You export a month of comments, open the spreadsheet, and start scrolling. Two hundred rows in, you have a vague sense that people are annoyed about shipping and a much stronger sense that you are not finishing this today. Your listening tool counted the mentions. It never told you which complaint deserves a meeting.
Brooke Sellas, CEO of B Squared Media, shared her screen for this episode and ran one month of a kitty litter brand’s social conversations through Claude. 1,902 incoming messages went in. Back came a themed breakdown, a ranked list of the complaints doing the most damage, and the questions customers keep asking that map onto next month’s content.
Run the same sequence and your monthly report changes shape. Rather than handing a client a pile of grievances, you give them three friction points ranked by how one-sided the sentiment is, with the verbatim comments attached so anyone can check your work.
What this episode covers
- Sentiment scoring flags a comment as negative and stops there, so it cannot separate a pricing objection from a shipping delay or a rude support reply.
- Claude stripped spam out of 1,902 incoming messages down to 1,829 real conversations without being asked, because Brooke has run that same cleanup on past exports.
- The biggest theme is often the least useful: odor and ammonia drew 249 mentions, 13.6% of all conversations, and split 49% positive against 40% negative.
- Ranking by one-sidedness pushes dust up the list at only 66 negative mentions, because 55% of every dust mention is negative.
- Asking Claude to show the verbatim comments and their row locations next to its own summary turns an unverifiable claim into a spot check you can finish in a minute.
- The FAQ pass returns questions ranked by frequency and matched to a format, including a competitor comparison page prompted by repeated “how does it compare to” questions.
Timestamp index
- 00:00 Why manual listening reports miss the real questions
- 00:51 Brooke’s undergraduate thesis, done by hand in a spreadsheet
- 02:21 A negative score never tells you whether it is price, product, or support
- 03:26 Natural language processing is what makes this work now
- 05:07 Pulling one month of conversations out of your inbox or analytics
- 06:09 Loading the CSV into Claude and writing the context block
- 07:45 What counts as a meaningful sample when you get ten comments a month
- 09:08 1,902 incoming messages, 1,829 left after spam
- 09:50 Odor and ammonia at 249 mentions, and why a coin flip is worth digging into
- 11:05 Starting with hidden buyer objections rather than the 82% praise
- 12:19 Ranking by one-sidedness instead of theme size
- 14:17 The discount double-check
- 15:29 Building the citation request into the first prompt
- 16:02 Running a what we heard loop weekly with the team, monthly with the client
- 17:39 Zero-click marketing and the carousel that answers a dust question
- 19:05 The FAQ ranking, and the competitor comparison page hiding inside it
- 20:57 Why Brooke ran this in a plain chat instead of her Claude project
- 22:01 Skipping context, accepting the first answer, and hunting for agreement
- 24:12 Six months of hand-coding, now a couple of hours
- 26:08 Where to find Brooke
Your listening tool counts. It does not explain.
Brooke did social listening before anybody sold software for it. For her undergraduate thesis she sat in the library with a laptop, watched three brands she had no access to, and copied every public conversation into a spreadsheet. Then she went row by row asking what the person said, what the brand said, and what sat underneath the exchange.
Tools arrived and took away the copying. They did not take away the reading. You set up keyword alerts, mentions land in a dashboard, and someone junior works through several hundred comments hunting for patterns. Brooke is blunt about the ceiling: the tools count how many times somebody mentioned your brand, and they stop before telling you why that person was frustrated or whether they were close to buying.
Sentiment scoring narrows it a little and then leaves you stranded. A comment marked negative might be a pricing objection, a shirt that arrived with no buttons, a shipping delay, or a support agent who was short with someone. Four different problems, four different owners inside the business, one label. Brooke’s phrasing for the gap: you end up with a mountain of raw conversational data and no scalable way to get from the what to the why.
Export the month, then spend your effort on the context block
The export is dull, which is the point. Open your social tool or the native channel analytics, go to the inbox or messages report, and pull one month of conversations into a spreadsheet or CSV. Brooke’s example brand is a sustainable kitty litter company whose comments arrive on their Facebook page and on the ads they run. Every platform names the report something different. The shape of the file is what matters.
