What Happens When AI Gets Your Brand Wrong?

📬 You like this podcast? You’ll love the newsletter.
Join the weekly 3-2-1 on AI search + marketing: subscribe

What if the AI answer about your brand is wrong? Not just incomplete or outdated—actually, factually wrong?

That's the question Cassie and Franklin Rios, CEO of NextNet, dig into in this episode. NextNet has been in the organic search game for 20 years, and Franklin has watched the industry shift through every major Google update. Now he's tackling the next paradigm: how AI engines decide which sources to trust, which brands to cite, and what happens when bad actors flood the system with low-quality content designed to hijack the answer.

Spoiler: it's already happening.

Franklin shares a real story about a consumer who asked an LLM for the American Express customer service number, got served a fraudulent number, and had their personal information stolen. It's not a one-off. It's a glimpse into the brand safety risk almost no marketing team is auditing for yet.

In this conversation, Cassie and Franklin cover:

• Why AI accuracy is now a brand safety issue (and the audit nobody's running yet)
• How LLMs actually decide who to trust as a source
• What "vectorization" means in plain English—and why content teams should care even if they're never going to touch a vector database
• Why citation is the name of the game in AI search, and how that maps directly to the FSA Framework
• The real economics of ChatGPT ads—$200K minimum spends, $60 CPMs, and whether it's actually worth it
• Why becoming the answer organically is the long game that compounds

Whether you're a content lead at a B2B SaaS company, a CMO trying to figure out where to put your AI search budget, or a founder wondering if your brand is even visible inside ChatGPT, this episode gives you a sharper way to think about what visibility actually means in 2026.

Resources:

What is the FSA Framework?

Let’s connect:

LinkedIn → Cassie Clark | Fractional Content Strategist
Website → https://cassieclarkmarketing.com

Download Freshness, Structure, Authority: The Framework for AI Search Visibility:

Amazon

P.S. Is your brand losing its "Answer Authority"?

Most series A/B and enterprise brands are being "nudged" out of AI search results because of entity gaps and "stale" content. I am opening a limited number of specialized audit slots to help you reclaim your Share of Voice using the FSA Framework (Freshness, Structure, Authority).

Request your 7-Day AI Search Visibility Audit: https://cassieclarkmarketing.com/ai-search-visibility-audit/

Transcript

SPEAKER_01 1:13

Hey, welcome back to Founding AI. I'm Cassie Clark, a fractional content strategist and AI search optimization expert, and your host of the show where we break down what's actually changing inside AI Search so you don't get left behind in this new wave of user search behavior. Today I'm sharing a conversation with Franklin Rios. He is the CEO of Nextnet. Franklin and I chatted way back in February. Today is April 28th. Looking at my calendar, April 28th. I'm gonna be honest with you, in February, I thought I was gonna have a semi-urgent brain surgery. If you listened to a couple of the episodes, you heard me say, oh, I have brain fog going on. That's what that was about. And I recorded a bunch of episodes with some really rock star guests ahead of time. Because when they told me, ah, you might need brain surgery, my first thought was, oh no, my podcast. And so we recorded a bunch ahead of time. Major shout out to everyone that participated. You all are the best. Anyway, in this conversation with Franklin, if you hear us mention something like ChatGPT ads, for example, and then you think, wait, wait, that is old news. Yeah, you're right. Chat GPT ads have been rolling up for a hot minute now, but there's still some really valuable takeaways from this conversation with Franklin that I think are worth your time. And no, I did not actually need that brain surgery, not yet anyway. This just stays. We'll see what it's doing next year. But the biggest takeaway with Franklin, and the reason that I wanted to get this episode out, is the trust conversation. Franklin tells a story early on about a guy who asked an LLM for the American Express customer service number, and he got served a fraudulent number. And not only that, he had his personal information stolen. That story sits with me because it reframes the entire AI visibility conversation. We have been talking about how do I get cited, but the important question might be who else is getting cited next to me? And is the answer my brand is showing up in, is it even accurate? So that's what I want you to focus on as you listen. Franklin's gonna get into some of the very technical side of all of this, like vectorization, trust scores, how LLMs decide who to trust as a source. You don't need to walk away knowing how to vectorize a website from the technical side, but you do need to walk away thinking differently about brand safety inside AI engines. All right, let's get into it.

SPEAKER_00 3:50

Franklin Reals, I'm the CEO of Nextnet.

