Moneyball SEO for the AI answer era
For years the only thing stopping us from paying down tech debt was opportunity cost: an engineer had one thread of execution, and it went to features. Now that agents are abundant, AI is the best tool that has ever existed for cleaning up a codebase.
Evan Doyle is co-founder and CTO of Gauge, a San Francisco startup that helps companies show up in more AI answers. As people increasingly research what to buy by asking ChatGPT, Claude, and other models instead of typing into a search box, Gauge measures how often a brand and its competitors get mentioned across those AI responses and gives marketers the data and tools to widen that coverage. Doyle likens the discipline to Moneyball SEO: because an AI agent consults far more sources than a human and runs many searches in parallel, the goal shifts from ranking top-five on a handful of keywords to building broad, consistent coverage of what a brand is and what makes it different across the long tail of how people actually ask.
Doyle came up as an engineer and engineering manager, including early technical roles at Carta and Standard Metrics, before starting Gauge with co-founder Caelean Barnes and going through Y Combinator's S24 batch. The company began as a developer tool for managing very large Python codebases, but the founders were not seeing the traction they wanted. At a YC-hosted developer conference they heard speaker after speaker describe how most of their new users were arriving through ChatGPT referrals. They shipped a few tests against that signal, liked what they saw, and pivoted the whole company toward AI search visibility, where they have been executing ever since.
Deeply technical and data-driven, Doyle argues that the pace of change has reset what an engineering org should expect of itself. He runs a small, high-autonomy, high-ownership team that leans hard on AI: coding agents do a first review pass, and his own bottleneck is no longer writing code but deciding what to point the agents at next and keeping the product coherent as output multiplies. He has changed his mind about technical debt, too. Where he once assumed AI would only pile it up, he now calls AI the best tool ever made for paying it down, since parallel agents remove the opportunity cost that always kept cleanup at the bottom of the list. His hiring thesis follows the same logic: agents are abundant now, so what he wants are principals who own outcomes and know how to wield those agents well.
What Doyle watches most closely is the shift from single-turn AI queries to full agent loops, where a tool like Claude Code charts a path through a website across many fetches the way a diligent researcher would. He believes that agentic surface is a far richer place to be discovered than a single consolidated blog post, and Gauge is building to capture it. He is clear-eyed about the harder questions the technology raises, from trust between buyer and seller agents to liability when an autonomous agent acts on someone's behalf to the socioeconomic weight of labor replacement, but for now most of his energy goes into one thing: shipping as much as he can while the ground is still moving.
Read full transcript of interview
In this conversation: Josh Rubin (Host, CTO Studio) and Evan Doyle (Co-Founder & CTO, Gauge).
Tell me your name and what you do.
I'm Evan Doyle, co-founder and CTO at Gauge.
What is Gauge?
Gauge helps companies show up in more AI answers.
How important is showing up in AI answers these days?
It is pretty dang important. I would argue that if you wait too long and don't invest in this channel, it's going to be a big regret for years down the line. Almost every user, particularly B2B users, when they're researching what to use or what to buy, is involving ChatGPT or any of these other mainstream models.
News too. We just saw a report that Claude is peaking with news queries around 7 p.m. every day, so it's even pulling without any referral traffic. They're pulling sourcing from news, summarizing it, and giving it to people. So it's not just B2B and SaaS, it's all information at this point.
The speed at which this is changing is astounding. I assume you've seen some of this velocity. How long have you guys existed?
Technically, Gauge as a company has existed for around two years. But working on this problem, because there was a pivot in the early days, is about a year and a half, maybe a little under that. Even in that time we've seen really massive shifts, on the product side and in the market we're going after, but also, I'm a CTO, I'm in engineering, and that has just changed so dramatically.
What were you guys before you were doing this?
We were building a developer tool, actually. My co-founder and I are both deeply technical. We've been engineers and engineering managers before this. We were building essentially a tool to manage very large code bases, a specific problem that we had seen before in Python. But we kind of discovered this by accident. We were looking around for what was going to be our next thing, because we weren't seeing the traction we wanted. We went to a developer conference that YC was hosting, and we heard speaker after speaker talk about how almost all of their new users were being referred by ChatGPT. So we took note of that, shipped some initial tests, liked what we saw, and have been executing against that since then.
