Transcript for "SparkToro Office Hours: SparkToro's Biggest Update Yet": Stage. Oh, that was fast. It did I thought it would give me a do are you sure you wanna go on stage? But no. It doesn't there's no two factor authentication. There isn't. Are we live? I think we are. Welcome to office hours, Amanda. Welcome. There we go. You well, usually, it counts down. It says three two one. Yep. They they change it every time. Alright. Well, this is exciting. Yeah. Lots of friends in the chat. A couple of new names or new to me names. Hello, everyone. Welcome. Welcome. Alright. Alright, friends. Well, we have a ton of stuff to get started on. But, Amanda, we always get some questions before we kick off about weather and how people get the recording. Yes. So you will get the recording. You know, we long as you sign you're you're all signed up, therefore, you're here. And the recording or the replay of this webinar will be available as soon as it's over. It might need a couple minutes to process. And then if you if you want to share this recording with anybody else or this webinar with anybody else, you totally can. Just send them, you know, the registration page, and then they can sign up and they can watch the recording too. Jason asks, how do we get a SparkToro mug? Unfortunately, they are not on sale. It's possible that Amanda learned of copyright laws after she made the mug? Yeah. So we have we have Totoro on our mugs and, I actually, I had this made, like, very manually. Like, I had a designer make it and, like, like, make it make it because when we tried to send it to Zazzle, they wouldn't print it because of Totoro. Yeah. So, like, I act this is, like, handmade, but, like, this is actually, like, man very manually made, I'll say. I I mean, I think, clearly, there's, demand. And, realistically, I think we should take, for folks who don't know, Casey's daughter, Reese, put together a little Oh, yeah. Cat dragon that we use on our t shirts and, like, some of our SparkToro Together branding. So we should just we should get some Oh, we do. SparkToro Together cat stuff going. Yeah. Like sparky mugs. Exactly. That's the name of the cat dragon. His name is their name is Sparky. And they look at water bottles made or something. Amazing. Yeah. Okay. So, yes, friends. You'll have the recording. You'll get the slides and PDF afterwards, and that's the main step. So this, this episode of office hours contains sorry. Give me just a sec. I gotta close all the doors or there's an echo. This off this episode of office hours contains a bunch of LLM prompt suggestions. However, I have not linked to the finalized versions of those yet, and so, I will make a special effort to do that. And so when you get the PDF of the slides, it will contain links to all of those. I'll try and add a slide at the end that has, yeah, links to every one of them so you can you can see. Because it is a lot of work to as much as I we might, dismiss the prompt engineering as a as a practice. It's a lot of work to get the right prompts. So Yeah. It is a lot of work. Yeah. So all that to say, please please keep an eye out on that follow-up email. It'll probably come in, like, twenty four to thirty six hours or so. So please keep an eye out because it'll have all the you know, have the replay, it'll have the slides, and the links to the prompts, which is what you really want. Right? That's what we alright. Alright. I am gonna oops. Oh, look. Casey's backstage. How wonderful. Isn't that let's see. Okay. So I'm gonna be sharing my entire screen. This one, which I think focus can now see. Let's hide that guy. Great. Okay. So, friends, what we heard from a lot of you, who use SparkToro is that you have started to change your behavior around how you use SparkToro, and you're starting to export the data and then use it to have conversations with your favorite LLM. It still looks like most folks prefer chat GPT, which is what I've used today. Although, Anthropic's Claude has got some very strong proponents for it as well, and some folks are big fans of Gemini also. So what I did is I talked to a bunch of people from our customer community, and thanks to those of you who contributed and participated. I know some of you are here today, and look through your use cases and then try to distill those. I wanna get to those, but before I do, I think it's really important, especially when we have, this this era of AI to do a quick refresher on what large language models do and how they work. So, historically, I've, pointed folks to, Stephen Wolfram's article about this, but I actually think that Paige, from Nielsen Norman Group did a really nice nontechnical primer for folks who wanna understand what's happening. And what's happening is not magic. It is not artificial intelligence. I know it's called AI. It is sequence prediction. Right? It's raining cats and and an LLM is really good at filling in the word dog. It is it is spicy autocomplete in the best sense of of that, phrase. And this this is how these systems work. They're probabilistic. This is why when you, you know, 10 people ask an LLM the same question exactly the same way, 10 different times, you get a 100 slightly unique answers. Right? Because it it probabilistically builds answers. This uses neural networks, which sound really fancy, but are essentially just an input layer, an output layer, and then layers that progressively influence in between like that. And what these layers are doing is they're, you know, they're basically using if you if you remember, like, your your coordinate system. Right? Like, this is a four on the, y axis and a two on the sorry. Four on the x axis and a two on the y axis, you know, in the upper right corner there. They're modeling a whole bunch of those in multidimensional space, which is very hard for a person to think of. Right? Like, outside of three dimensions, we we struggle because we've never seen a fourth dimension even though we we experience time, obviously. And then, right, they are they're they're using these basically sophisticated software engineering systems. Transformers is one of them. Self attention mechanisms is another. Right? And to to relate things to each other so that word relationships and concept relationships, these can all be mathematically put together. This is why you might have noticed. Right? This is from from January, someone asking on the chat GPT forms, how come it can't do simple addition? Why why can't this large language model it's a computer. My calculator that was built in 1959 can do this. How come ChatGPT can't do it? And the answer look. That's not what a large language model does. It predicts words that come after other words, tokens that come after other tokens. Right? And why does chat GBT give answers it knows aren't correct? It doesn't know anything. It has no knowledge. Right? What it has are predictive probabilistic systems. If you want tokens that frequently come after other tokens, it's really great at that. This is why whenever I see you know, there's viral articles out there, Amanda, about, like, wow. When, you know, they built this AI and then they threatened to shut it down and the AI responded that it would kill them rather than be shut down. And that has nothing to do with reality. The only thing that has to do with reality is this means that the AI in its training data had seen that when threatened with getting shut down, words that frequently came after that work, I will kill you first. Right? Like, just like all of us, it's watched a bunch of movies and television shows and read a bunch of fiction about these kinds of things. And so I I just want us to keep that in mind. Yeah. I I wanted to add Iran last yesterday. Here's what actually happened. It wasn't Trump, but president Joe Biden's administration. Wow. That is quite shocking. This is really amazing. How did it how did it come up to that conclusion? Did