RISHABH JAIN

Welcome to the Intelligent Marketer, where we discuss how AI is impacting marketing and commerce. In today's episode, we have Ethan Smith, CEO of Graphite, who is well known as one of the top SEO experts in the world and is now also one of the best minds in gen AI engine optimization.

MIKE DUBOE

Ethan, thanks for joining us. This is our second episode here of the Intelligent Marketer, and we thought no better person than you to talk about this very timely topic of answer engine optimization.

Ethan, we've known each other for over a decade. You've always been one of the smartest and most thoughtful minds on all topics SEO, and you've built SEO into a dominant channel at companies like Thumbtack, MasterClass, Ticketmaster, and more. And now you're doing this at scale with your agency, Graphite, who I regularly recommend to our portfolio at Greylock.

We're assuming a pretty educated audience here, so rather than starting with a primer on SEO, let's just jump into it and ask how is the AI era making you rethink your priors on SEO? Maybe as part of that, you could do a little bit more of an intro on yourself as well.

ETHAN SMITH

Sure. Well, thank you for having me. Excited to be the second episode.

Yeah, so our company's called Graphite. We build out SEO strategies, content marketing, answer engine optimization, paid advertising. We started in SEO, but we're doing all digital marketing and we're also full stack, so we can build everything from design strategy to the landing pages.

Yeah, so what is changing with AI and what is challenging my priors, and honestly, nothing really, it's just sort of additive. I started in SEO in 2008, and every two to three years there's a new thing, and then people are concerned that Google SEO is going to go away and this new thing is going to disrupt them.

It's like TikTok Search. Everyone's searching on TikTok Search, all the Gen Z, and so Google's going to go away. And that's not really true, but it's sort of true. So it's true that people do search on TikTok, and then before that it was Instagram and then before that it was YouTube and Amazon, and people do search there, but Google's slice of the pie never gets smaller. It's just that the pie continues to get bigger and all these new search surfaces, it's just all additive.

And so the same thing is true for LLMs. People are searching in LLMs, they're doing discovery there, but Google doesn't go down. These new services are just sort of additive.

So the second thing is what works in AEO, and it's pretty similar to search. So most of the stuff that works in Google Search is the same stuff that you use to optimize for AI.

The big difference is, number one, is that the tail is much larger. So the average number of words in a Google Search is about six, and then the average number of words or tokens in a chat — according to Perplexity — is 25.

So it's way more specific and the long tail is just much larger and a lot of questions have never been answered. They haven't been answered that many times. So the tail is much, much larger, number one.

The second is that, for the most popular questions, so the head of the questions, you can't win with a page on your site, you need to get mentioned because the LLM is a summarization of many opinions. And so for some of the questions, you win by getting mentioned by others.

And then the third thing is that the data sets are different. Like social, UGC, YouTube video, things like that are more important. But then the core principles of SEO actually mostly apply: The same way you get a page to rank in search is the way that you get a page to rank in the LLMs.

So those are a few initial ideas.

RISHABH JAIN

I'm wondering, when you say it's additive, and I've heard the same data point a bunch, which is that it's not necessarily that the absolute volume coming from Google Search is shifting, but when people look at percentages, there's a lot of recent uptrend data that people will say, "Up to 5% of my search in four months is now coming from LLMs."

And so it doesn't change your point, which is like, "Hey, the absolute volume is not necessarily changing." But when you think about the total amount of available consumer attention and that that consumer attention, now some amount of it is going into AI search, how would you recommend someone think about where they prioritize their efforts?

Because stacking your spend is very hard to do and ultimately you're fighting for a type of consumer attention. So how do you think about how someone should invest against that, even if you're seeing some of the underlying tactics are the same?

ETHAN SMITH

Yeah, so a couple ways to think about that.

The first way is, what's the value of the channel in the first place? And it's actually especially hard to measure because many answers that mention a product, you can't click on something. Since January, the answers are more clickable, so there are things that you can click, but frequently when you are a product that is mentioned, there's nothing to click on it, so you don't actually know what the effect was.

So we recommend doing a "how did you hear about us" question, especially for enterprise companies, like "what's the best payroll management software?" It's going to be a hundred different touchpoints. And if you just look at last touch, you're going to say, "Well, everything came from paid search and all of my other marketing stuff is completely worthless," which is not true at all.

