So What? How to use Generative AI to analyze survey data

John Wall – 00:54

Katie’s here. She was just here too. I don’t know what’s up with that because we were just talking to her.

Christopher Penn – 01:00

Oh, there you go.

Katie Robbert – 01:01

Oh, my God. I’m here. I promise I’m here. I’m gonna do the whole show with signals or something.

Christopher Penn – 01:10

Mime.

Katie Robbert – 01:12

Let’s start again. Happy Thursday, everyone. Welcome to So What?, the Marketing Analytics and Insights live show. I’m Katie, joined by Chris and John. Take two. How’s it going, fellas?

John Wall – 01:21

Excellent.

Katie Robbert – 01:22

If you ever wondered if we’re really human, the answer is yes. Well, and Chris, you’re on the wrong mic. So we’re off to an excellent start today. This week—oh, my God.

Christopher Penn – 01:38

There we go.

Katie Robbert – 01:39

This week we are walking through how to analyze your survey data with generative AI. We did this a few years ago, but so much has changed. One of the things that we’re really striving for in 2026 is making sure we’re showing you both lower-tech, easy entry solutions, but then also the higher-tech, more advanced solutions, because there’s room for both of them. We know that our audience and internal to our team, there’s a need for both versions.

Chris was showing me this earlier and we actually on the podcast this week—if you want to go to TrustInsights.ai/ti-podcast—Chris and I spent a decent amount of time talking about the low-tech solution that you can use now with your Google Sheets, which I think is very exciting.

Katie Robbert – 02:26

I think it’s a really good game-changer for people who are like, “I have all this data, I don’t know what to do.” But then we can get into the bigger, more advanced stuff. So, Chris, where would you like to start this week?

Christopher Penn – 02:39

We should probably talk to Sue. So this is going to use Google Gemini and Google Sheets. We are not paying subscribers to Microsoft Copilot. My understanding is that Copilot fundamentally does about the same thing in the current version of Excel and it uses Copilot’s models on the back end. For what we’re doing today, that’s okay. I have a lot of ranty-pants things to say about Copilot’s models, but that’s for another show.

Katie Robbert – 03:08

That’s for another show.

Christopher Penn – 03:09

It totally is. Or at the bar at the next event. First things first, we are going to need to get our data, which comes in our case straight out of Gravity Forms in WordPress—it’s just literally a plugin for WordPress.

Someone comes in, fills out the form, and I’ve removed the personal identifying information so that we’re not sharing people’s private emails on the air. What we’ve got here is just a spreadsheet where every column and every cell in rows is an answer. Some of them are multi-paragraph, some number, single paragraph.

Katie Robbert – 03:44

Now, can I back up even a step further from that to share where this data came from in the first place?

Christopher Penn – 03:50

Sure.

Katie Robbert – 03:51

We have tried—and to be fair, I have been a little remiss in getting this done—but one of our goals is to run a one-question survey every quarter. The goal of that survey is to help us understand what’s going on with our audience, what’s going on in the industry, and what’s going on with people who are taking the time to read our content.

So it’s the Trust Insights one-question survey—still live. You can answer it at any time. Our question this quarter was, “If you could attend one deep-dive webinar in 2026 to solve your biggest business or marketing challenge, what specific topic would you want to cover?”

The reason we asked this question, not only for our own planning purposes, but the topic of AI is obviously very prevalent.

Katie Robbert – 04:46

Everybody’s talking about different versions of it and what we do with it. But it’s not enough for us to just say, “Okay, here’s a webinar on how to do AI.” That’s too broad. We wanted to know specifically if you could deep-dive into anything. We didn’t want to limit it to AI or just marketing, but anything.

We got a wide variety of responses—very helpful responses—because there’s still a lot of people wanting to know about digital marketing that doesn’t include AI, or measurement, KPIs, organic search, or a lot of topics that aren’t necessarily AI, which I found to be incredibly helpful. Unsurprisingly, the majority of our data is around, “What do I do with agentic?”, “How do I do this?”, or “How do we do that?”

Katie Robbert – 05:31

That gives us a really good understanding of how we as a company can better serve our audience. Making sure we’re giving them what they’re asking for, not just what we think they’re asking for. I just want to set that context of here’s where this survey data came from.

Christopher Penn – 05:47

Exactly. That’s a great segue because you can see this is a paragraph—this is a big text box. You can put whatever you want there; it’s not multiple choice. By the way, slight sidebar: this is the way we should be doing surveying now in 2026, because the days of a static dropdown that you have to check a box are gone.

