EP 24·Jan 14, 2026

The ultimate guide to AEO (Director's Cut) | Ethan Smith, CEO @ Graphite

Show notes

Graphite is nothing like your typical agency. While most agencies sell slides and strategic decks, Graphite builds things. While others work on hundred-item SEO audits, they obsess over the 5% of work that actually drives results.

They've proven this approach in SEO with companies like Rippling, Webflow, Notion, and Upwork, and now they're applying the same ruthless focus to AI search.

That's why I'm excited to have Ethan Smith, CEO of Graphite, on the show today.

Ethan has led Graphite to the absolute forefront of Answer Engine Optimization, one of the largest new marketing channels we've seen in over a decade.

He already gave the ultimate guide to AEO on Lenny’s podcast a couple of weeks ago, but I felt like a lot of questions remained unanswered. So I asked him to do a follow-up deep dive, and now we’re here.

▶ Let's connect! 🔗 Niklas on LinkedIn: https://www.linkedin.com/in/niklas-buschner/ Radyant on LinkedIn: https://www.linkedin.com/company/radyant/ Ethan on LinkedIn: https://www.linkedin.com/in/ethanls/ Graphite on LinkedIn: https://www.linkedin.com/company/graphitehq/

Transcript

Full conversation

via podigee

00:00:00Graphite is nothing like your typical agency.
00:00:03While most agencies sell slides and strategic decks, Graphite builds things.
00:00:08While others work on a hundred item SEO audits, they obsess over the five percent of work that actually drives results.
00:00:14They've proven this approach in SEO with companies like Rippling, Webflow, Notion, Upwork and many others and now they're applying the same ruthless focus to AI search.
00:00:24That's why I'm excited to have Ethan Smith, CEO of Graphite on the show today.
00:00:29Ethan has led Graphite to the absolute forefront of answer engine optimization as in AEO, one of the largest new marketing channels we've seen in over a decade.
00:00:39He already gave the ultimate guide to AEO on Lenny's podcast a couple of weeks ago, but I felt like a lot of questions remained unanswered, so I asked him to do a follow up deep dive and now we're here.
00:00:51So welcome to the podcast, Ethan.
00:00:54Thank you for having me.
00:00:55I could talk about this subject for days.
00:00:57So let's let's answer some questions that haven't been answered yet.
00:01:00Okay, let's let's try.
00:01:02let's see if we Will will end up with a twenty four or recording.
00:01:06Maybe let's start with a quick recap for everybody that has maybe also not seen Lenny's podcast.
00:01:11you and Lenny's podcast.
00:01:12I highly recommend for everybody to watch or listen to it.
00:01:16Why should companies care about AEO?
00:01:19I think that they should care about AEO because it's a substantial channel, meaning the amount of impact you can get now is substantial and it's growing.
00:01:27There's many channels you could focus on.
00:01:29This is one of many.
00:01:30So I wouldn't say that everyone should be focused on it, but it is one of the channels to consider, especially because the growth curve is steep.
00:01:38And what would you say?
00:01:39who should focus on it?
00:01:40because you said not everybody should focus on it?
00:01:42I
00:01:42think it depends.
00:01:44So let's talk about late stage versus early stage.
00:01:46And when early stage companies ask me what they should do for SEO, I say do nothing at all or just rank for your brand terms, but don't invest in it because it'll take a long time to drive impact.
00:01:58And if you're a startup, you don't have, if you're a seed stage startup, you don't have time to wait two years for impact.
00:02:04Whereas if you're late stage company, probably most people should be focused on SEO.
00:02:07For answer engine optimization, I think it depends.
00:02:10It probably wouldn't be my number one channel.
00:02:12It might be my fifth most important channel.
00:02:15So if I'm early stage, I would probably do a small amount of stuff, but it wouldn't be my top investment.
00:02:20For early stage and for SEO, I would do nothing for early stage, but for AEO, I would do something.
00:02:27And the reason why is because you can win answers.
00:02:30by getting mentioned by other websites.
00:02:33It's unlikely that your URL as a C-stage company will become suddenly part of the citations because answer engines have algorithms similar to search algorithms where you need a lot of authority to rank.