right? Like nope. What what happened here is ChatGPT is trained on certain data. When when super current events happen, right, this was asked a a few days ago, like, right after the the the bombing, then these sorts of things happen. This is this is now correct. If you were to ask ChatGPT now because it can crawl and index the web and grow and learn. Right? It it it now gets the answer correct. But FYI, it's pulling from historic data. So just just be aware of all these things. This is astrophysicist, Katie Mack, who, has written extensively and and done a bunch of stuff around, large language models and predictive text. I They don't know facts. Right? They're not a fact database. It's not like Google's knowledge graph. They're not designed to be able to answer factual questions. They mimic patterns of words, and sometimes, mimicking patterns of words is all you need. In fact, for all the use cases today that I'm gonna show you in marketing, I am gonna be focused on mimicking patterns of words probabilistically because that is what these tools are excellent at. When they're right, it's because correct things are often written down on the Internet. Right? Those correct things fit the pattern, and so the probabilistic answer, the CHAChEPT or Gemini or Perplexity or whoever gives you are correct most of the time. Just just be really cautious. I wanted to try and create a framework for thinking about this, like something you could reference or or show your boss or team or client around good and bad marketing use cases. And so this this spectrum, what I'm saying here is, generally, the stuff on the right, right, that is is bad, like, I would I would caution against that or I would very carefully hand review every single thing. And the stuff on the left probably needs less review, tends to be quite good. Getting content ideas, great. Creating content you're gonna publish, terrible idea. In the middle, editing your content, Often good, needs a review. Generating insights from data sources? Great. Generating real data? No. But let me give you an example. In the this is probably two years ago, someone said, hey. I went to ChatGPT, and I did all my keyword research there. I was like, no. No. You didn't. You went to ChatGPT, and it gave you words that frequently came after other words, which look to you like keyword research. But none of the volume numbers were correct. None of the CPC numbers were correct. None of the trends were correct. Many of those keywords, nobody searches for. But it sounds quite credible because they're words that frequently come after each other, and so they they look right. And and then running data analysis. Some data analysis, it's good at others, as we saw it. The math problem, not so great. The only way I do references. Regarding the data sources? Web for v one answer. High risk. I wouldn't do it. Crawl the web without any misses. In the middle. Make predictions for further review. Perfect. Great. Give the right or best answer. Make prediction. You you get the idea. Right? All of these conceptually fit along the spectrum. There's certainly many more. And this is this won't be true forever. Right? I I think what will never be true, in my opinion, is that create publishable content will be I mean, depending on your publishing standards, I suppose, will be a great use case. But editing content might used might move to a very good use case. Could be that in the future, even even today, I think ChachiPT is getting way better at at math. Some of the more, advanced reasoning models will call calculators and that kind of thing, and so it'll do it correctly right. Some of the more advanced models, I think, in the future will call current event news before trying to answer. So it it'll get better and better. In general, AI in marketing, really good if you want conformity, if you want fast, if you want stereotypes. I know stereotypes are often thought of as bad, but sometimes they can be really useful. Like, what are the things that are often associated with this in the minds of people? It's great for asking those kinds of questions, and marketing needs answers to those sometimes. Giving answers based on existing human knowledge and words, tasks, numbers that often come after other words, tasks, and numbers. Really good for those. Bad at novelty, bad at quality of ideas, bad at illuminating complexity, bad at creativity, coming up with totally novel things, right, like new stuff, and considering any information that's outside the training corpus or inputs or that doesn't already exist in retrieval augmented generation, right, searching the web. Not good at that stuff. You wanna ask a survey of your customers? Terrible, terrible idea. You wanna interview people at terrible I didn't use AI for that. Okay. So one thing that I think is really cool is that AI output improves with real data. So when you give it data and then you say, hey. Consider this information in your answers to me, it's quite good at this. I think this is why so many people have started using SparkToro exports plus ChachiPT or, or their favorite LLM to solve these kinds of marketable problems. I'll show you I'm gonna show you a bunch of examples like this, including the the 10 that I'm I'm sort of recommending that I that I like. So my audience is these people. Right? They're middle to high income travelers from The US. They want they're going to Italy. They like unique cultural experiences. Tell me which social networks they visit for most of the least, which websites they're most likely to visit, which YouTube these are essentially the things SparkToro does. Right? Can ChatGPT answer these? It will confidently absolutely give you an answer, but it doesn't have this data. Right? The AI could not possibly know the correct answers to these because they don't have, you know, clickstream panel data, and they don't have social network use data. They they just have a corpus of probabilistic sequences of tokens, right, words words and phrases that frequently come after them. And so it says, you know, Instagram is the most used network followed by Facebook, then YouTube, then Pinterest. Still widely used among Genex and Boomer. I you you can see the stereotype machine playing out. Right? It doesn't say 41% of your audience visited this, you know, social platform in the last quarter. It says still widely used among gen x and boomers. It's it's it's stereotyping. It's creating words that it frequently sees around these topics in the past. It sounds plausible because it's based on the corpus of stereotypes, which are based on research reports. Right? Like like Pew Internet and American Life Project has been publishing the fact that Gen x and Boomers use Facebook more than other other groups. Likely websites videos visited. Again, these sound good. They look pretty darn good. I don't even think they're that terrible in terms of answers. You'll find many of these on the list of websites that are actually visited that you can see in SparkToro, but it's not answers. It's merely predictive text. Right? It's not the truth. It's just spicy autocomplete. So be be very cautious. Right? YouTube channel subscriptions will you'll start to see these gaps. There's no way to confirm the accuracy. There's no relative degree of use that it offers here. No contextual information. Some of these are problematic. They they don't seem like they I I don't think they're right. You know, we don't see them in the SparkToro data. So search engine versus AI tool usage, it gives a non answer here, which I sort of prefer to it just giving you an incorrect answer, but you'll you'll see a lot of these in there. Conversely, if I say, can you give me five ideas for social media posts on Instagram that would likely to be, you know, like be likely to appeal to this audience? These are good. These are pretty darn good. I I don't see anything wrong with them. They