Thank you for registering to download Building the Data-Driven, AI-Powered Customer Journey Map. Here are the session materials:

Video/Screen Recording

Can’t see anything? Watch it on YouTube here.

Audio

Listen to the audio here:

Audio Player

Download the MP3 audio here.

[linklist]

Machine-Generated Transcript

What follows is an AI-generated transcript. The transcript may contain errors and is not a substitute for watching the video.

All right, good morning, everyone! Welcome to Inbound! This is the deep dive session. We’re off to a fabulous start with the internet—uh—this is “Building the Data-Driven, AI-Powered Customer Journey”. This is a deep dive session, so I would encourage you, to the extent that your device and your—and the Wi-Fi stands up to it—to follow along because we’re going to be going through the process of building one of these using generative AI. My name is Christopher Penn. I’m the co-founder, chief data scientist of Trust Insights. We’re an AI consulting firm. Show of hands, how many people feel reasonably comfortable with generative AI tools like ChatGPT? Okay, good. So, so we don’t have to explain what AI is. Good. Um, never know. Sometimes this happens. Um, how many of you are currently using generative AI for customer journey mapping? Okay, three people. Great! Now you’re in the right session, then. Um, how many of you feel like—how many of you have done things like generate working code with generative AI? Okay, a few more. Great!

All right. So, today, what we’re going to do is we’re going to talk about building the data-driven customer journey. What does this mean? Uh, first thing, data-driven means that you make decisions with data. So, not with “this is how it’s always done it”, not with “I’m guessing”, not with “this is what the boss said to do”. You make decisions with data in terms of the customer journey itself. The customer journey that we’re going to use today is a pretty stock standard one, and it looks like this. By the way, I am recording what’s on my screen because there’s—we’re going to go through stuff pretty quickly. If you would like a copy of my screen recording at the end, just leave your business card on the stage, and we’ll just take a broom later.

The stages of the customer journey are:
– Awareness
– Consideration
– Evaluation
– Purchase
– Ownership
– Loyalty
– Evangelism

What this means is: awareness is people know that they’re a problem and there’s probably a solution, but they would like to figure out what. Consideration is the—is the stakeholder said to the exec, to the intern, “Hey, intern, make me a short list of companies that solve this problem.” Evaluation is then people going through that shortlist and saying, “Does this company solve my problem?” So, if you are that company, B2B or B2C, you want to be in the evaluation phase. Purchase is self-explanatory. Ownership is the data that you get about people’s experience owning your product or service, things like your customer service inbox, call centers, complaints, etc. Loyalty is getting more money out of people once they’ve—they are customers. And evangelism is when they are so delighted by you that they had decided they want to be part of your—your unpaid marketing team.

So, these are the seven stages. We want to find data for each of these seven stages and assemble it into a customer journey. Now, to start, we’re going to start with an AI model. I’m going to be using Google’s AI Studio. It’s free. It’s aistudio.google.com. You can use ChatGPT. You can use Anthropic. You can use regular Google Gemini. It doesn’t matter which one you use. They’re all very capable. We’re going to be taking a lot of screenshots today, and I’ll explain why in a little while. But if you want to follow along, now is the time to get your device out, hopefully, the internet holds up, and have a prompt window start open. Generative AI is, when you boil it down, really, is a word prediction machine. It’s very sophisticated, a lot of really cool math, but at the end of the day, it’s a word prediction machine. So, we want to give these things lots of words. The more relevant, specific words that you give to these things, the better they perform. If I said, “Make me a customer journey map,” it’s going to give me garbage. If, on the other hand, I give it a boatload of data, I give it all my data, I give it screenshots, I give it information, LinkedIn profiles, you name it, it will do a much better job of putting together anything we ask it. So, the more data you bring to the party, the better these tools will perform.

We’re going to start by priming the model. And what this means is we’re going to ask it what it knows because priming is the easiest way to get a lot of relevant words without you having to write a book yourself. You can, if you want to. So, our first prompt is going to be, “What do you know about building—about best practices for building data-driven customer journey maps?” This process is going to have the Gemini model tell us what it thinks it knows: “Define clear objectives, gather data from multiple sources,” etc., etc. This is a good start. However, this is insufficient because this is still pretty generic. So, the next thing we want to ask the model to continue the priming process, to load up its knowledge, is to say, “What are some common mistakes less experienced marketers make in building data-driven customer journey maps?” So, this is prompt two. Now what we’re doing is we’re basically trying to force the model to think a little more deeply than the generic. Here we have it saying things, “Mapping a generic customer moments, lack of clear objectives, failing to iterate. Avoid these common mistakes.” Third prompt, to continue pushing the model’s boundaries and asking it for more and better information, “What are some commonly believed things about customer journey maps that are believed to be true but are actually false?” So, here, we’re looking for those mistakes. These are—we’re looking for more mistakes that people like us make. “One journey map fits all customers. Building a journey map is a quick and easy process.” Those are all true. Our next follow-up is, “What are some things that are believed to be false but are actually true?” Fourth prompt in the series. “Surprising truths about customer journey maps. They can be too granular and overwhelming. Journey maps can be subjective and biased.” And so on and so forth. All this seems reasonable as it’s foaming at the mouth.

The last prompt in this priming sequence is, “What are some expert tips and tricks about building customer journey maps that we have not discussed yet?” Asking AI models what you haven’t talked about yet is finding the blind spots in you and in the models. Now, this first step, this priming step, remember that in generative AI, when you’re working with a model, every word that is in the conversation becomes a part of the next prompt. So, everything in the conversation history is now part of the prompt. So, we’ve gone from a short sentence as our first prompt to about 3,000 words of knowledge about this. Now, the most important thing for you to do at this point would be to read through this and correct it, to say, “That’s not actually true. That’s—” I’ve run this several times, and it, in general, it’s—it’s good enough for demonstration purposes. But for today, for when you do this back at the office, you will want to read through it and make sure that it’s not lying. Models lie. It’s called hallucination, but it’s called lying because they have three imperatives: harmless, helpful, and truthful. Model makers want these tools to not tell you how to build pipe bombs, so they’re harmless first. Second, they try to be helpful. And third, they try to be truthful. It’s kind of ironic that truthfulness is the last on them. Thanks, AI. Um, what this means is that they will try to answer the question no matter what, even if they have to make things up. And that’s partly because, if you think of these things like libraries conceptually, when you’re prompting, you’re talking to the librarian. And that, in turn, means that if you ask the librarian for a book they don’t have, they’ll try and get you the next closest book on the shelf. You may say, “Can you find me The Joy of Cooking?” And the librarian walks off, comes back, and says, “Here’s The Joy of Sex.” You’re like, “Whoa, that is not what I asked for, but it’s close.” It’s not true. Um, so we want to prove—we want to check the work of the machines.