The work starts in the prompt. Brooke keeps a prompt library and pulls a saved one for this job rather than typing a fresh instruction each month. Her context block names the industry, the product, the buyer, and the date range before it asks for anything: the data comes from a kitty litter brand selling to discerning cat owners, and it covers comments, DMs, and replies from July 2026.
Only then does she give the order. Categorize the conversations by theme, and for each theme report how often it appears, the emotional tone, and what the customer is trying to accomplish. That last clause is doing the heavy lifting. Asking what somebody wants is a different question from asking how they feel, and it is the one that produces something a product team can act on.
If your brand gets ten comments in a good month, Brooke’s advice is to ask Claude what a meaningful sample looks like for the specific thing you are trying to learn, then tell it what you actually have. It will work with a small set. It will also tell you when the set is too small to support the conclusion you wanted.
Read the breakdown, then distrust the biggest number
The first pass took a few minutes because the file was large. Claude reported 1,902 incoming messages, stripped the spam without being told to, and came back with 1,829 real conversations. Brooke flagged why the cleanup happened on its own: she has run this analysis often enough that her instance has picked up the habit. On a first run, you should ask for it.
Odor and ammonia performance topped the themes at 249 mentions, 13.6% of the conversation, splitting 49% positive against 40% negative. Brooke called it the biggest conversation driver and a coin flip in the same breath. Some customers credit the litter with finally killing the ammonia smell. Others say it made the house worse, especially once it got wet. Several of them compared the product directly to Fresh Step and World’s Best inside those same comments, which hands you a competitor analysis you did not ask for. Ingredients and sourcing came second, price and value objections third.
Then Brooke asked for the top three negative complaints worth addressing, and the ranking rearranged itself. Claude sorted by raw negative volume and by how one-sided each theme was, not by theme size. Price and value objections led with 100 negative mentions, 17.2% of all negative volume that month. Dust came second on 66 negative mentions, because 55% of every dust mention is negative and no theme in the data is more lopsided. Clumping came third on 53, flagged as the most reputationally dangerous of the three.
Sorted by volume alone, dust and clumping sit well below a 249-mention theme that is nearly half positive. Sorted by one-sidedness, they are the two problems you take into the next client call.
The discount double-check
AI finds patterns that are not there and reads sarcasm as sincerity. Brooke’s answer is a spot-check she named after a grocery store game: the discount double-check. She asks Claude to bring up the actual comments it flagged, displayed alongside the top three complaints it just ranked. It returns each message verbatim with its location in the file and a note on what to know about it.
She can tighten that further by naming the row and column she wants. Her export carries hot links out to the live conversation on social, so verifying a flagged comment takes one click rather than a search. Mike’s follow-up is the version most people should copy: put the citation requirement in the first prompt. Ask for three real comments per category with their row references, and the evidence arrives with the analysis instead of a round later.
One small warning from the demo. When Claude surfaced the verbatim complaints, it offered to draft a reply to the customer, and Brooke turned it down on the spot to stay in analytics mode. Helpful suggestions are the easiest way to lose an afternoon.
Questions customers keep asking become the content calendar
Nobody is clicking through to read your explainer. Brooke’s route around that is to answer the question where the question was asked. She pulled the FAQ ranking for July straight from repeated questions in the file, and Claude labelled it a strong zero-click content list.
The list reads like a content brief because it is one. Does it work in my automatic self-cleaning litter box, which Brooke recognised on sight as a question her team fields constantly? How much does this actually cost? How does the subscription work? Does it clump? Does it track outside the box? How does it compare to the competition? Does it actually control odor?
That comparison question is where the pass earns its keep. Brooke’s immediate read was that a competitor comparison landing page listing every brand customers name could pull traffic away from those competitors. The FAQ pass turned a support burden into keyword research, and the customers supplied the competitor list themselves.