are a little stereotypical. Many people in the travel space have done them before if you paid attention to, like, you know, Instagram, reels about traveling to Italy. You're gonna see a bunch of these because it's giving you things people have done before, and that's where AI shines. Right? Content ideation, stellar use case. What what SparkToro's customers figured out before, you know, Casey and Amanda and I, is data can make it even better. So, I'll give you an example. This is a feature Casey built. It is not live for everyone, but if you have the, agency or enterprise tier of SparkToro, you're able to use and test this now. I know a bunch of people have been testing it, playing around with it. We really appreciate your input. So you can say, my audience is middle to high income travelers is in the, advanced search section, and then SparkToro will essentially call an LLM. It'll go back and then, return to us with stereotyped data, meaning meaning websites. Remember I showed you that websites list, and I was like, gosh. These look pretty good. They align with what's what we see in SparkToro. And so we use that and then calculate off of those audiences, or the the combination of those website audiences, an audience report based on a description. Right? Natural language description. Just super useful for a bunch of things. And then there's a bunch of logic that Casey's written in there. Why did I just get a partial service degradation message? Okay. Hopefully, go Goldcast is okay. You guys still seeing and hear me alright? Okay. Amanda does. Okay. Well, who knows? That's the first time I've seen a message like that from Goldcast. Fingers crossed it's gonna be fine. Behind the scenes, right, Casey built this translation system for the outputs. We have tested it extensively. I know a bunch of you have tested it. We're very much liking the results from this, and so you might see this roll out more broadly pretty darn soon. Here's what happens. When I search for that described audience of of, you know, wealthy US travelers to Italy who like cultural experiences, SparkToro puts together, this report, and then you export the XLS file. This is what the this is the all data file. You could do this, obviously, in any specific tab that you wanted to. You upload that to your LLM of choice because most folks who use SparkToro for this are using ChatGPT. I'm gonna be using all my examples in ChatGPT today, but I've heard really good things about anthropics cloud as well. You upload it to the LLM of your choice, and then you can sit ask questions. Like, based on this data, give me those same five social media post ideas for Instagram. Check out what what it can do is, in my opinion, significantly better. It is not significantly better because the ideas are better necessarily. It is superior because ChatGPT is drawing from, the data and then giving the idea. And in my opinion, that is vastly superior for anyone who needs to make the case to their boss, team, clients, even themselves. So rather than saying, oh, why is Italian street food, spotlight useful and why these three cities? Check UPT is saying, well, we saw that in the interest section, Italian street food guide was very high affinity, authentic Italian recipes also very high affinity, and so we're recommending these. You can then double check, and you can see that Rome, Naples, and Palermo are places where hidden gem types of stuff or or, undiscovered or unique suggestions. There's lots of search demand for this. So it's it's taking data from the report and then giving you suggestions based on that. Yes. They're stereotyped, but they're stereotyped on data that you can show your work on. That's what I appreciate. That's what I think is useful. I think that's why so many people have been telling us that this is what they're doing with the product now. These still require human review. If you think AI is gonna put you out of a job, let me tell you, not anytime soon, friends, certainly not for this. Every single one of these, I would I would not dare make this my social media posting scheduled content or even my content plan without hand reviewing and making sure that everyone was a good match for my brand, the message I'm trying to send, what I think was gonna resonate. For example, Florence in a day must see spot secret cafes and sink views does not match up with day trips from Florence. We humans know day trips from Florence means not in Florence. So the topic, Florence in a day, this doesn't make sense. It suggests, in fact, the opposite that you are that a a traveler is staying in Florence for an extended period of time. They don't wanna see it in one day. They want to see other places that they can travel to from there. You you can see the how the stereotype machine sort of got to the answer that it's giving and why this human review is so important. Please, for all 10 of these that I'm about to show you, make sure you remember this human review aspect is absolutely crucial. You cannot just, whatever, hit publish or put it in a report without significant danger. Alright. You ready to rock and roll? Let's do it. Okay. Use case one, positioning. Positioning is the is the practice of essentially saying what's unique about our product or our brand that makes us, special and valuable compared to the alternatives that someone might consider. And positioning ideation, not necessarily saying you should take these as the final version, but it it gets pretty good. So I'm gonna be showing b to b and b to c examples from a wide variety of folks. But here's Clinically, which provides, clinical research software. And they might be right. Their target audience is essentially people whose profession is in clinical research. They're clinical research associates and coordinators and, clinical trial, development, you know, pharmacists, those kinds of things. And you take the SparkToro audience report, upload it there. Right? So here's here's my website. The ignore the naming convention of the spreadsheets. Casey, made a fix for these, so they they're they're all correct now. But the website is clinically given this attached audience research report for my target audience. Can you brainstorm? I want you to brainstorm 10 ideas for the positioning statement on our landing page because I don't think clinical research software that puts you in control is great. I don't know. Maybe they've AB tested the heck out of that headline, but I didn't think it was great when they saw it. No offense to the clinically team. And so let's let's brainstorm some positioning statements. Here you go. Yeah. Okay. Okay. Well, my colleague, these these sound a little dry. There's no problem with them. I can see where the data is coming in. Right? It uses the high affinity job titles from the bio data, connects with the drug affin drug process, which is which is in the search data. It's good, but we can do better. Anytime you are, crafting you know, going through this prompt process, I hate to say it, but you've gotta do prompt engineering iteratively. Like, it it it is a process, and I I have never successfully, not in one of these examples or any of the ones I've tried, have I done it right the first time. Like, it's always, oh, yeah. That's good. Maybe I can make it better this way, this way, that way. Okay. So now what I've done I had to do a bunch of iterations, but what I'm done here, excuse me, is, you know, brainstorm these ideas and then provide details on how each idea relates to the affinity data from the audience research. Essentially, I'm saying, ChatGPT, show me your work and show me how it aligns to that. And now this is great. I'm actually seeing directly applies to the top skills and plain pain points that are in the audience research report. Better. You know what? I looked at these two. I don't think they're bad. Smarter, faster clinical trials without the regulatory