SPEAKER_01 3:53

Okay, so what is it that Nextnet does? Like what who do you serve?

SPEAKER_00 3:58

Look, uh traditionally speaking, I mean, Nextnet has been around for about 20 years, and the company has had its roots in backlinks and content creation in order to be discoverable via site via uh you know Google search, you know, not PayPalClick or any other paid media, but a truly organic, you know, search search engines. So that's what Nextnet has been do had been doing for the over the last 20 years.

SPEAKER_01 4:22

So you've had a pulse on how AI engines have can't come in and change the whole game.

SPEAKER_00 4:27

Yeah, well, even the Google has been changing the game over the last 20 years, every every every quarter, right? When they had new rules and new things that were popping up here and there, and and we had to stay agile and adjust, and we had to stay on top of the game and adjust. And so it's a situation where you know we're used to having our cheese moved, for lack of a better word, because uh, you know, we had to stay up to speed with everything that's going on in the marketplace. And then, yes, and with the newest iteration of Chat GPT, perplexity, Gemini, and Claude, you know, um, it is a is a whole new paradigm. It's a it's a it's a paradigm shift, you know, uh throughout the industry. One that it is almost like a perfect storm situation because you have you know the agents coming on board, uh, you have the content still being somewhat relevant, but at the same time, you have several entities in in the space do doing it, not just Google being the predominant player, but having a wide variety of entities being involved in the situation.

SPEAKER_01 5:29

Are you finding that to optimize for chat GPT, cloud beings, any of them? Are you finding that there's a different strategy for each one, or is it working all the same?

SPEAKER_00 5:41

It's well, they all have their own little idiosyncrasies and little you know components that that differentiate, but I would say that uh uh 80% of the journey is is quite the same, and 20% is is different from model to model, right? You know, from agent to agent. And they they traditionally speaking, everything has been happening where where uh LLMs had been scraping websites and translating that scraping into uh into uh into citations and into models that can be used for answers in order to increase the relevancy of the answers. That leaves a lot of room for interpretation by the LLMs. The interpretations of how is the content being being interpreted, how the content is being construed, what are what are the topics that are being discussed and how do this fits within the questions that are being asked by folks like you and I using our using our phones or using our computers when when we use the LLMs. And most importantly, the trust factor. The trust factor of what is the level of fidelity in the answer that corresponds to the source, that corresponds to the answer being true to form. I mean, a very recent example occurred where a person had asked a question in one of to one of the LLMs about the American Express Customer Service number, a number was served to that person, and that number turned out to be a fraudulent number that was hijacked by a by a by a bad entity, and that person and others got their personal information you know uh stolen as a result of that. Um, and that's not because the LLMs were not were not true to form, it's because this particular bad actors were feeding the agents content that made that content relevant to the answer that the people were asking for. Right. And so um that then concluded in in in in in uh consumers being exposed to fraudulent uh you know responses.

SPEAKER_01 7:40

Yeah. I've not had that happen with fraudulent like that, but I did have I was trying to call a doctor's office three weeks ago, and the number that Gemini gave me was for a private citizen and not the doctor.

SPEAKER_00 7:54

So I'm just curious how many people Well, I consider that more of a hallucination instead of a fraudulent response, right? So the difference is the hallucination is they found a person and it has hallucinating that that's the person that you're looking for, the doctor's office you're looking for, and he served up because it's the one that is he felt more relevant to the answer. Whereas I think Claudia or someone uh put out an article that it takes 250 pieces of content for the LLMs to be persuaded to use it as a source.

SPEAKER_01 8:22

Oh wow, okay for citation. Uh-huh.

SPEAKER_00 8:25

So that creates that creates a uh a delicate situation where if if you have enough bad actors out there putting content out there to hijack the answers and to be able to intercept you know the conversation or intercept the the the yeah the communication and the conversation happening with the agent, then you know that's that's where things end up being bad. Yeah. Right. And and in certain industries, uh it can be uh if that trust factor is so important and it can be so uh detrimental to the consumer, especially in areas like like the example that you and I just spoke about, financial services, uh in areas like pharma, or you know, or even whether it's prescription drugs or over-the-counter, right? And then in other areas like you know, consumer packaging goods or or travel and things like that, right? Where the answers may be may not be you know so true to form or might not be as trusted or or as as uh as relevant, but still they can have consequences, right? So the level of the of the of the consequences you know can vary.

