When AI Gets Your Brand Wrong With Bluefish AI

[00:00:03] What happens when an AI assistant confidently gives a customer the wrong price, the wrong product specification or even medical dosage? For brands that error could influence reputation, trust and eventually a purchase before anyone realises it's happened. And my guest today is Alex Sherman. He's the co-founder and CEO of Bluefish AI and a great guy may I add too.

[00:00:29] And today he will explain why AI misinformation often begins with ordinary internet content rather than a spectacular hallucination. And sometimes a detailed Reddit post or a small YouTube creator could sometimes outweigh a company's entire website. So I want to learn more about AI visibility, accuracy, agentic marketing and the coming challenge of selling to both people and their digital agents.

[00:00:59] And Alex will also raise a question that should concern every information business. When AI gives us the easiest answer, do we gradually lose the habit for finding answers ourselves? But before you answer that question, let me introduce you to today's guest right now. So thank you for joining me on the show today, Alex. Can you tell everyone listening a little about who you are and what you do?

[00:01:26] Yeah, absolutely. Thank you so much for having me on the pod. I'm excited to be here. So, yeah, my name is Alex Sherman. I am the CEO and one of the co-founders here at Bluefish. And we are an agentic marketing platform. So simply put, our job is to help the largest brands in the world essentially manage AI as a new marketing channel.

[00:01:46] So no surprise to you or to your listeners, you know, today consumers use AI for every aspect of their shopping journeys, whether it's exploring a new category, comparison shopping, getting personalized recommendations. And so the largest brands in the world are looking to make sure that they're showing up the right way in those AI conversations, that their brands are being represented, you know, positively, accurately, that they're actually, you know, visible in those AI responses.

[00:02:16] And so we have a platform that helps those those brands gain visibility into how the major AI models describe their products to consumers and then a collection of tools to help them optimize and actually influence those outcomes and create the sort of content and data that those models need to represent their products and services accurately and positively to those consumers. So that's what we do. Awesome. It's such an exciting space to be in right now. I mean, the way we find information has changed so much.

[00:02:46] I mean, we go back, what, five years and all anyone wanted was to be on the front page of Google. Now we don't go to or many people don't go to the front page of Google. I recently, as I was saying, before we started recording today, I was recently on a trip to New York. And I must be honest, my whole itinerary was through AI. And even each day was perfectly matched with everything that I wanted to see, even created a Google map with nine different stops. The way we access everything now is completely changed, hasn't it?

[00:03:17] Yeah. You know, when we when we started the company a few years ago, you know, ChatGPT was kind of the only game in town. And so for a lot of marketers and for brands, you know, of course, they wanted to make sure that they were performing well in this new kind of media interface that ChatGPT was becoming. But it was relatively small potatoes at the time in terms of total volume and consumer adoption.

[00:03:40] But what happened is, is that ChatGPT kind of kicked off a sort of space race across all of the other big tech companies. And as soon as ChatGPT launched within months, you were seeing AI show up in Google. Right. They launched AI overviews that was powered by Gemini. You saw Microsoft launching AI into into their co-pilot system. There was meta AI. Amazon launched Rufus. Walmart launched Sparky.

[00:04:06] And so in a very short period of time, the sort of agentic Internet went from, you know, ChatGPT to effectively the entire Internet. And today, just like your trip to New York, consumers don't really care about the difference between the AI Internet and the non-AI Internet. It's just the Internet for them. And now there is this better, faster, smarter way for them to do research and get information and ultimately transact.

[00:04:31] And so we're just seeing AI sort of infused into every digital kind of experience that consumers have. And, you know, from our point of view, we're marketers. So we have been, you know, we've been working in sort of enterprise marketing for decades. The fact that there's a new channel on the Internet is not new. Right. We have that every five or six years, whether it was the rise of social, the rise of smartphones and mobile, whether it was the rise of, you know, of social media.

[00:05:00] Every couple of years on the Internet, there is a new technology that ultimately changes how consumers interact with information and content online. And so as marketers, you're kind of you're accustomed to that. What was different about AI is the speed. It happened so quickly. I mean, there was sort of an 18 month deployment cycle compared to smartphones, which took like six years to reach a billion users. And so I think there's this there's this sort of velocity that is really remarkable for this this channel in particular.