Generative AI language models are very capable of parsing language. There’s no excuse anymore to say, “For data analysis, we can only have a dropdown with these eight choices.” No, let people say what they want. You can use language models to understand that language. It drives me up a wall. “How’d you hear about us?” There’s five choices and they’re all choices from the 1990s.

Christopher Penn – 06:33

I haven’t gotten a postcard that I remember in quite some time.

Katie Robbert – 06:37

News, slash print, email, referral, radio. I miss radio ads. They were good. Anyway, let’s go on.

Christopher Penn – 06:49

We take those responses and put them into a Google Sheet. Now here’s what’s new in pretty much everybody’s office software: you can now invoke a language model like Google Gemini right inside of a sheet on a cell-by-cell basis. The formula for this in Google Sheets, unsurprisingly, is Gemini.

You have a prompt and then you give it a range. Because we’re processing individual survey results, we don’t want to use a whole range; we want to do it cell-by-cell. But before we do that, we have to know what we’re analyzing. We have to figure out what are the things that we would want to know. So, Katie, what are some of the things that you might want to know out of these survey results?

Katie Robbert – 07:31

I want to know the large buckets of categories of how people responded—topic, category. But then within those, I would want to know sort of what the subtopics are. Let’s say the topic and category is agentic AI. Well, that’s a pretty big topic. I would want to know within that if people are wanting to know how-tos, so that we can make sure we’re appropriately answering those questions.

Christopher Penn – 08:03

Right. Let’s call this “Topic” and let’s call the other “Subtopic.” What else would you want to know?

Katie Robbert – 08:16

Maybe I think you had put in—which I thought was interesting in an earlier version—sentiment. Sentiment is basically, at a high level, “is this a positive thing?” or “is this a negative thing?” In this context, I think it’s an interesting data point. I don’t know how necessarily useful it is, but at the very least, we can say, “Oh, this is what it is,” but then if it’s not necessarily useful, we can always take it out.

Christopher Penn – 08:46

I think the reason I put that in is because I might want to know if people feel differently about agentic AI than they do, say, regular AI. Do people feel differently about analytics? Like, “Oh, I hate Google Analytics 4.” Okay, clearly this is still a very sore topic with people. John, is there anything you would want to know out of this data?

John Wall – 09:07

The thing that I just love about this is because I’ve been on the other side of it so many times where you do these surveys where you get 10,000 or 20,000 results in, and most of the time it just goes to a bunch of bar graphs. Nobody pays any attention to the individual stuff.

Sentiment is the number one thing for me because there’s never been an easier way than it is today to just say, “Hey, categorize them as far as how many are the top-tier champions that love us to death and show me the five most horrible ones.” All those kinds of querying tools are more powerful than ever before. I love all that stuff.

John Wall – 09:47

The other one is to find all of those things and issues that you just can’t do if you manually go through the results. For software stuff, you can say, “Show me the three features that people have the biggest problems with and show me the four features that need to be fixed or added” to be able to mine it for that kind of stuff. All that is just… that’s the most exciting thing going in AI for me right now.

Christopher Penn – 10:15

Yep. Here’s a question. As our biz dev guy, would you be interested in a zero-to-five scale of urgency where somebody’s like, “Nah, I’m just curious about this thing,” whereas someone else says, “I want to buy something right now”?

John Wall – 10:29

That’s a good point. Katie has done a good job with the fact that we don’t just send the survey out and then I start pummeling people the next day.

Christopher Penn – 10:37

Cold calling.

Katie Robbert – 10:38

I know it’s tempting, but please don’t.

John Wall – 10:40

There’s a lot of, “Oh, you want that? What does it take to get in this car today?” But we don’t want to jump all over them. The big thing is to… yeah, on the other side of my mouth, when you get big enough, there will be four or five people per thousand that it just happened to hit them on the day that they’re looking to buy right now; they’re in the middle of a purchase. If somebody raises their hand and says they’re actually looking to buy, those people do get a call back right away.

John Wall – 11:11

Knowing their heat—I guess you could qualify all of BANT, too. It would be great to say, “Hey, segment these results by budget.” Let me see them by company size. Is there stuff that the bigger companies like or don’t like that’s totally different from the small ones?

Authority, too. Are we talking to a bunch of weaklings or are these the decision-makers? And then need—how much do you like this stuff? And timeline, too. How many people are talking about doing something in zero to three months versus over the next four years? All that stuff is worthwhile.