Claude closed with a content-ready summary that pairs each question with a volume and a best zero-click format: a compatibility graphic for the litter box question, a cost breakdown carousel for pricing, a short demo video or GIF where a picture settles it. Mike also points listeners toward a companion episode with Christopher Penn of Trust Insights on pulling the same kind of signal out of Reddit, which lands in the same batch as this one.
Where the strategist still earns the fee
B Squared runs what they call a what we heard loop, weekly inside the team and monthly with the client at reporting time. They collapse the themes into good and bad, name the top three of each, and bring suggestions for every friction point rather than the friction points alone. Brooke is clear that clients often take none of the suggestions. Arriving with problems and possible fixes is still what makes the agency worth paying.
She also declined two shortcuts on camera. She ran the whole demo in a plain Claude chat rather than her social listening project, because a project would have brought its own assumptions to a live analysis. And she pushed back on her own profession: marketers reach for AI to confirm what they already believe instead of finding what they do not know, and the insights worth having are the ones that surprise you.
Those two ideas sit in tension with her first rule, which is to load the prompt with context or accept slop. Every sentence of context you add is also a steer. Holding both at once is the skill.
On results, Brooke gave the number she could stand behind. Hand-coding a study like this used to take her six months. It now takes a couple of hours, and the demo you watched ran in minutes because it covered a small slice of a deeper dive. Downstream, she points to faster responses to emerging issues, social care playbooks built on conversations that happened rather than the ones the team assumed, and ROI showing up as fewer repeat complaints and better CSAT and NPS scores for the brands they run care for.
Full transcript
Mike Allton: Welcome back to the Lab. I’m Mike Allton. Manual social listening reports often miss the real customer questions. They’re buried in the comments, in the care tickets. Today, Brooke Sellas, CEO of B Squared Media, shows us how to feed raw conversational data into AI to uncover net new content strategy and audience insights. Hey, Brooke, welcome to the show.
Brooke Sellas: Hey, thanks for having me. So happy to be here and talk to you about one of my favorite subjects.
Mike Allton: I am so thrilled to have you here. Brooke, for years, I think social listening to many folks listening might have meant spending hours reading through comments manually or staring at keyword dashboards that just didn’t really tell why people were buying.
When did it click for you that you could use AI to do conversational research at scale?
Social Listening Before the Tools Existed
Brooke Sellas: Well, about a thousand years ago when I did my undergraduate thesis, I actually did social listening before social listening was a thing, and it was literally, literally, y’all, I want you to picture this, me sitting in the library with my laptop, like, looking at three different brands and taking every conversation that I could see publicly, right, I didn’t manage these pages, and putting them into a spreadsheet, a very long spreadsheet, and then trying to go through and find out what did the person say?
What did the brand say? What was the why or the intent behind what the person said? All of those things. Fast-forward to a few years later, we got social listening tools, which are amazing, but ultimately, it was still a big volume game. So, you know, in these tools, you can set up keyword alerts; you can pull mentions into a dashboard when someone mentions your brand; you can set up keywords that are, you know, when people don’t mention you, are talking about you, then someone on the team, usually someone junior, ’cause this is a very, like we’re saying, tedious task- yeah… would manually read through those hundreds or thousands of comments and try to spot some of those patterns, intent patterns, sentiment patterns, things like that. And basically, the tools can tell you how many times someone mentioned your brand, but they couldn’t tell you why someone was frustrated or what they actually wanted or if they were ready to buy.
So you’d have a mountain of this raw conversational data and no scalable way to make sense of it or to get from the what to the why, sentiment scoring helps a little bit, which we can always talk about if you want to, but just because the AI inside of the tool scores something as negative, that doesn’t tell you if the customer’s angry about pricing or if your shirt that you sent came without buttons or if it was a shipping delay or if it was the way they were treated by support.
The nuance has always been stuck behind a massive manual effort, and that is just no good.
Mike Allton: Yeah, and I think that’s why a lot of times, I think when people think about mining social data, they’re usually thinking about really expensive enterprise software, data science teams, ’cause mountains of data seem like it’s just too much for one person.
So what makes this conversational approach different, and why can everyday marketers do this now?
