What It Actually Looks Like to Build AI Visibility for a Startup

• Stitch AI's baseline share of voice was 1 mention out of 30 tracked prompts across ChatGPT, Google AI Overviews and Google AI Mode, measured with a 3.3% margin of error from ten runs per prompt.
• You cannot track AI-driven leads through click attribution because AI answers create a zero-click reality; the fix is to add "AI chat" as an option to every "how did you hear about us" form and CRM field.
• GEOForge's knowledge base module, BaseForge, took in 1,335 sales call recordings in about 24 hours, and scores each file's information gain (IGC/IGM) to predict how much of it AI models haven't already learned.
• To find category language, the team used G2 to identify a confirmed competitor (WATI) in "conversational support software," then mined 500 of its reviews with Claude for the actual words customers use.
• Neither Google nor AI platforms penalize AI-generated content by default; what matters is information gain, meaning content grounded in real, first-party experience that teaches the model something it did not already know.
• GEOForge tracks a fixed set of 30 prompts per brand and cycles out any prompt that locks near 100% share of voice for an extended period, replacing it with one still being contested.

Stitch AI, a WhatsApp communication platform and official Meta business partner, is mentioned in one out of thirty prompts tracked across ChatGPT, Google AI Overviews and Google AI Mode. That number, not a polished case study, is the actual starting point of this build-in-public series. Paris Childress of Hop AI and GEOForge sits down with Simon Young of Question.Marketing to run the first session of a GEO build for one of Simon's clients, live, with nothing edited out.

Generative engine optimization runs on a different measurement problem than SEO ever did: there are no clicks to count. The first work on any GEO project is not content, it is definition: what to measure, how, and where the brand actually stands before a single page gets published.

Why can't you track leads that come from AI the way you track search clicks?

Because AI answers create what Paris calls a zero-click reality. A person gets their answer from ChatGPT or an AI Overview without clicking through ten blue links, then arrives at a business later through a branded search, a direct URL, or a branded ad click, none of which trace back cleanly to the AI conversation that generated the interest.

Paris's fix is procedural, not technical: ask every customer, prospect, or lead how they heard about you, as early as possible, with AI chat listed alongside word of mouth, referral, Google search, and social media. Feed the answers into the CRM to set a baseline; it won't be perfect, but it produces a trend line over time. Simon confirms Stitch already has the drop-down live on its site, though it does not yet break AI chat down by engine.

How do you find the category and the language your buyers actually use?

Stitch AI sits in a hard-to-name space: WhatsApp messaging at scale, plugged into a CRM, for businesses that have outgrown the WhatsApp Business app on a phone. Rather than guess at category language, the team used G2 to find where a confirmed competitor, WATI, is listed: conversational support software. That gave them a defensible category and a lens for prompt research.

From there they mined WATI's 500 G2 reviews with Claude, flagging where and how customers actually described the problem WhatsApp tools solve. The goal was voice-of-customer language for the prompt set, not competitor intelligence for its own sake: finding out what people actually say, in sales calls, in reviews, in any customer interaction worth recording.

What does a 1,335-file knowledge base actually buy you?

Simon uploaded 1,335 files, mostly sales call recordings, into BaseForge, GEOForge's knowledge base module, in roughly 24 hours, the largest single upload the platform had seen. Every piece of content GEOForge generates is grounded in this corpus, which Paris credits for keeping output free of hallucination and what he calls AI slop.

The key metric is information gain, scored per file as IGC and IGM, with IGM predicting how much of a file's content has not already been trained into AI models. One file scored 84%.

That effectively means that we have an opportunity with an 84% information gain score to train AI models with new content they haven't yet slurped up through their crawling. That's great, because then they will cite us for that content. That's the reward: we give it the training, they give us back the brand recognition.

On whether AI-generated content gets penalized, Paris draws a line to the classic EEAT framework: experience, expertise, authoritativeness, trustworthiness. Neither Google nor the AI platforms care how content was produced, only whether it contributes information gain, something the reader did not already know, grounded in real experience. AI-written content from unique, first-party source material clears that bar; generic AI output does not.

How do you choose which prompts to track, and how many do you need?

