에피소드 138: How Contextual AI Could Change Advertising with Brendan Norman

I'll say that like I have learned so much in the last year. How I would describe things a year ago versus now. And then also I realize that I'm going to just keep learning like the whole commitment to doing this job as somebody who wants to see change and is trying to change things in the direction of travel that I think is best. I'm just deeply committed to realizing that there's so much that I don't know and I don't even know that I don't know and I'm just trying every day to get a little bit better.

By explaining what we do has evolved over the last year and early on it was like cool we can help your brand show up against the right content and that's going to drive better results and better outcomes for your brand.

Hi my name is Demetri Bonichi and I'm a content creator agency owner and AI enthusiast. You're listening to the AI agents podcast brought to you by Jot Form and featuring our very own CEO and founder Idkin Tank. This is the show where artificial intelligence meets innovation, productivity, and the tools shaping the future of work. Enjoy the show.

Hello and welcome back to another episode of the Agents podcast. This one we have Brendan Norman. He's the co-founder and CEO of Classify. How you doing today, Brendan?

I'm great, Demetri. How are you?

I'm living the dream. Feels like the year's already off to a hot start. A lot of cool stuff to talk about in the world of AI and agents and all that kind of stuff, but I'm really excited to get to know you a little better and learn more about Classify. Just to kick things off, tell everyone a little bit about how you got into the world of AI in general and we'll go from there.

Happy to. I've had a couple different career progressions before joining the tech world a little over a decade ago, but I got really lucky and joined the startup that got acquired by Facebook. Got to spend five years helping to build out the audience network, which is where I learned everything about go to market, how to build a business, how to market that business, product development, and understand what you need to be building. When we started off, nobody knew that we existed, and by the time I left, we had a $3 billion business that was number two right behind Google's business. I also learned a lot about advertising and learned that when you're trying to market an ad to somebody, if you know a lot about them and also know about what they're looking at, like a web page or a video, it's the best time to distribute that ad. There could be many different versions of ads and many advertisers looking to reach you, and the when is the really important question. That's why Facebook ads convert so well, and Instagram ads convert so well.

When I left, I realized that the whole open web didn't really know anything about that part of the equation. The whole open web had been organized into a really basic version of taxonomies, categories, and keywords that don't really meet the needs of how advertisers should be reaching people in content at a giant scale. Fast forward to AI, I pulled together a group of really smart data scientists, engineers, and product people, and we built a much better way to understand content and organize it in ways that make it really searchable and easy for advertisers to reach people. That was a little over a year and a half ago. We were building in the ad tech space and have been helping to service clients who are brands looking to reach the right people at the right time.

But the evolution you mentioned, things have gone by really fast this year already. In the last three months, we've taken a step back so that we can take many steps forward and realize that the product we built is incredibly AI native at the infrastructure layer. It's this contextual intelligence, and I'm happy to go into more depth because we've discovered that it's a really important ingredient to how AI agents understand web content and it makes them really useful. So that's kind of the current version of where we're at.

Just out of curiosity, you've seen everything going on in the last couple weeks with Claude, well, whatever you want to call it now. I think it's got four names in one week, which is a record. It's the fastest growing open source project and also the most names in the first week probably ever. You're working in a very specific space that's also cool. You work on giving contextual intelligence for agents, advertisers, and publishers. With the world of agents analyzing information in this ad gold rush, I'm very curious how you guys fit into the story of what's going on with AI right now.

If you'd asked even six months or three months ago, I would have told you that everybody was kind of sitting back and looking at agents, especially in the ad tech space, a little skeptically. There's part of the ad tech industry, and I'll go through the lens of how we've been thinking about it as a company, but because we're so ad tech focused, we've been helping to design some of the protocols for how more B2B agents interact. One of them is built on top of MCP called ADCP. Brian O. Kelly pioneered this idea and built a protocol that allows different selling agents, buying agents, and creative agents to talk to each other using natural language. It standardizes communication between supply side platforms, demand side platforms, and creative signals providers like us that do contextual intelligence to allow all these agents to talk. That's a very B2B component.