[00:05:31] It really is. And there will be millions of consumers like you and I that are listening to this conversation today. And maybe they're nodding their head in agreement. They ask AI systems about companies, products, purchasing decisions, where the nearest subway is. There's so many different things there. So when those answers are wrong, though, how big a problem is that becoming for brands? And how are most companies even aware of what AI is saying about them if they're not using it to their advantage?

[00:06:01] It's such an important question because I think, you know, AI sort of unfairly developed a reputation for being this kind of, you know, omniscient, you know, sentient thing that just had perfect access to information. And and we know now that that's just not the case. You know, ultimately, you know, large language models are effectively trained on a sort of download of the Internet. And the Internet is very often wrong.

[00:06:29] And there's a whole host of low quality content out there, some of which, by the way, was was built by AI itself. And the models are training on that stuff. And just like anything else, there's kind of this like garbage in garbage out problem. And so, you know, at Bluefish, we track AI accuracy as sort of a, you know, as a foundational part of our of our platform. And we've been doing that for for over a year now.

[00:06:54] And and what we find is that there are AI inaccuracies and hallucinations throughout AI responses to consumers, whether that is, you know, an inaccurate pricing. Maybe that's an inaccurate dosage for a particular, you know, medicine or drug. Maybe that's just a, you know, a product spec or a detail that is, you know, inaccurately portrayed.

[00:07:18] But what we find is that in AI responses, there is easily, you know, 10 to 20 percent, you know, hallucination that we track across all AI responses. And I think when people say hallucinations, I think they think, you know, something cartoonish. But usually we find most hallucinations to be in kind of this gray zone where it isn't, you know, it isn't exactly inaccurate, but it is kind of a misportrayal of the brand.

[00:07:44] And so if you're a company that is spending billions and billions of dollars, you know, across, you know, your marketing materials, across your paid media, those inaccuracies can add up to massive problems for your bottom line, for your brand. It can create brand safety issues. And so one of the key things that we do with our customers is, A, help them understand what are the inaccuracies? Like, how are you being misportrayed by these models? Two, why are you being portrayed that way?

[00:08:12] So how are the models learning about your brand such that they're portraying you in an inaccurate way? And then three, what can you do about it? So once you've sort of identified the problem, you've done kind of a diagnostic analysis, then you transition into optimization. And what we find is that today, brands actually are in the driver's seat more often than not, that they can start to create that high quality content that ultimately the models want to train on and want to learn from so that they can portray the brands more accurately.

[00:08:42] There's sort of a convergence that we're seeing in terms of the incentives of the AI companies and the incentives of the commercial brands and frankly, the incentives of users, right? Like everybody wants a high quality AI experience. Everyone wants that experience to be accurate, to be grounded in fact and truth so that, you know, you can continue to use the ChatGPTs and the Google Geminis of the world as a reliable source to inform your, you know, your travels.

[00:09:11] And so what's happening right now is that there is sort of a new kind of data foundation that's being built on the internet between these large commercial brands and the AI models that ultimately is making these, you know, consumer experiences smarter, faster, and more accurate. And I'm curious from everything you've just said there, where are you seeing misinformation about a company actually enter the AI information ecosystem and why can connecting it be surprisingly difficult?

[00:09:41] What kind of thing are you seeing there? Any examples just to bring that to life a little? For sure. So, you know, I would say on a, on sort of a foundational level, there is a lot more really low quality content than high quality content on the internet. One example is, is that, you know, for a lot of brands, the way that they describe their products are through product detail pages, right?

[00:10:05] So if you've ever been to an e-commerce site, you've seen a, you know, a PDP, that PDP is probably 250 words, 300 words. Well, a model probably needs something closer to 3 million words in order to really train on that product, on all of its different dimensions, how it fits with different consumer needs, right? Like you and I might have different needs from, you know, from a running sneaker, from a smartphone.