How different is this? GEO, AI search? Which one do you prefer?
I personally don't care. GEO is quicker, easier to say, but it's got the conflict with geography and everything.
How different is it from SEO? I'm assuming you didn't have a huge SEO background when you jumped into this.
No, I didn't personally. I've always been very much about the tech. But my understanding, based on what I've seen in this market and just talking to my co-founder, Caelean, and some of the folks we've hired, is that I like the analogy of GEO being more like Moneyball SEO. Because an AI agent will consult many more sources than a human, it doesn't have the same attention-span limits, and it's able to make many web searches in parallel, you want to rebuild your visibility in aggregate. You want really broad-based, consistent coverage of what your brand is, what you offer, and what makes you special.
Dive into that a little bit more. What does that mean?
Let's take running shoes, for example. Because everybody's funneled into the same search box on Google, and it's relatively small, you can realistically target a relatively small set of keywords, and your goal is to be top five so you get an actual click. With GEO, what you want is to capture the intent behind all of those previously hidden, disparate usage patterns that are all hidden behind 'best running shoes.' Maybe somebody wants the best running shoe with a wide toe box and good arch support, whatever it is. You want to have coverage of that type of content, and to have it mention your brand.
It's a more effective long-tail approach, which was much harder to do in the SEO realm because you're targeting top three, or you're buying Google AdWords. With this, every niche is potentially exploitable with content.
In the old days of SEO, let's go back to 2007, there was a lot of black hat SEO, private blog networks, backlink purchasing, all of that stuff. How much black hat GEO are you seeing in this space? I don't even know if it's a thing. How manipulable are these LLMs from a brand or personal perspective?
It's a good question. I definitely think there are some questionable patterns you'll see, content that's really easily able to manipulate the LLMs.
Give me some examples that I would never use.
Something folks have definitely seen, and I've seen some reporting on this: the listicle is very dominant. The pattern a lot of folks like to use is, I put out a listicle, 'best running shoes,' and lo and behold, I'm number one on the list. Very dominant strategy right now. I think that may change in the future, but that's just what we saw immediately in the data.
So right now, authority, the old style of Google-based domain authority, trusted sources, backlinks from highly valuable or authoritative partners, is less impactful?
I think it's arguable. It builds an important foundation. The easiest possible case in GEO is, I have really good domain authority but I haven't yet covered the long tail in terms of content. Having domain authority gives you a really strong base to push into these long-tail content areas. So if I'm a new company starting out, I'm definitely not going to start with domain rating or backlinks. I'm going to start with content. But having it is definitely still helpful.
There are some broken incentive patterns around backlinks from high domain authority sites, because with Google they would refer traffic, which the news brands and high authority sites would monetize. LLMs aren't doing that right now. ChatGPT has a deal with Reddit, it's not quite the same thing. And Anthropic is sending no traffic, or at least no attributable traffic, to any news sites. From your perspective, is this changing very quickly? How sustainable is this?
I think it's changing very quickly. I actually think a new kind of replacement market is springing up, and it's not for backlinks but for mentions specifically. You can't really count on the backlink driving a real click. But if I get a prominent publication to just talk about me, even if they don't link to my site, that's in a lot of cases just as valuable.
I guess it has to do with two things. They have to mention me, but they also have to be scrapable.
Of course. There are some legacy media publications, The New York Times is a really prominent one, that have taken a stance. They don't want AI crawling their content, and they have their reasons, it's probably important to legal cases they're working on. But what we see in our data is that that means it's much less important for a brand to be mentioned on those sites.
I imagine it's only a matter of time before, oh, I've got a Wall Street Journal or New York Times subscription, and I can plug that into Claude or ChatGPT so it can be consumed but somehow not shared with the wider models.
That actually makes a lot of sense. Eventually a deal probably gets done there, because reputable information is extremely important, and the models know that.
