Can You Use AI to Write Blogs? It's Not As Easy As You Think
- Irene Schouwenaar
Artificial intelligence is supposed to take much of the work out of generating marketing copy.
So surely ChatGPT + experienced marketing writer = success. Or does it?
It’s a weird feeling when your job requires you to use the very tool the internet insists is here to replace you. As part of TPM’s ongoing AI experiments, I was tasked with using ChatGPT to create a blog post. Can you use AI to write blogs? This experiment was eye-opening.
The Experiment
The Players
Two passable-but-unoriginal blog posts and several valuable lessons
How Does ChatGPT Work?
Before we get to the head-to-head AI experiment, I should provide some context for those readers who have never attempted to generate text with ChatGPT. Getting started with AI blog writing is as simple as promised.
- Create an OpenAI account (sign up, verify your email, done)
- Ask ChatGPT to write something (this is called a prompt)
- Keep prompting until it’s produced what you want (or as close as it can get) And when I say, “keep prompting”, I really mean that.
That part was easy enough. Maybe this wouldn’t be so bad.
How to Fairly Compare Skill Sets
To kick off the AI experiment, Andrew and I set some parameters.
- Prompts: With unlimited time and the power of indefinite prompts to edit and tweak the text ChatGPT produced, I had a clear advantage as a writer. We agreed we could each use a maximum of three prompts to generate copy, and up to two additional prompts to generate metadata for our hypothetical blog post. We also decided we’d save all our prompts to compare as part of the experiment.
- Subject matter: We decided on the relatively high-level topic of content marketing metrics. As an inexperienced prompter, it seemed likely I would struggle to produce something really technical or niche with just three prompts.
- Editing: We agreed neither of us would directly edit the text. The final product in each case would be what we could prompt ChatGPT to create – no polishing or correcting.
Experienced vs. Novice: The Results
We sat down together to review the results with our boss, Dean. It was clear at first glance which prompts were Andrew’s and which were mine, and his blog post definitely got a lot of technical details right that mine fell down on. But when it came down to reviewing the copy the way we’d review something produced by a member of our writing team, the results were very similar (in some cases, too similar.)
Can AI write blog posts? The jury is still out. But we walked away from the whole adventure with two relatively mediocre blog posts and four vital lessons about using ChatGPT to produce marketing blog posts.
Lesson 1: Practice with AI Makes Perfect (Sort Of)
Andrew’s experience creating prompts was immediately evident when we compared results. Let’s look at the specific instructions we each included in our initial prompts:
Andrew’s Prompts
- Topic
- Target audience (much more specific)
- CTA
- Keywords for SEO
- Guidance on language/tone
- Guidance on the headline
- Instructions about specific things not to include
- Request for link placement suggestions
- Request for metadata with specific instructions about URL, character limits and what not to include
Irene’s Prompts
- Topic
- Word count
- Target audience (admittedly, a pretty broad one)
- CTA
- Request for metadata without specifics
As a result, the post Andrew created included
- Useful metadata (mine invented a URL and name for our company)
- An engaging headline
- Keyword-optimized content
- Clear, informative subheads.
Mine did not.
I assumed if that many people are using ChatGPT to create this kind of thing, it would probably know rules like character limits and stuff like that. It does not. Lesson learned.
Lesson 2: The Value of Experience with Marketing Copy
While Andrew’s post was clearly better than mine in terms of technical components, the differences in the actual text of our pieces were a lot smaller. My introduction was definitely more engaging and personalized, but his conclusion bested mine with a strong CTA and compelling copy. The body between those two was weirdly similar.
It feels like an existing blog already. Like, somebody already wrote a blog on this topic and it just parsed it together. Both are OK pieces in and of themselves, but I don’t know if we published this… there could be a blog that looks exactly like it out there.
But our respective backgrounds were apparent in the text.
Andrew’s Edits
With a background in digital tools, Andrew directed ChatGPT to provide guidance on what tools readers could use to measure various KPIs cited in the post
Irene’s Edits
My post clearly drew on my understanding of marketing copy best practices — I had instructed ChatGPT to expand on why each point should matter to the reader
I think it’s interesting that here, as a writer, Irene understands the sort of things a reader would want to know and how to engage them, whereas I, with my background in digital, went a totally different direction. I was trying to add validity to the content while you wanted to boost the quality of the content.
Lesson 3: AI Tools Still Have a Lot of Limitations
The speed at which ChatGPT can produce a blog post that I’d normally spend at least a few hours on is impressive, but this AI experiment highlighted some of the limitations a novice user wouldn’t necessarily be aware of.
- It can’t count. The reason Andrew didn’t specify a word count was because he knew the tool would ignore it.
- It can only produce a maximum of 1200 words at a time. If you try to force ChatGPT to create something longer than that, your text will be cut off mid-sentence at the 1200-word mark. (Andrew successfully created a 1500-word piece by prompting it to expand certain sections of text without repeating bits like the intro or conclusion.)
- It lacks originality. The 10 content marketing metrics described in each of our blog posts were nearly identical, and the only variations in their descriptions came down to differences in our prompts.
- It can’t handle links. ChatGPT is still unable to read current sites and draw information from them, but it also fell down when Andrew asked it to recommend anchor text and topics for cross-linking to other posts on related topics.
- It doesn’t know marketing rules. If you already know how frequently your keywords should appear in your text, how many characters your meta description should be, or how to break up copy with subheads, this is not important. If you’re counting on AI tools to provide content that follows accepted guidelines, you’re in trouble.
The big point here is that it’s vital to get someone who understands AI but also marketing and writing to do this. ChatGPT is like a friend who’s somewhat one-dimensional; sometimes they listen to you and sometimes they do not.
Lesson 4: Writing With AI Requires Some Particular Skills
The overarching theme of this AI experiment was that using ChatGPT to write marketing content is a lot harder than it looks. Even with Andrew’s AI expertise and my writing experience, neither of us was able to produce a blog post that didn’t require significant editing.
Leveraging the right people within your agency is the only way to get top-tier quality. Just one person, unless they were experts in all aspects of content marketing, would really struggle.
At TPM, we understand that many of our clients don’t have the skills or expertise needed to make artificial intelligence work for them in a valuable way. That’s why we’ve launched our AI Tech Stack, which combines ongoing experimentation with AI marketing tools with our knowledge of best practices and years of experience. Contact us to learn how you could benefit from our AI efforts.