SPEAKER_01 9:31

Right. It's just something that we all have to think about when it we come to sit down and create our strategy. Like I've not really thought too much about the bad actors. I have seen some reports about it, I've read some reports, not given it much thought, but now it's gonna be something I'm gonna be thinking about probably.

SPEAKER_00 9:51

And then that that that trust factor, then so what what we do with Nextnet on the platform is then we what we do is then we create an uh a score, an error score for the content. We verify the source of the error score, and therefore, when we are embedding and creating the vectors for LLMs to consume, they they have already not been properly organized. They speak the natural language of the of the LLM, but most importantly, it has a trust score associated with it so that the LLMs know that this is a high fidelity source of of the answer and the content.

SPEAKER_01 10:26

Right. So when you say vector, can you explain to us what you mean by that exactly?

SPEAKER_00 10:31

Yeah. So uh, you know, a website has multiple pages, and within any any any particular page of a website, uh it contains semantic information, contextual information, it can contain video, it can contain audio, it can contain multiple things. Uh and we then then take that and convert that into data. We convert so vectorizing is embedding information into a data format. And and they what the reason why we use a data format is because that's the natural language of the LLMs. They consume mathematics, they consume data, right? And so when when an LLM is scraping a website, it's truly scraping an HTML and converting it to a vector, right? Versus when when uh when when the Nextnet platform is is aligned to that to that to that site, whether it's a publisher, a CPG, e-commerce, what have you, you know, we have already done that work. And therefore, when the LLM is asking the question, it's already pointing them to the right area for the answer. So the chunk, the chunking of every page, you know, by the topic, the semantic, the the relevance, the validity, and the trust score, it's all in in the mathematical model. That way, when the LLMs are asking the questions, we you know, the answer is ready, is is ready to is ready to be served.

SPEAKER_01 11:47

Yeah, which brings me back to agenda commerce, which is I think on the horizon, but a bit closer than we're expecting.

SPEAKER_00 11:56

It is, it is in the horizon. And look, agenda commerce is uh is gonna be a handshake of agentec to agentec, right? So more more and more. Um so that that will be a good high fidelity way of agent-to-agent conversations, but at the same time, the agents have to also have uh uh a way to interact in a very fast manner, right? So speed is equally important because um any any any of the main agents, like the ones we've been discussing, you know, ChatGPT, Perplexity, uh, Claude or Gemini, they're going to be talking to multiple agents at the same time. So, what agent are you going to trust? What agent's information is the one that is most relevant, and what agent's information is the one that has already been uh uh prepared in order for the LLMs to easily consume? So that goes back to your original question. Why is why do some uh uh publishers or or or content generators or what have you need to have need to get early into this, right? Or not wait. And and why is it important for them to have the information organized in a manner that can be consumed by the agents?

SPEAKER_01 13:57

Yeah, yeah. So I want to talk about because you've mentioned before we started recording, we were talking about ads. So when you say, when you tell me build a funnel inside LLMs, what does that look like exactly?

SPEAKER_00 14:12

Is it is basically for in any any content to become trustworthy to be and and to be uh to be vectorized and embedded and to be organized for it to be ready for consumption for the LLMs.

SPEAKER_01 14:25

Okay.

SPEAKER_00 14:26

So that is that is call it the natural fit. And there, and that way you can you can get cited as part of the answer. So citation is the name of the game. Getting cited as part of the answer is the name of the game.

SPEAKER_01 14:37

So if I'm listening to this conversation and I hear the word vector, and how do I do that with a content strategy? In simple terms, how do we do that?

SPEAKER_00 14:48

Uh basically, our application vectorizes the site for you and it gives you a proximity score that says how close are you to the answer in a mathematical manner. So none of this is panel data, none of this is synthetic data, none of this is uh qualitative, it's actually very quantitative. Right. And that's why we're using NVIDIA because we have to do this in a quantitative manner, because we need the GPUs to truly be able to process the tremendous amounts of data that we are creating by vectorizing sites. Um and so by by doing that, then by this whole vectorization component, uh, it creates uh a mechanism whereby, you know, the the not only agent to agent but agent to website, whatever website it is, it creates it creates the natural handshake to it. Now, if someone, if if another party wants to be part of the answer and sponsor or have a mention of a different, that's not a true call it natural citation, that is a sponsor citation, which is fine as long as this as long as that citation is um relevant to the answer, right? Uh right. And so it must be relevant to the answer. And it cannot just be a random, a random uh uh mention or an or random or random exposure, you know, for a brand that has nothing to do with the conversation that is being that is being uh that is being asked of the agent. So the the big difference is do you want to be naturally cited for what you are and and and the high level of content and the high level of information that you have on your sites that is relevant to the answer, or are you going to pay to try to influence the answer?