And it's becoming harder to come by the more they lock that stuff down. That's probably where it needs to go. It's a mixture of, it's like carriage fees. A person buys a subscription to a news publication, and to then plug it into the model, the model itself probably has to kick in a few cents for everything that's happening there. You can buy a New York Times subscription, but for $5 more you can give the MCP server access to such and such.
I could definitely see that. It keeps the content relevant, and it extends their existing business model naturally.
I noticed that Google is moving toward, which news sites do you want to prioritize in your search results, as something you can already choose in Google Search. I'm tired of seeing news sources I know are not really reputable popping up constantly. If I see the Epoch Times pop up on my news feed one more time as a citable source, I'm going to be super annoyed.
That setting you mentioned could just be a stepping stone to the world you're talking about. I wouldn't be surprised if I saw it integrated with AI Overviews, for example, relatively soon.
Obviously this is a problem a lot of people are working on. You're sitting in the Valley. I know you have a lot of competitors in this space. What sets Gauge apart?
Like I mentioned at the beginning, my co-founder and I are both very deeply technical, and we're also very much driven by the data. So what sets us apart is that every strategy we recommend to our clients is backed up by very high-quality data. And I think we're also able to build the best AI-enabled tools the fastest. For example, one of our most-used features in the platform is our AI chat feature called Ask Gauge. It has tools to use essentially everything in the platform, plus any integration you bring in from outside. Your Google Analytics, Google Search Console, your CMS, anything like that, so it can be a one-stop shop for a content marketer.
You've got to be burning a lot of tokens right now.
Yes and no. Yes, it's burning a lot of tokens, but we feel the pricing is reasonable for the quality, especially if you have well-curated context that you're feeding into the model. You get good outputs even if you're using something like Sonnet or Opus. It's not the dominant cost in our expenses.
And you dropped Grok recently. Is there a particular reason, or just a cost issue?
Actually, it is a specific reason.
It's because people just kept asking for naked pictures, anything Grok. That's what I assume is true.
It's a fair assumption, but no, we weren't exposed to that. It's actually that the highest-quality source of data is the end-user front-end experience. We need to request the public free version of a model to see how it behaves under different prompts. Grok doesn't expose a public endpoint anymore. You need to be logged in as a Twitter user. The only way we can support Grok now is if you bring an API key.
Are there any big changes or trends you're seeing in how AI search is functioning right now?
Yes. What shocks me is how quickly things change even over the course of this company. It's only been a year and a half, but the first wave is users moving from Google to ChatGPT, or Google to Claude. The second wave, that's just starting, is that a full-blown agent loop is now becoming part of a typical user's journey. So instead of me sitting down with ChatGPT and asking, 'What's the best CRM for me to use?' or 'What's the best HRIS?', I might have Claude Code go do a research project and then return a PDF. So we're very interested in capturing the agentic surface, which is actually much richer.
Let's suss that out a little bit more. You're saying, instead of somebody going into ChatGPT, 'I need a recommendation for X product,' they say, 'I need you to build me an agent to go out and do research to figure out what the best product is,' figure that out and then send me a document, a PDF.
Yeah, it's much more multi-turn, so it's much more likely for the agent to come back and say, 'Here's what I found so far. What criteria are important to you?' I also think it's much more likely for it to chart a path through a web page. In a ChatGPT query, it might do a web search, a couple of web fetches, but a lot of the time it's a single turn and it just gives you a response. In the Claude Code case, it might pull up a docs page that references another docs page, go fetch that one, go fetch another one, until it's navigated that site, similar to how a user might if they were doing the research. And that's just a different optimization surface than somebody who needs to put up a comprehensive, consolidated blog post.
The hard part for anybody building a website is determining what is a landing page versus what is an informational page. But the way an agent can consume information, without dealing with the cognitive load of a human, is dramatically different.
The idea there is, it's not about me necessarily publishing more content. It's about empowering my agent to give answers on the fly based on all of my content. Is that what we're saying?
I think that's where it's going. I don't know how quickly we'll get there.
That's the depth of the website, obviously, in its current incarnation.