It's important. And the higher the consideration, the conversion, the more important it is to look at all those touchpoints and do that with "how did you hear about us?" Or if you can do multitouch, that would be great. You can also do tracking of "am I appearing in answers?" and sort of back your way in, but the measurement thing is hard and whatever your last touch says is going to be undercounting.

So step one is just what's the size of the channel? And then step two is, given the size of the channel, how should I split my resources?

One argument would be, "Well, this is going to be huge and I'm going to have first-mover advantage if I do it now so I'm going to overinvest for the first-mover advantage," which I actually don't agree with. Whenever this channel becomes big, you can launch a strategy — if you launch a strategy today versus two years from now, it doesn't really matter.

What matters is do you have the landing pages, and if you have landing pages in two years, they'll start ranking right away. So there's really no first-mover advantage. It's sort of a false concept.

So whenever the channel's big enough, invest in it. It's not big enough right now, don't invest in it and when it is then go ahead and invest in it.

I'll say that, for early-stage companies, when I advise early-stage companies, a VC will say, "Please come talk to my seed-stage companies and advise them on SEO." And my first advice is, don't spend any time on SEO because it's not going to work. Do something else.

However, for answer engine optimization, interestingly, it actually can work well for seed-stage companies. And the stuff that gets you from zero to one is very different from the stuff that gets you one to two.

One to two would be Meta, Google, SEO, but zero to one can be these more unconventional channels. And AEO actually can work for early-stage companies because to rank in SEO might take years, but as long as you're getting mentioned by the citations for the question that you want to rank for, you actually can rank pretty quickly.

So I would — depending on the stage of the company — I would decide whether to invest in it, but the first-mover concept I think is generally not applicable, so I just say invest in it whenever it's big enough for you.

MIKE DUBOE

Yeah. So Ethan, let's stay on this. This is interesting. So stage is one vector that might determine whether this channel could work for someone early on. What are some of the other vectors? I would imagine there's something around skew complexity, categories. What are the examples of companies? Feel free to anonymize, but what are the other attributes that determine whether this channel will work early on or not?

ETHAN SMITH

The category matters a lot, and first movers — or sorry, early adopters — in chat are not the average person. So engineers are definitely early adopters. So if you have a developer platform, then it's going to disproportionately work well for you. Vercel, which owns Next.js, said that 10% of their signups come from LLMs, which is not the case for something with general consumers.

So early adopters are people like engineers, people like us, investors. So if you're in that category, [it] can make a lot of sense.

Really high-LTV categories also. So an example of this is, I was talking to a company that does asset management software, which is really high LTV, and part of how you can rank for best asset management software is to make a YouTube video talking about asset management software platforms.

Now, how many YouTube videos are there about asset management software? Maybe one, maybe not. People make YouTube videos about traveling and cooking. They don't make it about asset management software.

What that means is that there's this untapped opportunity where I can make videos on these obscure topics that most people wouldn't be interested in, but it's super-high LTV and there's no YouTube videos on them. And so I can come in and saturate that market.

So category is one thing, really high-LTV categories is another, early-adopter categories like engineering and venture investors.

MIKE DUBOE

And so maybe going back to the top on ... One comment that you made, and I might paraphrase, is it actually is not ... The actions you take to actually influence results is not a whole lot different for this new world than the past. And we've roughly split off-page SEO into content generation and backlink generating links, and then on-page into technical architecture.

The latter, let's spend a minute on the second bucket because I would assume that the website that is optimized to convert and attract agentic traffic probably looks different than that which is designed for humans. Push back on that. Do you disagree?

ETHAN SMITH

I disagree.

MIKE DUBOE

Okay.

ETHAN SMITH

So there's two, there's LLMs and then there's agentic. So agentic, I think by that you mean, plan a trip to Miami and there are decisions made with no human in the loop. So pick my hotel, restaurant, and my flight with no intervention from me.

That I do think is different, but I don't think that that's happening that much. I think that'll happen eventually, but that I think will be different.

But what's happening now is just I'm asking Chat questions and how does the answer, how do I get my URL in the citations? It's basically the exact same thing.

