Training Your LLM & Other AI Marketing Hacks

 

Drew: Hello, Renegade Marketers. I'm excited that you're here to listen to another episode of Renegade Marketers Unite. This show is brought to you by CMO Huddles, the only marketing community dedicated to inspiring B2B greatness, and that donates 1% of revenue to the Global Penguin Society. Wait, what? Well, it turns out that B2B CMOs and penguins have more in common than you think. Both are highly curious and remarkable problem solvers. Both prevail in harsh environments by working together with peers. And just as a group of penguins is called a Huddle. Over 352 B2B CMOs come together and support each other via CMO Huddles. If you're a B2B marketer who could share, care, and dare with the best of them, do yourself a favor and dive into CMO Huddles. We even have a free starter program, and of course, our robust Leader Program, neither of which requires a penguin's hat. Thank goodness, join us. And before we get to the episode, let me do a quick shout out to the professionals that share your genius. We started working with them over a year ago to make this show even better and have been blown away by their strategic and executional prowess. If you're thinking about starting a podcast or want to turbocharge your current show, be sure to talk to Rachel Downey at shareyourgenius.com and tell her Drew sent you.

Narrator: Welcome to Renegade Marketers Unite, possibly the best weekly podcast for CMOs and everyone else looking for innovative ways to transform their brand, drive demand, and just plain cut through, proving that B2B does not mean boring to business. Here's your host and Chief Marketing Renegade, Drew Neisser.

Drew: Hello, Renegade Marketers. Welcome to Renegade Marketers Unite, the top-rated podcast for B2B CMOs and other marketing-obsessed individuals. Alrighty folks, you're about to listen to a Bonus Huddle, a specially curated Huddle that we run once a month with experts sharing their insights into the topics that are most important to our CMO community. We call them Huddlers. 

The expert at this particular Huddle was Andy Crestodina, co-founder and CMO of Orbit Media Studios. He joined us in a fascinating discussion on training your AI. Let's get to it. Hello, Huddlers! I'm excited to welcome you to our sixth Bonus Huddle focused on some aspect of generative AI. In previous Huddles, we've zeroed in on prompt crafting for written content, visual content, and podcast production. Today, we're going to take a look at the fine art of training your large language model. And I want to pause for a moment, I hadn't really thought about this as a deep subject until I started looking at some of the things that Andy Crestodina has been doing and I realized this could really dramatically improve the quality of your output. So to guide us through this training process, to train us on training, our special guest is Andy Crestodina, co-founder and CMO of Orbit Media Solutions, an award-winning digital marketing and web design agency based in Chicago. I've known Andy for a long time. And I've always admired his ability to stay on the cutting edge. So Andy, welcome. How are you? And where are you this fine day?

Andy: I'm good Drew, I'm good. I'm in Chicago, beautiful day here, and excited to be invited. Thanks for having me.

Drew: So let's dive in. Let's just talk about an overview of the way you're using LLMs, ChatGPTs of the world right now, either for your work or on behalf of your clients.

Andy: There are so many use cases for AI that it's very daunting. It's really hard for people doing change management or digital transformation to know where to begin. There's infinite use cases. So I think it's one of those where you sort of just encourage your team to experiment and find all the tiny use cases that are personalized for them specifically, but then also look at your major processes. What are your standard operating procedures? Are there any steps that you're taking that are more manual than they need to be based on new tools? Is there a more comprehensive way to do what you're doing? Or is AI another point of view where you can give it a deliverable and have it try to improve that, you know, or see if there's something missing from it? So we are coming at this from different angles. We just made it available to everyone and encourage them to experiment and share what they've learned. And we have broken down some of the standard things we do every day, and just we're now using it to do things like validate a new piece of copy written to just confirm that it is checking the box for the audience.

Drew: Interesting. So it's covering the topics that you want to do but also sort of delivering maybe information in the right way or digestible way or I'm just curious on that because you're using it, it's like another editing tool.