It's important to take a step back and look at the different types of agents and bots. Bots tend to do autonomous things like drive traffic and click on ads and are kind of the not good part of the internet. Agents tend to be more connected to a human and do something autonomously for humans. There's a broad swath of different types of agents. We were building in the direction of travel to support more B2B type agents like signals, buyers, creative seller agents. Recently, we took a step back, especially with cloud bot, open claw, and what's happening with multbook, watching agents talk to each other and realizing many agents are scraping web page content, finding content to grab, making that content usable inside an LLM with a human on the other side. Because it's more probabilistic than deterministic, it's not always the same input and output.

There are many opportunities to advance the industry and standardize that through control plane, which means taking contextual intelligence, containerizing it, and making it accessible for agents to retrieve information in a standardized, repeatable way. When you go to a chatbot and ask for research, like finding the best pair of shoes, it's fairly open-ended and you'll get different answers each time. The next big phase of infrastructure development, in addition to better inference and larger models, is the contextual intelligence layer that helps standardize how retrieval and understanding of content works.

A lot of that stuff is changing. I don't know what got you to the point of creating Classify. It was you and a co-founder, right?

Actually, there are four of us. I had been building something totally different. I thought I left ad tech a couple years ago and was building a platform to serve some self-interest, connecting backcountry ski touring guides with people like me who wanted to meet them and book guiding services. I tried to market that platform but couldn't find any advertiser technology companies that could help me find the top 5,000 web pages that just talk about backcountry ski touring. I realized there was a huge gap in insight into understanding web content at scale. I pulled together some friends from the industry and we built a much better way to scrape content, classify it, which is where the name comes from, and organize it into databases that make it accessible and searchable quickly. It's useful for advertisers, publishers, and now agents.

When you look at your website initially, it's kind of hard to find something like this before. I interview like three founders a week and still feel like there's a new category of thing you can do to help out with AI. Would you say this is a decently niche service and how does it compare to other AI tools?

It's an interesting question. I think niche isn't the right word. It's a very specialized service. It's specialized in the context that not many people do what we're doing, but the applications aren't niche. It's about context classification and contextual analysis of semantics, which is what we do. The industry hasn't really thought about how important standardizing the contextual understanding of content is. That means the words on a page, the technical understanding of the page, and eventually video and audio too, like semantic understanding. A human analogy would be telling a friend a five-minute story about a date. A human understands all the nuances, not just tagging it as dating or Italian restaurant. Machines haven't done a good job of understanding and organizing content in a way that makes it searchable.

We took a different approach to classify content in a much deeper semantic way like a human would and put it in a vectorized database that we can organize and make searchable quickly. It's very useful.

On your pricing page, it shows how a human sees content versus how an agent sees it. Agents or models don't quite know what words mean. Without proper context, they can completely miss the mark. Like with the date example, there are baked-in assumptions in humans versus models. For example, someone might convince models that lawyers using M dashes are all people. Models might overdo some things that don't actually make it sound personable. Is there a better way to phrase what I'm trying to say?

It's really cool because it's developing so fast. We're all figuring out how to describe this stuff, even the people building it. Every human looking at this, from my mother in Kentucky sending me articles about agents talking to each other on Reddit to people building it at OpenAI, is trying to understand it. These are software designed by humans getting better at understanding stuff. Three years ago, chat GPT was novel, letting you talk to a computer program to write a song in the style of Drake. Now it's advanced to writing legal documents, though it can't legally be a lawyer. With more data, parameters, and infrastructure, contextual intelligence will make everything smarter. It's still probabilistic, so outputs vary. We're investing in infrastructure to understand digital content contextually so agents can work with it repeatably.

What are the main improvements happening on the back end that people haven't seen on the front end, and how do you help improve the gap in context understanding? Models like Opus 4.5 seem to bake in context better or make better probabilistic assumptions.