head case. It's it's on point. You know what it's missing, though? It's missing it's missing that compelling emotional element. That's what I want from these positioning headlines. So so hang on. We can do even better by saying, hey. My website is clinically. Given the attached report, do the same thing. I am particularly inspired by Talia Wolf's new book, emotional I have so many copies. Emotional targeting. Can you use the framework concepts from that book to come up with positioning headlines that are backed by audience research data and emotionally resonant? Oh. Oh. You know what? I'm not saying I'm gonna use these exactly, but this is good. This is, like, very good for the ideation process, and it's so fast. Let me tell you, friends. What I absolutely love someone told me this, and it it just stuck with me forever. Have you been in a brainstorm meeting with let let's say Casey and Amanda and I get in a brainstorm meeting, and Amanda suggests a few ideas. And I think they're like, good and lead me to something better, but I don't love them exactly. It's hard. It's hard to be like, yeah, Amanda. Those are pretty good, but I don't love them exactly. Amanda has thick skin, and we've worked together for a long time, and it's fine. Right? But, like, it's very easy to take offense, and it's very it's also very easy to be like, oh, yeah. I don't wanna I don't wanna shoot down everyone's ideas. You can shoot down ChatGPT's ideas for a 100 suggestions, and it will never lower the quality of what it gives you. It does not care that you don't like any of its ideas. You can you can berate it and tell it to tell it to, like, change and it it will do just as good a job tomorrow. This is a very nice thing about programmatic software that human beings, you know, struggle with. I work a lot with Geraldine on ideas for the video game, and when I give her feedback that's negative, it's tough. But ChatChiBT can keep giving ideas until you find things that you like, and it's a this is a solid framework for this process. In general, when you can give it more context and suggestions like, hey. Include the framework for emotional targeting, the results get better. They they really do. And this is again, you have to have a human brain to do that. ChatGPT will not suggest to you, hey. Would you like to, call to mind this book that, you know, is very resonant and sort of, you know, whatever best selling in the category right now that might be useful? Nope. It doesn't suggest things like that because it that stuff isn't in its, you know, knowledge base or, you know, prompt system. Alright. Use case number two. This is a hard one. Show me news articles and stories that my audience is likely to be consuming. We've actually we've wanted to do this for a long time at SparkToro, like, to offer this inside the product so there's a section where you can see news. Still very difficult because of the speed of web crawling and how much data we'd have to store and process and all that kind of stuff, but you can eventually get there. I'll show you my process. So this is, CUNY, which is the the, what you call it? City University of New York, which my grandfather went to way back in the day for a couple of years. And they offer an MA in forensic psychology, which is not just used by the FBI. I know that's, like, what all the TV shows think, but forensic psychology has has lots of use case applications in health care and higher education, law and government. So I'm gonna say, hey. For that program, that MA program, I'm starting a newsletter, right, for applicants to the grad program. Given this, can you can you go find, like, news topics from the last thirty days that those people are likely to be interested in that we could cover in our newsletter? This this is a great use for for use case, and it'd take a bunch of time, right, to manually sift through. Okay. It'll give them to me. I I don't you know what? These are kinda I don't see the the connection is faulty. It's not it's not quite right, on these. And so instead, what I set what I did is since we're looking for temporal web crawling, I switched to o three, the reasoning model. It's more expensive, but, you know, if you're if you're doing manual queries versus, programmatic ones, the the the cost is not big. Right? It's not it's not many dollars. And after ninety one seconds of thinking and processing, it'll it'll show you its logic in here. Right? So, as it answers, it'll show me here are five timely story lines past thirty days that map directly to the high affinity topics in your SparkToro report on mental health assessments and advancements in ethics and forensic practices and emerging public policy debate. This is good. This is good. Yes. Yes. I bet a lot of grads grad students to the MA program in forensics psychology are very interested in the supreme court's decision around banning gender affirming care. This, Graham Flatman investigation from the, it featured a bunch of forensic psychology stuff, in the, the court's decision, AI assisted deception and large language model risk confront. This is this one I think this is the highlight for me. Right? Like, I would absolutely take this journal of American Academy of Psychiatry Law and and talk about that story in my you know, if Amanda and I were writing this newsletter, like, it's just ding, ding, ding. Bang on. And I love how they, you know, they tie it to here. They they, they'll give you the the the link in there. You can download it. Really nice. Product copy ideas. If it's good at positioning ideation, it's probably gonna be good at product copy ideas. I want you to get the sense here. I'm always asking it to give me ideas, not things that I would publish or use directly, and that's because of all the stuff we talked about at the beginning. So here's French florist. They actually make some they do a really lovely job. Like, they have a very selective process of which florist they work with locally, and they're very generous with their florists and blah blah blah. I like the I like a bunch of things about them. They they do really nice packaging. So I'm looking for people who search for the best flower delivery service. I upload that. I wanna create more compelling copy for this particular landing page, the the the berry pie. Because I don't know about you, but a beautiful way to bring a sweet mood to the room or table, the berry pie inspires a soft explosion of radiant feelings. It sounds AI written or at least, like, really cheesy human being written. I like French florists. They've been a long time spark to our customer. I don't wanna rain on their parade. Maybe they copy tested this, and I'm full of it. But, Amanda, like, that copy lands bad with me. Anyhow. Alright. Here's the attached research report. These are the web you know, all I I upload the whole thing. So it gives me a bunch of suggestions around features and benefits to highlight. And, honestly, while I think more prompt tinkering might get even better, this was really solid. Local florists you can trust, right, scores a 92 affinity. So we know that that's a deep interest of this target audience. And so they say hand arranged and hand delivered in LA by expert florist, never boxed or chipped. Oh. Oh. By the way, I don't actually know if that is true. Again, another reason is very important to review this copy before putting it on the page. But if that is true, if that's how, you know, French florist in LA works, hand delivered and hand arranged. Oh, yeah. That is that is compelling. It speaks exactly to the problem statement, right, that people are expressing, which we can see in the SparkToro data. This is good. Like, it looks really the emotion driven gift giving, all of them. Okay. Identify audience's likely objections to purchase. Gumshoe dot ai, is actually, founded by a a a friend of mine. Some of the guys that I play Dungeons and Dragons with