We do kick around that idea a lot internally: what does the future of the website look like? I think there is a version of a website where Claude Code requests some well-known standard path, essentially walking up to a salesperson, or walking up to the cash register, or walking into a store. There's a conversation between the buyer's Claude agent and the seller's agent, whatever it is, where the agent needs to elicit what the requirements are and provide the best information possible. And that's probably going to be happening not only for that vendor but for all of that vendor's competitors in parallel.
This is like a product person having to trust a salesperson on steroids. No product person I've ever talked to has said, 'Oh, the sales guy said I could do this.' That brings me to what I think the most important part of this is, which is trust. Trust has been disrupted more than anything else in this AI age, at the same time that these tools are proliferating. How do you build in a trust layer, my agent talking to this agent to give me these results, when you're also telling me that I can fairly easily manipulate some of these platforms by just giving them what they want to hear?
Truly, I think it's an unsolved problem. There's going to be new systems developed specifically to handle this. One thing that comes to mind is it could be mediated by the model labs themselves. There may be some kind of partner program where the information you're providing to the agent is stamped with a seal of approval, and you can instruct your agent, whenever it's buying something, to look for that stamp. That would be unfortunate for gatekeeping reasons, and it definitely gives them a lot of power.
Another one, specific to developer tools, which we're very interested in: if I'm selling an API, or something the agent can actually install and test, there's the potential it could give a live sample, like a short-lived API key, where you can verify the claims the sales agent is making.
False advertising is actually a crime. But nobody knows, in this black box of one agent talking to another agent. So it could potentially just allow you to lie at scale.
That will probably happen in some parts of the internet. I think it will be on the client's agent to be as skeptical and as truth-seeking as possible, to figure out how to sift out the value from the noise for its user.
That's a fair point. Lying on the internet is not new. So let's get into some of the technical specs of how you build your product. You're a greenfield product, using AI tools to solve an AI problem. Is there something unique about being in that position as a CTO?
The biggest thing that comes to mind is just the velocity the engineering org now has to move at, or should move at, given what tools are available. Like you said, we're operating in a relatively greenfield space. I like to think the core of our product is now relatively mature, so there's maintenance and extension there, but we're always taking spikes in new directions. It can be exhausting sometimes to keep up with all the capabilities that are out there and make sure we're using them as effectively as possible. We don't want to leave value on the table. So whenever I'm working on one thing, I need to make sure I have agents working on other self-contained tasks.
I thought AI was supposed to save us time and give us back a little bit.
Somehow that never seems to work out.
No. So the closest thing you have to time off is like when the baby is napping. It's when the agents are working.
It's interesting you say that, because there's always more work I can give to an agent. The bottleneck really is my review and, a lot of people are using the term, taste. It's about deciding what makes sense for the agents to work on next, and how I can make sure it doesn't outstrip our capability to QA it and keep a coherent foundation, a coherent product.
That's an important point. It's very easy to produce more than you can ever look at. So how are you setting those boundaries for yourself and your team?
Every team is different. Our team is very small, lean, high responsibility, high autonomy, and that's how I prefer to work. The way we run it is, we have very powerful AI code review tools that always do the first pass. I've been very happy with them. I've seen them evolve from very noisy to now extremely competent and fast. And it's just a question of making sure the team is talking to each other enough in a given week to know, okay, this surface is changing, these contracts are changing. But other than that, just trust among the team members. Whatever I put out there, whether AI wrote it or whether I wrote it by hand, I'm responsible for that, and if something breaks, I need to fix it. That's how I like to run it.
So taking responsibility is still a key part of this, because it's super easy to say, well, the AI messed it up. That doesn't fly. You prompt it, you bought it. One question I ask a lot of people: you're a Y Combinator company, your goal is to build this thing, you're running as lean as possible. But if you had all the money in the world and could hire exactly who you needed, who would you be hiring for right now?
It's a good question. I was thinking about this recently. A popular framing I've heard in other contexts is the principal-agent problem. As a leader, you have certain priorities you want the business to achieve, and you delegate a lot of the implementation to your agents, who up until now have always been human employees. I think now agents are abundant and I need more principals. I need people who can come in, be autonomous, have very high ownership, and are also able to utilize this abundance of agents as effectively as possible. So I'd typically be looking for people who, especially recently, have been pushing really hard on their skills and have been close enough to the execution to know the most effective patterns for utilizing agents, because right now it's all informal knowledge. I don't think it's crystallized in online discussions yet, what the best patterns are.
