By An overwhelming number of marketers are using AI tools to bolster their campaigns, their tactical output, and even their strategic ideas. I don’t actually recommend all of those (AI can seriously mislead you on strategy, and using it to write anything you’re going to publish directly is a recipe for unsubscribes), but there are some clever, high-quality marketing use-cases out there.
But… I think we can do even better.
On this week’s 5-Minute Whiteboard, I’m showing how we can use higher quality input from real data to get better, more specific, high quality output for all sorts of marketing-based use cases.
Transcript
I see y’all doing some pretty clever stuff with AI tools. And I don’t wanna stop you from doing that, but I do wanna try and bolster and make you even better at the output you’re getting. So I go to a lot of conferences and events and, you know, in my LinkedIn feed, I see people doing this. In subreddits, I see people talking about it where they’re using these prompt libraries of, hey.
Here’s a an advertisement that’s performing well for us. Give me some variance to that ad or write some high click through rate titles for my upcoming blog post, my webinar, my YouTube channel, you know, my latest YouTube video for my channel, my research white paper, may maybe this video you’re listening to. Give me some ideas for visual meme based content that I wanna put on my social channel that’s gonna help promote, you know, whatever it is that I’m doing this week. There’s nothing wrong with these.
I’ve seen huge libraries of them and people who are like, you know, what what we used to call prompt engineers who are giving optimized versions of these. But we know that the way large language models work is essentially through its pattern recognition based on the corpus, its context understanding based on, model fine tuning and input, and its its knowledge application based on sort of knowledge graphs. But the problem is that AI is gonna be constrained, constrained by specifically three things. Those three things are, one, the training corpus.
So whatever is found on the web. Essentially, all if you imagine the index of the Internet, as being a bunch of words that come after other words and what words are found on documents together, that’s gonna be a constraint.
Second, the model’s design and tuning is is a constraint. And third, the input that we give is the third constraint. You can impact these two. Like like, throw in the towel.
Don’t we can we can count on eventually, you know, the model’s getting better. They’re certainly better now than they were three, four years ago. Sometimes they regress, but mostly with they make progress. But we can impact this one by giving better input to the AI tools, asking it better questions, providing more context, providing more data.
We can improve it.
For example, this was a prompt that I found, that that folks were recommending and and works reasonably well in the world of conversion rate optimization, landing page optimization, and testing.
So using what you know about visitors to this particular website, this is balsamic dot com, which is a wireframing tool in the software b two b world.
Give me ten ideas for headlines and key messaging to create a new homepage.
I want the strongest converting page possible. And and, apparently, according to the person who posted this, test quite well. I tried it. I don’t think these are bad. Think in wireframes, not in pixels.
Okay.
I I don’t totally love it. Your ideas deserve a sketch, not a spec stock. I like this one. It resonates with me.
I might I might change up the language a little bit, but it makes sense. I can see why it’s compelling. What ChatGPT is giving me here is things that it already knows about balsamic dot com because that’s what I asked it for, and that’s coming from the web corpus. And it’s giving me ideas with the greatest probability of language matching patterns.
Right?
That output is informed by the input that I gave. So if I use this prompt that was suggested, that’s what ChatGPT is doing. But what if we could give a better prompt? What if we could tell ChatGPT more information about the audience so it could be more specific?
What if we could deliver a greater variety of topics, interests, you know, things that people search for, things that other websites that they visit.
And from that, essentially, provide the higher quality input that we know large language models benefit from. Well, it turns out I I know a guy.
Yo. This is my guy. He’s a I was born in Jersey, so I’m allowed to, I think, use this accent.
You you can by exporting all the data in SparkToro, right, from from a research report about what SparkToro knows about people who visit Balsamiq. Right? There’s there’s websites they visit and keywords they search for and topics that they’re interested in. And from that, my output, my perception of this output was that it was considerably better.
So what I did is I I took that export and then I uploaded it, and I used a slight variant on the suggested prompt. And now ChatGPT is giving me what I think I would call them better, but there’s no way to know for certain. What these are for sure is more specific to the audience. So rather than thinking, you know, not in pixels but in wireframes, which is a little abstract, it’s instead saying this is the easiest way to create wireframes, no design degree required, which speaks to engineers and non engineers because it knows from the from the SparkToro data that fifty five percent of our audience is engineering centric.
Now is that the absolute best variant you could possibly come up with? I’m not making that argument. Only conversion rate optimization testing could could actually tell us what’s gonna work. But what I know is that my perception of the output is that it is more specific, more based in real audience data, and I can get far more variation and more interesting useful variance from my perspective that are worth testing than if I just asked the AI, the inputs you know, just inputted the recommended prompt.
So I would say if you’re gonna use AI, it doesn’t have to be SparkToro. Similarweb gives you great data. Google Search Console can give you data about visitors and your audience. You could you could use some of the SEO tools to get a list of keywords. You could even ask Google itself. You could, use other tools, audience.
You could use, you know, discover tools, Google discover tools, whatever you want. The power here is in giving better input so that you get better, more actionable, more useful output.