GEOForge tracks 30 prompts per brand, spread across the buyer's funnel. For Stitch, that meant leading with the term people actually search, "bulk WhatsApp," even though it undersells the product's deeper CRM and coexistence features, because that term has the search volume that gets a prospect into the funnel at all.

Each prompt runs ten times, not once, to reduce noise, producing a stated margin of error of 3.3%. Once a brand locks a prompt near 100% share of voice for an extended stretch, it can be cycled out for one still contested, the same way an SEO team stops optimizing a keyword it already owns at position one.

The three surfaces GEOForge measures today are ChatGPT, Google AI Overviews, and Google AI Mode, excluding Gemini, Grok, Claude and other assistants for now. ChatGPT dominates consumer usage, and Google already shows AI Overviews on an estimated 60 to 65% of searches with informational intent, with AI Mode positioned to take over once Overviews are sunset.

What happens after the baseline: how do you turn a knowledge base into content?

With the baseline at one win out of thirty, GEOForge's next step is an AI agent that reads the knowledge base against the tracked prompts, finds where share of voice is weakest, and proposes topics the knowledge base can support with evidence. On this call it surfaced FAQ pages, blog posts, case studies and how-to content, including a Meta business partner FAQ and a case study on a client moving from individual WhatsApp accounts to branch-wide team inboxes.

Stitch's site had testimonials but no real case studies, a gap both hosts flag as common, since building one properly is real work regardless of tooling. Content generated this way strips identifying details: no customer names, no PII, only distilled results.

Notable moments

• The baseline reveal: Stitch AI is mentioned in only 1 of 30 tracked prompts, 3.3% margin of error across ten runs per prompt.
• Simon's 1,335-file BaseForge upload, almost entirely sales calls, in about 24 hours: a new volume record on the platform.
• A live tangent on Anthropic's AI watermark and whether it threatens ContentForge, resolved by the information-gain argument.
• Simon's story about a client meeting where saying "AI slop" got the whole room laughing, because everyone they talk to already uses the phrase.
• The GEOForge AI agent generating its first ten content topics live, including a Meta business partner FAQ and a branch-wide inbox case study.

This is the first of a planned series tracking Stitch AI's share of voice over 60 days. The next session covers the first content drafts and whether the baseline actually moves.

Paris: Okay, we are live. Hey Simon, nice to see you.

Simon: Yeah, and you. Good to be live, to be fair. We don't often do a live.

Paris: Yeah, these are fun. For those of you who are just tuning in, I'm Paris from Hop AI and GEOForge, and I'm joined by Simon Young from Question Marketing. Today we're going to kick off what might become a series, hopefully. It's going to be about helping one of Simon's clients, a company called Stitch AI, dramatically improve its AI visibility in as short a period of time as possible using our GEOForge platform.

So this is going to be a build-in-public exercise where we're going to dig into the good, the bad, and the ugly. We're not going to try to sugarcoat anything here. We're going to start to go through the journey, starting with how we understand this business and assess the challenges and opportunities ahead. So with that, Simon, do you want to say hi and introduce yourself?

Simon: Yeah, hi, I'm Simon. I met Paris a couple of months ago, so it's really interesting to understand his insight and the tools he's got and everything that's going to go on in the AEO and GEO space. I've been in SEO since day one, so I have a lot more gray hair than most people. I've been through all of that: the good stuff, the bad stuff, done the black hat stuff, and I've been talking about the way search is changing for six or seven years now. I've been working a lot on this, hence my LinkedIn URL is forward slash AEO. I've been talking about the answers and the questions a lot, and I ended up at Question.Marketing as my new brand and how I'm going to push forward as a business.

I've got a few clients I'm working on in terms of pushing people into the AI and LLM search arena, because I've seen a lot of them dropping off in terms of clicks over the last year, and those clicks aren't necessarily ending up on the website anymore. That's the bit I love and like. Stitch, one of my clients, works in the WhatsApp space. They're a communication platform and they don't have a lot of share of voice in LLMs at the moment. So it's interesting to see if we can do it in the wild: show people whether it works, when it works, what works, what doesn't work, what tools we're using, and how that moves the needle.