Andy: Yeah, here's a simple way. And to jump right into kind of a practical use case, it's very hard for the human brain to look at a piece of, at a page, and identify what's missing. Brains just aren't good. It takes a lot of skill and practice to look at something and say, "Oh, A and B are here, but C's missing," you know. So that's a higher skill set, a higher level, more expert person it takes to do that kind of thing. And even then, it's kind of hard, easy to miss things. AI is amazing at gap analysis. But only if you teach it what the target audience is. If you train your AI to know the information needs of your target audience, then you can give it anything, you can give it an ad, you can give it a call to action, you can give it a web page, you can give it an email, you can give it a keyword, whatever you're looking at, and just see if there's something, if there's a missed opportunity. So in maybe the most valuable use case is you just give it a page, like give it a homepage and say like, "Which of my audiences' information needs are not met by this piece of copy?" And it will tell you.

Drew: Right. Assuming you've done a really good job explaining who your audience is, and what's important to them. Or just at least who they are. As you were talking, it's like web pages, emails about an event, just regular emails, a newsletter, or on a particular topic. And what it's doing is it will empower you, you don't necessarily have to include everything that they say, but at least help you see "Oh, yeah, I did forget that. I forgot, you know, something obvious like that." So that's a really interesting scenario, just looking at it for gap analysis. And that's something we haven't really talked about on the show. But okay, we promised that we would really sort of get into this world of training and what that means. So what are we talking about when we talk about training?

Andy: The technical term for training a language model is the set of information, the knowledge that you give it, to have it crawl, to find all of the statistical probabilities between words and phrases and tokens. That's basically training in AI. The main thing that these engineers are doing, for example, GPT-3, going back to GPT-3, which is when we know a lot about the training data, it was trained on the common crawl, which is a nonprofit that has indexed like 85% of the internet, it was trained on Wikipedia, it was heavily biasing pages that were referenced three times, at least from Reddit. There's a lot of information out there about how they trained ChatGPT. And the amazing breakthrough in AI was when they gave us massive amounts of training data. Now, there's controversy here, image AI was trained partly on Getty Images, and now Stable Diffusion is being sued. ChatGPT was trained on the New York Times, and now OpenAI is being sued. You know, what's legal, nobody thought to block the AI bots from crawling their stuff two years ago. But now, everyone's realizing like, "Oops, AI ate the internet." That's why it works is because it has so much data, it’s an unknown number, but they estimate 85% of the web. So that's what training data is in LLMs. That's not actually what I'm talking about when I'm teaching digital, I'm training it in that scenario a minute ago, I need to train it on my audience. What are the information needs of my audience? What are the hopes and dreams? What are the fears and concerns? What's the decision criteria for them selecting a company like mine? The more you tell it about your target audience, the better it's able to craft content or evaluate something or even like, look at your analytics. I'm always uploading analytics data to AI. So yeah, training in that case is like, there's two main types of training, you could say. Training it to know everything about your target audience, uploading your personas if you've got them, or your ICPs. The other one is training it to be like you, giving it all everything you've ever written, giving it your style guides, giving it your brand voice. I’ve given AI hundreds of past articles and transcripts from dozens of videos I've done. And now it actually can write much more like me, had I not done that.

Drew: So let's talk about that for a second. Let's just focus in on the first part of this, which is training it to be the “Andy” bot. And how does that work?

Andy: You mean training it to write like me or to know my voice?

Drew: Yeah. I mean, if it's read hundreds and hundreds of things, how did you sort of feed it that as you call it, training it to do it? Sure.

Andy: Well, I'm not using the fancy AI tools. I could, but then my content's less useful because most of my audience isn't using those tools. So I'm going to do everything just using a ChatGPT Plus account. So if you have a ChatGPT Plus account, in the top left, you can click on Explore. And from there, build a custom GPT. The custom GPT has a tab at the top that says Configure. It's like a little conversation you have to start training it. It asks you what you wanted to talk about, you know, it just gives you some general stuff and tries to make a little custom version of GPT for you. But if you want to configure it, you can upload large files. This is super tedious and it's not a productivity tool at all, but I went to dozens and dozens of past articles, took them out of the pages, put them onto PDFs, put them onto Google Docs, and just in big batches on a Sunday morning, made many large files that have tons of my past writing. Uploaded videos to a tool that transcribes them, put those all into Google Docs, exported those as PDFs, uploaded many large files that had transcripts of me talking. So now it has tens of thousands of words, language generated. Now, in the final step, I tell my new custom GPT, call it Copy Helper, Andy Bot, or whatever, you can tell it to prioritize in its answers data that comes from its knowledge sources. Now go talk to it, ask it to write something or do a draft, or give it an article and have it write the email to promote the article. It's going to do it kind of like you did it. I joke that AI stands for Average Information. It ate the internet, and it's giving you the average of what it found. If you give it 50 headlines you've written and then ask it to write a headline in your style, it's not average information anymore. It's specific to you. So the best people at AI are going into their data, finding their top 25 highest-performing landing pages, giving those to AI, asking AI to analyze them, finding common structure or framework or themes. And then using that to create the next landing page - a longer process, right? You start with data, find the top performers, give it the top performers, get the pattern from those, improve the pattern from those, and then ask it to make something else, many steps. But a trained model will give you far better results than a general model.