Christopher Penn – 11:48

Okay, so I think—

Katie Robbert – 11:51

I think for today’s purposes, urgency is good. But I think even just picking up on almost like a yes/no column of, “Did they say in there, can you help me or can you do this?” Looking for those cues because we have a couple of hundred here that we can go through manually.

It would be helpful to sort down to a smaller list—did they say immediately, “Oh, and do you do this?” or “Can you help me?” or “Is this something I can buy from you?”

Katie Robbert – 12:28

When you’re getting a lot of data, being able to pull that out and immediately give it to someone like John—who’s like, “Give me more people to talk to”—is great. If they raise their hand and say that, that’s great.

We’ve done that on previous surveys where we’ve added an additional question of, “And do you want us to reach out to you for the service?” 100 percent of people said no, which is why I stopped asking that question. It was pretty clear: they just want to give us their feedback. They’re not looking to get pitched.

Christopher Penn – 12:58

Exactly. Okay, so we’ve got two categorical columns: Topic and Subtopic. We have a binary column: Intent (yes or no). And we have two numeric columns: Sentiment and Urgency. Now we have to actually build this thing.

First things first, we need to figure out what those topics and subtopics are. For this, we’re going to use regular Google Gemini. The reason we’re going to use regular Google Gemini is because you can specify the most powerful model, which is the Pro model, and you can give it the entire data set all at once to think through broadly what those things are. First, we need to give it the file itself, which would be a useful thing to have. Let’s go ahead and drop in the file.

Christopher Penn – 13:45

Even though this is a CSV, it’s only got one column and that’s the text of the responses. That’s perfectly safe to use. Here we’re going to give it a prompt. We’re going to say, “Read through the corpus of survey results attached. We want to understand the five to six major topics in the results, plus two to three subtopics per major topic. Present your results as a markdown list with the major topic on the first level and the subtopic on the second level.”

We want Gemini to try to think this through. We do not have to get exact counts here. In fact, that would be a terrible idea because a straight language model like this can’t count—they’re incapable of doing math. That’s why we’re using Google Sheets.

Christopher Penn – 14:29

But we absolutely can explore the text space around all these freeform responses. This is what people should be using language models for: to take that unstructured freeform text and distill it down so that you can then apply categories to it.

Let’s see how Gemini is doing here. Oh, actually, it looks like it was writing some of its own code along the line. You can see it because it’s popped a little code bubble there. But based on the survey responses, here are the major topics and subtopics identified in the results:

AI Strategy, Organization Adoption, Change Management, AI Governance, Strategic Planning, AI Agents and Automation, Building and Deploying Workflow Automation, Technical Execution, Marketing Strategy and Analytics, Attribution, Reporting, Content Strategy and SEO.

Role-specific systems: Practical AI Implementation (AKA Vibe Coding) and App Development. Tool Integration, Practical Use Cases, Major Tool Ecosystems: Copilot, Workspace, NotebookLM, and Business Operations on Scaling. Small Business Knowledge Management and Operational Efficiency.

So, Katie, does that seem to fit with what you’ve read in the results?

Katie Robbert – 15:35

It does. Even at a high level, this would be incredibly helpful to me. I could take just this and build out a content calendar. But I’m excited to see where we go next with this. So far, because—unsurprisingly—I’ve been paying attention and read every single result that has come in and have been mentally banking them. It’s really nice to see that’s what Gemini is also coming up with in terms of categories.

Christopher Penn – 16:01

Right. Now what we want to do is start assigning categories. To do this, if you want to save your sanity, please start a new sheet. We’re going to call the sheet “Prompts,” and in the first column we’ll have the prompt name, and the second column will have the actual prompt. We’re going to have a category for Topic and Subtopic. We’re going to have one for Sentiment, one for Urgency, and one for Intent.

We should rename that and rearrange that so it spells the word “SUITS.”

Katie Robbert – 16:41

No acronyms today. We’re not developing anything new.

Christopher Penn – 16:45

Surprise. Oh, my goodness.

Christopher Penn – 16:48

Now, here’s the catch with Google Sheets: if you’re used to writing very long prompts, they have to fit in a cell. Cells can occupy a lot of text, but they can’t have things like line breaks in them, which is highly problematic.

What we’re going to want to do here is have Gemini help us write the prompt that will fit inside one big paragraph, and that is what it can then reference in the sheet. Let me give you a concrete example of this. I’m going to start off by saying, “Let’s classify the text in the cell based on the major topics provided. The major topics are the topic headings. Generate and return just the major topic.”