Next, we want to tell the machine what our customer journey is. So, your customer journey at your company may be different than mine. I listed mine off—the seven steps. Yours might be different. For example, some B2B companies, your customer journey is like 28 steps, “Proposal one, proposal two, opportunity, opportunity nurturing,” and so on and so forth. To accommodate this, I recommend taking a tool. This is an example of one called Otter. They have a booth down on the floor, but any transcription service will do. They don’t want to hear that. Um, and spend some time in the voice memos app on your phone and, literally, say, “Here is what our customer journey is.” So, you would talk through your customer journey and—and what phases there are. So, in my transcript, I recorded on the way in this morning, I just foamed at the mouth and said, “Here’s all the things I think are true about our customer journey.” You want to do this yourself with any voice memo app, any transcription app. What I’m going to do is I’m going to take that and say, “Let me tell you what I think the customer journey is,” and I’m going to attach my transcript. Maybe there—there we go. Now, me foaming at the mouth on the drive in worked out to about 600 more words, but now this is tuned to my customer journey. So, you write your own, and you can—yeah, if you want to actually physically write it, that’s fine. I just find it’s easier to talk, and it—it’s going to go through and say, “Here’s what you think your customer journey is.” So, now I’ve given it the framework for customer journey mapping by giving it my stages and what I think each stage means. You want to do the same thing. Here’s what awareness means to you, here’s what consideration means to you, here’s what opportunity 3A means to you in your pipeline. Load your customer journey into the tool.

The next thing we want to do is we want to tell this thing we’re going to build a customer journey map. “Let’s get started building a customer journey map with our data. So, first, I’m going to tell you what we do for marketing, so you understand what is and is not available.” And again, I took 10 minutes while I was doing rehearsals this morning to talk through everything that my company does for marketing. I am not going to subject you to listening to that. That’s not fun. I mean, it might be fun, but probably not. So, I’ve instead loaded this transcript. And what you should already be seeing is that we’re bringing a lot of data to the party. We are not just hoping the machine understands what we do. Instead, we’re saying, “Here’s what we do. Here’s what I do, how I market your marketing.” You need to do these interviews if you want this to work really, really well. Get your transcription app of choice, get a case of beer, talk to your sales team and say, “Tell me what you’re seeing and hearing from prospects. Tell me what your roadblocks are. Tell me why people aren’t buying.” And take those transcripts, and you’ll feed them all in here. When it comes to choosing an AI model to do these things, a tool you want to work with tool generally speaking that has the biggest conversational memory. The technical term is context window or context length. A tool like ChatGPT can remember about 90,000 words, give or take, the size of a business book. That’s good for smaller companies. Uh, Claude, from Anthropic, can remember about 150,000 words. And this version of Google Gemini can remember a million and a half words. So, which is two copies of Shakespeare’s works. So, this is the one I prefer to use. But again, they’re all roughly capable.

So, we’ve now got what we do for marketing. We’ve now got what a customer journey map is. We’re ready to have this tool start ingesting data. What you want to do is go through each section of your marketing with all the tools that you have and start getting data out of them. You don’t necessarily need to say, “This is an awareness thing.” In fact, it’s probably better if you don’t. Instead, what you want to say is, “Here’s the data. Where does this fit in?” So, let’s go ahead and do this. I’m going to start with Google Search Console. Is everyone familiar with Google—is anyone not familiar with Google Search Console? Okay, a couple of people. This is a free service from Google that tells you how Google sees your website. Um, I’m going to—so, for context, I’m going to be using my personal website today, and I’m going to be using my data because I don’t care if you see it or not. But there’s—there’ll be some parts of this process I can’t do because we’d be showing other people’s info, and that would be bad. I’m going to go into search results here. I’m going to go down to queries. Queries is how people find your website, the words and phrases they use to find your website. Um, for that first stage, the awareness, I’m actually going to choose impressions rather than clicks because I want to know, in general, what are the words that Google thinks I’m relevant for. We’re going to take this, and we’re going to do a very simple screenshot. Now, why am I doing a screenshot? Well, these tools are very good at processing image data and plain text. They can take things like CSVs, but particularly ChatGPT goes really off the rails and starts trying to write code when you give it a CSV file. And that can get very, very, very frustrating. So, you can try it. Try all the data you have. I—for today’s demo, we’re just going to use screenshots. I’m going to go here and take about two pages worth of screenshots.

And now I’m going to go back to here and say, “Let’s start loading data. These are the words and phrases that Google thinks I am relevant for in Google Search Console. By the way, um, capitalization and grammar do not matter. Spelling does matter because of the way models work. If you misspell things, they will not perform as well. Here are the top 50-ish or so words and phrases that I get impressions for in Google search. Take a look at the data and then provide your analysis of it and where you think in the customer journey this data belongs.” Oops. And now I’ll copy that real quick. I’m going to upload my two screenshots. There’s one. There’s two. Gemini, ChatGPT, and non-Anthropic Claude all are multimodal, which means that they can accept image data. It says, “Hey, here’s the your data. Christopher Penn. We’ve got brand awareness, content relevance, AI marketing focus, mapping the customer journey stages. We can tentatively map them to different stages of the customer journey. Search terms, customer state, consideration, evaluation, information,” and so on and so forth. Now, let’s go ahead and see what we can do with this data. Evaluation, etc. So, already, just using search data, I’m starting to populate this tool’s knowledge of my customer journey map. So, if you’re following along, you would want to use your Google Search Console data now to start populating this in your customer journey. And I want to emphasize, everyone’s is going to be different. It should be different. If it’s not different, something’s gone horribly wrong.

Our next channel—actually, you know what? Let’s stick here for a second because we might want to know what pages of mine are getting good—good search traction. So, I’m going to click on pages. And I’ve got to get rid of that lighter fluid blog post. Let’s go ahead and take another screenshot of—oops, I want impressions. There we go. What are the pages that are getting great impressions? One and two. We’ll go back, say, “Next. Here are the top pages in Google Search Console by impressions. As before, take a look at the data, provide your analysis of it, and where you think in the customer journey this belongs.” We’re going to let the model decide. So, here are my screenshots. I’m guessing the Wi-Fi went on vacation, maybe. We’ll try again. Let’s see. Yep, the Wi-Fi went on vacation. Oh, wait, there it is. So, now top pages, “There’s a wide variety of content diversity. The lighter fluid—” I’ve got to get rid of that blog post.

As you’re going through this, you’re going to get recommendations for—for what the models see. Again, remember, by doing that first part where we prompted for five pages of text about customer journey mapping and data analysis, the tool now knows very clearly what this conversation is about. So, its recommendations are going to come up with things like, “Hey, is this actually what you want to be doing or not?” So, some of these pages, it’s saying, “Yeah, you, um, your content’s a little all over the place,” which is absolutely true. Next, we’re going to go—we’re going to leave behind Google Search Console. I’m going to go to my YouTube channel. Same exact thing. Take a screenshot. I’m going to say, “Here is data from my YouTube channel.” Same exact thing. “Take a look at the data.” Oops. Wrong. “Take a look at the data and provide your analysis of it.” If you don’t have a YouTube channel, get one. And I—I say that somewhat facetiously, but YouTube is the second largest search engine on the planet besides Google. It is also the search engine—is also the data source that every AI company is training their data on, whether they have permission to or not. Which means that if you want your brand to be known in AI tools, you better have content on YouTube and a lot of it with closed captions so that AI models can understand it and—and, essentially, scrape it. In OpenAI’s tool, they have scraped so many of my videos that when you use their transcription service and it sounds somewhat like me, it automatically puts my name in the transcripts. Which is great for me, awkward for a whole bunch of other people who are like, “Who is Christopher Penn, and why is he in my transcript?”