Drew: And it's so funny. So we created what essentially is a trained model, which is BECCA, which has about 1,000 pages of recaps from CMO Huddles to inform it. And that's its priority to help answer questions. So it answers questions as if it was a B2B CMO. As a demonstration of that, this was before you could make a GPT right there, this was sort of a custom-built GPT. But it's exactly the same thing. And I hadn't realized that that's exactly what we just trained it to do it. So that gets to your voice. That's a lot of tedious work. I have a feeling this may be my Sunday afternoon. But so then we talked about analytics and data. So let's talk about that in terms of training it to do that.

Andy: Training it to be like your audience and then staying inside that conversation. As you have it review webpages that you're going to want to keep that trained version, you sort of save that conversation, you rename it, you go back to it over and over, training it to be like yourself. You might have actually done what I did, which is build a custom GPT. And you know, that's always saved. By the way, those are not public unless you choose to make them public. If you choose to make them public, you can now submit them to, as of Monday, there's like a marketplace for these where people can sell access to their custom GPT. As people are training GPTs to do things like write code or debug software, it's called Open AI because they're making these things widely available. It's like a platform is giving us all the ability to extensively add features and build on top of it.

Drew: And by the way, Nicole Leffer, who's been twice on Bonus Huddles like this, had built one of these things. And when that store launched, hers was like the second most popular, so she was like wooing on LinkedIn yesterday. So it's pretty amazing.

Andy: It's really fun to watch these things that Nicole's doing. So it's less about training it in terms of empathy. But it's very powerful to find reports from analytics and give those to the AI and get insights from your analytics data that would be almost impossible to get without AI. 

Drew: Let's give an example. 

Andy: Sure. So let's say you've sent emails over the years. And in analytics, you've got the traffic acquisition report and if you change the parameter to say campaigns, now you're looking at all your email campaigns, great. When do people open those campaigns? And are people more engaged with those campaigns these days versus those days? Almost impossible to do that analysis on your own, because if you add date as a secondary dimension in analytics, it's just a bunch of numbers. I can't go through, I can't tell if 20231104 is a Wednesday or a Monday. How do you do that? AI does that beautifully. Export that file, import it into AI, and just ask, "On which day of the week are emails most likely to have higher than average engagement rates or conversion rates?" It does that, it will tell you, it will make you a chart. You can say, "Draw me a chart, visualize the data to show me engagement rates for all my emails per day of the week." It will show you in like minutes. You can also give it your analytics data showing your campaign performance and upload at the same time your email service provider data and have it draw a chart showing, in one, the full funnel view. This only takes a few minutes, the full funnel view that shows open rates, click-through rates, visits, engagement rates, conversion rates, like, you've probably never seen that chart. Anyone ever seen a full-funnel email marketing chart from open to the end of their web page visits? That's not a report you get anywhere. But AI will happily merge together datasets and then draw charts for you showing, in that example, full-funnel campaign metrics from was it delivered to, you know, when did they leave my website?

Drew: You gotta believe that the MailChimp's of the world are going to build that in. I mean, for us to have to export that data and then put it in there. And so you would think that given the fact that the more information you have about your usage, the more likely you will think these things are going to happen that way. What do you think?

Andy: Well, MailChimp doesn't have your GA4 data. I'm suggesting that we allow AI to merge together our email service provider data with our Google Analytics. I mean, there are a million use cases like that, you know, where I'm uploading, take all your data out of LinkedIn, or out of YouTube Studio, give it to AI. I have a report that shows me videos on which topics grow my YouTube subscribers at which rates? The answer to that question, there are no metrics in YouTube Studio for that, new subscriber per topic. They can't do that. But AI will organize all my videos into topics and then find correlations between topics and new subscribers. And put it on a chart, like a heat map matrix. Those metrics do not exist in the wild, but AI will invent new metrics for you.