Why AI Reads Conversation Better Than a Dashboard
Brooke Sellas: I think that AI is certainly one of the reasons that has helped us, right? Because AI, think about how AI exists. AI is trained on NLP, natural language processing. So what does it do the best? Natural language processing, which means all of these conversations- that’s natural language.
Think about Reddit, right? Which is literally the number one, uh, place where AI trains because it’s just conversation after conversation after conversation happening there. So AI can go through and, in seconds or minutes, depending on how large your data set is, can go through and get to the why behind the what.
Like, yes, they were angry, but let me pull out all of these trends and patterns, and now all of a sudden you’ve got all of your negative conversations, and the AI’s being like, “Hey, here are your top three things. You gotta fix the website. People are pissed about that. They don’t like that the sale was, you know, a 24-hour flash sale, and, uh, yeah, your customer support sucks.”
So when you know those three things, now we don’t just have a mountain of data. Yeah. We have a mountain of actionable data. We can do things that create outcomes with this information versus just spending hours and months and days filtering through things.
Mike Allton: Love it. Love it. So let’s actually take a look at your screen so you can walk us through some of this stuff. And while you’re getting that set up, uh, I love that you mentioned Reddit, ’cause we do a deep dive with Chris Penn into how to actually pull ideas for content, for Q&A, that sort of thing, out of Reddit, what other people are talking about, and how you folks listening at home can actually use AI to help you process all that data.
But Brooke, walk us through how you pull and prep your actual, like, raw conversation, customer care data before you ever, you know, actually give it to an AI tool.
Brooke Sellas: Okay. Don’t mind all of those tabs that are open. But you’re going to go pull your conversational data. So whether you have a social media tool like Agorapulse or if you just use your channel natively, you’re going to go into your analytics, and you’re going to pull a report.
It might be in your inbox; it might be your messages. It, you know, every platform says something different. But you’re going to pull in this example. Fine, but it could be “I’m going to use one month’s worth of data. You’re going to pull… all of those conversations into a spreadsheet or a CSV, and I’ll just open mine up real quickly and show you.
We’re not gonna convert. So I’ve got all of the information that I’ve been able to pull from this particular client here. You can see these are all of the conversations that happened, whether it was a comment on their Facebook page or an ad comment, so on a comment on an ad they were running, right?
Loading the Export and Setting the Context
Whatever platform you use, they have different names for it. I’ve pulled all of the conversations. You can see they get a lot. This is like a pet brand, a kitty brand, and I’m going to save that. I’m gonna pull that into Claude, and then what I’m gonna do is I’m going to give Claude as much context as possible. I’m using Claude.
Whatever LLM, large language model, you’re using, you’re gonna need to give it context. We like to save prompts that we use over and over again, so I’m gonna pull something that we have in our prompt library around this. Um, you can see the prompt here. I’ll just read it real quick. “You’re analyzing customer conversations from a kitty litter brand’s social media channels,” right?
We want to give as much context as possible. “This brand sells kitty litter to discerning cat owners.” It’s a sustainable kitty litter, maybe. All right, so I’ve added that context in. “The following data includes comments, DMs, and replies from July 2026. Your job is to help me understand what customers are actually saying, feeling, and needing.
Categorize these conversations by theme.” Now, so I’ve given all the context; now I’m giving it the action, the execution order. “Categorize these conversations by theme. For each theme, tell me how frequently the emotional tone appears and what the customer is actually asking for or trying to accomplish.” See how I’m trying to go from the what to the why with this kind of prompt?
So let me just… I’ll just copy that in. So hopefully you have a prompt library. You would take that from your prompt library. You would copy it in with your, um, data, your raw data. We’re gonna hit enter, and we’re gonna see what Claude says.
Mike Allton: Now, Brooke, quick clarifying question again while we’re thinking, while we’re waiting for Claude to think. I know with this brand, as you demonstrated, they have a ton of engagement, and that’s probably just one channel, which is great. That alone might have been a quick takeaway for folks listening. You might not have realized you can download your social engagement, your comments, and your chat history, and use AI to help you with it.
