Building company blog posts with GPT-3

Thought leadership is dead, long live thought leadership.

I’m on the AI hype train. I’ve wasted spent countless hours goofing around with the new wave of generative AI technologies, exploring Stable Diffusion’s wacky dreamscapes, and feeling generally full of awe at a wave of technological progress that is both astounding and incredibly tangible.

Somewhere in there, I got to thinking about the potential impacts of generative AI on creative workflows within companies. To explore this idea further, I turned OpenAI’s GPT-3 on a topic that I both deeply understand and eternally loathe: thought leadership blogs. I’m talking about the “How Finance Leaders Can Adapt in a Downturn” type of stuff that prioritizes SEO without managing to provide a novel perspective.

Webmasters host it, and sales teams send it, but can an AI write it? How long would it take, using GPT-3, to create a thought leadership piece that hits the sweet spot of mediocre and publishable? I thought I’d find out.

Building the foundation

First, some constraints: I’ll pretend I’m working for Snyk, a cybersecurity company. I don’t know a thing about cybersecurity or their marketing beyond what’s on the homepage, so no home team advantage here.

The blog post needs an outline, here’s roughly what I’m aiming for (messaging taken from Snyk’s homepage):

• Intro: when building a cybersecurity program, companies should focus on three areas:
• Area 1: automation
• Area 2: visibility across every cloud environment
• Area 3: faster vulnerability remediation
• CTA: try Snyk today

I’m being intentionally high-level here to give myself some wiggle room with the model’s various outputs. Now that I have an outline for the piece, I can start to construct prompts for the model.

From outline to prompt

Spoiler: OpenAI’s playground app limits the token size of your combined prompt and response, so it’s not feasible to generate a full-length blog post based on a single prompt. I’ll break up the outline to build the piece one section at a time.

OpenAI’s API guide to prompt design notes that GPT-3 benefits from plain language, and specific instructions with examples when possible. Here’s a first pass I wrote, with a focus on creating an introduction for my piece.

My input is plain text and the API’s response is highlighted in green.

This prompt uses zero-shot learning, meaning I’m providing a prompt without examples and relying on GPT-3 to understand what I want as an output. Given the vast amounts of data used for training the model, there are many tasks it can complete without much direction at all (e.g. summarization, language translation).

As far as my blog goes, this is fine! It could be more compelling but this is solidly mediocre copy for a B2B blog.

I don’t love how sales-y this sounds, so I’ll remove the company name and see if that creates a response that lends itself better to strictly “thought leadership”.

This is way better. I’m getting a more advisory tone instead of a sales pitch, plus it’s picked up my instructions to create an introduction to the blog post.

Weirdly enough, it also specified that the blog would have “three pillars,” despite that concept missing from the prompt. I have to assume the structure of intro -> three points -> conclusion is saturated enough in the model’s training data to show up here. People love threes, after all.

I’ll keep this as my unedited introduction and move on to the first body paragraph, expanding a bit on the “automation” theme to create an input that will become the paragraph’s thesis.

Again, decently mediocre copy based on a zero-shot prompt. GPT-3 doesn’t know that I’m stitching together a blog post, so the output lacks a natural transition (this is fine).

I’m finding that the real strength of the model in this use case is simply turning a single topic into more words. GPT-3 is taking my ideas about cybersecurity and expanding upon them in layman’s terms, which is sufficient for this application.

I’ll eventually mash all of these outputs together and do some editing, but this really feels like a shot to the heart for, like, the bottom 50% of all B2B writing. This feeling becomes especially pronounced when you consider that my total time investment in the model is roughly 15 minutes thus far.

To the finish line

If I actually worked for Snyk’s content team (who I’m sure do a much better job than I’m doing here), here’s how I’d proceed:

First, I’d use similar prompts to generate my two additional body paragraphs. Then, I’d add a closing CTA (this is marketing copy, after all) and take a quick editing pass for paragraph transitions, company voice, etc. I might even add a quote from an industry analyst or other thought leader, for a bit more substance.

Here’s what I actually did, where you can see the GPT-3 outputs, my edits, and the “final” piece.

In total, putting together the “final” post took me less than an hour for a topic I am totally unfamiliar with.

Of all the cognitive load involved in content writing, GPT-3 excels at what might be the heaviest part: taking outline-level ideas and expanding them into chunks of fairly coherent material that you can smash together and edit later. It’s not going to finish your thoughts with spontaneously excellent ideas about your company’s unique product, if only because training it to do so is a massive investment.

What it can do is churn through the hardest parts of content creativity faster. If you have to make content for SEO, might as well have a machine do it.

Note: I didn’t talk through any of the settings of GPT-3 (temperature, top_p, etc.) in this post. Check out the OpenAI API docs for a summary.