All right, “An internal error has occurred.” That is fantastic. Let’s try that again there. Yeah, that’s the YouTube one. Try again. All right, Gemini, the internet appears to not be having a very good day today. So, like a good cooking show, let’s switch over here real quick. Like magic, here’s a pre-baked turkey. Like magic, here’s a pre-baked turkey that has, um, all the information. So, you—the next step would be to load, for example, my email subscription data. How are people subscribing? I want to take every single source that I can possibly get and load it in. So, this one, it said, “Here’s what we see for email subscriptions.” Let’s see if this is feeling better. Yet there we go. Um, let’s leave behind YouTube while it thinks. Next is Google Analytics. Who loves Google Analytics? Put your hands down. Nobody loves Google Analytics anymore. It used to be good. Now it’s a pain in the ass. However, it is still one of the better data sources that we have to know our audience and to know what’s going on. And there’s a couple of things—there’s a couple of reports in here that are actually useful. So, we’re going to go into reports. The first report we’re going to go to is, we want to go to traffic acquisition. Set yourself a reasonable time frame. I tend to like 90 days. Well, set yourself a time frame that is twice what your reporting interval is. So, if you have to do a monthly report on your marketing, choose two months. If you do a quarterly report on your marketing, two to six months. Basically, you want a period of time for a report and then a comparison period in the past. Now, I’m—I’m going to use 90 days just because there’s a decent amount of data. And again, no surprise here. I’m going to take a screenshot of my traffic acquisition. How did I get traffic in the last 90 days?

Okay, “Next. Here is some traffic acquisition data from Google Analytics 4 for my website showing how I earned traffic—traffic by channel in the last 90 days. As always, take a look at the data, provide your analysis of it, and tell me how it fits in.” Prompt list fail. That’s fine. Let’s see if it will accept that. Okay, “Organic search is the primary traffic driver source. Email, awareness, consideration.” Good. So, it’s providing an analysis of my traffic data from Google Analytics. This is—this is decent. This is telling us how people came to the website from a marketing channel perspective. We also want user acquisition. And the—the difference between the two is, essentially, last touch versus first touch. Your traffic acquisition is last touch, how they got to you most recently. First touch is the—the user acquisition. Same exact exercise. Take the screenshot and go back here. “Here is my first touch user acquisition data from Google Analytics 4. By channel in the last 90 days as well—channel showing the first touch for users. Take a look at the data, provide your analysis of it, tell me where in the customer journey this belongs.” While it’s thinking about that, you go and do that thing.

The third thing we want to look at in here is, we want to look at user attributes. We specifically want to look at demographic details. And the demographics I actually care about are interests. This data comes from Google Ads, which they also have a booth down on the floor. I don’t know if I’m supposed to be plugging sponsors or not, but whatever. What this tells you is the general interests of what people who are on your website also have, the other things they would want to know about. This interest data, by the way, is super helpful for you if you’re doing content marketing because you can pretty clearly see, here’s the clusters of information, the things that people care about. So, we have avid investors, shutterbugs, travel buffs. Even up here, if I was speaking to people who were solely on my website, I would want to use examples like travel because clearly that is of interest to my audience. So, you want to do this with your data as well. Let’s go ahead and take a screenshot of this. I don’t know where this would fit in the customer journey. It might or might not, but I’m going to provide it.

“Here is the interest and affinity data from Google Analytics 4 for my website showing what customers who visit my site are also interested in. Take a look at the data, then provide your analysis of it and where you think, in the customer journey, this belongs.” You could, if you wanted to, load up all of your Google Analytics data. You shouldn’t, but you could. But if there are things that are relevant to you in your Google Analytics data that you think are really important, like maybe top pages, maybe there are certain sections of your website, like the “About Us” or the “Services” page, whatever, you could load all of that into the model and have it processed. Keep going. And this is true of every tool that you have. So, if you’re a HubSpot customer, you can take screenshots or even exports from your HubSpot CRM or your HubSpot marketing and say, “Here’s a list of the top 50 companies that have visited my website as shown in HubSpot.”

So, we’re going to leave Google Analytics behind. Let’s take my email data. So, this is from Substack. Again, if you’re using HubSpot or any marketing automation software, you can take this data. I’m going to take two things. I’m going to take a screenshot of my traffic. I’m going to take a screenshot of the top sources of my traffic and for my website—for my email list. “Next, here is subscription and email newsletter subscriber performance data from Substack. My email newsletter. This shows a graph of engagements over time, plus the channels by which subscribers interacted with my newsletter. Take a look at the data. Tell me what you think. Tell me where you think this belongs.” I’m using newsletter subscriptions as my conversion, as the thing of value. Whatever your company’s thing of value is, maybe it’s booked appointments, maybe it is requested a demo, maybe it is someone came into the showroom, maybe it’s an e-commerce swipe. Whatever data you have that represents the conversion you care about, you want to have that available and feed it to the model to see what it has to say about it. Let’s see. “Analysis. Email is the primary traffic source.” I would hope so. It’s an email list. “Substack app usage. Limited social media impact.” Okay.

Well, let’s talk about social media. Where does this fit in? We’ve done YouTube. We have stuff like—oh my goodness, go away. We have LinkedIn. I use LinkedIn. Now, if you are not on LinkedIn, not active on LinkedIn, that’s fine. Use the social media channel of your choice. You could use it if you’re active on Instagram or Facebook. You could use it on Facebook or on whatever we’re calling Twitter these days. You would use that. I’m going to take a bunch of screenshots here. I’m going to take my broad LinkedIn performance. I’m going to take my content performance for the last 90 days. Let’s grab that. Let’s grab my new followers for the last 90 days. Remember, back in the old days, we used to say in social media, “Stop counting followers. Followers is a bad number. It’s a vanity metric.” That’s kind of true, kind of not true because if your followers is zero, you’re not doing a good job with social media marketing. It’s one of those things that you should include. Where we went wrong was people “Followers is the objective.” It’s not, but it’s like the one orange—lonely orange sitting off to the side in the cereal commercial, and they say, “Part of a nutritious breakfast.” Followers is part of your marketing mix. It is not the whole thing, but it’s part of it.

We’re going to take the followers, and then we’re going to take the job titles of those followers as well. That’s helpful. Let’s look at my last 90 days for my newsletter, which I also cross-post on LinkedIn because I go where the people are. Let’s put that, grab that, and let’s look at subscriber demographics. Who are these people? I don’t want to show the actual people. Let’s go to who’s viewed my profile. That’s not super helpful, but maybe it’s worth something. How many times have I shown up in search? Job titles you’ve found for? Attendant? God, people are weird. “Next, here is my social media performance data from LinkedIn. I cross-post my newsletter on LinkedIn as well as post daily on it. This data includes information about my profile performance, my content performance, and my followers. Minor tip, Google’s AI studio can accommodate 20 documents at a time. If you’re looking for more information, you can take a whole bunch all at once and put them in if the internet holds up.” You’d want to repeat this for Facebook, for Twitter, Instagram, wherever it is that you are being active on social media. If you use a social media management tool like Agorapulse, for example, you can just export all of this at once from there. You don’t have to do each channel individually. I do use that, but I’m using my personal profile for today’s example, which means that I can’t use it because it only works with business profiles.