So when you ask an LLM, "What's the best payroll management software?" frequently it's doing a search and the search is either Bing or it's a search engine similar to Bing, the same principles apply to show up in the citation set. Then the answer is a summarization of that, but it's essentially the same thing.

And then part of your question is, "Well, do I need to make my page specially parsable by the search engine?" And the answer is no, you don't. The things that matter technically are ... The most important thing is the title tag. Title tag tells the search engine most important stuff about what your page is about. So put your keywords in there.

The second thing is make sure the page can be read, so don't have weird JavaScript, and then crosslinks and schema, there's nothing else.

There's misinformation from 2006 about Google Crawler has all these issues where it doesn't understand your page, but Google and Bing, all these search engines are pretty good at reading pages. It's not that hard. And so I have to specially make my page with alt tags and this sort of stuff is just sort of a legacy from 20 years ago that's not really applicable here.

RISHABH JAIN

Ethan, I'm curious, if it's true that the average search is more nuanced — I'm using slightly different words, so it's six words versus 25 words or whatever the Perplexity stat is — shouldn't that suggest that while the same techniques work, the amount of coverage that you need in the content on your pages in order to be responsive to that level of query nuance is actually larger? Or does it not suggest that in your view?

ETHAN SMITH

It suggests that.

So we don't actually know what the 25-word question looks like because no truth set, but we'll make assumptions. So the assumption would be something like, "I want a product with all of these attributes. I want a payroll management software with this integration and this attribute and this language and this region." That's probably the kinds of things that it is or follow-up questions.

In many cases, there's no answer anywhere on the internet. So an example of this is you ask, "We want to be able to track our sales calls and our client calls and see who showed up and what did they say, and do they say good things or negative things." So we want a chart.

And I asked, "What meeting transcription software integrates with Looker?" and the answer is none. There's no one that does that.

However, there kind of sort of is because Otter has a Zap where you can send the meeting transcription via Zap anywhere, and you could send the Zap into BigQuery and then have Looker do a chart there, but there's essentially no answer anywhere on the internet.

So if I made a page saying, "Actually, I can integrate with Looker," then I will show up for that. And there's all these other integrations with meeting note transcription with this integration.

And so I need to have that information accessible by the LLM. So I need to, number one, know what questions people are asking and then answer them. But yeah, so in order to rank for something like that, I might have something like, "Here's a giant page with all of my integrations." So here's all of the things that Otter actually is capable of doing — the use cases or the languages of the regions and things like that.

So I would do that, but it wouldn't be a technically different structuring of the HTML. It would be, "I need the information to exist on my site and to be something that the bot can see."

MIKE DUBOE

So maybe just to ask the provocative question, why is so much money starting to get spent on monitoring tools for AEO?

ETHAN SMITH

It's a new ... Okay, so it's similar to why was there money spent on SEO tools?

Let's contrast lifecycle with SEO. So SEO is as big as lifecycle if not larger, but the amount of money people spend on tools in SEO is $100 a month, whereas in lifecycle it's thousands of dollars a month.

So why is there this massive spend difference for channels that are the same size, if not smaller? And the reason why is because it's easy to build SEO tools and it's easy to build AEO tools. It's a commodity market and it's easy to build it internally.

So with SEO, there's Ahrefs, Semrush, those are the two main incumbents. And then there's Surfer, which is also great. Semrush, Ahrefs, I think is 1 billion, 2 billion, and then Surfer is probably a hundred million or something like that, I'm not sure.

And then there's all these other startups and they say, "Mike, I have this new idea, we're going to do it 30% better than Ahrefs." And then you do that and then, "Well, I'm going to charge $500 a month instead of $100 a month." And you close a few deals, "Look at this chart, Mike, it's looking great." And then Ahref says, "Cool, I'm going to take that feature that you made and I'm just going to add it because it's not that hard to build."

And so with AEO, it's this brand new category, people don't know how to do it. So then I build a tool and I'm like, "This tool is amazing and you should spend thousands of dollars on this tool," which is basically just tracking. You just put a question in and you say, "Did you show up? Yes/no," and then you put it somewhere and you chart it.