every couple of weeks, are behind this company, and they it it's sort of like, it's not not that far off of what Moz was doing for SEO software back in the early days. Right? So they they try and show your brand visibility in all the AI tools, which is now a super hot space, but when they were when they entered it maybe a year ago, it was not that big. And, so for people who visit that URL, which I was, glad we had some some nice data around that, Can you identify potential customer objections? This is this is a common problem in b to b SaaS is identifying your customers' objections. We're gonna we're gonna talk about why AI might not be great for this in a sec, but just hang with me. How is it different from my existing analytics tools? How accurate is the AI generated insight? Can I trust it? Will this actually lead to business results? These are already useful. These are already things where if they haven't addressed this on their home page, I would consider it. I would bring it up in the meeting. I would confirm with, you know, whatever customer data that that these are things to solve. Oh, I think we can do even better. So now I'm going to, say, hey. I want you to specifically, in this case, focus on these four areas, topics, keywords, related questions, and YouTube channel tabs, which I do generally recommend if you can figure out which tabs which tabs data applies to the problem you're solving and upload only those one or two or just tell it to focus on those specific ones from the SparkToro data, the results tend to get better. We've done a bunch of testing around this for, for SparkToro itself and seen that consistently. I already have a tool for that. How is this actually actionable? Like, what will what will I change based on okay. Okay. Like, these I think this is equally solid in addition or maybe as a replacement. The that value add is between the real audience behaviors just like I've talked about on on examples one and two and, and the suggestions that it gives me. So we're get we're getting better. A big warning. Like, this should be flashing red warning. These are generalized objections. These are objections that a stereotype machine is stereotyping in the answers. So please, friends, right, if you get value from these and addressing them helps, that probably means that whatever hiring a CRO professional or working with your internal CRO team and actually interviewing and surveying your customers will get you more personalized, more salient, even more useful ones. We've done this process at SparkToro. We did it last year with, Asia Rangio from Demand Maven. Results have been quite good. So I would I would not I would not replace a human with this. I would add to or start the process, FYI. Brainstorm some content that's missing from my site. This one's a little tricky because it requires going and crawling the website and figuring out what's not there. But, this is Referral Rock, which is Josh Ho's company. Slight plug for, Spark Together. I met Josh, Amanda, because he came to Spark Together, and he was at my table for the mastermind session that you ran. And so, like, we got introduced and, yeah, really hit it off. He's a just a lovely lovely human being, great founder, runs this cool company. And so, anyway, I I ran their data through here, and I want ChatGPT to crawl through the site and find content gaps. And I I gave it some more. This should be informed. I I learned from my mistake. Right? So these answers should be informed by, in particular, the keywords related questions, topics, and interest tabs in the report. Alright. Yeah. These are these are not great. These are not great. And I think the problem actually is with the model choice because it did not it was not iteratively, going and seeing whether those topics were actually addressed on the pages. And so instead, I switched to o three, and then I had to switch to a two pronged approach, which was first, I need two tasks to accomplish. First, identify the most important content opportunities from the audience research report, and second, crawl through our website, figure out which of these items are already addressed by existing content. Oh, damn. This is this is dope. I I am impressed. I am impressed. Trona, I would not be surprised. I've heard from people that Claude is equally good or better, than ChatChaPT is for this, but this is this is solid. And what again, what I love is SparkToro is giving me the data about what my audience actually cares about and then, you know, making that sort of a fast process. The LLM is then surfacing those and and grouping them together into insights I can see. So referral program ideas and examples, choosing referral rewards and incentives, how to promote and keep a program top of mind for your affiliates, an affiliate program versus a referral program, what the difference is, blah blah blah blah blah. Right? This is coming from keyword data and interest data. Now, hey, here's the URLs on which that content exists, and here's the ones where it's missing. It's just nice. It's really nice. Okay. This is a common use case of SparkToro always, and it's always been done by hand, which is identify websites that offer opportunities for for some direct reach. So this is like, you know, PR person goes to SparkToro, and they're looking for, you know, the right journalist or or, analyst to to pitch their story to or pitch coverage to. Whereas a link builder who's focused on SEO might be looking for a place to get links, whereas, an advertiser might be looking for, you know, sites that offer ads. So I, Savannah Bee Company. I do you guys know about this? So this is, Tupelo Honey, which I I my only familiarity with it is from the Van Morrison song, probably like a lot of people. Apparently, it's a real thing. It's from these two there's only two small regions where, particular kinds of bees flower, particular kinds of plants that produce these. They sound ridiculous, especially if you are not American. The two regions are the Okefenokee Swamp, which I had only heard in Mickey Mouse cartoons, and the Apalachicola River Basin, which sounds like where Pepsi sells in Appalachia, but, no, they're real places, and they and they really make this incredible honey. I ordered some for Geraldine, and she was she was over the moon for it. It it is very unique because the sucrose to fructose ratio is different from every other honey in the world. Anyway, you can the science of it is fun and fascinating. What I like is I was able to take the URL, the specific product page URL for this product and plug that into SparkToro and then ask it, hey. I'm trying to do PR for my company, Savannah Bee. My goal is to pitch publications for inclusion of our brand or publication of guest pieces about Honey or even sponsored articles. So can you sift through this attached webs list of websites, crawl them, and classify them as follows? Ones that offer guest contribution, that have contact info for stories, that have offer advertorials, etcetera. And, yeah, then I'm gonna show you what can happen. Those of you who played with LLMs a lot know that this probably have have seen this before. I can definitely help you categorize those. But before I begin, could you please share the list of websites? What what what the f are you talking about? I I clearly shared the list of websites. Whatever. I don't care. I'll do it again. Please use this attached list of websites. Thanks. I currently don't have access to the actual URLs from the file. What? What is going on? I went and started a new chat. I did it again. I got the same response. What's happening? Well, I have found that four o does this a lot when it is overwhelmed by a task, especially in cases of crawling. This this has been my experience, like, probably half a dozen times or more. If you want four o to crawl, you have to give it very, small, you know, little ones. But