But my question, real quick for you, is: what’s a good sample size? I mean, with this one, one month was fantastic, but most people, like myself, I don’t get that many comments in a month. Sure,
Brooke Sellas: Sure.
Mike Allton: Should I be looking for 50 to 100 comments to get any kind of reasonable, you know, data set, do you think? Or is there a threshold you have in mind?
Brooke Sellas: I mean, I would also prompt my AI, right? Sometimes whatever you have is what you have. Yeah. But I would also ask the AI, “What’s a meaningful sample? This is what I’m trying to do. What’s a meaningful sample?” And then it can say, “Well, I need 1,000 comments.” And you can be like, “Oh, you’re so cute.
I have, like, 10.” “And in my best month, I got, like, 15.” And the AI will work with you, right? And honestly, it’ll be a lot easier. This is taking a while because, uh, this is a lot of data. This client happens to get this many comments. But one thing that I would say is you can take this above and beyond by then, once you see these themes, and we can prompt this, uh, Mike, if you want to, and show everybody, and we can say, “What, what’s the good stuff, right?
The Theme Breakdown, 1,902 Messages Deep
How do I create more of that? What’s the bad stuff? How do I make this actionable and fix this friction for buying for this audience? So here are the results real quick. I dug into the full inbox, so they did have 1,902 incoming messages, and after stripping out spam, which it did for me, we try to clean up the data as much as possible, but the nice thing is, without prompting my AI, it stripped out the spam.
Now, I do this all the time, so it’s learned through machine learning that I want to strip out the spam, so you might want to have that conversation with it. So it brought it down to 1,829 real conversations, and then it’s gonna give me the breakdown. Odor and ammonia performance: 249 mentions. That’s 13.6% of the conversation, and the tone is split to 50, or 49% is positive, 40% is negative. This is your biggest conversation driver, and it’s a coin flip. People either credit the litter for finally killing the ammonia smell, or they say it made their house smell worse, especially once wet. Right? So that’s something to dig into.
Several compare it directly to Fresh Step or World’s Best on this exact point. Now I have competitor things that I can do with social listening. I can go look at Fresh Step. I can s- download and grab what they have publicly and run an analysis on that, right? Then they tell me, number two, ingredients and sourcing.
Three, price and value objections. So it’s gonna give me the percentage of these conversations weighted out, and look at all of this. I mean, this happened in- it did take a few minutes, right? ‘Cause it was a lot of data, but do you understand how long it would’ve taken me personally to go through 2,000, those 2,000 conversations and pluck this out?
Mike Allton: So what’s next? So you’ve got this data now inside your AI tool. What are some of the specific questions or follow-up prompts that you’re running that help you uncover maybe, like, some hidden buyer objections or content ideas?
Brooke Sellas: Yes. I love that you said hidden buyer objections, because that’s the first place I like to start, right? It says at the bottom, right, “The pattern underneath it all is that your product love is real and loud. 82% of these conversations are praise buckets,” right? It’s, it’s great. We’ve got it. I’m gonna mine that for UGC and all of that love, but I wanna start with where I can actually make change, and that would be, like you said, what’s the intent?
So, digging into the negative, what are our top three negative complaints worth addressing, right? This is the number one thing. So it’s gonna tell me the top three things that happened last month, the top three topic themes that were complaints, to help me address those complaints. Now, it’s gonna come back with an answer.
It’s gonna run this command, but it’ll take a little bit of time, and you want it to take a little bit of time. Like, we’ve gotten a little greedy with AI. I think we expect answers in, like, one second. This is not an easy prompt. We are… It’s going through and looking at, again, all 1,902 of those conversations to give this back.
Ranking Complaints by One-Sidedness
So then it comes back, and it says, “Ranking by raw negative volume and how one-sided the complaint is, not just the theme size, here’s where I’d focus. Price and value objections, 100 negative mentions, which was 17.2% of all negative volume for this time period, which was July.” And it goes, and it gives me some data. Dust: 66 negative mentions. 55% of every dust mention is negative, the most one-sided theme in the data, and then clumping performance: 53 negative mentions, and this is the most reputationally dangerous of these three. So now I’ve taken all of these conversations, which by the way, a lot of you don’t do anything with from social data. But now I’ve taken it and now I can do something with this, and I can certainly do it on the positive side, but I know that change and trust and all of those things, loyalty, comes from fixing the friction points.