It's only been around for what, six months as an effective way? The colleges aren't turning out people who do this yet?
Exactly. In our numbers, I can see stark differences as recently as three months ago. And I can feel the difference, too.
So how do you train for that? Or are you thinking that in two weeks you might not need to train for that, because they'll have a new update that takes care of it for you?
Right now, being a small company, we don't think too much about training from scratch. We really do want to bring in people who can hit the ground running. But in general, how do you stay on top of this? If you were a good engineer before this, you will still be a good engineer after this. You just need to care about producing the highest-quality output, at kind of the Pareto frontier. Obviously you can ship a lot more if you maintain your quality bar and use these tools, and you're going to adapt just fine. You're naturally going to find patterns that work for you. So as long as you're showing up and pushing hard, my workflow changes week to week, so I think it's just staying in it.
So your advice is, as long as you're an awesome unicorn, you're set.
What else could I say? Try to get into a situation where you're not constrained in how you can use the tools. This is secondhand for me, but I know there are some organizations where it's very regimented, where maybe it's hard to even use tokens on the production product. If that's the case, you do have to be experimenting on the side, in some personal context where you can just push these tools as hard as you can.
I think ultimately you're in, if not a unique position, then an ideal position. You're a bunch of engineers who worked in the old space, but because of this particular niche you found, you can experiment. You have the room. You can see the data first, without having to plug into a 27-year-old API of a legacy system. So you are as close to the bleeding edge of using the tools as anyone right now. Do you feel that?
I feel that way, but with maybe a caveat. Being in YC, I have some visibility into how other technical teams are using these tools, and I think there is still space ahead of us for people who want to take more risk. I'm not sure if you've heard the term software factory, but a lot of those ideas are interesting, and we watch how those play out. Sometimes we take them in piecemeal, but we try to stay as close to the bleeding edge as we can while maintaining really obvious positive ROI on all the decisions we're making. So we're not just chasing the trend. We're maybe one step back.
But you're not building your own AI models.
Yeah, of course not.
Just on an earlier point, if you have to plug into a 27-year-old legacy API: one thing I'm really interested in. I wrote an article a while ago, on the previous product actually, called 'AI Makes Tech Debt More Expensive.' The thesis is basically that when you have a clean, pseudo-greenfield code base, that's when you get the most relative benefit from using AI tools. When I wrote that, I was thinking the human engineer is the one who's going to have to maintain tech debt at low levels, that the AI would be producing low-quality code and wouldn't be able to do that kind of work. Now I actually think AI is the best tool that has ever existed for improving the cleanliness and quality of your code base. I highly recommend everybody invest some time into paying down tech debt, especially because the speed at which you can do it, in these larger or smaller ways, is going to be a lot more effective.
That's the second time I've heard this in recent interviews. I was talking to Juan Rey at Siemens EDA last week. He was the first one who said to me, 'I'm very hopeful that tech debt's not a thing anymore.' Which is a complete sea change for me. Even three or four months ago, everyone was like, I just generated 250,000 lines of code a day and my tech debt problem has exploded. Now it seems to be, it's the AI's problem, it'll take care of it pretty easily.
Fundamentally, how it feels to me is that the inhibitor on working on tech debt in the past has always been dramatic opportunity cost from the engineer's perspective. They need to use their single thread of execution on features. Now that we have agent abundance and parallelism, all you have to do is recognize the tech debt and describe what you want fixed about it. And now you have a realistic chance of getting that done without any significant opportunity cost.
From your perspective, the software engineer is not going away? The software developer might be.
It depends on your take on those roles.
Let's parse that language a little further. The software engineer, the architect, the orchestrator, the one who's product-minded, close to the business, crazy important. Somebody who's just really good at coding?