I don't know about you, Paris, but I've seen so many people talking about this subject on LinkedIn. Even yesterday, I called somebody out on a set of results. There was a set of results published that showed hundreds and hundreds of referrals from AI onto their site. I said to the guy, it would be really interesting if you showed me which client that is. Personally, had I made the results up, I might have said I'm under NDA and I can't tell you the client. But the person then said, who said they were real results? And I'm like, wow, we're in that sort of era of search now.

Paris: It feels like SEO around 2010, I'd say, or maybe even a little earlier: the very beginning of SEO. It feels like the Wild West, where no one is really sure yet how to do it or how to measure it, but we're all sure that this is a real channel. If there's one thing that's established, it's that this is not a flash in the pan. This is not going to be something that goes away in a few months. This is the emergence of a new channel that comes around maybe once in ten years. We had the internet and search, which brought SEO, and then paid. Then we had social media, and then of course mobile shook that up. I think this is as big; I'd lump it into the whole AI movement.

We're going to be calling it GEO, primarily because that's the name we went with for our product, GEOForge, but a lot of the industry calls it AEO or AI search. It's the real deal. Right now what we have is a lot of practitioners, a lot of former SEOs and other people jumping into this game and trying to act like they have a long history with it, because they realize this is the real deal. But no one really knows yet what the exact game plan or the playbook is to succeed and win. We're all feeling our way through it.

And also, how do we measure it? Because the measurement here is a lot harder than with SEO, because we don't have the clicks anymore, frankly. So what are we going to do?

Simon: We end up with, hopefully, a proper marketing stack, in that people continually see your brand. But then, like you said, we don't have the attribution layer. Not that everybody uses that attribution layer properly anyway, but that's another bugbear of mine. We see people ending up on a site or placing an inquiry, but attributing that is a lot more difficult.

Paris: Yeah. Well, let's stay with that for a minute, just so our audience can understand it a little better. With AI, we're increasingly dealing with a zero-click reality, which means people are getting their answers from AI without having to click through the ten blue links from Google. What that means is they still will eventually make their way to your business, most likely to your website, and they still will probably convert downstream, but you won't be able to track that through a click from a search result page.

It might come through an indirect visit. Somebody might Google your brand name and come through that way. They may Google your brand name and click on the ad you have against your brand name or the related keyword. They might type your URL directly in if they know it. There are many ways they can arrive that are not easy to trace back to that LLM chat conversation.

So in my view, the most important way to measure success with this new channel of GEO is to do your best to ask every customer, prospect, or lead that comes through the door, as early as possible, either in a form or over the phone: how did you hear about us? How did you find out about our brand? Give them a set of options. One of those should be AI chat, with examples like ChatGPT and Gemini. The rest could be the classic word of mouth, referral, Google search, social media, or an email. See what they select, then flow all that information into your CRM, and set a baseline.

What we're going to be doing right now is baselining Stitch's starting share of voice from our tool. It won't be perfect, but you want to see the trend line over time. We're going to be talking about share of voice, and that's the primary KPI right now in this industry. But despite that, I believe every new marketing channel's ultimate ROI is about how much revenue and sales pipeline it can deliver. Right now, the way to get that is to ask the customers and prospects how they heard about you, and hope that they remember.

Simon: We've set that up already on the Stitch site. As your introduction suggested, we've got the drop-downs. It'll be interesting to see if people are selecting AI search. We haven't gone so far as to listing out the different engines, so I think AI search as a baseline will do for now.

Paris: Yeah. So, Simon, why don't we look at Stitch and try to get an understanding of the business before we dive into the GEOForge setup? I'm going to share their website on the screen here. Let's see if that's coming up. All right, so what do you want to give a quick overview of, in terms of what Stitch does and the competitive landscape here?

Simon: Yeah, so these guys are a WhatsApp communication channel. We struggled a little bit with what the name of that vertical is, whether it's business communication solutions or a WhatsApp-specific channel name. But essentially, we're moving into a world where Meta bought WhatsApp and paid nineteen billion dollars for it, so they don't intend to leave it unmonetized. I've been doing a lot of work with clients using WhatsApp messaging as a comms channel, especially now that email is something that just doesn't get opened, or is very hard to get into somebody's inbox. So a lot of companies are turning to WhatsApp.