That’s our first part of the prompt. Our second part of the prompt is going to be what those major topics are, and we’re going to copy and paste that. To make this work in Google Sheets, I went ahead and created a Gem called the “Google Sheets Gem.” All it’s going to do is optimize the prompt.

Christopher Penn – 17:44

It’s basically going to read through this and figure out how to put this all in one paragraph so that Google Sheets will not blow up when we get it back. It will have to think this through to make sure that it is intelligent, but once it’s done, we can then put that in the sheet and reference it.

It’s thinking through the taxonomy and all the pieces. It should—if I did it right in the Gem—come up with two different options, and then we can decide which of the two options makes the most sense. As a pro tip: don’t ask any AI to just give you “the” thing because it will only give you the most high-probability answer. Ask it to give you options so that it forces the model to think a little more broadly.

Katie Robbert – 18:56

One of the pro tips that we had talked about last year—which perhaps we’ve sort of debunked—is asking, “What questions do you have?” or “Ask me one question at a time until you have a satisfactory amount of information.” Is that still a tactic that you would employ?

Christopher Penn – 19:16

For this specific use case, we’re asking it just for the options because the Gem itself is four pages long. I use those tips like “Ask me one question at a time” to build the Gem itself.

Katie Robbert – 19:30

Got it. That’s a helpful distinction because I think when a lot of us hear a pro tip, we’re trying to employ all of them at once, and it gets very overwhelming. The prompts themselves become very cumbersome.

Christopher Penn – 19:49

Okay, so now we’ve got our topic prompt, and we now need to generate the subtopic prompt, which is going to look very similar. We’re going to say, “Let’s classify the text in the cell based on the major topic in the first column and choose which of the subtopics is the best fit for the survey response text.”

We’re going to take the exact same thing, start a new chat with our Gem, and let it do its thing. While it’s doing that, I’m going to wire up this first one. We’ve got a prompt in cell B2. Let’s go to our topic. We’re going to go Gemini prompt—which is in “Prompts!$B$2″—and then the range is just E2. Yes, I know Gemini can make mistakes.

Christopher Penn – 20:53

And so it has now executed that prompt. This is why we want the prompts in a separate sheet, because if I were to drag and drop that prompt all the way down this entire sheet, it would have that prompt repeated over and over again. Nobody wants that. It would take forever and a day for it to do that.

Instead, we reference the prompt on the other sheet and it does this whole thing for us very nicely. Here on our first pass through, we have the topics themselves. Now let’s go back to Gemini, and it has come up with a second set of instructions. We’re going to put this massive, gigantic blob into the subtopic prompt. Let’s get the third prompt, our sentiment prompt, baking while we get this second one in place.

Christopher Penn – 21:55

So, go into our optimizer and this time we’re just going to give it something relatively straightforward. We’re going to say, “Do some sentiment analysis on this block of text. How would you score it from minus five to positive five?”

While it’s doing that, I’m now going to use Gemini prompts—this is B3—and then the range for this is A2:F2. I believe that’s how you do range and closing parentheses. No, it’s not how you do that. Oh, dear.

John Wall – 22:44

Is it a colon instead of a comma?

Christopher Penn – 22:53

It’s running into an issue where it’s self-referencing. It’s creating a circular reference.

Katie Robbert – 23:01

All right.

Christopher Penn – 23:01

Let’s ask Gemini, because one of the nice things is Gemini is built into itself here. “How do I fix this formula so that it references cell A2 and F2?” and then I’ll just give it this… Oops, stop. We’ll see what it does. This, by the way, is available in Copilot as well in Excel. Oh, look, that was easy. It says you just need to put it in commas. Shows you how good I am at this. But this is yelling at me.

Christopher Penn – 24:15

Clearly, AI functions cannot be used with other functions.

Katie Robbert – 24:23

So what does that mean?

Christopher Penn – 24:25

Basically, you can’t have a sub-function in there. So here’s what you do: we’re just secretly going to put together a column that says “concatenate.”

John Wall – 24:40

This is like spreadsheet judo today.

Katie Robbert – 24:44

What’s interesting about it, John, is that this is the low-tech version because we’re not writing code. We’re literally just trying to summarize and categorize the qualitative data, which is where a lot of people get stumped. There’s so much information—how do I know?