SPEAKER_01 16:31

Yeah, I have been having this discussion across both the podcast and over on LinkedIn. It's you have two options either be ready for AI search optimization or pay to play.

SPEAKER_00 16:43

Exactly. And and look, and the play-to-play part, Cassie, Casey, is it's going to be quite expensive for a foreseeable time. Um, I I think I read that uh while you know ChatGPT saying that a minimum span of$200,000. And and in one of the articles, I also read that that translates to about a$60 cost per thousand CPM.

SPEAKER_01 17:06

I I did see the six sixty dollar CPM.

SPEAKER_00 17:09

Yeah, and that's that's quite expensive. Not even not even not even the early days of CTV uh uh you know uh ad insertion were sixty dollar cps. Yeah, they were they were never that high. And see, and you can measure CTV down to the audience and down to the specific screen and down to the specific demographic and device, right? Because of the data that comes from the AC, the ACR data that comes from the smart TV that is being shared in order to truly measure uh the efficacy of the impression. But you know, but to charge$60 CPM and not have any attributes that you can point to as a result, it's is uh it's it's expensive in so many ways, right?

SPEAKER_01 17:52

So uh just for the listeners, we are recording this on February 19th, and Chat GPT rolled out those ads two-ish weeks ago, maybe at this point. Have you talked with any brands who have already jumped on the train of we are doing ad for sure?

SPEAKER_00 18:09

No, uh, we have not spoken to brands that are that that are like eager to jump into that into that scope yet. Um they will eventually when things settle down, uh, because right now the agencies are the ones that are embracing this because the agency want to test it. And they and and specifically the large whole codes, you know, are are they want to test it and they have probably budget allocated you know to test things like this. Um but whether where they're going to have a hard time is okay, how do you measure what happened? How do you attribute the results? And that's going to be just part of the evolution and the cycle of the ecosystem.

SPEAKER_01 18:47

Yeah. Without knowing what Chat GPT is giving for data, it's kind of hard to figure out. I I mean, I think again, I haven't looked at it, so I don't know to figure out what we're actually measuring when it comes to this.

SPEAKER_00 19:00

Right.

SPEAKER_01 19:01

I'll be I'll be curious to see reports within the next couple of weeks of someone who has tried this out, but I have not yet. So I guess it comes down to the question if it comes down to expense,$200,000 is a lot to put into just Chat GPT ads. Then we have working to become the answer. What what would you suggest for someone listening to this? Like top tip to start to become the answer if they wanted to get in on this right now.

SPEAKER_00 19:33

Yeah. The first thing I would recommend is that you need to you need to partner to vectorize your site immediately, and in order for the LLMs to easily consume the information coming from your business, whether it's a CPG business, e-commerce, uh, and and and the good news is that the way that we build the platform, it it can be useful for a plumber or an HVAC company, you know, at a hyper-local level. And it can also be useful useful for a national brand or a multinational brand. Because the use cases, the day at the end of the day, math is math and data is data.

SPEAKER_01 20:09

Right.

SPEAKER_00 20:10

So when we vectorize a piece of content that is for I'm going to use St. Petersburg and St. Petersburg Plumbing Company, we vectorize it with all those data points and all those pieces of information so that when someone is a St. Pete asking for a plumber, you know, or asking what do I do with my clock pipes, right? It it cites the content coming from that plumbing company as an answer. Like you should do X, Y, and Z to your pipes, and you know, XY and Z plumbing company can help you with it. And that's part of the citation, right? But that's at a hyperlocal level. Versus if you were to say, you know, I am looking for a new air conditioning unit that still has a hyperlocal component to it, because there's still a source of it, right? That's where you and I are. Um, whereas if we're we're as having a different question of I would like to take a family vacation, what are the best places to take children between the ages of 10 and 15? Right, then the answer is more is more of a global nature, right? And and it gets more narrow and narrow and narrow as the conversation continues with the LLMs. And so uh it's so the in either of those three examples, that's it when you when your site is vectorized, embedded, and it's and it's got a trust score associated with it, uh the the probability and the chances of of of your content being cited as part of the answer are tremendous, mathematically tremendously high. So we're not we're not again, we're not being subjective about this. We actually have a tool that's that that vectorizes your site and says, and here are the recommendations that we that the you should take on in order to make your content better. And if you do X, Y, and Z to your content by page, by vector, by by by chunk, you will have a score increase of X. You know, you could have a scoring a score between zero and one hundred, you could have a score of 50 or 60, and and and we can give you mathematics with a mathematical center certainty that if you do X, Y, and Z, you can go from a 60 to a 68 or 70. Right? And those scores continue to improve uh as as you make more changes to the site.