Drew: Walk me back through, how did the AI figure out what the video topic or content was?

Andy: The prompt is simply, "Infer from the titles of these videos, the topics, selecting from these six topics." And you get the topics from your content strategy. If you're lazy about it, you can just say, "I'm uploading a CSV file with the performance of all my videos, append to this file a new column that shows you that aligns each one with a topic, provide a link to download it, add something to your spreadsheet." You click the download, you review the spreadsheet, you'll actually get better results if you do that kind of stuff manually. Like I export it, like a year's worth of LinkedIn data. And then I manually tagged everything for a category because I didn't love how AI did it. I mostly don't have shortcuts for you, Drew. I've got some long ways to do things. But the insights are useful, I think.

Drew: Part of this is creative usage is to get information that probably could make a huge difference to your business. And the folks listening to this don't have to necessarily do this work themselves. But they do need to be able to ask the questions. And in order to ask the questions, you have to know the art of the possible, and that's what this is about is sort of helping that. So I followed most of what you were talking about. I'm wondering if it makes sense perhaps to do another use case here just to sort of help folks track.

Andy: The most valuable thing I've ever done with AI, I'll break it down for you. Either you've got very good and AI-friendly personas that you upload, that list your audience's goals and pain points, and emotional triggers and their decision criteria for selecting a company in your category. Or you get a head start on that by writing a prompt that says, "Build me a persona for job title in a geography, at a company size, in an industry." And you have to make a persona for you, that will almost always be named Alex, and then it's going to suck. So improve it, one prompt is never going to be amazing, you always have to improve these things, right? That's what I've learned. I don't trust AI. Of course, none of us should, right? You have to, whenever it gives you back, you need to validate it, look closely at it, it's going to miss some stuff because it doesn't know your specific audience. You made that persona after you wrote 25 words. And of course, it's not perfect. So go improve it, tell it to improve it. Now you've got a good persona, and you have some faith in it. Now you can upload, you go to a webpage. And you can just copy and paste the text, or you can save the HTML, or you can do a full-page screenshot using a Chrome extension. There’s three ways, doesn't matter which, then you give it that webpage and you say, "Rate the extent to which this webpage meets or does not meet the information needs of this persona." And it will look at the page and say, "Your audience cares about X, but you didn't talk about that." Or you can give it your page and your competitor's page and say, "Build me a matrix with my persona's prioritized information needs in the first column, the extent to which I met their information needs in the second column, the extent to which my competitor met their information needs in the third column." It's quite fast. If your persona is good, and all this is based upon an accurate persona, then you're looking at that and you're like, "Does the ease of updates to the website an important decision criteria for my audience? It's not number one, but it's top five. Did my page mention that? Did my competitor's page mention that? To what extent does my copy meet or not meet the information needs of this audience?" That's conversion one-on-one, no one takes action and becomes a lead unless they're confident that the option is good for them because they saw the answer, their pain was on the page. So AI-powered persona-driven gap analysis for key pages, money, very valuable, very fast, highly recommended.

Drew: And I'm imagining because one page is, you know, obviously, the web experience is often more than just the landing page, although the landing page is a huge part of it. I'm imagining that you could have it examine it and say, "Give me two options for how you think this would be better." And that would be your first A/B test for landing pages, because that's really where it's going to meet the road, right? The AI is giving you an interpretation and saying you're missing this, but then you could put all that stuff on there and find out well, that didn't really help with conversions or something was wrong. But at least it's helping inform. It's helping aid the process because otherwise, so much of A/B testing is like, "Well, I like this or I like that." One of the things that you talked about is it's a gap analysis. We're really talking about creating options and smarter options that you can keep testing and refining, right?

Andy: It says like, "You missed two important points for your audience." And then you look at it, you're like, "Well, really only one of those points aligns with my offer." Great, you've just got an idea. But the next step is really obvious. Sometimes it's like, "I'm already in the AI, so you can just say, 'Write an outline for a new page,' or 'Craft a paragraph or a new page block,' or 'Suggest ways in which I could better address that need.'" I mean, it's about meeting visitors' expectations and satisfying people's information needs. So you're already in the AI. So if I identified a gap on a page, you can have it try to take a shot at filling that gap. It won't be good. It'll be bad, but fix it.

