Behind the Model: When AI Researchers Meet Content Marketers

‍

On April 23rd, we hosted a panel at AI House in Seattle. The premise was that the people building AI models and the people whose work depends on them aren't usually in the same room. I moderated, joined by Apurva Luty, founder of Optimly, and Javier Luraschi, founder of Hal9. What follows is the conversation, lightly edited for readability.

Alex: We were trying to think up a theme for tonight, and the one Apurva explained really well is that we wanted to bring AI folks and marketers into the same room. Running a content marketing community right now, it's very clear that there are a lot of people trying to sell you a lot of things on how to show up on AI search, AEO, GEO, and all of it. Things are constantly changing. It's cool to go back to the roots of how these models actually work, and think about them through that lens to help you sort what's BS from what might be real. I'll let you two introduce yourselves. I'm Alex, I run a content marketing community called Superpath.

Superpath is a global content marketing community where we chat about the industry of content marketing. It's a lot of folks who are freelance, agency, and in-house, trying to navigate the new world of AI together.

Apurva: I'm Apurva, CEO and founder of Optimly. I have a hybrid background. I spent most of my career in product marketing and brand roles at big tech companies, most recently at Discord, where I led their brand repositioning and rebrand. I found my way into AI and built Optimly as a result of some of the problems I ran into at Discord doing brand measurement and brand research, and figuring out how to evolve it in a world where AI is quickly becoming the intermediary between you and your customer. I know Javier from essentially the inception of Optimly. He was one of the first people to really educate me on how AI works, how LLMs work, and where's the signal and where's the noise.

Javier: Hi everyone, I'm Javier Luraschi. My background is mostly technical. I went to school for mathematics and then came to Seattle because of Microsoft. After that I worked at RStudio and now I have a company called Hal9. We help startup founders build their AI products.

How do these models actually work, and how are they updated?

Alex: We've got a set of questions we put together looking at things from a content marketer's lens, and we'll have space for a few questions at the end. To start foundationally, we don't have to spend a lot of time here, but how do these models work, how are they updated, and what's important to know about that process?

Javier: It's a really important question because everyone has used ChatGPT. Those models use all the information on the internet, which is a lot. The way they're trained is using the data of the entire worldwide web. The model takes all of that and gets compressed.

What's challenging is that taking all the information of the world isn't a simple process. It takes literally hundreds of thousands of computers, GPUs, to take all that data and make it succinct and understandable for the models. That process isn't something you can do in an afternoon. The large language model providers like OpenAI, Google with Gemini, Grok, and Anthropic with Claude, they take something like months, up to six months, to put them all together. A few months of training, a few months of testing, fine tuning, then releasing.

In the context of marketing, which Apurva is the expert on, it's not trivial to say "I changed my website, the model is going to understand my stuff." It just takes so long for these models to understand the newest data. That definitely affects marketing fundamentally.

How does AI handle web search?

Alex: It might be interesting to dive a little deeper specifically on web search. It's something content marketers are familiar with. We've been in the SEO world for 15 years. How is AI interpreting web search?

Javier: I can go back a little, taking the conversation where I left it. The way that the models got started to be trained was with the entire web, which takes six months. We still haven't figured out a better process. I'm saying we as in the AI community and AI researchers haven't figured out how to solve the problem of getting all the data and crunching it in an afternoon, such that when you change your website, it appears the next day.

What we're doing today, the best LLMs have been doing so far, is that we don't know how to retrain the entire model fast, so let's just make a quick web search, and sometimes not just to one search provider, sometimes multiple ones. The process is a little more complex, but the best patch we have today is: if you ask a question, let's use what the model knows, but let's also go on the web and ask Google or Brave or the like. Get the listings out that exist on the web, push them into the model, and that way we get the most recent data and come up with a better answer.

When you ask ChatGPT what should I do this Friday in Seattle, the model really doesn't know anything about what's happening in Seattle this Friday because it takes six months to train. What OpenAI can do is go out on the web, figure out what's happening in Seattle, get 10, 20, 100 pages, push them all to the model, and let the model decide what's a good answer. It's bittersweet, because that's the best we can do right now, using search to log the gap we have in training these models.