SPEAKER_01 22:11

So I do want to go back to something that you did mention that somewhere you read it takes 250 pieces of content to become the source for an answer.

SPEAKER_00 22:22

To become citable, yeah.

SPEAKER_01 22:23

Yeah, is that something that you're also recommending too when someone asks you about this? Hey, go out and create 250 pieces of content?

SPEAKER_00 22:31

No, we don't. No, actually, we we we we believe in the quality of the content and the and the and the trustworthy worthiness factor of the content that will become more relevant and more impactful than just having the a quantity of content out there. Okay, because if you are if you're a pharmaceutical company, you know, you're not one is not going to create the pharmaceutical company, it's not going to create 250 use cases for the over-the-counter medication or what have you, right? But if you go in depth into what's happening with the medication, into the good, the goods and the pluses and the benefits of the medication, whether it's OTC or or otherwise, um, then you are be that relevancy increases and that trust factor increases. And the more that you and the more that those answers appear as part of the LLMs, the more that the LLMs will trust it as a so as a true source of high fidelity information. So it's not it's not a matter of quantity. What what what the bad actors are trying to do is not have the quality, not having the trust score, and try to override it with massive amounts of noise and massive amounts of of content.

SPEAKER_01 23:39

Okay, I know that was technically heavy. So let me translate that conversation into three things I actually want you to do something with. One, AI accuracy is now a brand safety issue. So the American Express story that Franklin told is not a one off thing. There are bad actors flooding LLMs with low quality content that is designed specifically to hijacking. Answers. OpenAI calls this prompt injecting. They had a release about this a couple weeks ago. I'll cover it here on the show. They know this is a problem. As far as I can tell, there's really no fix for this. But if you're a brand, we need to be paying attention to it because I think we can fix it on our side. So if you are in a high trust category, say financial services, healthcare, anything regulated, you really need to be auditing not just whether your brand shows up in AI engines, but how accurately it shows up. So pay attention to things like outdated pricing or where the links go to form fills, if competitors show up in your braided answers, and those wrong phone numbers. I also have a story about that in relation to calling the neurosurgeon about my brain that I mentioned earlier. When Jim and I gave you the phone number in an answer to the hospital to the neurosurgery department, I ended up calling some random lady. I didn't leave a voicemail, but I'm just it happens. This happens, and we need to be auditing whether that information is correct and then working to fix it. Number two, citation is the name of the game, and that game hasn't changed. Franklin said it from the engineering side. I have been saying this from the content side. It's the same conversation. Getting cited as part of the answer is the win. And the way you get cited is by being the most trustworthy, structured, extractable source on the topic. That is the FSA framework. Freshness, structure, and authority. If you've been listening for a while, you've heard me beat this drum. Franklin just gave you the technical validation for it. And if this is your first time listening, first of all, thank you very much. Um, there's a link to the FSA framework, what it is, and how to apply it to your strategy in the show notes. I'd encourage you to go check it, check it out. Three, pay to play is still a bad bet for most brands. A$200,000 minimum spend on chat GPT ads with a$60 CPM and no real attribution. Listen, the math ain't mathing on this one for the vast majority of B2B brands. I said this when the Chat GPT ads roll out, I will stand 10 toes on it. Becoming the answer organically is slower, but it does pay off. And that's where I'd put my money on every time, instead of just paying to be near the answer, work to be the answer. If you want to dig into how your brand is currently showing up in AI engines or whether what's being said about you is actually accurate, head over to CassieClarkmarketing.com to get started with an AI visibility audit. Alright, that's it for this one. I will see you in the next episode. Until then, stay visible.