One of the challenges is that if you've got WhatsApp Business, that's not the same thing. WhatsApp Business on your phone is probably for a smaller business. If you want to use it at scale, across an enterprise or plugged into a CRM system, that's what Stitch does. It plugs into a CRM, enables you to send bulk messaging, enables you to do AI chats between customers, and respond to leads really quickly. They work across a lot of different verticals, principally things like estate agency, automotive, and trade counters. All sorts, but it's a comms channel.

Paris: Yep. And one of the ways that I learned about it was actually looking at the pricing page and digging into this comparison. This is really the key features; let me make that a little larger. One of the things you said that was interesting, Simon, when you introduced me to this brand, is that the really killer feature here, the one most people are looking for when they start this journey, is bulk broadcasting. What they want to do is send WhatsApp messages out to customers or prospects at scale. Is that right?

Simon: That's normally one of the first searches people make. If you're a marketing manager and somebody's saying, well, I'm using WhatsApp and I'm communicating at scale with thousands of people, the search term that comes up is bulk WhatsApp. It's not normally what Stitch would market themselves as; it's one of the features. If we were running a funnel, this would be the bottom of the funnel, saying to people, this is one of the features within it. And it's something that gets searched a lot; there are loads of searches for that.

The more interesting thing is when we get into the CRM integrations and the ability to have what's called coexistence. Coexistence is the fact that you can have it on your devices as well, as in the app or the CRM. It's a complicated one to get some people to understand exactly all the features it can do, but Meta are actively moving a lot of businesses into using WhatsApp now.

Paris: Okay, so if we put this in the context of the prompt discovery exercise we're starting, our goal is to come up with a set of prompts that we're going to track. We're going to track the share of voice, which means the mentions of the brand or citation appearances relative to all the other competitors. What you've described is that most people's initial awareness of the problem, when it comes to using WhatsApp for business and for sales and marketing, is this bulk broadcasting feature, and that they have a lower level of awareness of the more intricate features, like the CRM integration and those things.

So what I'm thinking is that a lot of the prompt discovery should start with bulk WhatsApp as a hook, just to get people onto the website and make them aware of Stitch. Then, later, they can get more educated once they're really in the funnel. As we were talking about on one of our previous calls, we were trying to define the category: what category do they play in? We went to G2 to try to find this category, and we landed on a competitor called WATI. Do you remember these guys?

Simon: Yeah.

Paris: WATI, if we look at their... let's see, this is the company. I'm going to try to back into the category they're in here. Here we have...

Simon: So, I think they're trying to list them under customer engagement, aren't they?

Paris: Yeah, I believe so. These are their reviews. Oh, here it is. Okay, so I'm going to back up one level in the hierarchy, and we've got conversational support software. So this is the category WATI is in, and we're pretty sure WATI is a direct competitor. Now that we have this category, I strongly suggest Stitch sets up a G2 profile in the same category. Let's get back into WATI, because we were going to look at their reviews. They have 500 reviews here, and we were going to try to use this as a place to mine some prompts.

What our team did, Swati from our team, is she went through these reviews using Claude and read through all 500 of them, paying specific attention to the instances where WhatsApp is mentioned. As we scroll through, you can see WhatsApp is featured really prominently in almost every review, because that's the core of this tool. What we wanted to uncover is the voice of the customer: how people describe the value, or how they even pose questions as they write these reviews, so we can reflect those same questions and that same sentiment back into the prompts. That's what gave us this. Now I'm going to move into GEOForge. Let's go there.

Simon: Just before you carry on, this isn't new, right, in terms of how you should be talking to customers. Go and find out what people are saying in those sales calls, or in the reviews, or in any interaction you're having with a customer. We should all be recording it, right?

Paris: Absolutely, it's a great point. That actually reminds me of something. Before we look at the prompts, I want to show the knowledge base, and the sheer volume of knowledge base that you, Simon, built here in one or two days this week. Let's jump in. This is BaseForge, which is the knowledge base module of GEOForge. This is actually the foundation for everything, because all content that's generated is grounded in this knowledge base. This is how we ensure there's no AI slop, no hallucinations, and that the content that comes out and gets published has high information gain for AI crawlers. And I think you set a record here, Simon, because...

Simon: Did I?