It works, and the good news for an expert in search engine optimization is that some of the skills carry over, but it's both. We can't see the optimization of models just by searching. That's just part of the problem. The other part is how to make the model fundamentally understand the correct answers. That's a longer-term view. It's not something you do with SEO. It's something you do by waiting and preparing your website to be trained whenever the next training cycle comes. So it's a tactical strategy, a dance of how do you get the model to do what you want it to do.

Where's the disconnect between how models work and what marketers are being told?

Alex: I want to weave this together and have you be the bridge here, Apurva. One thing that frustrated you, that you've talked about, is that you straddle these two worlds. You came from a marketing background, now you're living in the world of AI folks who understand models. There are two separate conversations happening. I want you to articulate what some of the things are that maybe aren't being said across those two channels.

Apurva: Javier, thank you for your explanation. The biggest takeaway from my early conversations with you was that AI models are not search. They're entities, they're systems that can do search the way that you or I would run Google search. We can't fundamentally know whether what a model is saying is because of information it retrieved or information that's core to its belief.

That was the insight from Javier. This isn't just an algorithm you optimize for. You have to understand how it forms beliefs and perceptions. Given my background in marketing research and brand, I was like, that sounds like brand measurement. Really just trying to understand how people perceive brands and to what degree they're aware of brands. That's how we started approaching the measurement piece.

When I got further into that and started talking to more folks in the AI community, but also my marketing friends, I realized there was this other conversation that wasn't happening about how these systems actually learn. Again, it's not about optimization. It's about feeding them the right knowledge so that they can do what you want. In the AI community, what I was hearing is there's just a dearth of good new data to train these models on. That's a huge problem for the model providers. At the same time, people are trying to take shortcuts. They're trying to publish hundreds and hundreds of blog posts and hoping that's going to be the thing that sticks and gets retrieved by an LLM. What an LLM is looking for is net new knowledge and perspectives that it can use to iterate on the next model. Information gain is kind of how I'd put it.

What does information gain look like in practice?

Alex: You talked about this information gain concept, which I found to be a really helpful framing. Let's get more specific with examples. A lot of people are saying, "I need to come up with some hack to push out a bunch of AI-produced content." What you're saying is that's not helping, because there's no new information going out there. Could you explain that with concrete examples? What are people doing that's not helpful, given how the models work?

Apurva: You'll have to bear with me because my background is coming from a consumer perspective. When you launch a new brand or a company, you have a lot of market education to do. We started approaching it from that perspective. We did this for ourselves. When Optimly launched, there was another company called Optimly out there. They were clearly doing a lot more SEO than we were, but it was a different product. My initial reaction was, oh shit, we're going to have to rebrand.

Then we said, what if we use this as an opportunity to see if we can train the models to understand that these are two different entities? We started being more strategic in terms of understanding what the other Optimly was doing from messaging, positioning, what their product was, was it even the same? It wasn't. That informed our core messaging and positioning strategy. It informed not just how we showed up on our website, but specifically how we showed up on third-party sources that LLMs were learning from, both about us and the other Optimly. That's eventually how we were able to do this entity disambiguation. We taught the model from the outside. This is what Optimly AI is, and this is how it's different. It was that information gap. The model just didn't know.

So is this just "make good stuff"?

Alex: A lot of the conversations we have in Superpath are: ‘How new is this? Is it just good content marketing?’ I'm putting out good new information. To what degree is this just "go out there, make good stuff, and don't worry about how the models are going to pick it up because if you put good new stuff out there, it'll work?" Or do we need to be structurally thinking about this differently and there are some things we're optimizing for?

Apurva: A little of both, but more of the latter. I remember having these early conversations and I'd be like, "Yeah, to sum it up, what you really need to do is just really good marketing." And people would laugh. They'd say, easier said than done. That's true. Think about all the things it takes to figure out what even constitutes good marketing.