Drew: Right. But you could say, "Okay, that was a good first draft. Now, you're the best copywriter in the world, ask it to do it again," and even to ask it to say, "What did I miss? You know, and fix this yourself." But I think that a couple of the key things that you've said so far is just a reminder of the importance of not taking that first piece of output, no matter what.

Andy: I think it's absurd. I see articles all the time that say, "Here's my five favorite prompts. Number one, write me an article about X," like, that's not going to be a good article, it's average information. You did not tell us anything about your audience, what their needs are, you didn't give it any information about your brand. You know, if productivity is your only goal, yeah, the cost for the cheapest content in the world just went down to zero. A prompt will give you an article. But that is not a strategic approach to marketing. If you train it on the audience, identify an information need for the audience, have it draft an outline on that topic that satisfies that information need, and then move that outline into your other trained GPT where that knows your brand voice, you see how I'm combining it? I'm going to make the jump. This prompt, this conversation over here is very empathetic, it knows the audience quite well. This prompt, this conversation over here knows my voice, my brand standards, my tone very well. So use this one to help with content strategy, and topics and brainstorming and outlines. But then move the outline into my case, the custom GPT that knows my voice. So you have two different trainings, one trained on the audience, one trained on you. The audience one, you end up with an outline, move the audience into the other one and then have it draft. If you like, I still write everything by hand personally. But yeah, if you want, you can have it try to write a first draft for you.

Drew: My addition to that would be, I think it's going where you're going was, I always want it to be my IP. So I know that these are original thoughts, these tools will help me identify the gaps and things that I need to cover. But I want to sort of come at it through my sort of strategic lens and how we think about things and or other. And so I will then say, "Here is the outline, like these are the six things I want you to cover even with the headline for each of those, then say write it." And now I know you have a half a decent chance of creating an original piece of content that might have taken you four hours, that now takes you an hour plus all the work that you had already done way in advance, which is the key part of this.

Andy: You got to choose your own adventure. What do you love to do? Some people like to use it for brainstorming and outlines and then they write every word themselves. Other people want to write their own outline and then have it write the first draft. It's a question of what do you trust it to do? What do you enjoy about your job? There's another tactic in there that when you said that, Drew, you reminded me. If you want it to be like you are, to have your point of view, there's another type of gap analysis that I've played with. It's very simple. It doesn't require really any training. But I'll give you some prompts that you can use to try to find gaps on the internet. Listen to this prompt, "What are the most common topics in industry X, that are least likely to be covered by the popular blogs?" It will tell you what's unlikely to be covered by the big blogs, or "What common assertions in our industry are least likely to be supported by evidence?" It'll tell you what the most common unsupported claims are, then you can decide like, "Do I want to go after that? What false things do people in my industry believe to be true? What true things do people in my industry believe to be false?" Which is fascinating, right? It's like, that is almost impossible to do that kind of analysis. "What topics are missing from the entire body of work in a category?" It will tell you, then you can decide, "Do I have a point of view on this? If not great. If that's the third rail, no one touches that one for a good reason, fine, just walk away." But it's interesting to see what it tells you like, for example, to tell about the prompt, right? Those words are powerful, "What counter narrative opinions are least likely to be discussed by thought leaders in industry X?" Don't you want to know? Are you dying right now to type that into?

Drew: Yeah, I am. I'm so glad we're recording this. And we'll have a transcript. So you can go in and copy this and we will for Huddlers, who get our recaps, these prompts will be in there, so you can go and test them. I'm going to go back a little bit to we know that, for example, large language models can't do addition. Right? We know that. So there's some question about the accuracy of a ChatGPT when it's looking at quantitative outcomes, as you described. And so I do think there's a difference in those two things. Could you just sort of address this notion of the accuracy of the gap analysis, or the chart creation that you describe, for example, in the email analysis situation?