This is the low-tech version of that. We could bring it into Colab or Claude Code or something and have Python code written, but what I wanted to make sure we did was show that if you’re savvy with Excel, great, you’re already ahead of the game. Now you can still do that and get the benefits of generative AI. It’s that nice blend where you’re high-tech in Excel but you want the low-tech AI solution to benefit from it.

Christopher Penn – 25:38

So, let’s see if our little magical hack worked. There it is. By just doing a concatenate in a column along the side, we now have the subcategories for each of these things.

Katie Robbert – 25:53

Nice.

Christopher Penn – 25:54

Now let’s check in on our sentiment prompt here. Version A is robust, so we’ll use Version A and paste that in while I’m going to have it do the urgency one next.

Katie Robbert – 26:18

For the purposes of the livestream, obviously we’re on a limited time, so it’s that magically-already-cooked-in-the-oven kind of thing. We still highly recommend human intervention—making sure that the data is coming out correct, making sure that the prompts being written make sense to you, and making sure that sentiment analysis is reading positive or negative the way you want it to.

For the livestream, Chris is going through this a little bit more quickly, but in real life outside of a time constraint, you would want to make sure that you’re really double and triple-checking everything that AI is giving you back.

Christopher Penn – 26:59

Exactly. So here we have our prompt for urgency. We’re going to do the Gemini prompt for the urgency level. Now what we’d want to do next is figure out, okay, we’ve got all these different levels. What do we do with this?

At the most basic level, we probably want to tabulate by topic and subtopic to see what things people ask about and how many of those things there are. So, we’ll start a new sheet in here. I don’t have the foggiest idea how to do this—no clue whatsoever. But I don’t have to. I can literally say to Gemini, “I need help tabulating the answers in Column A of Sheet 1 by count.”

Christopher Penn – 28:25

Can you do that for me in Sheet 2? Let’s see what it comes up with. This, to me, is the most useful thing about having your AI here in the system: to say, “Help me with how to do this.” This sheet appears to be empty—of course it’s empty. Let’s call this “Answers” and this “Counts.” I’m going to say in Column A of the Answers sheet, can you do that for me in the Counts sheet? Let’s see if it recognizes the names we’ve given it.

Christopher Penn – 29:44

I’m going to cheat here and just go over to the formulas we used in the previous version of the sheet, which is we’re going to do a unique. Answers, and this is A2 through A999. I think I’m missing a closing quote. Parse error… why does this not work? Unique… just do this. There we go.

Now in the next column, we have to do the count-if. The formula for this is “Answers!A2:A999.” All right, so now we have our basic chart. We have our basic results. We have eight uncategorized and we have marketing strategy and analytics. Let’s put this into a nice bar chart. There you go, Katie. There’s the answers from our survey.

Katie Robbert – 31:33

I like it. To satisfy my OCD, can you go to the Answers sheet and freeze the top row? People who don’t do that are maniacs. I can’t believe people can work that way.

Christopher Penn – 32:02

Okay, I think that did it.

Katie Robbert – 32:05

No. Okay, cancel that.

Christopher Penn – 32:07

Okay.

Katie Robbert – 32:09

Go to View. Go to Freeze. Freeze row one. Yep, there you go. Now can you do me a solid and sort the topic so that we have… yeah, because everything was sort of mixed together. So that’s more helpful to me is like, “Oh, here’s everything under AI agents and automation,” and then the subtopics next to it.

Katie Robbert – 33:07

Basically, I’m trying to look at the clusters together. Then if you go to the counts, it’s going to adjust based on what we have on answers. Or you can sort and filter the count column, which is Column B, biggest to smallest.

Christopher Penn – 33:07

Right. And we could actually do this with subtopic, too, to drill down into what the subtopics are.

Katie Robbert – 33:11

But to your point: yes, this tells me what I’m looking for. The first thing I would go looking for is what’s unassigned. What’s in those that can’t be assigned?

Christopher Penn – 33:23

Right, which we go down to the bottom. “How do you pick winning lottery numbers?” “Tell me everything Katie knows so I can be just like her.”

Katie Robbert – 33:31

So it’s the junk stuff.

Christopher Penn – 33:33

Yes, exactly.

Katie Robbert – 33:34

Got it, which is also helpful to know.

Christopher Penn – 33:37

Yep. Now here’s the next thing: what do we do with all this information? One of the things that we want to know—that you could do in here, but you can also do elsewhere, and this is where I would personally spend some time—is take this into a system like Colab to have it start doing that statistical analysis.