00:02:43So the time for your website to rank might be... a long time, but you could get mentioned tomorrow by a citation.
00:02:49So for seed stage, it could be interesting because you can get impact quickly.
00:02:53Now, I wouldn't focus on thousands and thousands of prompts.
00:02:56I would focus on a small number of them, but I'll probably focus on a few.
00:03:00So I would do a light investment if I was a seed stage or the stage company.
00:03:03And then the larger the company, the more interesting it is because typically you will have saturated.
00:03:08at least the obvious things in search and ads and stuff like that.
00:03:12So it's a new channel that's incremental to what you're already doing.
00:03:14So the later and larger you are, the more interesting I think it is.
00:03:18You talked also to Lenny about who should be responsible for AEO, and you gave a good answer already there, but I'd like to dive a little bit deeper into what you think how companies should set up their teams.
00:03:33So if they should rely on their in-house teams, like maybe the marketing team, the SEO team, if they should partner with an agency, what are your thoughts on that?
00:03:42I think there's two different strategies, and it's off-site and on-site.
00:03:47Could be the same.
00:03:48Person or group it could be two different ones, but the strategy the the skills to do both are fairly different.
00:03:54for onsites That's just traditional SEO.
00:03:56so make pages the target prompts have it the content be good technical SEO technical a you know things like that.
00:04:00So that's pretty straightforward.
00:04:01the off-site stuffs the stuff that is somewhat new and the strategies are actually not that different from SEO, but usually you wouldn't have.
00:04:10Usually, at least today, you wouldn't have a team spending a ton of time on backlink building.
00:04:13You might do a little bit of that, but it wouldn't be a core focus of a team.
00:04:17Whereas with answer engines, getting mentioned by citations is actually probably more worth the time than it would be for link building.
00:04:24So essentially, when you say chat you between what's the best credit card, you want the citations to say that you're the best credit card.
00:04:30And if you are mentioning the citations, you're suddenly the best credit card.
00:04:34That's not really true, or you could suddenly rank.
00:04:37you know, outranked Chase and Wells Fargo for the best credit card in Google, but that is true for answer engines.
00:04:42So having an offsite strategy, I think is quite impactful.
00:04:46The kinds of strategies, they vary quite a bit.
00:04:49Reddit, YouTube, LinkedIn, Instagram, G-II, these are pretty different from... Key we're targeting and content optimization things like that.
00:04:59that doesn't mean that it needs to be a separate team But frequently those are two different types of skill sets.
00:05:04So it may or may not be to separate a separate team.
00:05:06Now your question was about agency versus in-house.
00:05:09I think it could be either.
00:05:10I don't think that there's a real distinction.
00:05:11the one thing that I'll mention is that there's not that many people who are masters of Reddit and G to and LinkedIn and these new channels.
00:05:23so You could either find an agency who does have experience there, or you could find a smart resourceful deals with ambiguity type person, which typically you would get in-house more so than an agency.
00:05:36Agencies don't tend to be suddenly inventing brand new strategies, you know, in the middle of a project, whereas somebody in-house would.
00:05:45So if you have an agency who's already good at these things.
00:05:48Great, but if you don't, I would expect that an in-house person would be more resourceful, more adaptive.
00:05:54And how do you usually collaborate at Graphite with companies?
00:05:58So if we're thinking about like your collaboration with Webflow, which is pretty public.
00:06:03So you've published a lot of things about how you increased the number of signups from LLMs.
00:06:09And also, I think Josh Grant, VP growth at Webflow, posted a couple of times on LinkedIn about it.
00:06:16How does a good working relationship between an in-house team and an agency look like from your experience at Graphite?
00:06:26I think the main thing is.