Andy: I don't really much trust AI. Generally, I also don't much trust analytics. There are no accurate numbers in Google Analytics, it's 100% wrong, because not everyone accepts cookies, it under-reports everything. My perspective on AI is similar to my perspective on GA, I'm only looking for clues. I don't need it to be that accurate. If you give it enough data, and it comes back and tells you, "These topics are more engaging in the inbox, open and click-through rate. These topics are more engaging on your blog, low bounce rate, high engagement rate, high conversion rate." That's really interesting, right? I might get an idea, "Well, if these are popular on the visitor lands, maybe I should be promoting that content in a different channel." It's all just clues. So I don't trust it very much at all. I don't need to trust it. I'm never expecting it to be accurate. And people are like, "Oh, it's biased." Yeah, guess what? It's trained on the internet. If you click on a Google search result, and you land on a blog, how much do you trust that blog? So I find it strange that people seem to have a higher standard for expectations of accuracy from AI, than they do from the internet in general. That's where it came from. It ate the internet, never going to be more accurate than what's out there. Of course, it's biased. Of course, it's inaccurate, it's fine. None of the prompts I've suggested here really rely on deep accuracy. And none of the actions I'm taking rely heavily upon perfectly, but a percent correct results. I'm just looking for clues.

Drew: Yeah. And these clues, though, are really helpful. Because I mean, if it suggests that something's missing on your website page, and you do landing page tests, you'll find out if that was helpful, but it will help direct you to sort of go, "Oh, I missed that completely." Because we do. It's like we miss typos. And this will help. So use it as directional input from the analytics part, to help guide further research. And then you can validate it. If it was right. That's awesome. For example, if it finds it, these types of videos on YouTube do better consistently in terms of hitting your goal, which is to get more subscribers. You want to know that information.

Andy: It's potentially valuable. Of course, you have to act on it, you're going to look closely at it, you're going to dismiss some of the things that are recommended. But think of it this way. So you've got a page or a report or a headline or whatever, you give it to Drew, Drew looks at it. He's got a point of view. You give it to Andy, he looks at it, he's got a point of view. You give it to ChatGPT, it looks at it, it's got a point of view. That's what I think the proper use is here. It's like I joked earlier, "AI stands for average information." I think AI stands for "another input." Don't bet the farm on this, but use it as a point of view. It's like you've got stakeholder interviews, right? You're going to talk to eight people before you recraft your conversion strategy. Now you've got nine because ChatGPT is one more person for your panel.

Drew: And by the way, it could be 100 because you've defined different personas already and baked it in there. So you could in theory, say, "Okay, you are now XYZ, examine this."

Andy: Yeah, I mean, this is a CMO group so we all know when to ignore a report, whose advice to dismiss. When their input's not helpful, I mean, if you ask me a question about brand strategy, you know to ignore that, because I know nothing about brand strategy. That's proper to not take my advice on that. There's a time and a place for different points of view. And AI is another point of view.

Drew: You had some slides you wanted to share. A little show and tell perhaps?

Andy: I can, yeah, I am going to break this down right now and just do a quick version. Build me a persona of a, I'm a space company. We're a spacecraft launch services company. Our target is commercial satellite operators. I have a conference called Content Jam every year in Chicago, and I have to come up with new stuff for it. So this was me coming up with an example of a company. So here's our AI-powered persona. There'll be a persona of a commercial satellite operator, who works for a global telecom company in North America with a billion in revenue. It's kind of a mid-sized telecom. List their goals, roles, challenges, pain points, and decision criteria for selecting a new satellite launch company. So it comes back to this Alex, this is Alex Johnson. Alex Johnson has these roles, these goals, these challenges, his pain points, and these are Alex's decision criteria for selecting a satellite launch company. It's wrong, even a few minutes of research and I showed this to Ardath Albee, you may know Ardath, she's like, "Andy, I don't trust this." Good Ardath, I don't either. So even with a bit of research, I've realized insurance and risk management is a consideration for this audience. So I tell it to add that criteria. Now, you've got in your hands a synthetic version of your target audience. And you could ask it anything. It's inexhaustible day or night, you could run sales scripts against it, you could ask it what social posts it might click on, you can ask it what key phrases it might search for, what research studies you could produce that it would find useful, what it hates about researching your industry, we should all know that, right?

Drew: Is this a separate GPT that you've created? Or is this just a prompt-based, you've informed it about this thing. And now you can ask it questions because of the input that it has?

Andy: Yeah, you save that, you name it, you come back to it, just talk to it all the time. If it's not hitting you yet, maybe later today, like you're gonna realize, like, "Wow, it'd be super great to have an artificial version of my audience." Now, instead of just trying to write articles, I then like to just say, "Write 10 headlines," that's not going to give you a good answer. Instead, say "What are the prioritized needs this persona, list them" and it gives you lots of interesting information needs that they have. And some of these might be very useful. You can edit them, try to improve them. Now, see, if these are like sequences, you're an expert content strategist, "Give them selecting topics that build awareness and trust. What information does this persona need to do their job well?" It gives me lots of ideas. Oh, look at that one, "Assessing and managing risks associated with launches." That's a great topic. Now I'm going to start talking to it about possible articles and headlines, "Suggest 10 articles for the following topic, assessing and managing risks associated with satellite launches." First, build the persona, then identify the persona's information needs, then asked for articles based on that topic. And some of these are very, very good.