And so there are these tools doing that now that ... We have a page with 60+ AEO tools and they all do the same thing: They all do tracking.

The reason why there's 60 is because it's easy to build. It takes a few weeks to build that — just like SERP tracking takes a few weeks to build that — and then somebody's got a new feature and then everyone's going to copy that feature because it's easy to build also.

Content scoring is easy to build and question research is easy to build, lists, all of these are easy to build.

So everyone's going to copy each other, and then six months later, Ahrefs and Semrush will add the exact same feature for $100 a month.

If I'm spending $1,000 a month, then I'll just eventually churn and I'll go use Ahrefs and Semrush, and then all these AEO tools either go out of business because they raised too much money or they're sort of a nice lifestyle company.

What I recommend to — and as a fellow founder, I'm rooting for all these founders, so I take no pleasure in people not doing well — but my recommendation to anyone building an AEO tool company or an SEO tool company is to raise no money or raise almost no money and try to make it a lifestyle company where you make a million dollars a year and go to Jamaica and have a nice time, but don't try to build a venture-scale company. I think it's impossible.

RISHABH JAIN

Ethan, if you believe that's the view and you're serving these customers already, right, for SEO and AEO purposes, how do you as someone who runs an agency, what are you telling the merchant or the end brand who is saying, "Hey, Ethan, which one of these should I purchase?"

You hold this view and it is truly what you believe. And so if that's truly what you believe, then what do you say to someone when they say, "Hey, I hear" — I'm just going to pick random names — "I hear Profound is good for this reason, and Evertune is good for that reason and Scrunch is good for this third reason and whatever is good for this fourth. I hear these things." And you're saying, "Hey, these things are all going to consolidate and they're all asking you to sign annual contracts." What do you say to someone?

ETHAN SMITH

If somebody says, "Which tool should I use?" I say, "Use the tool that does what you want that is the least expensive tool. And don't sign an annual contract." That's my recommendation.

We're actually building our own tools as well, and we offer them to our clients for free. But if they're trying to choose another tool, I just say, "Pick the tool that does what you want and has the lowest cost. And two years from now, you'll probably be using Ahrefs anyway."

MIKE DUBOE

As a former user of Ahrefs and Semrush, I'm not 100% sure, but I hear the point that any SEO team uses these already, and so they can and should and already are attaching AI offerings to this because it's really a product extension.

Going back to your assertion, Ethan, if you end up being wrong on this and a company that has built AEO monitoring as a wedge ends up being a multibillion-dollar business X number of years out, why will that be?

And I'll share, my view is I think actually it is known that this is a hair-on-fire problem — or perceived to be a hair-on-fire problem — for CEOs, and thus CMOs, who are, these are probabilistic kind of searches. And so you're seeing yourself not show up in the results, pay whatever it takes to help me get in those results.

Figuring out where you are is step one. Your point is a real observation, which is like, "Hey, no one's actually helping do anything novel there to actually go and influence the results," but perhaps monitoring earns you the right to do that.

I think your view, what I heard is your view is like, "Hey, if anyone's going to do that, someone has already built a good services offering." But if you're wrong on that view, what will have happened?

ETHAN SMITH

Ultimately, people will spend money to cause outcomes, and knowing what I rank does not cause an outcome. I already know that monitoring cannot possibly be a wedge because there are 60 tools that all do monitoring, so that can't be the wedge.

But the wedge would be causing outcomes that are significantly better than anyone else. So how would you cause outcomes?

The first is, you would have tools and you would say, "If you want to rank higher, this is what will cause you to rank higher." And there are SEO tools that do that, there will be AEO tools that do that.

The second would be something like, well, the second would be doing all this work for me. When I started with SEO consulting, I would come up with an entire strategy, all of the details, and more often than not, maybe 20% of that was launched; 20% of that was launched because "engineering resources are constrained and we have many priorities, we can't launch all these things," which makes sense.

So I could have the most perfect tool and the most perfect information, but actually to cause the outcomes, somebody needs to build all this stuff.

I think that the wedge would be implementing these things and doing that quickly, doing that cost-effectively. So that could be services, that could be a tool, it could be both. It's probably somewhere in between, it's probably hybrid services.