All right, let’s see what the model has to say about my social media capabilities or lack thereof. Close all these other tabs. “Significant profile views. Viewers from companies. Declining article engagement.” That’s sad. “Awareness, consideration, decision,” so on and so forth. Oh, I forgot. I forgot my referrals from Substack. So, one of the things that you can have is people can recommend your newsletter. So, I’m going to take my leaderboard here, and I’ll feed that in as well. “Here are my referrals and recommendations from Substack. This represents the number of people who are recommending my newsletter to others. Take a look at the data. Provide your analysis of it. Tell me where it fits in. And go.” This feels pretty disorganized, right? There’s just—we’ve just been throwing data out. “This model thing, figure out what to do.” And that’s the beauty of using generative AI tools to do customer journey mapping is you don’t have to sit there and think about the sequence out. Like, how do I fit all this data? And instead, you say, “AI model, you have like 255 PhDs. You have a PhD in everything. You can figure this stuff out way better than I can.” Think of these tools as the world’s smartest interns. These are, literally, the world’s smartest interns. And if you give them really great directions, a lot of information, they will create really good outputs.

So, we’ve loaded up a whole bunch of data in here. If this was—if I was doing this commercially, I would be loading in my CRM data, my sales, my top accounts, and so on and so forth. We’re not—I’m using newsletters as an example for today, but you would load all of this data. One note on data privacy because I’m sure it has crossed some people’s minds. Here’s the golden rule. If you’re not paying, your data is the product. So, if you’re using the free tool for any of this, your data is being read by other human reviewers and probably being used to train someone else’s model. So, don’t put confidential data into a—anything free, ever. All right. We have gathered up a lot of data about my marketing. The goal of my marketing is to earn and retain subscribers to my newsletter, Almost Timely News. Based on everything we have talked about and all our analysis so far, create a consolidated, comprehensive, unabridged, verbose, complete outline of my data-driven customer journey map, following the seven-stage outline I gave you earlier. We are now at about 14,000 words of stuff.

“Christopher Penn’s Almost Timely News Data-Driven Customer Journey Map. Awareness. Individuals are becoming aware of Christopher Penn. Google search, YouTube, LinkedIn, or some organic social. Customer state, individuals seeking information about Google’s keywords, YouTube videos, website traffic metrics,” and so on and so forth. “Purchase, subscription, free subscriptions. Loyalties, developing ongoing interest, retention rate. Key insights and opportunities. Strong SEO performance, optimized email marketing, and targeted content strategy,” and so on and so forth. Now, this is a decent map. This is a decent start to a customer journey map. This is still not necessarily actionable. So, the next step might be, “Great, I need to turn this into a presentation for my stakeholders. Take the outline so far and create a slide outline with slide title, key points, key point, and supporting data for each stage of the customer journey. And show me the comprehensive presentation outline.”

This cannot create the slide deck for you. Believe it or not, creating a slide deck is extremely difficult for machines to do. And here’s why. It’s several different tasks. It is coming up with a logical narrative. It’s coming up with something that is structured and orderly. Coming up with something that is creative. So, order and logic and reasoning are actually diametrically opposed to creative because you want things that are high probability in reasoning, and you want things that are low probability in creative. For example, in a high-probability text, you might say, “The patient suffers severe gastric distress.” That is a logical, reasoned, factual statement. On the creative side, if you wanted to express that creatively, you might say, “They look like they power washed their toilet with Nutella.” Those two things are completely different in how you express them, and machines can’t do both at the same time. Also, slides are a combination plus of words plus images. Those are different models. They don’t talk to each other. An image generation model cannot read. That’s why when you ask Dolly or Midjourney for a picture of a cafe, none of the words make any sense. It looks like someone’s face rolling on their keyboard.

So, what it has done is come up with a slide presentation outline of the general journey. I like this, but it’s not great. “Revise this outline to include actual numbers. We have to show the changes in data over time or over channel so that our stakeholders understand what’s going on.” With all these tools, their first try is generally bad because it’s high probability. So, what you want to do is you want to go through these and, essentially, ask it to revise. If you want to get really clever, you might have a scorecard or a scoring rubric saying, “This is what constitutes a good report at our company.” It might be something as simple as, “Hey, we always have to have the logo or the picture of our raccoon mascot,” or whatever. If you have—if you know who the stakeholder is, you might spend some time with a voice memo app and describe, “This is what my boss likes to see. This is what my boss doesn’t like to see. These are the things that trigger my boss. These are the things that will get me fired.” And tell that to the model so that it understands. So, we have better numbers here. So, now I’ve got my data-driven customer journey map, essentially, laid out. I can go and make slides out of this. I can make slides out of this now and present it to somebody and say, “Here’s how we’re doing. Here’s the steps.” This is good. This is helpful. This took 40 minutes to do. Previously, this would take me a few hours to do because you have to do the inference yourself. Now the machines can help us do this.

However, this falls under the category of “So what?” Like, I’ve got a customer journey map. Here it is. Looks like it says, “Great, we’ve got our customer journey map. Now go back to doing things exactly the way you’ve been doing them.” That’s not helpful. So, instead, what do we do? We just spent 40 minutes giving this thing all of the knowledge we have about our company, about our marketing. I told it what we do for marketing. We can now start asking questions of this. Oh, one thing. I’m going to get this set up. I’m going to copy this whole customer journey slide presentation. I’m going to call it “Customer Journey Map”. I’m going to go to a tool called Notebook LM. This is a free Google tool. It is a research AI. What it does is it can only—I might go there. I might not. It can only return information you give it. So, regular AI, like ChatGP, you can make things up. This one, because it’s a research tool, is so heavily constrained that it will say, “Hey, you didn’t give me the answer to this question in the data you uploaded. Therefore, I can’t do anything with it.”

Let’s go ahead and start a notebook here. I’m going to upload my text file that I just created. It’s going to read through it. And what I’m going to do is I’m going to hit the Generate button. And it’s going to make me a podcast of my customer journey map, which is kind of fun. This will take a little while. So, we’re going to let it bake while it’s doing that.

So, we’ve got our customer journey map now. What do we do with it? Let’s start asking questions because these tools, once they have the data, are phenomenally good as advisors. “Great. Let’s start at the beginning of our customer journey: awareness. What three things does Christopher Penn do well in building awareness? What three things does Christopher Penn do poorly in building awareness? What three concrete recommendations would you make, based on all of our analysis, that Christopher Penn should do well in building awareness? What three things does Christopher Penn do poorly in building awareness? What three concrete recommendations would you make, based on all of our analysis, that Christopher Penn should do to increase awareness of the Almost Timely Newsletter?” So, we’re giving it very clear instructions as to what we want it to think about. We’re not saying, “Hey, am I doing a good job?” That’s not helpful. Instead, we want to give it a clear set of instructions. And it’s going to take all of our data and give it some thought, perhaps a lot of thought. You’re still working. “What Christopher Penn does well: strong SEO performance, LinkedIn profile visibility, diverse content performance. What it does poorly: low YouTube click-through rate, limited social media promotion, lack of paid advertising. Optimize YouTube for click-through rate, conduct A/B tests on video titles, thumbnails, and descriptions, expand social media strategy, go beyond LinkedIn, experiment with targeted paid advertising.”