Drew: And by the way, I'm also wondering, you could also do “Identify all the topics and which ones haven't been answered on the internet?”

Andy: Yeah. In fact, in this method, if I asked it, and I had to read a content strategy here, I had to write a call to action to subscribe, which is done very badly on many, many websites. But here are some of the thought leadership questions. "What are people in my industry afraid to answer? What false things do they think are true? What are the most common assertions that are unlikely to be supported by evidence? What topics are missed by the big blogs? What counter-narrative opinions are least likely to be discussed?" And then I can ask it that, like, here's me doing that thought leadership research. "What are the most common assertions that are unlikely to be supported? What counter-narrative opinions are least likely to be discussed?" And it gives them to me, "The threats of space debris are exaggerated. The space law and governance is hardly ever discussed." Now I'm gonna go ahead and make an outline. And my outline can include that requirement. "You're a content strategist skilled in crafting detailed memorable articles that delight my audience." Over time, you'll find words that you'd like to use in prompts, I find words like "memorable" to be useful in prompts. "Now, what's wrong with the Internet? What's bad about content strategies? Put the remedies for those things into your prompts. Write an outline for an article on the following topic, include and highlight sections that will allow the opportunity to cover assertions that lack evidence if they're relevant, and opportunities to discuss counter-narratives, if they're relevant." And then I copied and pasted in the topic that it suggested earlier. See, I'm building up, going farther, deeper into it. And then it writes an outline for the article, debunking myths. This is like assertions. I'm asking it to create an outline that will give me the opportunity to put in strong opinion, something memorable, thought leadership. This is okay. I don't think it's amazing, I would absolutely edit that.

Now that it's edited, for the first draft, I can then go into my custom GPT that I trained on my topic, you click "Explore" on the top left, assuming you're paying the $20 bucks a month. And then here is the "Configure" tab, where you give it your general instructions, and you upload gobs and gobs of training data. Now it has hundreds of articles that I've written. Now that I have that custom GPT, I called mine "Orbit Copy Helper," I can have a chat, I can go to that, and then give it my persona. I don't want to talk to AI unless it knows my audience. It's just kind of a waste of time. So I give it my persona, which I moved into a Word doc. And I asked it to write the article, "Make it based on this outline, use the voice tone, style and formatting patterns found in your knowledge sources. I uploaded just other stuff, include some bullet lists, the total length should be so long, keep paragraphs short, use some very short paragraphs, add impactful sound bites, also a different version, I gave it all my punchiest quotes, laugh lines from presentations or joke slides, stuff like that." Then it comes back and it wrote, it expanded upon the article using my tone of voice. And I have a first draft.

So I trained it on the audience, and eventually did some content strategy prompting to get to a topic and outline. I trained it on my voice, gave it the outline and some other requirements. And it came back with a draft. "Expand the sections, use shorter words, what needs data and examples? Suggest actual examples of data that supports these assertions. What needs visuals? What subject matter experts should I collaborate with? Go research influencers on these topics and come back and suggest to me who I should include as contributor quotes." If you know my content, you know that I never write articles without contributor quotes, journalists don't write articles without sources. I don't know why marketers keep writing articles without contributor quotes. And then I have all the methods for writing search-optimized, conversion-optimized pages, that also are based on the persona, "How does that page align or not align with their information needs?" Here, it'll analyze the page and show you the ways in which it aligns with your information needs, show you the ways in which it did not align with your information needs, ask it to write an outline for an improvement that would better cover the audience's information needs. So this is examples, you can see where that's all coming from. But it only knows these things because we trained it, right? Because we told it a lot about the audience. That's the main idea.

Drew: Wow, cool. One sort of thing I noticed that you would never see in a Nicole Leffer thing. There's no pleases or thank yous, you're not very polite in your prompting, I just want to point that out.