But I think that the wedge will be causing outcomes which is caused by executing and implementing the stuff. It's not by really great monitoring.

MIKE DUBOE

Maybe let's jump to running an agency right now. You have a very different set of tools at your disposal than you had a few years ago. How has this impacted how you're building Graphite?

ETHAN SMITH

AEO or AI or all?

MIKE DUBOE

Sorry, I'm saying AI products and building agents into your workflows and just in general how running Graphite looks differently now than it did a few years ago?

ETHAN SMITH

So a few years ago ... We actually built the company from the beginning — we started it five years ago — we built the company from the beginning building AI and thinking that AI was going to be large. We built a 10-person AI team and overinvested in AI specifically because we thought that services would be disrupted by AI. This is before things started trending.

And that's been true. I'd say that we've evolved our thesis. So one thesis with AI or with any new technology trend is to do what you're doing right now and to do it cheaper, which I don't think is actually the solution. I think the solution is to do it better.

So right now, when people are doing AI agencies, they take a B-minus workflow and then they automate it by chaining ChatGPT queries, and then they make it from a B-minus into a C-plus, but 80% off.

I don't think that that's interesting. I don't think that Greylock or a large company — I don't think Rippling is interested in an 80%-off C-plus work. What they're interested in is better.

And ultimately, to build an agency, the renewal rate and the retention rate matters. It's the same with an enterprise software company. What really matters is the retention rate.

If the retention rate's 100%, then nothing else really matters. Your acquisition doesn't matter, your margins don't matter. That'll all work itself out. If you have perfect retention, then everything's great.

So to build a really big, large, successful agency, the only thing that matters is the renewal rate; everything else is secondary.

Given that, if you're implementing AI, the AI should be built based on what will cause a higher renewal rate. What will cause a higher renewal rate is not cheapness. It's stuff like, I've launched things way faster. I was way more efficient so I got better impact. I gave you analysis and told you what your competitors were doing and what you're doing, and I found an actionable insight that you didn't know and therefore I applied it to this strategy.

It's also things like the relationship with the company. Back to my meeting transcript example, "How do I know whether or not a customer is becoming disengaged?" So it's that type of stuff. It's not cheaper, it's better.

RISHABH JAIN

I'm curious then, a lot of the prognostication on other podcasts or on social media is we're going to have a 10-person, billion-dollar company, or whatever those different prognostications may be.

Do you think that that is just the wrong conversation? Meaning, that point, while it may be true, is not the relevant point?

When you think about the building of your business and then also when you think about the servicing of your clients and what they are expecting of you, how does that idea interact with what you just said? Or do you believe it's orthogonal because it's not even the worthwhile goal?

ETHAN SMITH

It depends on the size of the company.

If you're talking about small [to] medium businesses, I think that the 10-person agency can work really well.

If you're talking about Rippling or a large company, you're going to have a human involved in that. It's the same with ... Meta wanted to have completely self-serve automated ads — and Google as well — and have this giant account management team.

They have a giant account management team because where you make the money is the whales and the whales need a person to talk to.

The same thing with services. The whales are where you make your money and they need to talk to somebody.

And if it's a 10-person agency working with 1,000 whales, there's no way that it can work. It doesn't matter what your technology does.

Ten people can work with thousands of small businesses, but the whales need to talk to someone. So if you want those people, you need a big team of people like me who will talk to them and manage the relationship and bring them interesting insights and ask them questions and understand what's unique and special about them and then apply that to the strategy.

That's the only way you'll keep them. There's no AI solution that automates that away with 10 people.

MIKE DUBOE

I'll ask a question that might be on Rishabh's mind. Agencies in the past — and this might expand beyond the scope of Graphite — but in general, agencies have proven to be an interesting distribution channel for software players. And I would imagine, Ethan, over the years, many vendors have approached you with either white label or kind of rev share agreements.

Now that it's become so much easier to build software, how do you think about that? Does that channel start to degrade?

ETHAN SMITH

Does self-serve marketing MarTech software degrade?

MIKE DUBOE

Does the agency distribution channel for MarTech software look different now as a result of agencies building in-house software?

ETHAN SMITH

Yeah, I don't see agencies building software in-house. That's actually pretty rare.