o three, which is the reasoning model, the more expensive one, great job. Done a great job. Filing several sources and how to classify them. Boom. There you go. It gives me alright. These ones accept guest contributions and freelance pitches. These ones have story idea and newsroom contact pages. These ones have advertorial or paid content or media kit info. Thank you. Perfect. Exactly what I wanted. Obviously, I'm gonna have to do a whole bunch of handwork again, but this saved me a ton of time and custom crawl building and, honestly, amazing. It's not as good as a human being doing it manually. I will bring that up. Right? Like, if if I did it personally, I would get even better at this. Like, you know, I would do iterate more on the searches and, like, get more accurate contact pages. I'd probably find some that have opportunities that didn't fit exactly what I wanted, but still get the job whatever. If you're looking for a quick solution that will get you most of the way there, that's really nice. Design content outlines. This one's very simple and straightforward, for team sports, which, Geraldine and I are investors in. It's it's run by this great guy Francisco Baptista out of, out of The UK. And, you know, Francisco's built this amazing platform that helps coaches and players, mostly in amateur leagues, to manage their progress and track all sorts of things. You might have seen if you saw a picture or a video of me doing push ups without a shirt on, which is very embarrassing, that was Francisco filming me with the Team Sports app. Right? And it measures, like, your you know, looks at your arms and sees if you do it correctly and all that kind of thing. Anyway, so I'm writing an article for this website about goal setting for athletes and coaches. Can and and and I uploaded the data about, people who search for coaching app for athletes specifically in The UK because Team Sports is a UK focused one. And it gives me some this this is one of the worst responses I got. This is one of the most straightforward things that people do all the time with ChattGPT, and this is crap. It's practically a parody of an outline. That is I just I just thought it was so bad. And I I tried to figure out why that was, and one of the reasons is it didn't seem to do the tie to the data. So instead, I said, hey. Take into account those specific particular areas, and for each subsection, tie it to the relevant audience data from the spreadsheet and include an explanation of why it will help answer the audience's interests. Right? Or show me the connection between input and output. Outline with audience data justification. Uh-huh. Uh-huh. Wow. I I don't wanna say this perfect, but I think it is quite useful. If I were writing that article and I wanted to cover the topics, this this is really useful. It still requires review. Right? Like, no doubt. Like, I, you know, I would take this best coaching certification and not necessarily include, an answer to that question in my outline, but I would go research the certificate programs that recommend doing it and how they recommend doing it and then include that in my article. So helpful. It just, yeah, it doesn't get you all the way there. Does it give me a content outline that is a a a great fast start? Absolutely. And way, way better than what I was getting in this sort of generic AI slop plot. So sometimes you gotta ask CheckGPT to prove to you and connect up the data and double check it. Sift through where we got? Okay. One of those. Sift through and prioritize keywords and channels. So this is Jane Friedman. Hopefully, some folks here are familiar with her. She is an absolute legend in the world of book publishing. I I did a consultation with her before I published before I wrote lost and founder. She was hugely helpful in helping me, like, understand the world of publishing, how it works. She's got this great newsletter. She speaks at conferences. I've I've met her at a few she's wonderful. I think I wanna get her to spark together sometime. Like, she's just she's just fantastic. So I I looked at people who visited her website and specifically the search keywords, and what I'm asking for here is prioritization. Like, help me prioritize the ones that matter most. And so I said, hey. My website's Jane Friedman. I wanna identify the topics I could write about from these keywords that are gonna bring me the most potentially qualified folks to the site. Can you sort through the list of keywords, determine which ones are relevant to the audience that I'm targeting, and then prioritize that list from most to least likely to produce ROI? Here's a ranked list. Yeah. Yeah. This is not terrible. Not terrible. As I went through it, though, the individual keywords felt too chaotic. What I wanted was grouping, and I know from past work that LLMs are quite good at classification and categorization of words. Language classification is like the thing they shine at. And when I asked for this, great. Like, now I really get publishing and industry topics, poetry and literary forms. It it has a bunch of these different ones. I I was very I was impressed. I was quite happy with that. Give me ideas for ad creative. Should be good at this too. Well, you know so, this is the Seattle Pinball Museum. If you've been to Seattle, the pinball museum is a really fun place to take, either, like, aging gen xers who love it, who are, like, obsessed with pinball machines. You know, Bill got a Teenage Mutant Ninja Turtles pinball machine in his basement, and he cannot he cannot catch up to Rio's high his son his oldest son's, high score. He's like, I swear to god. We stayed with him for a few weekends and he was like, oh, I gotta beat Rio, and he can't he's, like, never been able to get close. Rio's just like a master of the the pinball machine. And, fun fact, by the way, Adrian, all Gen Xers and human beings are aging. There's there's no there's no reverse aging human beings, so take no offense. Alright. So I look for people who search for Seattle Pinball, and then I ask them about, hey. I'm gonna run some online and offline ads. Please relate these things back to the data points. Right? Do all the things that I that I know in the past. Unlock the classics, play pinball, win Seattle, retro postcard style. You know what's great? It's like, as you do this and you get better at the prompting, better at asking for all the right things, you get you get you get good stuff. Like, it this is pretty darn good. Often better than blank page brainstorming. That's that's, I think, the, the best use case I can I can describe? Alright. Last one. Audit an existing piece of content. Should should be quite good at this. Right? I I think I I might suggest that, you know, specialized tools can be even better. A lot of folks like Grammarly. But, our friend, Scott Heimendinger, who runs Seattle Ultrasunks, have you guys seen this? I kinda wanna show you. Can I show a video on here? I think maybe I can. Let's find out. Hopefully, you can. I don't know if you can hear me. On it. Oh, no. Quantified. Okay. So cool. So there there's this page. You guys can see that? Yeah. Oops. What happened to oh, no. Did I close Goldcast? Oh, Rand. What did I do? Amanda, are you still there? I'm here, but you can't hear me. We're all here. Everyone else can hear. This is so awkward. Where's Amanda? Oh, my speaker's cut out. Wow. Wow. Can you hear me now? Panic. The panic is real. Wait. Can you hear me now? Yes. Now I can hear you. That was why I wasn't talking. I gave up at some point because I was like, you know what? This is fine. Okay. Okay. Watch this part right here. There's no audio on the video though. But we can we don't need the audio. Right? Sorry. There's no audio? No. But the video works. Okay. The video works. Alright. So what he's showing here here. I'll mute it, and I'll I'll walk you through it. So what Scott is showing