[00:10:28] And so ultimately 300 words is probably not the right amount of data for a model to really train on all of those different use cases and applications and benefits. And so the models go looking elsewhere. And there are a lot of people who are writing 3,000 word essays on, you know, community platforms. So like Reddit would be a good example. And they might say, hey, here's a 3,000 word essay on why I hate my Ford truck or why I love my Ford truck.

[00:10:57] There's very few people who go on to like a website like Reddit and write a sort of balanced, you know, right down the middle assessment. It's usually skewed to the extremes. But the models train on that because there is ultimately a lot more content coming from those sorts of platforms than from the brands themselves. And so, you know, we definitely see that. And there are other examples, right? We know that the models are very heavily influenced by YouTube. YouTube is one of the top sources.

[00:11:25] We see this over and over again in our platform. And, of course, there is a wide quality range in terms of, you know, YouTube creator content. And, you know, we can definitely see that there are influencers, for example, that might have 100 viewers that have a sort of outsized influence on how a particular LLM thinks about a particular dimension of a brand.

[00:11:46] And so what's happening right now is I think there's a kind of rebalancing where in the beginning when the models were first released, they were sort of they treated all content relatively equally. They were trained on all of it and sort of were left, you know, to their own devices to figure out what is true or not.

[00:12:04] But I think we've now realized that there might be some downsides and there's a sort of rebalancing that needs to happen and probably, frankly, more high quality content that commercial brands need to create in order to give the models the training data that they ultimately need. And so we're kind of, you know, mid stride there as an industry. And at Bluefish, you work with some massive companies, including Adidas, American Express, Hearst, to name but a few. And again, I'm quite curious here from all the conversations that you're having.

[00:12:34] What are large enterprises most concerned about right now when it comes to how their brands not only appear but are represented inside those AI platforms? Yeah, so we track a number of different AI KPI. But maybe to simplify, I would say commercial brands look for four things. They care about what we call AI visibility, which think of that as kind of your share of voice. So when an instructor is prompting AI within your category, are you showing up?

[00:13:03] Are you actually visible in that response? The second, once you've kind of checked the box on making sure that you're actually in the response, is, okay, well, how are you portrayed? How positively? So we track a number of metrics there, but AI favorability is sort of a critical KPI to our enterprise customers. Three, we've talked about AI accuracy. So what is the fidelity with which you're portrayed? And then the fourth, we think of as kind of AI influence.

[00:13:31] So think of that as how influential is your brand? Do you have kind of your hands on the steering wheel in terms of driving the way your brand is portrayed? So do the models treat your brand as a reliable, influential source that is credible and can be counted on as a source of truth? Or are they more focused on that angry rant on Reddit, for example?

[00:13:54] Of course, what's happening now is that brands are starting to think about a fifth dimension, which is conversions, which is, okay, how many actual sales has this new AI channel driven for our brand? And how does that compare to all of the other channels that we track? And so those are kind of the five dimensions that brands are focused on. And as you said, Bluefish is unique in sort of our category in that we only work with enterprise brands.

[00:14:22] We support about 10% of the Fortune 500 today. And so we work across dozens of vertical categories. Almost always, we're working with, you know, the top leaders in that space. These are brands that are spending billions of dollars a year on marketing. And so for them, all of this change that's happened over the last, you know, 24 months has really shifted the way that they think about marketing holistically.

[00:14:49] If you think about, you know, most marketing, it's based on kind of the customer journey. So how do our customers, you know, discover, engage and purchase our products online? And every aspect of your marketing function, whether it's, you know, content marketing, paid media, social media, it's all based on that customer journey, right? What's the right way for us to reach customers at each stage in their purchasing journey? Well, now the models are training on all of those different marketing touch points.

[00:15:16] And so as a CMO, your job right now is to figure out, okay, how should all of our marketing teams start to reorient themselves and rebuild the way that they run marketing based on this new agentic purchasing journey and AI funnel that we need to manage? And so there's a big transformation that's happening across, you know, certainly across corporate America and sort of the large kind of enterprise brands. But frankly, same thing is true of, you know, SMBs and mid-sized companies.

