s a Pittsburgh-based nonprofit tracking more than 2,600 data sources.

What’s Changed Since 2023

Some of what I said in 2023 aged well, and some was overtaken fast:

I was wrong about the timeline for data journalism, and I’m glad. AI agents now let any brand operate like a data journalism outlet, producing original investigations that used to take weeks. The parts that held up matter most. Break work into steps, ground AI in verified data, keep people in the loop, and invest in what AI can’t copy.

Where AI Fits in a Content Marketing Program Today

Beyond data journalism, artificial intelligence now supports almost every part of digital marketing. These are the most common uses across content operations:

Ideation and research Map topics, questions, and data sources before anyone writes a word.	Content creation Use AI chatbots to outline and draft, with writers owning the facts, angle, and final copy.
Content optimization Compare drafts against search intent and competitors, then tighten structure for search engines and AI answers.	Content personalization Use customer data and audience segments to tailor email marketing, subject lines, and landing pages, with clear data privacy rules.
Repurposing Turn a study into social media posts, email copy, and video scripts, with a person checking brand voice and brand guidelines.	Performance analysis Pull patterns from engagement metrics and conversion rates, and use predictive analytics to choose what to make next.
Workflow automation Connect marketing automation and social media management tools so routine steps run on their own.

AI adoption is no longer the question. The teams that stand out pair GenAI tools and AI assistants with original data and human judgment. Fractl Agents packages several of these workflows.

Build Your AI Content Marketing Around Original Data

The tools I described in 2023 look primitive now, but the approach still works. Give AI clear steps, feed it verified data, and keep experts in charge of what gets published. Fractl uses that approach to create data journalism that earns coverage across search, social, and AI platforms.

Reach out to talk about what your data could become.

FAQ