here, this is like a knife influencer showing how crazy sharp his knife is. But as Scott explains, the way that this is measured is by, passing a knife through a filament. And yet, the big problem is no one gives a shit how well their knife cuts a filament. Right? That is not that's not what's important to you in a chef's knife. What's important to you in a chef's knife is how it cuts food. And so Scott built this, arm, this robot arm that measures and tests by cutting through real pieces of food. You you can see the robot arm here, and then he's got a, like, a device that that measures that he he hand built this all in his laboratory. This is this is in his apartment, which is incredible. He uses the second bedroom as his lab. And he yeah. So he's he's measuring the the, density. This is like the best content of this kind on the Internet by a mile. Absolutely the best. And yet, this is the article about it. Are you seeing the difference between how much work went into him producing the absolute best thing and no offense. But this is the result. Right? Like, this is the result that he's got a food cutting rank. It shows edge retention. It shows best sharpness, which is the the filament, and then the the price. And you can see, right, these are all the best selling chef's knives on the Internet. Holy gods. The amount of work that this man put into this project, and then this this is the output. Like, I I'm worried. I'm deeply worried that this no offense to Scott, but that this is gonna feel really more anyway, so here's what I said to, Is Rand gone? Yep. I just hit the button on my browser rather than hitting, play on this presentation. Okay. That was fun. That was fun. I'm trying to make can you can see and hear me. Yes. You can see and hear me. I'm an agent Gen Xer friends. Like, you I think I'm actually at the very youngest end of the Gen Xer cohort, but it doesn't matter. I'm trying to make my article about the scientific process that we use to measure noise more compelling. Here's the article. Here's the audience research report for our target group's behavior. Using the data, can you make recommendations on things that are likely to improve the resonance of the piece with this target customer audience? And you get I not gonna lie. Like, in this case, I don't I don't think the prompt needed all of this. But look, key takeaway is like, oh, oh, yeah. I really, really appreciate this. And in fact, in the future, when I ask chat g p d, I'm gonna specifically say, before you do give me recommendations, please give me the key takeaways from the audience data. I think that's hugely huge. Oh, you can't see that? Yeah. We're not yeah. Your screen is I didn't share the thing. Yeah. Yeah. Well, you know, look, I I'm a aging gen x or I need to share the entire screen, that screen. Someone said, at least this proves that you are not AI. Oh my god. Right? Well, I mean, maybe it proves I am AI because as you can see from this this presentation, it makes as many mistakes as a human, if not more. So, like, I just my god. Okay. Here we go. I'm gonna hide this. We we we get this really useful thing, and then the recommendations to improve the article, why it matters to, you know, various audiences, including chefs, which are one of the the more, salient audiences from the audience research data, nice skills and usage, context, like when when would I be using a knife to do these kinds of things, when the QKP shows a 12% edge retention, what does that help me accomplish? I actually agree with a bunch of these video and social enhancements, which Scott has done some of these thirty seconds, sixty seconds shorts, that have done really well. If you follow him on Instagram, you can see a bunch of these beautiful videos of the knife kind of do you know he had to design a camera design and build his own new camera technology to be able to visualize the knife blade edge successful? It That guy's insane. By the way, Amanda, he's speaking at Spark together. He's gonna bring yeah. He's gonna bring the demonstration and show it. There'll be an overhead camera on top of his setup so that you'll be able to see it from the audience and, like, look at the anyway, it's just That's gonna be cool. Yeah. Friend, I was gonna say, the nice guy, Scott Heimendinger, will be at Spark together this fall. So if you haven't bought your ticket yet, buy your ticket and come in person. It's gonna be awesome. Yeah. Yeah. And also, because it's in person, not recorded, he's gonna share a bunch of very interesting stuff, including about the tariff situation Oh. Which I'm sure will all be fascinating. Oh my gosh. Yeah. So these 10 things are this is far from the all the list of things that you could do. You can do so much with this data plus request of things. You have to be careful. You can see how many times in my process here, and I didn't show you close to all of them. But in this process, like, how many mistakes you're getting made, how many imperfections, how many things require hand review, these LLMs as advanced as they might seem, they are nothing close to, hey, Amanda. Can you take care of this? Like, you or I assigning work to each other, we will just do the right thing. And LLMs are not that. They are not human beings. I don't the like, there's it's not close to AGI. It's very far off. If you wanna take next steps, what I would urge you to do, Britney Mueller has this actionable AI for marketers course. It is extraordinary, well reviewed. Yeah, bunch of people who have taken it, loved it, and and are raved about it. Not very expensive. She's offering it again in, at the June and July and then October. And, hopefully, Casey has left because if he's still here, he might not be thrilled that I'm showing up. So this is the new take action section of SparkToro, where we're taking the prompts that we think have produced the sort of best things, and and we will be in the future probably, launching this section called take action where on any given report, you can do a bunch of the things that I showed right inside the product rather than having to export, go to ChachiPT, tell it which things to upload. Casey's done a really nice job of making sure that it sends the passes the correct data from oh, we only want keyword data here, or we want audience interest data here, or we want website data here, or we found that website data doesn't perform as well as hidden gem websites, so we uploaded that. And so you can get he's I see Katie's comment. Shit. Oh, no. Anyway, I I don't know when this is launching. I won't even promise it, you know, by the end of the year. But, eventually, this is a direction that we know because a lot of you are doing it. We should provide it for you, and that, yeah, that that should be quite quite powerful when we get there. Still in the early stages. Please don't share. Alright. Q and a. And I also wanna mention, I know we talked about SparkToro Together briefly, but, we will have an in person SparkToro training session that's coming in, the the afternoon before SparkToro Together. So if you arrive in Seattle early, you can come to that training session, and join us for it's about two and a half, three hours of, like Mhmm. Here's a bunch of use cases of SparkToro. We've we had a lot of people ask about that, and so even though I feel deeply uncomfortable talking about product stuff, I think it it probably makes sense to do that. And, we're gonna It'll be good. It'll be good. It'll have the training, then we can all walk over to happy hour together. Exactly. Exactly. And, and obviously, you know, SparkToro is offering a lot more than it has in the past, these days. So lots lots to take those through. Yeah. Friends, I think we have time for a little bit of q and a, which I'm gonna let Amanda run for us. Yeah. So, yeah, friends, if you enjoyed the, chaotic energy in this webinar today, spark together is kind of that, but