They're mostly stitching together other people's software. And maybe that becomes more fragmented, but agencies are not really building their own software, and when they do, it's usually inferior to whatever is publicly available.

So agencies will try to build software and either they'll copy or they'll try to build a better version of something that's available in the wild and do it worse, or they'll just stitch together other things or they'll do workflow tools.

But I don't see agencies replacing HubSpot or replacing Iterable or replacing Ahrefs.

RISHABH JAIN

I guess the follow-up question I was going to ask was, how do you square that with you guys building your own AEO tracking tool?

There's a version of the world where you partner with one or many of these companies that exist for AEO tracking, right? And then Graphite becomes the distribution, and I'm sure any of them would be thrilled to have you be the endorsing distribution arm of their company.

And at the same token, you have decided to build it internally. Help us understand what drove you to decide to build it.

ETHAN SMITH

So a few things. I don't think that we'll ever make a lot of money from AEO tools. We are just accepting that.

However, we can be first to market. I know exactly what the AEO tools are going to look like two years from now and we can build them quickly and we might as well just build them quickly.

And I know that they'll be priced at commodity prices — it would be $100 a month so we can just charge $100 a month. We know that right now.

And then how I'm thinking about it is a great distribution mechanism is to get people, a bunch of people to use your tools and then some percent of them will want to hire your agency.

And then this is a great way to do product-led marketing, but we'll never make a lot of money on the tools. Maybe we'll cover our costs and then we'll get a bunch of ... like, Neil Patel has Ubersuggest, which is a free keyword research tool. I don't think he needs that now, but before, that helped a lot.

And so get everyone in the world using our tools and then half a percent of them will hire our agency and then we won't need to do any marketing.

RISHABH JAIN

Just hopping back, Ethan, to something that you said earlier where you said you don't believe that there's a first-mover advantage, but at the same token you think that there are categories today where AEO search is important.

How would you describe a forward-thinking brand in taking advantage of this opportunity?

Because usually most technology trends, it's like, the best example is being first on an app store. There are clear reasons why you want to be first on an app store. The first apps on the Shopify app store were very successful, the first apps in the Apple App Store, because it was a distribution mechanism that worked very well.

And I guess you're suggesting that, in the world of AEO, that dynamic may not exist. There's not a compounding advantage to being first.

And so what should a forward-thinking brand do if they don't fall into the bucket that you previously described, which is they already have audience there.

ETHAN SMITH

Training data and proprietary data sets. Whether that's AI search or AI generally, if you think 10 years out from now, who has leverage in AI?

There's the LLM companies, and you're not going to compete with them. I think of them as the ISPs. So there's going to be the ISPs and that's fine, but you're not going to make a better ISP.

So there's the LLM companies, there's going to be five of them, but no one's going to win because they have a better algorithm. I'm not going to make my own algorithm and I'm going to beat somebody because my algorithm is better. Everyone's going to have access to the best algorithms.

What they're not going to have access to is the best training data. The algorithms are only as good as the input. So if you have better training data and better proprietary data sets, then your output will be better than everyone else.

So buying and acquiring and building proprietary data sets is the forward-thinking thing to do. And the way to think about that is, how much better is your dataset than the second-best dataset, and how valuable is the information? And that's the value of your dataset.

So Reddit, for example, is perfect for AI training data. Another would be if you have a closed community. We work with this company called Sermo, which is a closed doctor community where there's actual patient outcomes and they talk about, "I gave my patient this particular prescription and this is what happened," which is really valuable. That doesn't exist anywhere.

What exists right now is drugs.com and WebMD, which is nausea, vomiting, and diarrhea. That's basically the side effect of everything. That's not that useful. So if you have actual thousands of doctor outcomes from real patients, that's significantly more valuable than the second-best dataset.

So if I'm Sermo, I'm just going to grow my dataset as much as I possibly can and then it's going to be extremely valuable.

You just go category by category. Google, for example, will do custom deals with weather.com and weather, shopping, sports. You just go category by category. I'm going to need a data provider, maybe I'm going to have exclusivity, maybe I won't.

Then the gap between the value of my dataset and the second-best — that and the value of the category — that's the value of the dataset.