They definitely do. Part of that is prompt engineering—being more specific in how you write prompts helps inform the model. It's like having an intern get you lunch: vague instructions might get you anything, but specific instructions get you exactly what you want. Over time, the intern learns your habits. Adding skills to Claude lets you say, 'Say this in my voice,' or analyze bodies of text to replicate tone for generative AI or coding. Another interesting area is building your personal context window. Chat GPT can only store so much conversation before resetting context. There's a memory challenge that's not sharable across different models. Some companies are figuring out how to containerize and personalize this, making styles and conversations interchangeable. Much advancement is being built using generative AI, speeding up prototyping and market entry.

I've found Claude skills very useful in my workflow with Claude Co-work and MCPs. I'm moving from tools like Notion to more basic delegation systems like Todoist because it's easy to make skills for projects or categories of work. You can assign tasks and update context easily. How does your tool help with analysis and contextual intelligence for agents like Claude Co-work, which is essentially a VM agent installed on your computer?

Anything that touches content benefits. For research or product lookups, any web-based content, our tool was designed for advertising and operates at big scale. It standardizes understanding when reading a web page quickly. Right now, asking the same question many ways yields different answers. We're building towards standardizing that understanding and translating between how an LLM reads a page more accurately, faster, and cost-effectively. Running LLMs to scrape whole pages is expensive. We've figured out more cost-effective ways. Everything costs a lot now, but costs will come down in six to twelve months, accelerating development. You'll be able to build, prototype, get feedback, and iterate faster.

Have you been using tools like Open Claw or Claude Co-work personally? Are you using cloud code?

I was very anti-AI six months ago, wanting to do everything myself. But then I started dabbling with Claude and ChatGPT more deeply, using tools like Bolt, a website creator integrated with cloud models like Opus and Sonnet. You can ask it to build something and it writes code with real-time previews. I've gotten deep into using cloud for front-end website development, even though my team focuses on backend and data science. It's enabled me, a nontechnical person, to do front-end development quickly and learn about Tailwind and React. I rebuilt our whole website over Christmas break.

I also wanted to show clients ROI. For big publishers or supply side platforms not getting good signal into hundreds of millions of web pages monthly, we show how increasing accuracy at scale improves incremental revenue. We built calculators on our website as sales tools to show potential earnings and industry value.

How many people are on your team?

We're four co-founders plus a couple of part-time helpers.

Not large at all.

The best part of having a lean team is speed. We had an offsite last week. We're remote, distributed, asynchronous but get together when we can. When we do, it's 16 hours a day of a firehose of ideas and building. We then build asynchronously and collaborate over Slack. A lean, diverse team with different perspectives and personalities pushes each other to get the best results and focus on the right priorities. I've worked at large companies and amazing teams, and also teams without that compatibility. With the right people and framework for prioritization and execution, you can get a lot done. Modern tooling lets us ship products faster than ever.

Do you think there's a workflow change so drastic for small teams that small business owners can outmaneuver big organizations? While some fear AI Armageddon, do you think smaller businesses will outgain larger companies with niche or focused products?

It's definitely changing quickly. Opportunity is shifting way down funnel. It's going to be a while before a single person can build something like Salesforce with AI, but anyone with a niche business can build quickly and easily and make good products. That's just going to get better. Recently, developers shifted from 80% manual coding with 20% AI to 80% AI with 20% manual review. Many developers now chat with tools like Claude and Cursor. There's going to be a lot more opportunity, like when Instagram democratized photography and created influencer economies. Small business owners with great ideas can now build, distribute, and test products quickly. Advertising technology on the open web has been limited to large advertisers, but new tools will enable smaller owners to get their word out. It's democratizing and increasing competition and saturation, but good ideas will win where they previously needed huge teams.

There's going to be more opportunity for small teams to succeed. Previously, you'd need many contractors to make it big, but now delegation and project management with MCPs and skills make it easier to get work done with fewer people. There will be saturation but also unique flares. Even I find it hard to keep up with AI developments despite interviewing many founders weekly. There's a lot of opinion without requisite understanding. How has it been dealing with customers and explaining your vision in a market doubling capabilities every two months?