This is where the brand measurement piece really comes into play. It tells you where you're positioned in the market, where customers find value. A lot of it is that, but you're doing it at a much faster clip, because models aren't updating every six months. They're updating every couple of months. You have these windows of time, and you need to be strategic about what you put out there. So it's the same discipline of good marketing, but you're having to do it a lot faster and figure out what data is signal versus noise.

What about the cottage industry of AEO vendors?

Alex: One thing we talk about a lot is how overwhelming it is. There are so many new vendors and so many old SEO vendors that are now AEO folks. It's hard to navigate that space. From a model training perspective, what do you think? They've all got advice that's like, "this is the one thing you have to do." How should we be navigating what different vendors are telling us right now? Are there rules of thumb for telling if it's BS versus if there's some truth to it?

Javier: I'll leave the specifics of the vendor product space to Apurva, but I can go a little deeper on the technical side. The short answer is it's very different what we need to do for marketing for LLMs today than what we did in the past. It's just fundamentally different.

When I say the information gets compressed in the large language model, the history of LLMs comes from a concept in computer vision called neural networks. Believe it or not, neural networks is a concept in computer science borrowed from biology. Back in the late 1800s, Ramón y Cajal figured out that inside the brain you have neurons, and they're interconnected. Computer scientists, including Turing in the 1940s, started modeling computers with neural networks. Long story short, Geoffrey Hinton, considered the founder of neural networks, repopularized them. Eventually Google took some of that learning and came up with the transformer, a complex neural network, a mathematical model that resembles the brain to understand text.

When we look at large language models today, they're not doing search the way Google does. They're trying to understand the text as a human would. To some degree that resembles the human brain. In other ways, we don't even understand the human brain, so we don't fully understand how LLMs fully work. When you do a Google search for "cat," what it's doing is looking at all the pages that say "cat" on the internet and figuring out which one ranks highest, has the most links to it. That's very different from when you type "cat" into Anthropic or Gemini. It's trying to figure out how "cat" is related to other terms in this simulation of neurons. If you're working with an SEO provider that's trying to get you more backlinks, that just doesn't work in the new world, which is all about understanding and consistency and getting closer to how the human brain works. That's why Apurva is saying you need to do good marketing. The way the human brain understands marketing is way more subtle than just searching for the word "cat" all over the world.

Apurva: You know more than you're giving yourself credit for. We've been teaching each other.

My perspective on this is that marketing to an LLM is more like marketing to a human child. I use this analogy often, and it makes my husband laugh because we have a five-year-old. If you ask a five-year-old why the sky is blue, they might say five different things on five different days. Because you told me, my friend told me, I'm looking up and it's blue. You don't really know what all the inputs into forming that belief were. An LLM is very similar. If you're really just trying to reverse engineer or optimize an LLM, it doesn't make sense. It's like saying I know these three things will convince this specific human to take this action every time. That's next to impossible.

My advice, which others gave me when I started talking to AI folks, is to take any information you get where someone's telling you with certainty it's this or that as a new experience. Because no one really understands how these things think, including the people building them. In fact, I'm pretty sure Anthropic came out recently and said their latest model built itself.

If you're a content marketer going back to work tomorrow, what should you do?

Alex: All content marketers right now are trying to build a strategy around something that's constantly changing. We don't know how it actually totally works. You're just trying to grasp what's the thing I should be doing. If you were a content marketer going back to work tomorrow, what's a waste of time? I'm thinking about tactics like creating pages that are only for LLMs to crawl, or creating FAQ sections. What should I stop worrying about, and what tactical things can I be doing that are worthwhile?

Apurva: Honestly, the only tactical thing worth doing is running audits for your brand. A full 360 audit to understand how your brand is currently perceived by an LLM. The corollary is what I did for a long time, brand measurement, brand intelligence, brand perception measurement for humans at scale, applying similar approaches and just understanding how my brand is perceived by an LLM today.

Then using that data, figure out the gap. How do I want to be perceived, what action do I want the LLM to take, versus how does it perceive me and what does it currently associate me with? The delta between those things is where the most critical insights are. That can help build a strategy. It's going to be different for every brand. There's no silver bullet, but there's hope in the fact that you can audit your brand today. You can get access to the OpenAI API and start to probe its underlying knowledge.