What is the relationship between sentiment and intent, or sentiment and urgency, so that we could ask those questions? Because maybe there’s some topics that—yes, it’s a frequent topic—but there’s no buying intent. Nobody is going to buy anything from us on it; they just want to broadly know something about it. That’s what you can do inside here, but it’s probably faster to do inside Google Colab.

Katie Robbert – 34:26

I think that’s where we’re really talking about the lower-tech versus higher-tech solutions. This, in and of itself, without that additional analysis, can take you pretty far. This gives you a content calendar. This gives you a direction to revise your services. It gives you a starting point for whatever your speaking topic is this year. Or it tells you everything that you’re doing is the exact opposite of what your customers want.

Fortunately for us, we looked at this and none of it was really a surprise. But for me, it’s helpful to know that it’s not all agentic AI; there are other things that people want.

Katie Robbert – 34:26

So, if you go to the counts, for example, Chris, you see agentic AI, but the biggest topic is marketing strategy and analytics. That is our bread and butter. That is super helpful to know. We’re not just going to dump everything we’ve been doing for the past eight or nine years.

We’re going to figure out where agentic AI fits into it. We’re going to figure out where business operations and scaling fits into strategy and analytics, and so on and so forth, so that we’re addressing all of those things both separately and as combined topics.

Christopher Penn – 35:44

Exactly. When you dig into the prompt for that, one of the things that makes that category so prominent is AI ROI is in the analyst category. That’s where those subtopics really start to help out is for us to be able to look in and go, “Oh, people want to know how do I measure the ROI of my AI efforts?”

Katie Robbert – 36:04

Which completely makes sense in that category, because AI agents and automation is just how to build the thing, theoretically. That’s not saying, “And I want to measure my progress.” So ROI, regardless of the technology, really does belong under strategy and analytics.

Christopher Penn – 36:24

Exactly. I’m going to fill in this last column here because I do want to get the Gemini prompt in place for intent. This will give me my intent scores. Then we can take this over into Colab and have it do the data analysis on those columns because we’ll have the quantitative. Although it looks like it’s spinning out a bit more stuff than I wanted there. We can have it just detect the integers.

Christopher Penn – 36:24

Interestingly, just eyeballing this, there’s a fairly diverse range here. There’s nothing here that’s like a zero, so it’s one, two, and three is the background of where this intent is. That’s actually pretty good to know for John’s purpose—there is some buying intention here. It’s not just looky-loos.

Katie Robbert – 37:36

You can tell by John’s face he’s excited.

John Wall – 37:40

Let’s make that money.

Christopher Penn – 37:42

And there’s obviously, as expected, look in the zero—this is irrelevant chatter, no buying intent. That too is useful to know as a diagnostic. If you put in some intentional junk answers, it’s a way of fact-checking these things.

All right, so our next step then would be to say we want to do a more thorough statistical analysis of this data file. We have survey data, so let’s drop that in and give it some instructions.

Christopher Penn – 37:42

We’ll say this is survey data that also contains sentiment, urgency, buying intent, topic, and subtopic. We want to analyze this for a few different dimensions. First, we obviously want to know the counts by topic and by subtopic. We also want to know if there’s any correlation between sentiment, topic, urgency, and intent.

The urgency and intent should be about similar in terms of their values. We want to know how things like topic, subtopic, and sentiment influence urgency and intent, if at all—and they may not.

Christopher Penn – 38:39

So that’s the prompt I’ve given it. For those who may not have watched previous sessions where we’ve done work in Google Colab, Colab is going to use the Gemini model to write Python code to do the data analysis.

First, it’s going to come up with its plan. It’s like, “Here’s my plan: I’m going to read it, I’m going to analyze and explore relationships, and then I’m going to produce a final output.” Read it before you hit auto-run. Sometimes things just go bad.

Christopher Penn – 39:31

All right, so we can see it’s already starting to build our Jupyter notebook, which is a Python notebook, and is looking at the condition of the data. It shows some formula-like strings and some weird stuff in there.

So it’s going to attempt to do its own cleaning on this to try and understand what it’s looking at. It’s already flagged that there might be some data quality issues in what Google Sheets spit out, which I think is interesting and helpful.

Katie Robbert – 40:09

What I like about Google Colab is that it is still technically a low-tech solution because you, the user, aren’t writing code. You’re just saying, “Hey, I need you to help me do this thing. Can you write the code for me?” That is specifically what this tool is meant to do.

