As many folks have observed, ChatGPT has pretty remarkable potential for changing the way many of us work — particularly those of us in marketing. For those who might not be up to speed on ChatGPT, it’s an AI chatbot that can generate pretty convincing content. Here’s an example.
When I first saw ChatGPT back in December, I couldn’t wait to give it a spin with my actual work — which, believe it or not, does not involve writing blog posts about kale smoothies using expressions lifted from the 1960s.
The Prompt
And so I started experimenting with having ChatGPT write basic Facebook ad copy. I simply told ChatGPT to write me some copy that emphasized certain value props of our software, and asked it to keep it within Facebook’s recommended 125 character-limit. For privacy purposes, I’m not going to use the specific ad copy itself, but I’ll stick with my hypothetical smoothie company.
Within about five seconds, ChatGPT spat back pretty plausible ad copy. Not amazing, and pretty generic, but considering that a real human could only start typing a sentence in five seconds, it’s not terrible.
But remember: the copy is generic, and if this were actual work produced by a human employee, I’d be disappointed. So I gave it some feedback.
Okay, these are a bit better. Again, not great, and not particularly clever or funny, but from an input-output perspective, it’s not bad. A few seconds of writing a prompt yielding five (or more!) bits of copy isn’t a bad work ratio. And in this instance, I might use 2-3 of these in real-world testing. Now, I could have gone further and asked ChatGPT to refine the copy even more, but for the sake of time, I’ll hold off.
👋 So let’s talk emoji for a second. From my testing, posts with emoji work fairly well (in contrast to email subject lines, which don’t always). So I asked ChatGPT to throw some emoji into the ad copy.
I didn’t clock it, but it took about a second, maybe two, for ChatGPT to add relevant emoji to each line. Again, a real human would have only located the emoji button in the same amount of time. 🤯
So, how’d the copy do?
Remember, I’m not running this actual smoothie copy, but we have been A/B testing some GPT-generated copy over the last few weeks. And the results have been pretty good. Over the course of the 3-4 weeks of our testing, we’ve seen an approximate 30% reduction in CPA on some of these AI-written ads vs. human-written ads. Before anyone’s head explodes, remember that these things vary lot from one week (not to mention ad set to ad set, or product-to-product), and it’s too early to definitively say that AI-driven ads will drive meaningfully cheaper conversions over a longer period of time.
I am not claiming that AI copy is better — it might or not be — but AI copy can be written, and tested, much faster. In some ways, it’s less about the ends and more about the means. And gosh, those means are fast.
Quality = meh. But speed + iterative capacity is 💯
The hard thing about ChatGPT is that there are myriad dependencies — I imagine you’ll get different results if you word the query differently, or ask it to emphasize one facet of your product/service rather than another. But the killer features here are speed and the ability to iterate on feedback.
As you saw above, the copy was kinda meh. A little repetitive, and not original at all. An editor would ask the writer to “punch the copy up” before throwing it into Facebook ads.
However, the fact that ChatGPT could spit out plausible copy in less time than it takes to tie your shoe is mind-blowing. I could have asked for ten copy variations, and it would likely have taken only a fraction of a second longer than asking for five. And probably 2-3 of them would have been good enough.
Moreover, the fact that ChatGPT could iterate – emphasize characteristics, lengthen or shorten lines, add relevant emoji – in a matter of mere seconds makes it feel like a scene from Minority Report. Heck, it takes me a good five seconds to realize that I’ve left caps-lock on.
📸 What about AI-generated images?
There are a bunch of AI image generators out there, with Stable Diffusion being the most popular. I’ve tried some text-to-image generators, but I haven’t found anything that didn’t produce vaguely eerie / uncanny valley-esque images (why are there always six fingers!?) and that precluded me from using anything in the real world. But it’s only a matter of time.
AI-Written Copy + AI-Driven Placement = the Future?
In recent months, both Google and Facebook have been pushing hard on AI-driven ad placement/iteration solutions to drive trackable conversions, partially as a result of Apple’s use of ATT as a mean of kneecapping ad platforms’ tracking capabilities. Google offers Performance Max and Facebook offers Advantage+ — both of which automatically mix and match copy, creative, and placement across their various channels (YouTube, Instagram, etc). We’ve been using both Performance Max and Advantage+ with some success.
In my test above, I essentially layered one AI solution – ChatGPT – on top of another AI solution – Facebook’s Performance Max – which I used to place the copy within ad sets. This obviously warped some of the test results (since we didn’t go through and add the new copy to dozens of legacy ad sets), but it also was a function of what I mentioned above: deploying + distributing copy at scale, and at maximum speed.
And to me, this kind of feels like the future.
I think that, in digital ads platforms, a chunk of the iterative and technical work will eventually get winnowed away, and platforms can leverage their datasets, computing power, and large language models to deploy copy or creative that performs. Now, I’m not saying better in a branded or artistic sense. It’s still hard to see that. But for some folks, performance might be good enough, even if the copy/creative isn’t particularly unique or clever. I don’t see brands using ChatGPT to write a billboard tagline, but performance marketing will, I suspect, continue to become more automated.
To wit: I sense that ad platforms like Google and Facebook may, in the near-future, offer some kind of licensed or homegrown version of ChatGPT within their dashboards. You might not need to even write the prompt. Since your site is obviously linked, with Google and Facebook’s pixels embedded in your code, and capturing (some) visitor data, they can customize messaging on offsite channels based on what landing page folks have seen. In fact, you might need only click a button (or not even that!) to spin up a dozen new ads, optimized for your ICPs, that will get delivered on the right channel, at the right time.
I’m loathe to make predictions, especially as the might pertain to people’s jobs, so I’m not going to do that. The arc of human history and technology is long and looping and complex, and I’d rather focus on improving the quality of my own work.
But for many of us, it’s worth adding ChatGPT into our tech stack, as one more tool to brandish. It’s another productivity multiplier to keep in a marketer’s back pocket, alongside things like a bookmark to Merriam-Webster, a list of Google Sheets shortcuts, and the knowledge that control-command-space is the right way to trigger the emoji picker. 😉
Silicon Spatula 