So, that’s the first place I’m gonna start is to fix these friction points.
Mike Allton: It’s fascinating. And, and to your point, you know, we, we joked about the volume and how maybe, you know, like me, we don’t have this kind of volume, but people do have, you know, representing brands or clients with hundreds of comments every single month, and it’s just too much for a human to comprehend this level of repetition, theme, categorization.
We just don’t have that capacity as human beings. You could have put these in a spreadsheet and tried to sort them yourself. This is probably what you were doing years ago, I imagine.
Brooke Sellas: 1,000 years ago, yes.
Mike Allton: 1,000 y- I did not say 1,000. This is what you were doing a couple years ago. But I also know, ’cause you kinda hinted at this a moment ago, that AI can sometimes find patterns that don’t actually exist, misread sarcasm in comments. How do you audit exactly all this output it just gave you and know that you’re making decisions based on real customer sentiment?
Brooke Sellas: We call it the discount double-check. And what our discount double-check is: we act like we’re in a grocery store or like in one of those grocery store games. It’s time for the discount double-check.
We go through and we spot-check. We go through, and we’ll say, like, “Hey, bring up the actual comments-
Mike Allton: Yeah …
Brooke Sellas: that you’re, you’re flagging here alongside my top three, um, complaints.” And then it’s gonna run that command. But basically, what I’m gonna do is when it brings this back, it’s going to show me, right?
And I can get more specific, too. I could say, like, which row, which column, all that kind of stuff, and I’m going to double-check that. I’m gonna go through, and I’m gonna make sure that what it’s flagging is right. So it’s like spot-checking. So I’ll go through; here’s the message verbatim. It tells me where it is.
It tells me what’s, what are some things to know, and it’s gonna give me some more information. So, and then it’s saying, “Want me to draft a reply for this person?” Well, no, I don’t want that. It really is helpful. It’s gonna try to bring you down so many rabbit holes, but no, we’re, we’re in analytics mode.
We’re gonna stay in analytics mode. We’re not gonna go down those rabbit holes right now.
Mike Allton: Love it. I imagine you could probably bake that into some of the initial prompts, like, you know, “Categorize this and give me three examples of actual comments from each category, and tell me what row, column, whatever is in there,” so you’ve got that right in front of you.
Right. Where does your own domain expertise come in? ‘Cause I, see you doing all this stuff. I know you’re bringing a lot of education. I mean, we’ve obviously tapped you so many times, uh, to help us with our social listening and help us educate folks on social listening. How do you turn these raw, synthesized takeaways into actual high-quality strategy?
The What We Heard Loop With Clients
Brooke Sellas: We come back to the client, and every month- we do every week, which was when we do the big reporting internally- we do what we heard, loops weekly with the team. So, like, here’s what we’re hearing. Here’s what we’re seeing. We’re flagging some of these top bad things, top three bad things, top three positive things.
And then in our what we heard loop with the client, when we do reporting, then we can go through, and we can collapse, uh, all of those themes into the good and the bad. And then with the bad, obviously, we wanna say, “Hey, how can we work on some things with you?” And for us, just if any of you are listening who might be an agency owner yourself, for us, the real value comes when we’re able to say, “These are the things we’re seeing, and here are three suggestions for each one of those that would help solve some of those friction points or solve some of those pain points or reduce friction to buy or whatever it may be.”
They may take none of those, but the fact that we come saying, “Here’s the problem, but also here are some potential solutions,” I think is a real value add for us as an agency.
Mike Allton: Oh, absolutely. So once you’ve got those insights, how does that change the way you build, like, monthly content calendars, pitch campaigns to stakeholders or clients?
How does that inform all that?