Paris: Well, so far for us, we've never seen this amount of volume. You uploaded 1,335 files, and I think you did this in about a 24-hour period. From what I can tell, these are sales call recordings primarily. Is that right?

Simon: Yeah. And I can get you a load more if you want, but I think maybe we'll start where we're at.

Paris: I think this is a good starting point. What this does is give us a phenomenal head start into creating content that reflects the voice of the customer, and trains AI crawlers, who are going to come in, crawl, and read that content, on how Stitch's customers speak, the language of this category. Hopefully, down the road, ChatGPT can effectively become Stitch's best sales rep, because it will recognize questions it's already seen and give the correct answers.

This is outstanding, and there's one other thing I want to point out, which is the concept of information gain. We have two scores here next to every file upload; you can see IGC and IGM. The one that's most important is IGM, which stands for information gain model, and that's basically the prediction of how much new knowledge this file contains that hasn't already been trained into AI's models. So an 84% information gain score means we have an opportunity to train AI models with new content they haven't yet slurped up through their crawling and baked into their pre-training. That's great, because then they'll cite us for that content. That's the reward: we give it the training, they give us back the brand recognition.

Simon: I wanted to say, before you carry on, I had a really funny experience yesterday. I'm an AI geek, I love all this stuff, but others may not find it as funny as I did. We were sat in a meeting yesterday with a client, five or six people around the table, and I was explaining to them the need to have unique content in their socials and their knowledge bases, and take everything out of their head and put it onto... well, I said put it onto paper, and then I said, well, obviously not paper, right? But then I mentioned AI slop, and everybody around the table just started giggling. I was like, what's going on? And one of the senior managers there went, no, no, you've not said anything bad, it's just that everybody we talk to is talking about AI slop at the moment.

Paris: Yeah, that is the buzzword. Just a second.

It's funny, whenever I'm going live, I always hear someone banging outside the window, so I've got to shut the window and turn on the AC, unfortunately. But AI slop is probably the marketing word for 2026. I don't know if there's a phrase I've heard more. If there was some kind of word-of-the-year dictionary for marketing, I think it's AI slop.

Simon: One hundred percent. Even LinkedIn has it as a drop-down now, don't they?

Paris: Yeah, that's right. That's going to lead to some serious abuse. You're going to let your entire LinkedIn user base try to determine what's AI slop? And what happens if 90% of it is actually AI-generated? What is LinkedIn going to...

Simon: Well, you would hope, you and I would personally hope, that LinkedIn has the ability to spot AI slop on its own, without people having to report it. But maybe not. I don't know.

Paris: I'm sure they have some tools, and I guess now Anthropic's new watermark, which was announced recently, is going to make things a lot easier as well. I don't know if you caught that, but...

Simon: Yeah, yeah. And I even posted something on LinkedIn yesterday about Google local business pages and places not now allowing AI-generated images. So there's some pushback coming across the whole space.

Paris: Yeah, well, we can definitely thank the EU and EU regulators for that nice move, for forcing Anthropic to be the first one to watermark. I think the others will probably follow. Here's my take on this.

Simon: Then you just need the tool that removes the watermark, right? But anyway, no one would do that.

Paris: And you still need a tool to verify that it is AI. And then, to what extent is it AI? Was AI used as a light-touch editor at the end? Was it used to generate every single word from the first draft to the end? Was it somewhere in the middle? There's a scale there; there are different degrees. If you're LinkedIn, where do you drop the hammer? Because LinkedIn also had its own AI-enhance button, up until recently, to enhance your own posts. So they were playing the game too. It's really interesting times.

My take on all of this is that Google and the rest don't care if you've generated content with AI. What they care about is whether you've contributed information gain. Have you added something new to the conversation that represents your real opinion, your real point of view, or an expert's point of view grounded in real experience? You can think of it as the classic EEAT acronym: experience, expertise, authoritativeness, and trustworthiness, in the lens of SEO, or in this new lens, information gain. If you contribute information gain that can teach models and humans new things, and it came from your unique perspective, and it was generated with AI because that helps you produce content faster or more efficiently, more power to you. That's where it comes down to.

Simon: Right?

Paris: Right, right. So at first I was a little bit worried that this Anthropic watermark might really spell doom for GEOForge, particularly our content module, ContentForge.