[00:15:45] If you are a company that sells a thing on the internet, this is a period of great transformation for you. The leading issue of agentic AI in businesses right now is ensuring agents act with compliance guidelines. And Denodo applies guardrails across your entire data estate by aligning your company's data infrastructure under one system. These guardrails perform consistently across your platform.

[00:16:12] So start scaling your business and start with Denodo. Simply visit Denodo.com to learn more. And as I said at the very beginning of our conversation today, for years, companies invested heavily in SEO because they wanted to influence what customers discovered through search engines. And as discovery has now moved towards AI-generated answers, even away from the browser. What changes here?

[00:16:37] What can organizations legitimately do to not only improve the AI accuracy, but also without simply trying to manipulate what these systems will say? Because there is a lot that try and game the system too, right? Yeah, it's so interesting you say that because, you know, I think in our industry we saw, especially in the early days, a lot of, you know, search teams act as kind of the canary in the coal mine.

[00:17:02] They were kind of the first teams because they've been so focused on SEO for so long to realize, hey, search behavior is starting to change. So when Neil is planning his, you know, his trip to New York, he's starting to use these AI tools to do research and figure out an itinerary and, you know, which flight or hotel or restaurants might be the best fit for him.

[00:17:22] And so we saw in the early days of our category, which, you know, we call agentic marketing, AEO, this sort of idea that this new channel was going to be sort of like SEO 2.0. It was going to be the next evolution of search. I think what's happened since then, though, is that it's become much larger than that. As I said earlier, it's sort of clear that if you're going to optimize for these LLMs as a marketing team, it's a team sport. It involves the search teams, but it also involves the content teams.

[00:17:51] It involves the social teams. Increasingly, it involves the paid media teams. Everyone sort of has a position on the field because the models are learning from the output of all of those different teams. And so I would say compared to other marketing channels that are sort of, you know, siloed off, AEO has really become this kind of all encompassing thing where every marketing team has a role to play.

[00:18:15] And so to answer your question, I would say the big difference between legacy SEO is that a lot of legacy SEO was solving for a search algorithm. You ultimately wanted to make sure that your content was kind of structured in a particular way, maybe included certain keywords. And ultimately what you were solving for was, are we showing up at the top of the results? You know, where are we in sort of the ranked list?

[00:18:42] With AI, you're solving a kind of different problem because AI isn't a list of links. It's a long form qualitative conversation that the consumer is having with the model. And so the model ultimately isn't just saying, here's a link, click on it. It's describing your brand in sort of lush detail.

[00:19:01] And so instead of trying to sort of hack a search algorithm, ultimately you're trying to build really high quality training materials and content for these much more complex large language models, which are a thousand X more sophisticated than a search algorithm. So it is like, we definitely see that a lot of the search teams kind of carried some of those behaviors and best practices into AEO. And to this day, you still need to do technical optimization.

[00:19:30] Of course, you still need to make sure that you're anchoring key phrases. But ultimately the problem that needs to be solved here from a marketing perspective is a bit more complex. And so we've seen AEO kind of break out of the search teams and really expand across the marketing organization. And before you join me today, I was reading how you've launched an AI accuracy tool that is designed to identify and help correct misinformation in real time.

[00:19:57] So can you walk me through maybe a practical example of the journey from maybe a brand discovering an inaccurate AI answer to understanding its source and ultimately then correcting the information? Yeah, of course. And so there's sort of, there are, you know, a few steps to that process. But what Bluefish is basically doing is for any of our large corporate brands, we are tracking their AI accuracy 24-7.

[00:20:23] So anytime that there is an AI response that portrays the brand in an inaccurate way across a variety of categories, so that can include brand safety, it can include pricing, it can include product details, etc., etc. Our engine tracks the response, flags it for the brand, and runs an automatic diagnosis to figure out what's causing this. What is the source, for example, that was cited in that AI response?

[00:20:49] And then our engine actually goes and visits that URL to see, hey, where is this coming from? What's causing this? What is the piece of content that the model is reacting to? And what's the link between that and the inaccuracy? And that forms essentially the foundation for how you fix the problem. So you might find that the sort of the troublemaking source is, you know, a blog that, you know, that posted something inaccurate about your brand.