Brooke Sellas: Yeah. So one of the things that we really love to do is take this information- the conversations that are happening on social and I’m happy to run another prompt for you here if you want me to share my screen again, and we start to create content that answers some of those FAQs or if they’re not FAQs, people are asking that month, like, what are the top questions that are asked this month? How do we create content about that? Because Mike, as you and I know, zero-click marketing is here. People are no longer clicking on those blue links to go to your website and read about something.
So if we could take those frequently asked questions or the frequently asked questions from July, and then create a carousel that says, “You’re having dust issues with your litter? Here’s what we suggest. You know, when you get the litter, open it like this, wet it down like that.” I’m, I’m making this up, but then we create that carousel.
It’s literally then addressing the majority of the questions we got or that we do get on a regular basis for the brand, and that’s content people want to see. That’s the content they want to engage with. That’s content that helps them get closer to buying.
Mike Allton: Love it. Love it. Yeah, and this particular example, you’re talking about a B2C brand, you know, that people are, you know, going into stores or they’re going online to purchase it, so giving them the opportunity to learn and be
Brooke Sellas: engaged with that content on social makes complete sense.
What I love, too, about using AI, it’s gonna tell you exactly what it’s doing. So I know she’s thinking about identifying trending FAQs for social media content creation. So she took what I said; she synthesized it down to her command. Now she’s running the command, which means she’s going back through all 1,902 of those conversations and looking for frequently asked questions.
The FAQ Ranking and a Competitor Comparison Page
Brooke Sellas: Now, there could be 1902 questions, and none of them are the same, but we’ll see what she says. Okay, here’s the FAQ ranking pulled straight from actual repeated questions this month. This is a strong zero-click content list. These are the exact things people won’t scroll past your ad or your click, blah, blah, blah.
So does it work in my automatic self-cleaning litter box? I am smiling because we get this question. This is actually an FAQ. Hey. This is a question we get all of the time, which tells me we need to create some pinned content, some carousel content that addresses that FAQ. How much does this actually cost?
How does this subscription work? I can see the post in my mind. Does it clump? Does it track outside of the box? How does it compare to? So this tells me, oh, what if we had a competitor comparison landing page on our website?
It sounds like if we included all of these brands, we could steal a lot of traffic from our competitors with this sort of, like, tool usage page that does some sort of comparison on the landing page, right? Does it actually control odor? And then it gives me a content-ready summary for my team. It tells me the volume all the way down and the best zero-click format, so a compatibility graph, a simple cost breakdown carousel, a short demo video, or a GIF, right?
Isn’t that cool? Like, this is how y’all should be creating content, by the way, on social. It should always come from a place of conversation.
Mike Allton: 100%. Now, a quick technical question. I know you’re using Claude. You’re not inside of a Claude project; we’re using a skill, or is it just whatever background information Claude already knows about you that it’s bringing into the conversation?
Brooke Sellas: Yeah. I tried not to cheat and do this in a project. I have a project, obviously, that is not cheating … for social listening, but I didn’t want it to, like, do all the smart things, so I just- just did this in a regular old chat. Yes, obviously, my instance of Claude has a lot of information about me, how I work, our clients, all of those kinds of things, but I tried not to go into a project to, you know, to have some bias built in or that thing built in. Yeah. I want to show you just from a chat how I’m prompting this and doing this.
Mike Allton: Yeah, so folks listening at home, if you’re using ChatGPT, if you’re just talking about a straight chat, you could create a ChatGPT project. You probably should. I would. And you could put in additional instructions and context and that sort of thing, which you don’t have to.
You know, Gemini, we’re talking about, you know, Notebooks or Gems. Copilot, we’re talking about agents and notebooks. You know, so every one of these platforms has ways that you can bake in a lot of knowledge, bake in a lot of these prompts in advance. For someone just trying to run their first social data analysis in AI this afternoon, what do you think is their biggest trap or mistake that they should avoid?
Three Traps on a First Run
Brooke Sellas: Okay, there are so many. There are so many. I’m like, there are 68 in my brain. Okay, I’m gonna try to, I’m trying to narrow it down. The first thing I would say is you cannot skip that context-setting piece that I was talking about. So you can’t just throw that spreadsheet in and be like, “Analyze this.”