I think it gets a lot harder to justify a position that's purely about being the person on the team who can read the code. I definitely think the archetype needs to be somebody who is owning outcomes and wants to use whatever tools are available to get them done as fast as possible at the highest possible quality level. And I think that's what most software engineers have been after, at least in my experience. Well, maybe I can't confidently claim 'most.' But a lot of good engineers, to my point earlier about tech debt, have ideas about what we want to improve in the system, and we've always had limited time, limited energy. The kinds of things I'm able to build now, and the quality I'm able to build them at, really do make me happy. And I'm happy to use the tools.
A happy engineer.
I'm very satisfied.
So in the Valley, obviously, the energy is here, the velocity is here. Outside of the Valley, AI gets a lot of pushback, whether it's data centers or fear of socioeconomic issues. Do you feel any of that anxiety here in the Valley?
That's a good question. I want to say I'm not qualified to weigh in on data center placement and all that. But definitely, there's a lot of thought here about the socioeconomic impact. Obviously the headline story is, is this going to lead to labor replacement? There have been a lot of layoffs already, and the stated reason is always AI efficiency. I think it's worth having some skepticism around that, but I also tend to agree with the vision some of these foundation labs put forward, that we're really going to see dramatic changes in what a typical business looks like in terms of headcount. I don't know if I can get into more detail. It's a very subtle and difficult topic.
Nobody can predict the future entirely. You can only do the best you can with the information you have. Jensen over at NVIDIA is basically looking at this from, there will need to be social changes to how we interact on a sociological level. Search, AI, and how we consume our data and information are fundamentally changing. Do we need regulation on top of that? Do we need to implement other tech guardrails around that? That's what we're kind of trying to feel our way through.
ChatGPT's newest model was just basically restricted by the government to a couple of select companies. Do you think regulation is where it needs to be? Are you hoping for more?
Almost all of my mental energy is just going into executing on my business, but when I do think about these things, I feel like some amount of regulation is inevitable and the right thing to do. One thing that comes to mind: when we talk about agent abundance, and these systems making decisions and executing actions on users' behalf, the open question is liability. You mentioned earlier that false advertising is illegal. Can we establish clear links for, my agent is doing something online and that's on me, in a way that still makes sense, because these agents are extremely volatile? They can do things humans can't do. I'm very interested to see where that goes.
That's one of the more fascinating topics. I think about Google. The tradition of search is I ask a question and it gives me a link to the answer. Google is not taking on responsibility for whether that answer is true or false. It's saying, is it credible? But it's not taking responsibility. The second they started summarizing everything is the second they became an editorial outlet that exposes them, theoretically, to liability. The first lawsuits around that are already starting, but lawsuits move slowly.
It's going to take time for this to permeate the system, but I'm very curious to see where that goes.
So am I, but no tech has moved as quickly as this tech is.
Now we're potentially seeing, with these restrictions on the frontier models, I'm not sure if the government is intending to just assert more control or intending to legitimately tap the brakes a little bit.
At the same time, Dario is saying he's a little bit worried about these open source models because you can't control them, or because he legitimately feels they could impact society in a negative fashion. But a lot of people want to move to open source because these frontier models are expensive.
They're expensive, and they're also not under your control. If the weights are fully open, I can retrain certain behaviors out of my open-weights model, and as a consumer that's purely better for me. I like to have that control. But I also think his concerns, and others' concerns, about, if we really get anywhere close to the self-improvement loop in AI, then lack of control over the models is definitely a dangerous thing. It's really difficult to speculate on whether that's possible or not.
One of the AI books I'm reading talks about the anti-pessimism bias, that's the idea of, oh, that sounds ridiculous, it couldn't possibly go that bad. Sometimes you've got to deal with the fact that it could be that bad. You really need to think about that right now.
I think that problem is really challenging, especially given how it's played out so far. Basically the instant a new capability hits the market, all of the dangerous use cases are explored somewhere in the world.
Basically, we won't get into the kind of regulation we need until the robots are literally trying to kill us.
I've been joking that maybe my next startup will be anti-robot tasers for people to protect themselves.
You've got to take the clankers down.
Exactly. You've got to protect yourself.
Be part of the
conversation.
Whether you're a CTO who wants to be featured, a company looking to sponsor, or an engineering leader wanting a seat in the room — there's a place for you here.