Simon: Yeah.

Paris: It generates content using Sonnet 5, and that could be problematic if that watermark follows the content through the API, which supposedly it does. But I still believe information gain rules. If we can source, just like what we're showing here, recorded sales calls where questions that have never been asked before are asked and answered accurately by knowledgeable sales reps, and we can flow that information gain into content, it does not matter if that content was generated with AI, because we're contributing new knowledge to the corpus, and we'll be rewarded for that and not penalized because of the method of production. That's my mini answer to that, before we move on.

Simon: If you went back years and years in SEO, and you said to somebody, well, you'd be talking eventually about building a knowledge base... we didn't used to refer to it as that, but, you know, creating good old blogs. If you were blogging regularly with interesting topics that people land on, then read, digest, and spend time on, that sort of signal to Google was great. And we're not far away from it; it's the same principle.

Paris: Absolutely, yeah. All right, so here is BaseForge. We're off to, I think, a phenomenal start with a huge corpus of knowledge. Now what we're going to look at is the prompts, because we were talking about how people with different degrees of awareness of this category are searching. Some people might not know there are tools that can leverage the WhatsApp Business API. Some people may not even know what WhatsApp Business can do for them at all. And there are other people who know about the tools that exist and need help finding the right one. So there's top of funnel, middle of funnel, and bottom of funnel.

We set out to have a pretty good spread across these 30 prompts to cover top, middle, and bottom of funnel. I'm not going to go through these in detail, but they generally represent all the stages of a buyer's journey, from a business owner or business user. And of course, we're limited to 30 prompts; we had to draw the line somewhere. In reality, there are probably an infinite number of ways to ask these types of questions, and what we want to do is have a representative sample, understanding there are always going to be limitations.

Simon: Having spent a fair bit of time with you, I wanted to cover that piece where you said we're limited to 30. I don't know the answer to this, but if we went back to the good old SEO days, we'd try to optimize for 30 different keywords or phrases. Once you'd done that, achieved it, and were in the top results, you'd then enter a maintenance mode where you'd continually just top those up. Is there a scenario in GEOForge where we achieve amazing results and then move on to some other prompts? Does it work like that, or not?

Paris: We can do that, but that would mean cycling out prompts. Let's say we have prompts that have sat at 100% share of voice for weeks or months, and we know we really own them. It'd be the equivalent of sitting in position one in the SEO world and having it locked down. We would recycle that and replace it with another prompt we want to win, where we don't have the equivalent of a number one ranking, or 100% share of voice. So right now we're limited in total to 30 prompts, and we can cycle them in and out.

The good news, I don't know if this is good news, is that the share of voice for Stitch is very low for these prompts. In fact, only one out of 30 has Stitch mentioned in the answer. We did 10 runs, so this isn't based on just one attempt at running each prompt; we ran the prompts 10 times each to level out the noise and the variability, and achieve a lower statistical margin of error, which is a very important concept for us. Our margin of error here is 3.3%, which is pretty tight. What this shows us is that we have a lot of work to do. We've won on only one out of 30 of these prompts, and over the next 60 days, Simon, our goal is to win as many as possible.

We want to get mentioned in the answer, where brands are being mentioned, and when that's not the case, we want to be cited as a source, where AI is showing the sources, typically in the right-hand panel, as you normally see in ChatGPT. And by the way, to make it clear, we're measuring in ChatGPT, Google AI Overviews, and Google AI Mode only. This does not include Gemini, Grok, Claude, and the long list of also-rans. But I think, for consumers, ChatGPT is far and away number one right now. And of course, Google is now showing AI Overviews for, the last statistic I heard, something like 60 to 65% of SERPs when the search query has informational intent.

Simon: And nearly always when there's a question, always, I would imagine.

Paris: Yeah, yeah. So these are the ones we care about most. Google AI Mode, I think, is waiting in the wings to take over once they sunset AI Overviews. We don't know when that's going to happen, but AI Mode is gaining market share. So realistically, we can't measure everything, but we're going to measure 30 prompts that we think is a good sample, and we're going to measure them across these three LLMs that have the large majority of conversation share right now. This is what it looks like: we have a single point, and once we revisit this over time, we're going to start to see a trend line. This data is going to update every 7 days from here on, and I hope it's going to be all up into the right, but we're going to find out the hard way.