You can write code in Gemini or ChatGPT, but that is not the single function of those pieces of software. This is the single function of this piece of software. Therefore, from where I sit, that makes it the best at it in terms of what people have access to and what it is you need to do.

This is the one that you want to use if you’re like, “Okay, I need to do something more advanced, but I can’t write code. So let me find the tool that’s going to do it for me.” This is a great low-barrier-to-entry way to do those things.

Christopher Penn – 41:08

Based on the work plan and the fact that it’s running into as many data quality issues, my guess is it’s going to take 15 to 20 minutes to finish baking. The process is think through carefully what you want that prompt to be—maybe even use your prompt optimizer to come up with it—and then let it go and churn through.

So, it’s starting to come up with topic distribution. It’s looking at subtopics. Interesting in the subtopics: practical AI implementation, coding, and app development showed up there specifically, which is a nice little interesting thing. Again, that doesn’t surprise me. We know just from our… because like you did, Katie, I’ve read every single response and AI agents were definitely a very hot topic.

Christopher Penn – 42:05

The distribution of sentiment is a bit of a mess in there. That’s going to need some cleaning up. When this is done, ideally we’ll have an answer for John and also for ourselves to say that when people are talking about this particular topic or subtopic, or they have this sentiment about it, they are going to buy something from Trust Insights.

When they have this or this, they’re not really buying. It’s good to create that content for awareness, but we probably don’t want to build a service around it because there’s not enough buying intent.

Katie Robbert – 42:51

If you just stopped at the counts in the spreadsheet, it does give you a content calendar. But it doesn’t give you that additional focus because, to be fair, Chris, we’re only two people. We can only create so much content.

We can create a lot of content if we want it to be really crappy, but if we want to create a lot of high-quality content that’s actually helpful to people, my first question is: where do I start? How do I prioritize what people want the most? Just having the number of mentions isn’t enough for me to say, “Okay, this is what’s going to translate into sales.”

Katie Robbert – 43:30

At the end of the day, we are looking to sell stuff. We’re not looking to pitch these people back and say, “Okay, you said you have this problem,” but we’re looking to create things that then help them trust us to solve that problem for them.

We’re building that relationship through this one-question survey. Then we can pass along those contacts to John and say, “This is what we did. Do you want to be the one to follow up?” Like, “Hey, we answered your question. Here’s the thing. Let me know if you want to start that dialogue.”

Christopher Penn – 44:04

Exactly. What’s so cool about using language models like this is we didn’t have to ask, “What is your level of buying intent?” right? Because clearly anyone who’s ever filled out forms is like, “None. I don’t ever want you calling me.”

If we rely on language models to parse out and think that through, we don’t have to ask that question. Part of the assignment for everybody who’s watching this is to say, how much mileage can you get out of a freeform response? How much can you infer from it? We didn’t even do feature engineering. There’s a couple of things you could do in feature engineering, one of which would be: how long was the response? The longer response is, the more care someone put into writing it.

Christopher Penn – 44:54

And that itself can be an intent signal. If you’ve got a one-word answer like “agents,” they clearly didn’t really have a lot of intent. If they gave us a page worth of, “This is the eight things I care about right now, I’m so involved in this and this,” either they have a lot of free time—

Katie Robbert – 45:15

Or they have a lot on their mind. Someone took the time to ask the question, and they’re like, “Finally, I have the space to say everything that I’ve been thinking about.” I did that to you earlier and you were like, “Oh, you have a lot on your mind.”

I see it more as that versus the free time. It’s like, “No, nobody’s asked me what I’m thinking, and now they are.” Let me go ahead and tell you. What I see with this is a really clear next step: once we get the analysis sorted out through Colab, we can then go back and say, “What do we currently have that answers these questions already?” John, you can go through—

Katie Robbert – 45:55

It’s not a pitch, but it’s a touch where you can go through and say, “Hey, thank you for responding to our survey. Did you know we have this content already that answers your question? If you still have questions, feel free to reach out to me. I’m happy to help.”

It gives you that opportunity to build a relationship in a meaningful way, not just, “Hey, you filled out our survey. Here’s the five things you can buy today,” which everyone’s going to just be like, “Unsubscribe, block, get out.” No more John Wall. We don’t want that. We want to do it in a more thoughtful way.

Christopher Penn – 46:27

Katie, do you want to talk about what else we used this survey for this week?