For what? What is it? Who is it? What’s the goal? What are we trying to, yeah … you know. So if you get generic, you’re gonna get AI slop. I mean, this is why AI slop happens, I think, for the most part, is we’re giving generic prompts. We’re not giving it a lot of that, uh, context. The second thing is don’t ever treat that first answer as the final answer.
I want you to keep probing, keep having the conversation with your bot. Ask follow-up questions, just like Mike and I did here. Challenge the output, right? So if you’re like, “Really? That many people were talking about clumping?” Show me exactly which rows those message boards were on, right? Because in that spreadsheet, I actually have hot links that click to the actual conversation on social, so I can go really spot-check something like that.
And the last thing I would say is, you know, this is just my personal advice, but I think we need to be very careful about using AI to confirm what we already believe, instead of using it to find out what we don’t know. Because the most valuable insights are usually the ones that surprise you. Like, it’s nice to have, like, a gut check, and I get that. But try not to use AI in that way. I think, you know, leaving the critical thinking to yourself is an important part of using large language models.
Mike Allton: Absolutely. Gap analysis: what are the non-obvious, you know, ideas here? Asking those kinds of questions. I also always tell my AI to challenge my assumptions.
Brooke Sellas: Yes.
Mike Allton: ‘Cause I’m always making assumptions, and I want it to tell me when I’ve got a bad idea or, you know, when something’s just wrong.
Brooke Sellas: Mm.
Mike Allton: And it will if you give it that permission.
Brooke Sellas: I love that. Yes. Agreed.
Mike Allton: Yeah. So last question, Brooke. What kind of measurable results, whether it’s client buy-in, engagement shifts, time saved, have you seen since moving to this kind of a workflow?
Brooke Sellas: Well, I think you probably just saw it firsthand, and the biggest one is obviously speed, right? What used to take me six months to hand-code in my little spreadsheets and my little colors and all that kind of stuff now takes hours. And I do say hours. We just did this in minutes, but I mean, this is a very small part of the deep dive that we go into, so a few hours, I would say, but six months versus a couple of hours is amazing.
I also think for our clients it means faster responses to these emerging issues. It means actionable data, right? So not just saying, like, “Well, here are all these complaints. Oh, no.” Like, “Here are all these complaints. Here’s what their theme is. Here are the topics. Here are our recommended solutions to try to get in front of some of these problems.”
Obviously, then creating content that answers real questions instead of just broadcasting to people or trying to sell to people, and then obviously building out social care playbooks built around the conversations that are happening for these brands instead of the ones we think happen. Man, I love you, marketers.
I’m a marketer too, so I’m doing this myself, but we love to do assumptions and gut stuff, and this takes a lot of that assumption and that gut guesswork out. You know, the ROI, the return on investment, shows up for our client, our social care clients, when we can reduce these repeat complaints; when we can earn them higher CSAT or NPS scores; um, when we can help them create content that actually converts on social, and it converts because we’re speaking to what customers said they needed, period.
Mike Allton: Boom. That’s awesome. You’re awesome. Brooke, this has been so fantastic. Such an important technique that people need to be doing and learning how to do it. If they’ve got more questions or they wanna reach out to you, where should they go?
Brooke Sellas: I hang out on LinkedIn. I’m huge on LinkedIn. Not that I’m huge on LinkedIn; I’m just a big proponent of LinkedIn. I love to use it to talk to people and meet new people, so come chat with me there, or you can check out some of the work that we’ve done for these clients and some of our other case studies at B Squared Media.
Mike Allton: Folks, Brooke is huge on LinkedIn, so go follow her there. We’ll have her link in the show notes, as well as digital downloads that you can grab all the prompts that Brooke talked about today. But that’s a wrap for today’s lab session. If you want everything, as I said, click that link. It’s right there in the description below. And if you found this breakdown valuable, make sure to subscribe on YouTube. Leave us a review on Apple Podcasts. It helps other serious operators find the show. See you next Monday.