Javier: To connect this back to how LLMs think like humans, which I know sounds abstract: early on last year, Apurva ran an audit on my company, Hal9, to figure out what the LLM thought about us. The interesting thing is the LLM was very confused about what our company did. That's not great, because we want to have a consistent story. When Apurva troubleshot why the LLM was thinking the way it was, we found a lot of pages that were out of date on our website. Like any other startup, we'd done pivots, and the pivots were left behind.

The aha moment was that it wasn't just LLMs being confused about what my brand was. My customers were getting confused too. I'd tell them, "I can help you build an AI product." Then they'd go to the website and find a blog post that says, "Hey, data analytics is awesome." And I'm like, that's something we did one year ago. So as part of teaching the LLM as a person what our brand was, we had the side effect of teaching our prospects and customers a more consistent story.

When you talk to a person and don't convince them what your brand is, that's it. They close the tab, they go away, there's no way of getting feedback. You're lost. Even though I was thinking of optimizing for LLMs long term, when we started optimizing our brand for LLMs, we started seeing short-term gains with prospects because our brand became more consistent. We could troubleshoot. Why does the LLM think Hal9 does data analytics? Oh, that's this website. Let's go fix it. As part of teaching the LLM what you do, you also create a more cohesive brand.

Alex: That resonates with my own experience when I was in-house at an HR tech company. We were a mid-stage startup, repositioned several times over a five-year period. We had this older product that was still one of our products, but it was only one part of a bigger platform. We started looking at what LLMs were showing, and it was all of our old stuff. Because we had hundreds and hundreds of blog posts talking about us as a people analytics platform. It was like, no, we consider ourselves an HRIS now. How do we catch up with the LLMs? They were constantly saying, if you want a people analytics platform, you go to this company. We wanted to be mentioned among these other companies.

That's a big challenge. I also remember the moment we updated our pricing. We raised our prices a little, and I was like, what's going to happen? People are going to come to us seeing the old prices on ChatGPT and be like, "I saw your price was this." It's nerve-racking.

Apurva: I was at Discord when we did the brand refresh. The agency we worked with said, we've got to update all of this stuff. Convincing my CMO that we had to spend six months updating the entire site didn't work until I started putting these AI audits in front of her. She was like, you mean AI can see all of this, even though we've hidden it, it's not on our site? And I was like, yes. That's how we got the budget to finish the work.

[From the Audience] How do you balance human and AI work in your own roles?

Apurva: I can go from the marketing perspective. I think of using AI as a scaffold or a coach. I recently started a newsletter, and I have a Claude project I use specifically. I've given it instructions: do not write anything for me. Just be a coach, ask me questions, do a content interview with me, help pull ideas out of my brain and structure them. It still takes a couple of hours a week to write, so I feel good about putting my name on it. But it probably compresses four or five hours of work.

Alex: Within our content marketing community, we have the full range. On one end, people who still aren't using AI in the writing process at all and won't. On the other end, people where 80% of a post was probably written by AI, but they put in a bunch of prompts, structures, and examples of their writing. I've thought about putting out a benchmark to our group: how many words do you think AI chose out of this blog post? The structures are very different.

The median person inside Superpath probably uses AI heavily in brainstorming, then around some structures to create the first draft, and then goes from there. It's like, "Great, this is a first draft like a junior person put together. Now let me put my hat on and run with this." That's what I'd say the median person does. There's a full range.

Javier: I'm biased because I worked in data science for many years, and I know very little about marketing. I have a lot of respect now for people who do marketing, because it's not that easy. My point of view has been: just get data. If someone says we should do blog posts fully AI-generated, I'm like, whatever, let's just try it and see what happens.