What I would do is just go buy up a bunch of closed, proprietary, user forum communities, or I would go buy a bunch of publishers, like Time Inc and New York Times, I believe they did deals.

Publishing is a really bad revenue category, as you know, because it's monetized by ads, it's not monetized by the value of the data. So if you monetize by the value of the data rather than banner ads, then it's worth way more money.

Go buy up a bunch of newspapers and publishers that are undervalued because they're monetizing on banner ads, roll 'em up, and then sell that as a proprietary dataset. And now you monetize based on the training data and the input.

RISHABH JAIN

That's pretty controversial. I also think it suggests that — you tell me if I'm inferring here correctly — but it suggests that marketplaces and communities will be much higher value than they are today.

ETHAN SMITH

Yes.

RISHABH JAIN

Because that's basically where you observe the most behavior. So you take a DoorDash interaction or you take an interaction on Faire, for example. Nobody knows. Those are all private data interactions.

ETHAN SMITH

Also, something like Apple Watch, Fitbit, any outcome information. There's so many ... There's billions and billions of outcomes of things in the wild and almost none of them are digitized.

So if you can digitize all of these outcomes and then feed that into AI, that's extremely valuable.

I would also look at really high-LTV microcommunities like a finance community, something like that, not monetizing well, buy that up, and now that's a super valuable proprietary data set. Access to the outcomes is a huge data moat.

RISHABH JAIN

So on the one hand you're saying like, "Hey, this proprietary data is going to be very important." When you think about the tech stack of a given company on the internet in the future. So today's tech stack is, you have a static website, and maybe the sort of pseudo-modern tech stack is there are API interfaces to your product or to your publication or your data underneath that.

What do you think the tech stack needs to look like as more and more of the information that people are receiving are going through these either chat interfaces or through agentic interfaces?

ETHAN SMITH

It doesn't really matter / velocity. So part of what you're saying, part of the premise of your question is, you have all this information, how do you make it so that we can transfer it efficiently to the LLMs?

It's not that hard. So if I have a shopping data set, product pages, category pages, metadata, product feed, schema, there's nothing really new. That's all you need.

The second thing is velocity. All the CMSs can make a product page and most of them are slow. So a common conversation we have is we come in and we scope our project and they say, "Okay, well we talked to the engineering team and this is going to take nine months if we get the engineering resources. Most likely the engineers are actually busy right now, but we'll get back to this in a few months." But much of this stuff could be built in weeks or days.

And I mentioned when I started doing consulting with SEO, I would have a whole plan and almost not a bit would get launched. So we evolved the agency to be a full-stack agency so we can do everything from writing, building, designing, end to end, and it's actually better and faster if we do everything ourselves.

It actually costs us less money to do everything than it does for companies to do part of what we do, usually, because we can do it way faster.

So what matters the most with CMSs is speed and velocity. And so there's a bunch. We like Webflow, so we built a whole Webflow team because we can do things really fast.

I'll add that a lot of times, engineers don't want to use Webflow because when you have a hammer, everything looks like a nail, or whatever the saying is. So I can't use Webflow because I need to write code. And then people who don't code don't know how to use Webflow efficiently, but Webflow, we can build programmatic pages, search, product pages, listicles, articles. We can do that in days.

There are other CMSs, but we built out a whole Webflow team because it's just super fast.

What I actually think matters the most is velocity versus "this CMS has a special feature that parses the thing in an LLM-readable way." I think it's velocity that matters.

MIKE DUBOE

Well, Ethan, you've given us a lot of good fodder for post-episode engagement and probably another conversation on this in a few months, so thank you.

This has been a great conversation and we're excited to keep chatting offline on this stuff. For anyone who wants to find you, where should they look you up?

ETHAN SMITH

LinkedIn. So I post a bunch of stuff on LinkedIn, so add me on LinkedIn, follow. We do a lot of original research and then I post it there.

We have a blog, so our domain is graphite.io, and then we have a blog called /five-percent, which stands for 5% of what you do is impactful and 95% is not impactful. So we only talk about impactful things. So that's our blog, and then we post research there.

RISHABH JAIN

Amazing. This was amazing, Ethan. Thank you.

ETHAN SMITH

Yeah, you're welcome. Thanks for having me.

‍

‍