Simon: We are.

Paris: And then we're going to start producing some content very soon as well. We have about a minute left, Simon, so I just want to show, I want to forecast for our audience where we could go next here. That's going to be to start creating content against this huge knowledge base, with all these sales call transcripts. I'm going to have the system generate the first 10 topics here, and I'll just eyeball them, and then we can wrap up with final comments.

What's happening right now is that there's an independent AI agent reading the knowledge base. It's also looking at those 30 prompts, and it's seeing that we basically have zero share of voice in 29 out of the 30 of them. Its goal is to recommend topics that will help us achieve higher share of voice in the prompts where we have gaps, where we're losing, which is most of them. It's going to look for supporting evidence or grounding in these topics from the knowledge base, looking into all those sales call transcripts and asking, do we have data in here, do we have first-party information that can support these topics, and if so, then it's going to suggest that topic for production.

I thought this was going to happen a little bit faster, but let's let it run for a little bit. It's thinking, and it's probably taking a little longer than I expected just because of the sheer size of that knowledge base. Oh, here it comes. This is a mix of FAQ pages, blog posts, case studies, and how-to posts. So here we are: is Stitch AI a Meta business partner? That's interesting. What is Message Box? That is a feature of...

Simon: That's the branded product, yes.

Paris: How's it related to Stitch AI? How Central Lettings moved from individual WhatsApp accounts to branch-wide team inboxes: that's a case study. That's going to be great. So, if you'll allow me, I'm going to generate this as a draft, Simon.

Simon: Yeah, cool.

Paris: We probably won't have time to see it; it might take a few minutes. But one thing I noticed from Stitch's website is that they didn't have any real case studies. They had testimonials and snippets from people, but not real case studies. So I think we need to start building case studies. We need to build comparison pages against competitors, showing where they win. And I'd generally start by trying to address a lot of the questions we saw in the prompts themselves, and I believe that's what the system is going to do.

Simon: But even with the wish, in most businesses, to publish case studies, when you give that task to somebody in a marketing department, it's a chunk of work. Even if they're using AI tools to help them, it requires a lot of thought. That's typically why you'd probably see an awful lot of websites with just a handful of case studies.

Paris: Yeah, so our system is going to hopefully generate a lot of case studies out of this massive mountain of sales call transcripts, because I assume some of those are also with customers, and they're actually...

Simon: Support.

Paris: ...real data. One thing I want to point out here with case studies, and in general, is that the content, by nature, is designed to filter out any sensitive data. So we're not going to be displaying the names of any customers here, or any personally identifiable information. We're just going to be talking about the actual results and trying to distill the knowledge and the learning from the knowledge base.

All right, well, Simon, I think we're not going to be able to get into the first draft; let's leave that for the next session. But we did cover a lot of ground today, and I'm really encouraged. I like the fact that we're starting from a very humble place here, with a 3% share of voice. It means we've got a lot of work to do; we've got a short runway. But we're hopefully going to generate a lot of content, and we're going to build a lot of off-site citations. We're going to see how quickly we can get AI to recognize this brand and give us love in these conversations, and we're going to measure it right here.

Simon: I don't think you can put enough emphasis on, we'll probably talk more about it, but the creation of that set of prompts, and what we want to be found for, that's so important. That's the piece that a lot of businesses will probably get wrong. And then, obviously, being able to measure where you sit, that's not something I've really seen anybody being able to do properly at all. That's what brought me to you, and hopefully we'll have a great result for these guys.

Paris: Yeah, well, let's see, and we'll do this again soon and let everybody in the LinkedIn community know. Before we sign off, I'm just going to take a look to see if we have any questions here. We do have a live audience, but I don't see any questions that came up during the live recording. So for those listening, I hope you tune in next time and ask questions in the comments, because we're going to be looking at them and try to address them in real time.

Simon: Awesome. Thanks for your time.

Paris: Yeah, well, that 30, almost 35 minutes flew by, Simon, so we definitely need to do this again soon, and we'll keep everybody updated.

Simon: Good stuff.

Paris: All right. You have a great weekend, Simon.

Simon: And you. See you soon.

Paris: All right. Take care.