I've learned so much in the last year. How I describe things a year ago versus now has evolved. I'm committed to learning and changing things in the best direction. Explaining what we do has evolved. Early on, it was about helping brands show up against the right content for better results. That led to realizing understanding content is foundational to many spaces. Clients, investors, and the market have evolved to see the value of a contextual intelligence layer that deeply semantically understands content like a human and makes it accessible to industries like advertising and agents. We're becoming a translation or education layer, helping agents deeply understand content and improve their game.

In the next quarter, we want more feedback loops. We've classified close to four billion pieces of content, covering most of the ad-supported web. We run pixels with advertisers to see campaign performance, giving us reinforcement learning to improve understanding relationships between content. Contextual targeting campaigns do 50 to 200% better than standard technology. Our classification is two to three times more accurate than competitors. We want to scan more content, test relationships, and get smarter to provide agents deeper context understanding, making tools easier and more consistent to build and use. We aim to be the layer underneath everything that standardizes this.

What draws people to your product over competitors?

It's a combination of having a really good product that works well, existing clients recommending it, sharing it directly, doing thought leadership and publishing content to drive inbound, and direct sales building relationships and integrating into communities we serve so people understand who we are.

On a personal level, what's the best tool you use daily to get more work done and make life more sane?

Claude has quickly become my backbone. It's an MVP in the sports term sense, most valuable player, not minimum viable product. I use it for everything from having an AI assistant listen to calls and transcribe them, getting articulate notes, seeing things I missed, bouncing ideas, building prototypes, and reviewing my work. It will continue to get better. I use ChatGPT more for technical things and focus more on Claude for those.

Do you have interest in autonomous work bots like Open Claw? Have you played with them?

I don't know the full story but I guess Claude bot had a cease and desist because it was too close to Anthropic's product. Open Claw was probably a fun play on words. The creator was pseudo-retired and made it as a side project, and it became the fastest growing open source AI project. Are you going to get into autonomous work agents for personal work? Definitely. I have friends building that kind of stuff. Personal assistance with the right guardrails is important. There have been stories about open claw agents deleting codebases or buying huge amounts of hamburger meat. Better guardrails are needed before autonomous agents operate at massive scale. But agents are already operating at scale in advertising. Every human will have many agents and sub-agents doing different tasks. I want to spend more time thinking about guardrails and training them effectively to save time. People are building layers to give humans better control.

I'm concerned about opportunistic fake stories on Twitter about agents doing crazy things to stoke fear. There was a fake story about a clawbot doxing someone, which wasn't true. There's going to be a lot of false info out there.

There's talk about Sonnet 5 coming out with leaks and questions about whether we're approaching the singularity because new models are doing novel math problems. What are your thoughts on AI advancement and how close we might be to infinite self-improvement?

It's a fun question. Watching tools, the industry, and listening to smart perspectives, this is one of the coolest times to be human in this space. We've had major technological leaps like the industrial revolution, calculators, computers, internet, and mobile. This feels like it will create the biggest change if done thoughtfully. People building it, from Nvidia GPUs to AWS cloud to application layers, have a vested interest in success. Governments want success for economic growth and quality of life. Techno optimists are excited about doing more with less. There will be disruption in jobs and tasks, and new ways to manage time and create value. I have no idea what it will look like in five years, but it will be more productive and efficient. Humans are great at finding new problems to solve. Podcasts didn't exist 20 or 50 years ago; now they're common. Advertising technology is new and will change day-to-day work. We're at the beginning of something exciting. Everyone involved has a vested interest in thoughtful development. There will always be skeptics and trolls, but we aim to build in a good direction.

With that said, Brendan, I really appreciate your time. Please hit the like button and subscribe. Where can everyone find you?

Our website is www.tryclassify.com.

Put the link in the description. Thank you so much, Brendan. Really appreciate the talk today.

Thanks, Demetri. Have a good one. Bye bye.