it's also quite purposeful, and it'll be a lot of fun. You can partake in the chaos. Speaking of chaos, I'm gonna open my door because my office is now one bajillion degrees. Okay. Okay. We have we have a couple questions in the q and a here. I'm gonna read Nicholas's question. Is this webinar low key how you're trying to optimize SparkToro to appear in LLM search? Getting hundreds of people to plug in SparkToro data so it learns that SparkToro is a great recommendation for audience research. Is that no. Fascinating. Yeah. That's a really good question. But didn't occur to me. I think you know what? I would have to con that's a fantastic question. I would have to confirm whether people inputting prompts with SparkToro's audience research data would get ChatGVT to recommend us more if they don't also see mentions of SparkToro, sort of on the Internet, right, in their in their corpus that they train off of? I think the answer might be no. I I believe ChatGVT is sensitive, and all the elements are sensitive to that sort of attack vector, right, spam vector. Mhmm. Because conceivably, you know, if you were trying to promote your, I don't know, Savannah Bee Company Tupelo Honey, you could just get people to be like, hey. Can you go find me the best price on Tupelo Honey from Savannah Bee Company on the Internet? And you get, like, a 100 customers of yours to do that, and then, like, you know, are you gonna rise up there? I think my understanding is that the LLMs have protections against that, and they'll confirm before they just start using that data in their training. Alright. Cool. Quick question from Kristen on recommendation of using ChatGPT's o three versus four o. Yeah. I mean, do you want, like, the the, like, why to use one versus the other? Yeah. I think so. I think Yeah. Yeah. So, actually, ChatGPT's website has a really great breakdown comparing the two models and telling you what they're good for. Mhmm. In general, you wanna do iterative crawling. You wanna do tasks that require sort of other tasks. And what what it calls reasoning is essentially, like, more layers in the network. So this is, send this response, figure out, like, essentially takeaways from that, then, do do other tasks based on the takeaways versus do the one thing that was asked for directly. And that's, you know, four o is great for most individual tasks. O three is that iterative process layer. And, again, like I said, I I have found four o really falls apart if you ask it to do lots of crawling, and o three tends to be much more capable of doing that iterative multi crawl process if you need a lot of different URLs to be visited and data confirmed from, multiple retrieval augmented generation you know, the the, basically, use ChatGVT to be your search engine, do stuff for you. And and again, both retends to be superior. I also wanna point out I I am I am not Britney Mueller. I am not a, like, deep expert on this stuff. I have used these, and I have learned from a lot of you all who have shown me amazing things in here. But, yeah, I, I I would trust people like that to have more comprehensive answers. Mine are gonna be very limited to the kind of SparkToro and marketing use cases. Yeah. And Britney's Britney's course is really great. It's also very highly rated on Maven. And she she's really, really good at making all this just really accessible, to people of all levels. So great course. Yes. Rob also pointed out that the cost the token cost significantly higher. I wanna say, I can't remember if o three I wanna say it's, like, two orders of magnitude, like, a 100 x the cost per request than four o. Might not be that much anymore, but, yeah, it's it's a lot higher. A lot more computing power, reasoning power. Yes. Yeah. Okay. Candace is asking about, past office hours, how to become a super user of SparkToro. Candace, if you're still here, check our website. We have, we put up all of our office hours on the office hours page. You can find it in resources. We also in our in the SparkToro dashboard, we have tons of, like, articles and video tutorials, on how to get the most out of the tool. So definitely hang out there. My favorite one is, if you go to I guess I have to go to a manual here. Hey. We should link to the video session from we're so good at this. How do I get to the help center? No. That doesn't have it. Product FAQ. Is that it? No. About oh, it's in resource. Okay. Yeah. Do you want videos? There we go. I've got it. This is the one. Yep. You beat me to it. Oh, yes. This presentation yeah. Well, we'll get this on the website too. A couple other questions. Marcela asks, when is take action going to be available? Casey, you wanna you wanna jump in? No. I honestly, Marcela, we no time line yet. There is so much to do. You know, obviously, since this makes calls to an LLM, we we wanna be very confident that we're giving good results. Right? I think even though people know will know that it comes from LLM connected data, I just we don't feel great unless all of those sections are giving you really good stuff. So I'm gonna be sharing all these prompts. You can iterate on them. I I fully expect that we're gonna be iterating for a while before this comes out. I would not expect it in the next two to three months. And I like I said, I don't know. It has to write all the prompts. I no. I wanna find I wanna find someone who's better at it. I like human beings. That's one. Okay. Okay. And then finally, Andrea asks, what level of subscription of a SparkToro do you need to complete the tasks outlined in this presentation? You you can do it at the lowest level. At the higher levels, you'll get more data. So, like, you know, if you're at the $50 a month plan even if you're on the free plan, you can export some of this stuff and upload it to LLMs and get some data. So free users, you you can get some value from this process too. The $50 month plan is probably where it starts to get quite interesting, and you'll have, you know, a certain number of rows per different data type. You can see all of those on the, on the plan breakdown page here. But, yeah, basically, you know, those top 50 results for search keywords, that might solve your problem in a lot of cases. Fifth top 50 websites visited. Again, it might solve your problem. And then, hey. I I'm getting a bunch of value from this. I want more. Great. You can you know, if you go to the the business plan, you'll get, 150 results per search and at the at the highest tier, the 300. So more data, but generally speaking, you can start to get value even at the 50 results. Yeah. And I guess worth saying too that if you want conversational queries, that is in the agency plan. Yeah. But although that one, I think, will roll out sooner. That's probably in the next two to three months. That will probably go to everybody to describe your audience type of feature. That will probably go to everybody in the next maybe maybe two to three months. Could be four. Could be five, but around there. Yeah. I don't know. Casey, what do you think? Oh, friends. This is great. Amanda, thank you for, nudging me to do this. I think it was it was a ton of work, but, it was very ton of work. So much work. Thank you, Rand. Yeah. My pleasure. My pleasure. And thank you all for joining us. Thank you. Apologies for my chaos Muppet energy. I'm in a place. I will, I'm sure I'll recover soon and and be better. Alright. And I'm gonna get I'm gonna get these prompts. Someone commented, I think, wisely, that it'd be also useful to have this as a, like, list with a bunch of screenshots. I can publish that on the blog, and then we can, get that. Yeah. I wonder if ChachiVT could turn my presentation into a probably not, sadly. But we get the transcript or I don't know. Actually, probably not. I don't know. Anyway, thank you, friends. Thank you for joining us. Thank you. Thank you for the follow-up email. Bye. Bye.