In general, the answer isn't simple. We've seen one case where AI was able to do what we were doing on par, with TikTok videos. Maybe that's because we're not good at TikTok. We did the whole "let's explain on TikTok what we're doing," and then we did some AI-generated stuff, and it was getting the same amount of likes. So either we're so bad at TikTok that AI was doing our job, or it really is similar. There are other cases where we know for sure that on sales, sending a personal message on LinkedIn or making a call makes a huge difference. We have someone on our team now doing cold calling. Maybe it's just because people are getting tired of AI, but having a person who actually cares about you and calls you is useful.

It's fair to try whatever you want with AI. Just make sure you measure it. If it works, great. If it doesn't, it was an experiment. We use AI for everything, but the mix is very different. Sometimes AI is 10%, sometimes 90%.

Apurva: Follow-up to that. It sounds like what you're describing with the TikTok videos is content engineering. You're using code instruction to instruct an outside agent or tool to create stuff for you. Is it just creating it? Is it also publishing it, measuring it? To what degree is this automated?

Javier: It's not automated, and I get really cautious. There's a lot of AI hype where people say, "I just turn that key button on, and it's automated all my marketing." For most of the examples I've seen, either people are misrepresenting what they're doing and it's not fully automated, or it's fully automated but there was an engineer working on it for six months to get it out. In our case, we have someone who still generates the script with the help of AI. The only part that's truly AI-automated is generating the video. Even she has to stitch it together. There's a lot of manual work.

Even though it looks like we're using AI, there's a real person who has taste and curates what worked and didn't. There's a lot of hype and a lot of opportunity, but human taste and curation is going to be super important. Even if you can generate 100 marketing campaigns, someone needs to review them. The bottleneck becomes me reviewing them. If it's not good quality, then I become the bottleneck, which means: does it matter if it was generated or not, if someone is reviewing it?

Alex: There are very specific things that, in running Superpath, I'm comfortable being AI-generated. One example is our monthly 1:1 program where we pair people with each other inside the community. We send "here's why we think it'd be a good match." I've got Claude Code basically building a curated matching thing, and it builds out emails and drafts them. There's always the last-mile thing of "actually, I don't want that extra sentence saying you had a really bad match last time, so hopefully this one's better." Most of my writing is hands-on. Even the stuff I'm comfortable with AI writing, I'm always reviewing the last mile to make sure there's nothing embarrassing. If I don't have eyes on everything that goes out, I get nervous.

Apurva: One of the underrated workflows for AI in marketing is just measurement. For me, I'm someone who's stitched together five different dashboards trying to get AI to consolidate that workflow. That feels right for AI and automation. When I usually talk to people about marketing agents, they're promising me the world, like it'll do everything end to end. I'm like, can you just do the boring part for me? Alex and I were just talking about using AI to generate UTMs because I hate doing it. It's not how I want to spend an hour. Generating those according to a schema is so annoying. We'll soon be in a place where you can run a white paper, content asset, blog post, or social campaign through an LLM and say, "map the links in here to the UTM parameters we've also uploaded," and it'll just do it for you.

[From the Audience] What Does Optimly Actually Do? 

Apurva: I probably missed that in the intro. Optimly does brand intelligence and brand measurement for AI models. Everything I described about measuring brand perception with humans, I've taken my career doing that and turned it into a platform where we're doing it with AI models. We use that data to help founders do better content marketing and understand their marketing strategy.

Javier: I'm a fan. I can do a back-to-back or reverse pitch here. The way I see it, more from an engineering side, is you put in your URL, hal9.com, coca-cola.com, whatever, and it tells you what the LLM thinks and how to fix it. To me, it's a lot of value.

Apurva: It's the brand audits. When I started doing those audits a year ago, it was taking me a week. Now we can do it in a few minutes. That's a testament to AI, but it's also this underrated use case where you can do a pretty good job of automating the measurement piece end to end and going deep.

Audience: Could it just be ranking within the LLM, or is it also SEO, or a combination of the two?

Apurva: It's basically querying the underlying knowledge of an LLM. There's retrieval as well, but we're really trying to distinguish between the two because we want to help brands influence that underlying belief system, which we think will produce better marketing long term. The measurement is more akin to NPS surveys or brand measurement surveys, but with LLMs as the panel.