Katie Robbert – 46:34

The other thing we used the survey for—which I know we’re starting to run up on time—was to help us figure out internally where the heck we’re going with the company. It’s not that we don’t know—I have a very clear vision and very clear plans on what we’re doing next.

But what this survey did was two things. One, it validated the plans in place for services and growth. Two, it gave us a more concrete plan of specifics which were kind of missing before. Chris, you’ve been working on something that I thought was actually really interesting—and I’m still wrapping my brain around it—where you brought it into Claude Code and deployed agents against it. These agents are basically stand-in team members in this case.

Katie Robbert – 47:28

So you have a co-CEO, which is me. I feel very confident because I helped build it, so it’s going to represent me. The results we got back very much did, which was fantastic. We have a CFO, we have a Chief Revenue Officer, and then we have the voice of the customer, which is always important.

We want to make sure we’re not doing these things in a vacuum, but actually doing it on behalf of the person who’s going to buy it. Rather than Chris, John, and I sitting and going, “What do we do with this? How do we move it forward?” Chris took everything he knew about us as Trust Insights and us as leaders, took the survey feedback, and said, “What do we do? How do we move forward?”

Katie Robbert – 48:07

And it gave us the start of a strategic plan and said, “These are the options of things that you could do. These are the likely probability things that are going to make money.”

Again, this isn’t a, “Okay, take those things and then sell them back to the people who filled out the survey.” It’s for us internally to say, “Okay, how do we take what we have and refine it so that it becomes a service that solves these similar issues for people?”

Katie Robbert – 48:35

So Chris and I are working through the results of that Claude Code strategy session. We have some really good viable options that aren’t going to take us a lot of extra bandwidth to put together, and then we can get them out there and John can sell the heck out of them.

Christopher Penn – 48:50

Exactly. And this is all based on things like the Trust Insights ICP service and the Trust Insights Casino prompt framework for deep research. One of the things we did was have Claude Code go through our sales playbook and say, “Fact-check this. What things in here are there not supporting information for?”

It came back and wrote us 40 deep research prompts that I now have to go and get commissioned and set up. It did a fantastic job of researching even our own assertions about who we are and what we do. Combine that with the survey data we have—all the real words of the voice of the customer—to come up with what Trust Insights should do in 2026.

Christopher Penn – 49:38

Unlike if we had done this the old-fashioned way, it ran through this entire process in about 70 minutes. Whereas if I were doing it as a human, it’d be like a five-week project and I’d be crying into my coffee at the end of every day.

Katie Robbert – 49:58

To be fair, it’s the kind of work that I thoroughly enjoy doing. But it’s really about bandwidth. I can take the survey feedback, our sales playbook, what I know about our audience and voice of the customer, and I can manually put all of that together.

I enjoy that work. But what I like about this is it gives me something to react to. I still have to do the critical thinking. I still have to put those plans together. I still have to be the one to set the vision for the company. I’m not going to let AI do that for me because it doesn’t know what I know thoroughly.

Katie Robbert – 50:37

It just brought me closer to getting to the point of an execution plan versus a couple of weeks of getting through everything to get to the same result.

Christopher Penn – 50:48

Right. It’s basically a team of analysts that brought you the analysis so now you can make decisions on it much faster.

Katie Robbert – 50:55

Right.

Christopher Penn – 50:57

All right, so that was survey analysis. The medium-low-tech high-tech way is to spend some time with a tool like Claude Code and say, “Here’s my survey data.” Give it even this example of what we just did. Say, “Write me some Python code that will autonomously execute this against the LLM of your choice.”

That’s what you would want to do if you have survey results. Say you’re McDonald’s and you have hundreds of thousands of survey results every day—you’re not going to do it in a spreadsheet. Once you get beyond 10,000 results, you’re really going to want to automate it fully. But that’s for another show. Any final words, John?

John Wall – 51:37

Katie, I’m just thrilled to dig into all the numbers. Like I said, this is the kind of stuff that people would be churning for eight months. By the time you get the results, the people are dead. So it’s great to have it actionable.

Katie Robbert – 51:51

Well, and I will say—quick plug. This is the kind of work that we do for ourselves but also for our clients. If you are buried in what we call “dark data”—that’s always been our mission since day one—to help shine a light on the dark data, this is still that.

Give us a shout. Go to TrustInsights.ai/contact. You get to talk with John, and John will figure out how we can help.

John Wall – 52:14

Yeah, light it up.

Christopher Penn – 52:16

All right, thanks for tuning in, everyone, and we’ll talk to you all on the next one.