[00:21:15] That might be an opportunity for your editorial team to reach out to that blog and do some, you know, some correcting. You might find that it's actually coming from inside of your own house that maybe, you know, and that can be fairly innocuous, right? Like maybe if you run an electronics company, a consumer electronics company, maybe you have a new version of a particular electronic device that you release every year.

[00:21:40] And so maybe this year, the battery life of that particular device is different from the battery life last year. And so maybe this is just an LLM that's confusing the latest model with last year's model. And there's a different sort of, you know, battery life across those models. OK, so maybe the work for you to do is to make that distinction a little bit clearer on your own content. And then the next step, of course, is measuring. Did it work?

[00:22:07] So I would say, you know, for Bluefish and in particular for our clients, they want to understand the problem. They want to see a diagnostic of what caused the problem. They want to take action to fix it. And then the last step is we need to actually attribute the sort of outcome to the change we made. So are we seeing responses change? Are they more accurate now? And when we see the citations in those responses, can we link that to the action we took?

[00:22:34] So our job at Bluefish is to give our clients sort of an end-to-end solution so that they can sort of move from start to finish within a single platform. And, of course, everything that we've talked about so far today becomes even more consequential as we move from AI, not just answering questions, to agents actually going out and recommending products, comparing services, and potentially making purchases for us. This is where we're heading. We're already halfway there at the moment.

[00:23:04] So what happens when the autonomous agent will make a decision based on an inaccurate or incomplete or outdated information and send one of your customers to maybe one of your competitors? Yeah. Yeah. You know, I think that's what's on everyone's minds. And I think that even though we are in the relatively early days of fully agentic commerce, you can certainly see where the puck is going.

[00:23:33] And I think our customers in particular are preparing over the next few years for a future where commerce really speeds up. There is this higher velocity because ultimately the agents are doing a lot of the heavy lifting. And so, you know, our customers, for example, are thinking a lot about, hey, what is the future visitor to our website? Is it a human? Is it an agent? Is it some hybrid where we have to solve for both?

[00:23:58] And so at Bluefish, we spend a lot of time helping our customers think about how do you market to the agents? What are they looking for in terms of information? What is the sort of interface that needs to exist between an agent and your website or any of your other site properties? How do you measure the effectiveness? So once you've engaged with that agent, how do you know that it was successful? How do you build attribution systems that can track, you know, did that ultimately drive a sale or did you lose a sale? And if you lost it, why?

[00:24:28] What caused that? So you have this whole industry that is popping up around the dynamic and trend that you just described. How do we market to the agents? And so, you know, at Bluefish, you know, we really think of a brand as ultimately having two stakeholders on the agentic internet. There's the consumer, of course, the human consumer, but there's also the consumer's agent. And ultimately, you'll see marketers solving for both over the next few years.

[00:24:54] And I think there's also an interesting question of accountability here. So if an AI system incorrectly describes a company, a product, or a brand, and that information influences a customer or indeed another agent, who should ultimately be responsible for accuracy? Is it the AI provider, the brand, the original information source, or a combination of all of these? Yeah. Yeah.

[00:25:20] So, you know, in sort of in recent court cases, increasingly, it is the AI providers who holds that accountability. In sort of the previous version of the internet that was much more kind of, you know, traditional search driven, where a search engine was really just routing a customer or a user to a link and say, hey, here are the links that are relevant to your query. Just go visit the link.

[00:25:45] Now with AI, the sort of AI providers are actually summarizing and synthesizing that information themselves. So they're taking a position on what is true, what is accurate. And so we've taken kind of the next step into, you know, the models actually having some accountability and playing a role. And so that's kind of what I was alluding to earlier, where you have this convergence of the incentive structures between the models, the opening eyes of the world, and the brands themselves.

[00:26:14] Everyone wants this higher quality customer experience. No one wants a model to hallucinate or be inaccurate. The brand doesn't want that. The open AI doesn't want that. And the consumer doesn't want that. And so I think what you'll see is a lot more partnerships between those commercial brands and the AI providers to make sure that this agentic chapter of the internet is ultimately a high quality, accurate one.