Okay. That’s no longer a customer journey map. Now it’s a prescription. If I’ve identified that awareness is my problem, I now have some prescriptive steps to take that say, “Here’s what I should do.” Now, of these three, YouTube click-through rate—yeah, my thumbnails need a little bit of work, so I could definitely do some work there. Expanded social media promotion, I definitely—I don’t do it at all. I’m actually a terrible marketer of my own stuff. I could, for example, take my videos, stick them into Opus Clip and have it make 60-second promo videos, and then fling them all over the internet. Paid advertising, I don’t really have a budget. I don’t have a budget, so I’m going to ignore that tip. Every step of the way, I’m going to take this exact prompt, and I’m going to say, “Let’s start the next—let’s go to the next step. Next step in the customer journey: consideration.” Copy-paste, copy-paste, copy-paste. And away we go. Let’s do the consideration phase. I’m really curious. I actually did not get this far last night. One caution with models, the more data they have in the conversation, the longer it takes because it has to, again, use every word that we’ve talked about. That might be—I’ll give you another 30 seconds, and then we’ll just assume that you’ve—oh, there you go.

“Content depth and diversity, direct website traffic, strategic cross-promotion, linking YouTube, lack of clear value proposition for your newsletter.” That sucks. Yeah. “Clarify the newsletter value proposition, boost social media engagement, feature testimonials and social proof.” Clearly, I don’t do that. I clearly should, but I don’t. So, again, the next step in our customer journey, it is clearly outlined. The things that I’m doing just flat out wrong. And I should be building on top of that. Let’s see. Are you still working? So, each stage in the customer journey, you should be building the action plans. And then you can take each action plan step and get grounded—get granular and dig into it. So, for example, let’s take that Google Ads. Let’s take the action step of Google Ads. Suppose I wanted to write some Google Ads for my newsletter. “What are the general best practices for writing Google Ads for a newsletter like mine?” As we did at the beginning, we’re going to follow every step. Now, I’m not going to do all seven steps at the front because we’ll be here for quite a while. But, what do you know? What are the best practices? What are the mistakes that you made? What’s true but is actually false? What do you think is false is actually true? What are expert tips and tricks? But you would do this step with Google Ads data—Google Ads as a topic—so that you can get to a point where this thing can help you make Google Ads.

So, let’s go ahead—so, let’s go ahead and let it foam at the mouth and say, “Create a scoring rubric.” Actually, I’m going to start a new session because this is going to get really unwieldy very quickly. So—and I’m going to repeat that exact prompt, but in a fresh session. Actually, I don’t have to. I’ll just copy this whole thing. Let’s turn this to light mode so that you can actually see what’s going on. I hate light mode. “Here are best practices for Google Ads.” Put in the stuff that it just spit out. Make sure I’m on the wrong model. Choose Gemini Pro. Turn off all the safeties because I like to live dangerously. All right. Next, “Create a scoring rubric.” That is a magic phrase that AI understands that tells it you want to do something very specific. A scoring rubric is from academia. It’s like a scorecard. “Create a scoring rubric that would assess how well a Google ad is written and effective as well as aligned to best practices. Each item of the scoring rubric should be scored as a variable integer: zero to three, zero to five, zero to seven, etc. The scoring rubric.” And it’s going to create a scorecard for me. It’s going to create a scorecard based on what we know the best practices are for Google ads. Scorecard is then something we can use to create and assess Google ads quality. And again, this could be Facebook ads, Instagram ads, LinkedIn ads, your newsletter, your blog posts, your email subject lines. Anything doing creatively, you would use this method for.

Say, “Next. Your data analytics and data science for marketers with over 280,000, but. Create five Google ads candidates, headlines and descriptions. And marketing insights, headline, headline, headline, description.” Okay, cool. This looks nice. Not bad. “Score each of the candidates using our scoring rubric. Order the results in descending order.” So, they’re bad ads. Well, we have the scoring rubric. So, it’s going to go through and score all of these things. It says, “Candidate three scored highest due to its strong focus on a specific audience and a clear value proposition. Candidate one performed well due to its AI emphasis.” Great. “From the top two highest-scoring candidates, generate a refinement of each candidate, then score the refinements.” I want you to keep working. I want you to keep working until the ads you make score better. So, here’s the original, here’s the refinement. Both refinements improve their scores. “Proven AI and data strategies for growth. Scale your marketing. Free weekly newsletter, Almost Timely News. Drive measurable results. Why this ad is likely to perform well,” and so on and so forth.

So, instead of just saying, “Hey, make me stuff,” generative AI, which is a recipe for disaster, you would say, “Make me stuff. Here’s the rules. Keep trying until you do.” You could, if you wanted to, say, “Continue this process until the ads reach a score of 90 or above.” So, for every recommendation that we get from our customer journey map, we can now have it build not just the strategy, but the tactics and even the implementation. We can say, “Write these ads, write this newsletter, proofread this newsletter,” and so on and so forth. There’s one more thing we could do with our customer journey mapping if we really wanted to kick it up to—our services. These are the great new things that we’re doing, and our customer’s like, “Yeah, but I just want you to fix my problems.” So, what if we could have a model instead build—help us build our marketing according to our customer?

“Next, let’s build a customer profile of—for the Almost Timely Newsletter.” I’m going to go to LinkedIn. Let’s go into LinkedIn. Let’s look at my profile. Every profile has a “More” button. That “More” button has a “Save as PDF” button. So, you want to find 5, 10, up to 20 people who fit those profiles, that ideal customer profile. Now, did I do this right? I did not. That is where—is last week’s profiles. Yes, this will work. Oh, wow, these must—oh, well. And what we’re going to say over here is, “Are aligned with our ideal customer profile.” And let’s see how we’re doing over here. “I just gave you these. Try again. Run.” What you want to do is you want to create a customer profile. You want to create a customer—an ideal customer profile that represents real people. We don’t want something generic like, “Sarah Sales Executive” and just sort of the very bland, generic profiles. What we want is—for specific business lines or even, depending on the size of your company, a specific product. Who is that customer, and what data do you have about it?

So, here we have the data-driven executive. Data-driven executive. 40 to 55, predominantly male, located in the United States. Values data-driven decision-making. Industries like gaming, entertainment, financial tech. Into here, and now with that customer profile, we’re going to make new ads. We’re going to make better ads. Ads that are targeted towards what our ideal customer’s action—care. Gemini thinks, but what these—what these people actually represent. We now follow the same exact process. Great score. These new ads—you’re actually programming. You’re actually writing code. It’s just that you’re writing code in English or Danish or Ukrainian instead of C or Java or Python. You start thinking like a coder, thinking like a programmer, and saying, “Can I create logical instructions that go step by step, like code?” These tools can execute it like code. Things like for loops, while loops, conditions. If this, then that. You write this out in prompts like code, and these tools perform really well.