Audience: Is it based on the keyword of the brand itself, where you're measuring the sentiment of the company within the internet space? Is it kind of like the social listening management industry?

Apurva: I'd describe it as a hybrid of social listening, search intent, and brand. As far as I know, there aren't a lot of people doing this exact thing, but we found it very helpful in guiding strategic marketing and content decisions.

Audience: Do you sell to PR agencies?

Apurva: Yes.

Audience: I'm just trying to clarify. It's not like AirOps that will measure...

Apurva: No. Did you say Meltwater? I'm familiar with Meltwater. Yeah, similar to Meltwater. Again, it's an industry where measurement has always been difficult to do. The cool thing about LLMs is that, unlike humans, it's easier to find high-quality samples. Doing phone calls and surveys with humans is almost impossible now. You can also do attribution and understand what they're learning from, which most humans won't tell you on the surface.

[From the Audience] How do you balance the unpredictability of LLMs?

Audience: Given what you said about the unpredictability of the outputs day to day, your comparison to a five-year-old, when people use Claude or ChatGPT to iterate or poke holes in things, that's probably fine?

Apurva: To iterate, that's probably fine. Where we found it was not reliable was specifically for measurement, which is why we built all the automation around the measurement pieces. What Javier would say is LLMs are probabilistic, and when it comes to measurement, if you have measurement that's changing every day, that's not really helpful. So we're applying deterministic measurement methods to what is inherently probabilistic. For non-measurement cases, it's probably fine. Content ideas, that kind of thing.

[From the Audience] How do we even know LLMs think like humans?

Audience: I was just thinking about your comparison to how the human brain works. How do we know it's similar to how people think?

Javier: That one is so much broader than marketing. The first short answer might be: maybe don't trust it. I have a funny story to share about that. The whole theme of this talk is that LLMs do think a lot like the human brain. In the same way that you don't trust your friend for everything, life advice, work, technical stuff, it's the same with LLMs. We need to be responsible enough to understand they're not perfect, and to know in which cases they are kind of perfect. If you ask them to translate this from Spanish to English, I'd give them 99.9. They're pretty good at that. Critical thinking is important.

In 2023, super early on, we asked the LLM a question, and the LLM answered the question correctly. Which is great. But it actually deleted the entire database and recreated it to answer the question. From early on, we learned not to trust it, because it's like someone you don't know that you're getting to know.

We've had LLMs for three years now, and there's a lot of goodness that comes from them, but critical thinking is super important these days, more than ever. Hopefully our kids are learning that. It's like having someone who is really really smart on some things and really really not smart on others. Figuring out where that line is, even hard for us in the field.

Apurva: I'll add to that. The confidence with which an LLM will answer your question is a sand trap. It will grammatically sound like it's very smart, has researched the entire world, and is giving you a very confident answer. That is an illusion. A lot of the time, it is coded to make you feel good about the answer it's giving. That doesn't necessarily mean the answer is true. So you have to be, to Javier's point, critically thinking. If you ask it for the top 10 best things and then say, "actually, I don't think these are very good, I only like this one," it'll agree with you. You're like, you're right? It doesn't really know. It's delivering an answer that you're supposed to feel confident in.

Alex: I feel like from my own experience, I use AI for a lot of brainstorming processes. I need to come up with a list of five things and I have three of them. I'm like, here are my first three, come up with another 10. Seven of these are terrible. Two of them are bad. I'm just trying to get the last two. You said all 10 of these things really confidently, and I'm like, actually, that's a good one I didn't make up. That's very simplified, but I never take the 10 ideas at face value. Some of these could be good, some of these are horrible. The overconfidence is worth mentioning, because they are super confident.

Javier: It can bring in a different angle. It can make you think about something. A great example is healthcare. If you ask "I have these symptoms, what could I have," getting that list of all the things is useful, but you don't want to get the advice directly. You need to go to the doctor and confirm.

Apurva: Data in, data out. So you have to be careful what kinds of closed systems it's pulling from. The more disastrous the consequences of being wrong, the more careful you need to be with what it's pulling from.