[00:26:39] And you're already seeing that, by the way, you know, play out on sort of the AI platforms where they are actively soliciting, you know, data from the brands. They're saying, hey, please, like, help train our model. Send us your product information so that these models can represent your brands in the most accurate way. Because if, you know, if Sam Altman ultimately wants consumers to transact on ChatGPT, you know, that angry guy on Reddit is probably not the right source of truth.

[00:27:08] And so there's increasing recognition and accountability there. And they've made an incredible amount of progress in a very short period of time. So it's only going to get kind of better, faster, and smarter.

[00:27:19] And if we did zoom out for a moment and look at this wider information economy, if people are increasingly receiving synthesized answers rather than visiting the website and publishers and reviews and company pages, how is this changing the relationship between brands, information creators, AI platforms, and ultimately consumers? And anything else that businesses should be doing to prepare right now?

[00:27:46] You know, it's so interesting you say that because there is kind of this like double-edged sword, right? It's sort of like, you know, the like Netflix algorithm where, you know, you pop open Netflix and Netflix is like, hey, these are your choices. Like I have chosen sort of, you know, and if you want to go beyond that first page on Netflix, it's really hard. It's really difficult as a consumer to say, actually, I don't want any of this. I'm going to go manually search for something. You're kind of wading into the darkness.

[00:28:13] You are seeing something similar happening to information on the internet where now it is so easy for AI to do the work. But if you want to go, you know, sort of manually override that and take the car off-roading, it is a little bit harder. You do have to do a little bit more work. You know, a lot of online consumers will take the shortest and easiest path to get to the information. But there is a little bit of muscle that you're using there, that you're losing there, right?

[00:28:43] Like consumers are trained on manual research on the internet. We're really good right now at finding the information that we need and doing the work to get there. What happens when we lose that, when AI sort of does the work for you? So there's a benefit because it is, you know, better, faster, smarter. You can get to truth. You can do incredible things with AI. But there is also something that's lost. And I do think, like, there's not enough conversation about that right now.

[00:29:10] I think everybody loves the innovation of AI and this sort of new access and capabilities that you have on the internet with these new, you know, agentic tools. But it is sort of a, you know, a step away from choice. And I think we're going to contend with that over the next few years. Well, and I think that is a thought provoking moment to finish on. And for anyone listening who wants to carry on this conversation, I suspect there are more than a few.

[00:29:37] Where can they find you or your team online and find out more information about Bluefish and everything we talked about today? Where should they go? Yeah, absolutely. So we're at bluefishai.com and, yeah, we, you know, we partner with enterprise brands, Fortune 500 companies, marketing teams from search to content to paid media. So, yeah, we'd love to chat.

[00:30:01] Well, we've covered a lot today around how AI is changing brand reputation, what Fortune 500 companies are worried about there, and also the future of the information economy, what that looks like. I'd love to invite everybody listening to feedback their experiences and what they're planning. But I'll include links to everything for everybody listening. I encourage you to check that out. And more than anything, thank you, Alex, for starting this conversation today. I'd love to carry on and stay in touch with you. But thanks for sharing this today.

[00:30:31] Neil, this was super fun. Thank you so much for having me. Appreciate it. I think Alex left us with a useful reminder today that AI accuracy cannot be treated as a one-time cleanup exercise. Brands need to monitor what models say, trace the source of an error, publish clearer information, and measure whether the answer changes. Now, that becomes even more important when an agent can recommend, compare, and purchase, all without sending a human to the company website.

[00:30:59] And that Netflix comparison also stayed with me because I think convenience feels wonderful until that first screen effectively defines the limits of our choice. Sometimes I want to go and watch something or listen to something that is against the algorithm. That's what's wonderful about being human. I like the idea I might have a guilty pleasure whether it be a TV show or a band. So, yes, AI can reduce the effort required to research almost anything.

[00:31:27] But we should also ask, what happens to our own research instincts along the way? Well, I'll post all the links in the show notes. But over to you. Would you trust an AI agent to choose a product if you could not see the information behind the decision? Lots to think about there. TechTalksNetwork.com. Keep your messages coming through. And I will speak with you all again real soon. Bye for now. Bye for now.

[00:31:54] Bye for now.