So, now we have a new ad. This one here is “Data-Driven Decisions for Executive Success. Optimize marketing investments. The AI Almost Time is Executive Intelligence.” This is a better ad. Why? Because it’s not generic. It’s now tuned towards my ideal customer. You will do the exact same thing. You will take your data you have, turn it into an ideal customer, and then email subject lines or heck, the entire email newsletter. That white paper that you’re working on. Say, “Hey, read through this white paper and tell me how well our ideal customer is going to like it or not.” Blog posts, topics, social media posts, anything you’re creating. You ask this ideal customer profile as though it was a real human being.

In fact, let’s do this. “Convert our entire conversation so far into system instructions for a large language model like Google Gemini. The purpose of the system instructions is to take input from the user about a topic and an ideal customer, both of which the user must provide, and generate Google ads for that topic designed to appeal to the ideal customer. You will get the initial input from the user and then no longer ask the user for input. After you receive the user’s input, you will autonomously, without stopping, pausing, or talking to the user, generate five ad candidates. Score them with the ad scoring rubric. Create one refinement for each of the topics. Score each of the top two winners and then score the refinements. Then produce the winning ad.” That’s a fairly long prompt. But what we’re doing is we’re writing code. And now we’re having the model write its own code for what it knows best about how to do this task. And so it’s creating these system instructions.

What do you do with this because that seems like an awful lot of words? I’m going to take this entire thing, copy it. Let’s take a look at how big this is. This is 3,600 characters, 494 words. If you are using Google Gemini, the regular version, there is a thing called Gems. Gems are Google’s version of custom GPTs. If you’ve used OpenAI’s GPTs or Claude’s artifacts, I’m going to create a new Gem. I’m going to call it “Google Ads Maker.” Paste. Save. I now have an app. I now have an app so that I don’t have to go through this whole process again. And I can give this to anyone on my team—on my marketing team—say, “Hey, now you’ve got this thing that will help you generate ads faster with our rules.” This is how you deploy AI. This is how you scale AI because you want to create apps. You might have a blog post maker. You might have an email subject line maker. Anything that you could possibly want to do, you will turn into an app. You might have the customer journey mapping maker. So, I could go back to our first session and say, “Take everything we’ve talked about, building a customer journey map, accept input from the user, and then follow these steps.” And it will do exactly as we did because it’s going to repeat the instructions that we’ve already walked through.

So, this is sort of the—I guess the capstone project for you. You’ve gone now through the entire process of making a customer journey map using generative AI and then building recommendations for every step, where you’re falling down the most, and then building strategies, tactics, execution, and content from this so that you can then activate that customer journey mapping and turn it into something useful and usable. So, we have some time left now in the deep dive. We’ve gone through and walked through a lot. How many people feel like they’re a little overwhelmed? Okay. So, what I want to do is invite you to come up to the microphones if you have questions about any part of what we’ve done—the parts we’re not confusing—I would invite you to come to the microphones and ask your questions. Otherwise, I will just tell dad jokes. Actually, no, I won’t. I will have Gemini do it. Oh, I should check in to see how our podcast did. Ah, it says our podcast did—it says our podcast was done. It might take a few minutes to load. Let’s see.

Go ahead and ask your question while I’ll see if this thing wants to load.

Audience Member 1: So having the tool segment customers, I would do that at the data collection level. So, I would say, “Here’s the segment I want to start with,” and get data about those individual segments, and then do it segment by segment rather than try to just hand it all over the machine and hope that it works out.

Christopher Penn: If you can’t do that, then—oh, by the way, if you want to copy the screen recording, you need to leave your business card on the stage for those folks who are leaving the rally—that’s where I would do the customer segmentation.

Okay, next.

Audience Member 2: Yes.

Christopher Penn: Yep.

Audience Member 2: Yes, that’s correct.

Christopher Penn: Google has not implemented sharing yet, so you would just give them a copy of the system instructions, tell them to paste this in. I would expect the Gems to be available as shareables at some point because it would be too stupid not to. Fantastic. Love it. Thank you.

Audience Member 3: Hey, Chris.

Christopher Penn: Hi.

Audience Member 3: Fabulous presentation.

Christopher Penn: Thank you.

Audience Member 3: Have you done any work bringing in data from Facebook at any point in the last few years? The modeling is very hard to add to make it more emotional for people’s lives.

Christopher Penn: Absolutely. So, again, with these tools, if you have examples or you can explain to us, say, “Here’s how I want you to do to write this ad to make it more emotional,” or things, you would ask, “How would you do this?” And then you might even go so far as to have it try writing a scale. So, let me show you an example of what that would look like. Let’s go back here and say, “This ad is good, but it’s a little dry. We want to use evocative, emotional word choices. Create five completely unhinged words—create five completely unhinged emotional variants of these ads.” Now, yes, you could bring in Facebook profile data set. Here’s what we can do with this. “You have a professional voice right now, and you also have an unhinged voice of the two. Create three variations of the ads. Use the standard professional voice for version one. Use a blend of 50% professional, 50% unhinged for version two. Use entirely unhinged for version three.” So, if there’s a particular style of writing you want, you can have the model attempt to do these different versions. And then you can say, “I want a blend of this voice and that voice.” And it will know, based on the token selection, how to write. So, you could have the different variations of this. So, that’s how I would approach that.

Audience Member 3: Thank you.

Christopher Penn: Yep.

Audience Member 4: Yes?

Christopher Penn: Yes. So, the number one thing you can do is ask the models for help, to say, “How would you rewrite this prompt to be more effective? My goal is to do this. How would you do that?” Because, remember, these things are like libraries. There’s a difference between the Boston Public Library and the New York Public Library. They both have probably about the same number of books. But if you take a librarian out of the New York Public Library and put them in the Boston Public Library and say, “Where is books on zebras?” they will have no idea. Every librarian knows their own library well. So, you would write a prompt and say, “Hey, librarian, how would I make this better?” And it will go into its own knowledge and write a prompt that is better for that particular model. Which, by the way, don’t use Chat GPD prompts in Gemini, et cetera. They don’t work as well because you’re moving librarians. So, I would start by asking the models themselves for help. “How would you make this more impactful? Here’s my goal. How would you do this?” If you want a more structured course and stuff, we have one for sale. There’s a bunch of them out there. Google’s got one for free. But the best things to do are, A, talk to the models themselves because they keep evolving. And, B, just use the tools a ton and keep saying, “Hey, you missed the boat. Try again. Here’s what you got wrong.” And it will keep doing that.

There’s a technique called contrastive prompting, where there’s like 51 different prompting styles. But the contrast of prompting is, “Here’s what I want you to do. Here’s an example of how to do it, and here’s an example of how not to do it.” And so you would prompt it and say, “Yeah, okay, that was a good try. Here’s how I want you to do it. And here’s the last thing you just did was how not to do it. So, don’t do that again.” Again, it’s like an intern. So, however you would manage an intern, do it with these.

