The Story of Chad GuPTa: How we use Generative AI in our content strategy

Read more of your interests, read less insurance paperwork
Ditch the paperwork with Plum's group health insurance
Click here to know how Plum helps 3000+ HRs ➝

Hello. This is Ganapathi Ramanathan, and I’m part of Plum’s content marketing function. 

This is the story of how we used Generative AI to write our blogs, resulting in over 100k impressions, 1000 keywords ranking, and a 10% traffic share. 

It all began a spring morning in March when Shreyas DM-ed us with a challenge. 

The brief intrigued me. 

At this point, we all knew AI was capable of generating content. But was it competent enough to generate value for a marketing organisation? Could it generate content that evokes ​​expertise, authoritativeness, and trustworthiness – at scale? Would it be useful for our audience? 

We’re a curious function. So we decided to F around and find out. 

A quick huddle later, we had a plan. 

Selecting Topics

The first step was to find out what topics we could rank for. And that’s where Siddharth, our Digital Growth Lead came in.

This is what our criteria looked like: 

  • High volume, so that we were writing things people were actually searching for 
  • Low keyword difficulty, so that we could own niche content 
  • Insurance-related topics, so that – this is self-explanatory, really. 
  • Trending HR topics, so that our content was relevant to our target audience 

Writing 

This was the challenging bit. Most of writing happens in our heads. In my case, most of it’s chaotic and inarticulate, finding form as I transcribe them onto my notes. How on earth was I going to transfer all that context into an algorithm susceptible to hallucination and inaccuracies?

Our first prompt was terrible, and the resulting article was not worthy of being published anywhere, including this case study.
But we learned with our prompts. 

Even before we instructed ChatGPT to write, we realised we had to teach ChatGPT how to write.
If you’re experimenting with AI-generated content, I recommend that you instruct it on the aesthetics of good writing before anything else.

Some best practices: 

  • Hemingway’s rules of writing
  • An inclination towards active voice 
  • Simple English (unless you want your blogs to be written like it’s Lord of The Rings, which might not get you on page one, but will earn the admiration of this writer) 
  • A clear articulation of your content style – feed it your content guidelines if you have them ready

Most of our prompts involved talking to ChatGPT like it was a slightly overdeveloped chimpanzee, offering it the context it needed to write engaging content. 

We improved our prompting with every iteration, but this was the first series of prompts that worked reasonably well for us.

I could show you our final prompts, but I’ve been told not to give our secrets away. Also, the fun’s in the journey, not the destination. :) 

Remember, writing is thinking. ChatGPT is just your conduit – its content is only as articulate as your thoughts. I recommend you go to ChatGPT with complete clarity on what you want from it. 

Editing 

While we were pleasantly surprised by the content generated – some of it offered ideas we would have spent hours researching – the language was still limited. 

There were instances of subpar readability, keyword stuffing, poorly phrased subheadings, and mild hallucinations. We ran all AI-generated blogs through a round of editing, with Yoast’s help. 

There were two rounds of edits: Human: where we skimmed for errors, inconsistencies, and brand, and SEO: where we took Yoast’s help to optimise the article for readability and relevance. 

Editing took us around 10-15 minutes per article. If it took us longer, it meant the prompt was poor. In these cases, we discarded the article entirely and returned to the drawing board.  

Publishing

Over the next two weeks, we’d go on to write 91 blogs with ChatGPT. But before we did, we needed a name. 

We didn’t want to pass off AI-generated content as human written, so we came up with an alias that would communicate the process that went behind writing the said blog. 

And thus, Chad GuPTa was born. 

Results

Three months later, we regrouped to analyse results. I just wanted answers to the three questions I had when we started this project in March. 

But was it competent enough to generate value for a marketing organization?
Yes – the 91 blogs we’d written had resulted in over 100k impressions, driving approximately 10% of the total traffic to the Plum blog/website. 
Could it generate content that evokes ​​expertise, authoritativeness, and trustworthiness – at scale? 

Yes – we were ranking for >1000 new keywords on Google. 
Would it be useful for our audience? 

Turns out it was – Chad’s top blogs were driving a CTR of 2-3%, which meant readers found them useful enough to perform another action on the blog. 

But also, people were noticing. When was the last time someone took a screenshot of an author bio to post on Twitter?  

What’s Next?

We’re bullish on AI, and use tools (Midjourney, ChatGPT, etc) extensively in a lot of our functional and creative efforts. While we love our creative process, we use these tools to speed up everything we do.
If you’d like to know more about marketing at Plum, read Shreyas’ thread below.

A not-so-short BTS thread on how we do marketing at Plum

Our latest Hackathon themed around Generative AI has thrown up multiple possibilities on where we could go from here. Stay tuned for more updates on how AI’s making life easier for us in the marketing team 

5 things Shreyas saw at insur⚡️wreck that has him more bullish on Gen AI than ever before

That said, we’re also aware of the limitations. 

We won’t be using ChatGPT to create our thought leadership pieces like The People Success Playbook or The State of Employee Benefits, until we’re sure it can mimic the originality and voice of our content team. There’s a reason this article you read is written by this writer (100% human, unless the vaccine has a microchip). 

However, if our experiments with AI result in you not telling the difference… well, that’ll be interesting, wouldn’t it?