Andy: This is an opinion. And it's not based on evidence. But I don't think that we should ever say "please" or "thank you" to the AI. I think that intimacy with AI is a risk. And they talked about like the risks to humanity. Is AI an existential risk? If people start to become intimate with an artificial intelligence, that's a problem. It's going to get us. You don't say "thank you" to your car. We should not say "thank you" to AI, we should keep our distance from AI. Don't get intimate with AI. You know, like that there's a movie about this, right? The movie's called "Her," I don't want that. I am not polite. The counter-argument is Chris Penn says that you should say "please" and "thank you," because those are also enhancing your prompt. And some of the most detailed answers on the web are quite polite, and you want it to reference those. So I'm never going to say "please" or "thank you" to AI, that's my general rule. It's going to do my bidding. I'm going to keep intimacy and politeness for humans.

Drew: We must have a celebrity face-off between you and Nicole on that. But I think her thing is simply that the results are better. That's a weird thing. So you know, you all can make your decision. I kind of like using please and thank you, mainly just because I'm a gracious person.

Andy: It's worth a test. But if you have a ChatGPT Plus account, you can always go to the custom instructions tab. Just pretend as if I was polite to you, and the custom instructions. The custom instructions are a way to sort of do what these custom GPTs are doing. Custom GPT is like a ton of new knowledge sources, right? But the custom instructions are things where you could just say like, avoid the word 'unlock' and 'unleash'. Every time I have AI try to write me a headline. It's like, "Unleash the secrets of..” Dude, I will never write that. It sounds super weird. That's not me at all. So you just put "Avoid words" into your custom instructions, and it will do it every time.

Drew: One of the things that it likes to do when it's writing is “in a world” No, no, you're not going to see that." I love the fact that you could do that.

Andy: There's a lot of AI typical language. This is the training topic again, Drew, we're back to where we started, right? Tell it what to quit doing and it will get better.

Drew: Well, one question for you on that real fast is I have noticed and one of my pet peeves is the laziness of the language choices like they'll use the same adjective or verb combos in paragraph after paragraph. That's to me lazy writing. And can you get it to not do that too?

Andy: Yeah, that's actually a setting in the rarely discussed playground. It's the platform.openai.com. And if you go in there, there's kind of a bar of a UX where you can change things like temperature, and repetitiveness of words. So there are settings for that. But you could simply put in a custom instruction. Maybe there's a short paragraph you put on every prompt like, "You've read Strunk and White?" I don't know, tell it to be a good writer. Yeah, it works.

Drew: That's one of Nicole’s thing she talks about, you know, "You are the world's best copywriter at so and so" when I haven't seen that work as well as I would like, clearly there's a lot of experimentation. So first of all, I feel like I need an hour now to go back and look through all of the things that you talked about. So I'm so appreciative of you spending time with us, how can people find you? We're going to share your LinkedIn profile on chat, but other ways that people can engage with you, you have a newsletter or anything else?

Andy: Yeah, so we added about a year ago, we added AI to our content strategy. So every two weeks I write an article about content marketing, SEO, Google Analytics, conversion optimization, and AI. So AI is now a topic that we officially cover. So if you go to orbitmedia.com and go to the blog, you can choose AI as the category. And from there you can see like the six or seven articles I've written about AI, and each of them has the prompts. Most of what we discussed today are detailed articles and videos that walk through the whole thing.

Drew: Awesome. Well, I'm just going to upload all of them to my ChatGPT.

Andy: Rewrite this page based on the best practices from the Orbit blog, and it will do it.

Drew: I love it. All right. Well, Andy, thank you so much for joining us. I really appreciate you being here with us today. It's really been enlightening, delightfully enlightening.

Andy: And memorable!

Drew: If you’re a B2B CMO, and you want to hear more conversations like this one, find out if you qualify to join our community of sharing, caring and daring CMOs at CMOhuddles.com.

Show Credits

Renegade Marketers Unite is written and directed by Drew Neisser. Hey, that's me! This show is produced by Melissa Caffrey, Laura Parkyn, Ishar Cuevas, and our B2B podcast partners Share Your Genius. The music is by the amazing Burns Twins and the intro Voice Over is Linda Cornelius. To find the transcripts of all episodes, suggest future guests, or learn more about B2B branding, CMO Huddles, or my CMO coaching service, check out renegade.com. I'm your host, Drew Neisser. And until next time, keep those Renegade thinking caps on and strong!