Hi. Next.

Audience Member 5: There’s so many tools, and they’re all bad. The least bad of a lot right now is one called Gamma. It does an okay job. But honestly, at least for the creation of that visual customer journey map, you are better off just using a human. They’re not real good. Yeah.

Christopher Penn: That’s right. Yes. So, if you have the skill to do so, converting a CSV into what’s called a markdown table will do a really good job with that. Oh, the podcast is done. Let’s hear how our customer journey did.

Podcast Voice: Ever wonder how content creators figure out what makes you hit that subscribe button? Today, we’re going behind the scenes with actual data to see how people become fans. And not just any data. We’re dissecting a presentation about data analysis. So, we’re going meta, folks. Our source is an outline by Christopher Penn, the data whiz behind the Almost Timely News newsletter. We’ll uncover the surprising ways people find him and what keeps them coming back for more. Yeah, this is going to be good. And this isn’t just about one newsletter, though.

Christopher Penn: That is a podcast. That is a podcast created from our talk here today. What is in your annual report that nobody reads? Put it in here, and it will turn it into a really nice conversation. Then just hit the download switch, download the audio, and now you have a podcast for your YouTube channel. Any document that you don’t want to read, put it in here, and it will turn it into an engagement—it will turn it into an engaging conversation. You can listen to it. That will be a lot more fun. So, that’s Notebook LM.

Go ahead.

Audience Member 6: Sorry. So with their emails, is it all like individual personal emails, like Gmail address, or is it like corporate emails?

Christopher Penn: Okay. For the corporate emails, what you can do is make sure they are in HubSpot, in the HubSpot CRM. You can use the free version. And then HubSpot will allow you to export demographic and firmographic data about all the company domain names in those emails. So, you can pull out like the industry that they’re in, the company size, annual revenue. So, you can do all that from within HubSpot, export the data, and then you will have a really good—really good data set for building that customer profile.

Audience Member 6: Somewhere buried in the bowels of HubSpot, yes.

Christopher Penn: It’s easiest to get at it through the API.

Audience Member 6: Yep.

Christopher Penn: If they don’t have a business, then the easiest thing you can do is ask them. Like, “Hey, how did you hear about us?” In fact, I would have that on all your forms. “How did you hear about us?” And then you can follow up and say, “Who are you?”

Audience Member 6: Yep.

Christopher Penn: Yes. So, if you have the skill to do so—

Audience Member 7: At the end, obviously, you can help them to turn into its own—

Christopher Penn: Yes. Agents come after Gems and GPTs and stuff like that. It’s basically the next logical evolution. If you have an ideal customer profile, you might want to turn it into an actual conversational agent. Let me show you an example. If I go into Google Vertex, which is Google’s relatively costly agent system, you can take the ideal customer profile, put it to the next level, put it to the next level, and you can turn it into an actual conversational agent. I’ll show you an example. If I go into Google Vertex, which is Google’s relatively costly agent system, you can take the ideal customer profile, put it to the next level, and you can turn it into an actual conversational agent. Let me show you an example. If I go to Google Vertex, which is Google’s relatively costly agent system, you can take the ideal customer profile, put it to the next level, and you can turn it into an actual conversational agent. Turn it into an actual conversational agent. Let me show you an example. If I go into Google Vertex, which is Google’s relatively costly agent system, you can take the ideal customer profile, put it into Vertex and say, “Make this available to me as an agent.” And then what you do from there, which is actually pretty cool, is if you have Slack or Discord or Telegram, whatever, your ideal customer profile can show up as a virtual employee. And then you can talk to it. Your whole marketing team can talk to it just like any other person in Slack.

Let’s see. We’re on Agent Katie right now. Let’s look at Agent Sarah Johnson. So, in here, I can choose—obviously, choose the model I want to work with. And then I would have all the information about who Sarah Johnson is and talk to fake Sarah any time of day or night. So, this is one implementation of using it in an agent format. And this is great for larger marketing teams where you don’t even want to have people doing Gems. Like, just talk to this person in Slack and ask them your questions. If you want to have it doing things, then something like Agent AI or AutoGen or any of the agent frameworks would be the way to go. I would look at Meta’s Lama Agentic framework if you have technical resources because it’s really cool and it’s also lower cost.

You’re welcome. Next.

Audience Member 8: I would have been bringing in other types of agent. So, when I’m probably leading to debt, I don’t know if I have the best recommendations for that in terms of selecting that data. It would be helpful, I think, just to Google the location.

Christopher Penn: Yep. Depending on the types of locations, you might want to do this exercise for each individual location if the locations are wildly different. So, like, if you’re a grocery store, and you’ve got one in Jersey City, and you’ve got one in Framingham, they’re going to be wildly different. But you can bring in all of that data for that location and have it do the analysis, including the Google reviews. For brick-and-mortar, especially if you have heat map tracking of where people walk inside the facility itself, you can even put in the location data and say, “Hey, this is the frozen food section stuff,” and incorporate that into the ideal customer profiles.

Audience Member 8: Yes, all of it. All of it. The more of it you can provide, the better.

Christopher Penn: Yep, no problem.

Hi!

Audience Member 9: Yeah, no, I would absolutely use that information. Any data you have that is correct, I would absolutely use. The hardest part is actually getting the data. Like, once you have that, absolutely put in because as long as it’s in a format that the models recognize: PDF, plain text, images, etc., they can process it. My other question is, if they need PCI-X, would you send that?

Christopher Penn: Yes.

Audience Member 9: Why would someone send PCI-X, and then it must, like, touch the printer?

Christopher Penn: Yes, it was basically because it has the largest short-term memory. It can hold a million and a half words, whereas ChatGPT and Composition can only hold about 90,000.

Audience Member 9: Okay. Thank you.

Christopher Penn: Hi! Sorry, you’re back!

Audience Member 3: So those voices are provided by Google’s TTS service, the text-to-speech service. That is Google’s Neural Engine Voices Studio Edition. Can we add our voices?

Christopher Penn: You cannot. You cannot. However, you could build that yourself if you wanted to. You would say to Gemini, “Generate a script, and then feed that via Python script to Google TTS with your uploaded voices, or use 11 Labs.”

Audience Member 3: So there’s a way to do it?

Christopher Penn: Oh, absolutely. There’s a way to do it. It’s technically complex, but there is a way to do it. There is a way to do it.

Hi?

Audience Member 10: So the whole process you had, I think, was the software and the customers that you had, the audience that you had. How would you recommend it—we’re still working on it—how would you recommend it to the next generation? Do you understand customers?

Christopher Penn: To understand customers, get as much data as you possibly can about them. Again, one of the best things you can do is stuff like focus groups, one-on-one interviews, conversations. If you have a tool like Microsoft Clarity, how they browse your website, heat maps on your website, anything you can do to get qualitative or quantitative data that is representative about the customer base, you want to get that into some kind of consolidated format that you then feed to a model. The models, again, they’re just interns. So, if you don’t have the data, you need to create the data first. If you have the data, you can then use models to create even more synthetic data that mirrors your existing customers, and you can fact-check them that way. But I would start by interviews, phone calls. If you have a sales team, you should be recording every sales call and then getting it transcribed by AI, and then feeding that in. I would be taking the contents of your customer service inbox, any emails you’re getting from customers or potential customers, that goes in. If there’s a Slack group, exporting data. If you don’t have any customers at all, go find your topic on Reddit, get a Reddit developer’s account, get access to the API, and start extracting data from the subreddits about a specific topic.

Audience Member 10: Leave me a comment. You may say that people don’t like using square sponges, they want only round sponges.

Christopher Penn: So, go to the sponges subreddit, search for the word “round,” and then export all that data, and you can use that to tune how the customer profile works.

Mm-hmm.

Hi.

Audience Member 11: Hi.

Christopher Penn: Thank you.

Audience Member 11: No, my question was, are you really—my question would be, how would you maintain—would you suggest, like you said, in order, you don’t go to—would you go around and just start to—

Christopher Penn: Really good question. In Gemini itself, you will notice, as you use this, that individual blocks are controllable. So, I’m going to scroll back up here. I’m going to scroll back up here towards the top. I can take this section out and just delete this portion and load new data in and then have it repeated. If you want to get super fancy, you can actually get the code for the entire session in Python and then feed new data in every month or quarter, whenever you have it.

Yeah, yeah, yeah. You can take this. This is just standard Python code. Get your API key, and then you feed it all in. It will rerun. It will be expensive, but it will do it.

Yep.

Thank you.

Any other questions? Going once, going twice.

Audience Member 12: What if the data we have today is not really an ideal? I take all the people who are looking at me on LinkedIn or our company, and that’s not reflective of where I think we want to go or system.

Christopher Penn: So you don’t have your ideal customer. You use your ideal customer. You know, “Hey, this person is going to be a great user. You know, hey, this person is going to be a great user. You know, hey, this person is going to be a great user. You know, hey, this person is going to be a great user. And then you’re going to be a great user. I think we want to go with Cisco.” So, you don’t have your ideal customer. You use your ideal customer. You know, like, “Hey, this person is CMO, but at a 50-person company. We really want a CMO at a 5,000-person company.” Use that profile instead. There’s no limitation on what data you put into these models. So, if you know that today’s customers are the wrong ones, then try to get public data that you can about who is the right one.

Hi.

Audience Member 13: Hi there. Thank you very much for this very informative talk about privacy of data.

Christopher Penn: Yes.

Audience Member 13: It seems to me that you’re thinking about it more like that.

Christopher Penn: Correct.

Audience Member 13: What do you think of companies that do have sensitive information, and they’re so sensitive that they can derive from having their data uploaded and get you to show them that it’s an information and privacy of data?

Christopher Penn: They’re as good as their legal agreements, which means that if they do go wrong, you can join the class action lawsuit. If you want absolute ironclad confidence that your data is not being used, you should run a local model. So, a local model is a piece of stuff where you download onto your laptop. If your laptop can play Call of Duty, it can run local AI really well. And then there’s a tool that I use called Anything LLM that talks to the local model, and it’s, essentially, you choose what model you want, and then you have—it looks just like regular generative AI. You can load documents, you can ask questions, and it is as good as the local model you’re running. But I can use this on a plane with no Wi-Fi because it is entirely self-contained. I know my data is not going anywhere whatsoever. Now, the trade-off is that the model you use is going to be constrained by the hardware you have. So, if you want to run purely local AI really well, ask for a nice MacBook. And the models—the smaller they get, the dumber they get. So, the more data you have to provide, the less fluent they are. The two I would recommend right now—three I would recommend is: if you have a moderate computer, there’s a model called Mistral Nemo that is a 12 billion parameter model. If you want something that’s computationally stronger, Mistral Small D2 just came out two days ago. It’s really good. And then if you play Call of Duty on a 60-inch screen, no lag, then there’s a model called Lama 3.1, which is a 70 billion parameter model, which is very robust and can do almost anything that a web-based model can do. But that’s what—when you think about three-letter government agencies that want to use AI, that’s what they’re using because “We don’t trust you at all. We are going to have our data run only on our hardware.”

Mm-hmm.

Hi.

Audience Member 14: Great session. So, reflecting on the question a little bit, some of my friends, I suppose, I think they’re saying they just know that data is kind of going at all. So, you’ve asked that question already. We don’t have the strong computers or anything to, I think, load the code for you to do. It’s only when the agent is scrubbing and compiling the data that then it pops in.

Christopher Penn: Yes, so you can actually have these tools scrub data. So, you can have it write a Python script that will anonymize data. And as long as you have developers who can run the code, then it can take data, remove named entities and stuff like that, create synthetic data from it if you don’t have the compute power. I would honestly think it’s cheaper just to invest in the compute power itself. I mean, you can pick up a nice Mac Studio for about 10 grand.

Yep.

All right, last call. You don’t have to go home, but you can’t stay here. Thank you, everyone. Thank you very much for your time.

---

Need help with your marketing AI and analytics? Get help with your AI strategy » Learn generative AI for marketers with our new course » Prove the ROI of your marketing with our advanced analytics services » Just got a "silly" question you want an answer to? Ask us here »	You might also enjoy: Instant Insights: How to Write an Effective ChatGPT Prompt Contact Trust Insights INBOX INSIGHTS, November 8, 2023: Solving Problems with Predictive Analytics, Testing AI Models The Trust Insights Leadership Team The Intelligence Revolution: Large Language Models and the End of Marketing As You Knew It
Get unique data, analysis, and perspectives on analytics, insights, machine learning, marketing, and AI in the weekly Trust Insights newsletter, INBOX INSIGHTS. Subscribe now for free; new issues every Wednesday! Click here to subscribe now »	Want to learn more about data, analytics, and insights? Subscribe to In-Ear Insights, the Trust Insights podcast, with new episodes every Wednesday. Click here for Google Podcasts » Click here for Apple Podcasts » Link for all other podcasting software »

---

Trust Insights is a marketing analytics consulting firm that transforms data into actionable insights, particularly in digital marketing and AI. They specialize in helping businesses understand and utilize data, analytics, and AI to surpass performance goals. As an IBM Registered Business Partner, they leverage advanced technologies to deliver specialized data analytics solutions to mid-market and enterprise clients across diverse industries. Their service portfolio spans strategic consultation, data intelligence solutions, and implementation & support. Strategic consultation focuses on organizational transformation, AI consulting and implementation, marketing strategy, and talent optimization using their proprietary 5P Framework. Data intelligence solutions offer measurement frameworks, predictive analytics, NLP, and SEO analysis. Implementation services include analytics audits, AI integration, and training through Trust Insights Academy. Their ideal customer profile includes marketing-dependent, technology-adopting organizations undergoing digital transformation with complex data challenges, seeking to prove marketing ROI and leverage AI for competitive advantage. Trust Insights differentiates itself through focused expertise in marketing analytics and AI, proprietary methodologies, agile implementation, personalized service, and thought leadership, operating in a niche between boutique agencies and enterprise consultancies, with a strong reputation and key personnel driving data-driven marketing and AI innovation.