00:06:27when we started with Webflow, it was about SEO, and the stuff that we're doing were not things that we agreed to, and was not on the SOW, it was not on the scope of work.
00:06:36And if we stuck to what was in the scope of work, we wouldn't have gotten the impact.
00:06:41So we need to be able to work with other people, other teams at Webflow, and we need to be able to work on things that we didn't agree to.
00:06:49and try out new stuff.
00:06:50And this is part of why I mentioned, you usually would not do that with an agency.
00:06:54You would usually just have a SOW and stick to the SOW.
00:06:57You're not adaptive.
00:06:59We try to be adaptive with Webflow.
00:07:01But we need Webflow to be open to us.
00:07:04working with other teams and working on Reddit, working on YouTube videos, stuff like that, that was outside the scope of our original work.
00:07:12And that comes from building trust with Webflow and also just building relationships with other... parts of Webflow.
00:07:18and for us to be adaptive and for our company.
00:07:21The way that I think about agencies is it's kind of like a boat.
00:07:25And the bigger the boat, the harder it is to turn.
00:07:28So if you're in a tiny little boat with one person, you can very quickly turn.
00:07:32But if you're on a cruise liner, you can't turn that fast.
00:07:37You can do more with a giant cruise liner, but you can't turn quickly.
00:07:39So we try to be as adaptive as we can.
00:07:42But that's how you can work well with someone like us.
00:07:46Sounds good.
00:07:47I like the boat analogy.
00:07:49And what would you say, no matter if it's someone looking for hiring people in-house or looking for the right agency or maybe even freelancer to collaborate with, what would you look for in people that are the right ones to push growth from AEO?
00:08:04There's two things.
00:08:05There's analysis and experiment design, and then there's dealing with ambiguity.
00:08:09And I would say that Most people in SEO, AEO are not great at experiment design and analysis.
00:08:17And that's why you have so many best practices that are not true.
00:08:21It's because either people never looked at it or they didn't look at it correctly.
00:08:25And they had a false positive where they thought that something had a causal effect and it didn't.
00:08:29So being able to do rigorous analysis and set up an experiment to figure out what actually works is very important.
00:08:35And we don't know most of what works in the answer engine optimization.
00:08:38We're still early.
00:08:40So there aren't a bunch of best practices you can pull from.
00:08:42You have to do experiments.
00:08:43It's the only way that you know what's working.
00:08:45Otherwise, you're going to waste a bunch of time.
00:08:47So that's skill one.
00:08:49And that's hard to learn on the job.
00:08:51If you show up at a job without knowing how to do experiment design and analysis, you could do that.
00:08:56But I don't know if I've ever seen that happen.
00:08:59We try to look for people who already have that skill and they don't have that skill.
00:09:03Let me assume that they are unlikely to develop that skill.
00:09:06And then the second thing is dealing with ambiguity.
00:09:10And I think what makes a good person in answer engine optimization and in growth generally is the ability to deal with ambiguous problems.
00:09:18Like if I said, why don't you go figure out a Reddit strategy?
00:09:22That's an ambiguous problem.
00:09:24Whereas I said, I want you to take these ten steps and follow this.
00:09:29rubric and and you know replicate what I just showed you.
00:09:33that's a different type of skill and I would say most people are the second type where they can follow prescriptive process but they cannot solve an ambiguous problem and some problems don't need to be solved.
00:09:49like many things have already been solved a lot of stuff around paid advertising for example.
00:09:54there's very established processes for exactly what you do.
00:09:57You don't need to reinvent how to do bidding.
00:10:00But with Figure Out Reddit, you do need to invent that.
00:10:03And so you need someone who can do experimental design and analysis and solve ambiguous problems.
00:10:09Let's talk about another topic that I found very interesting.
00:10:12You also, I think you posted some stuff about it already in reference to your work with Vimeo, which is video.
00:10:23in AO.
00:10:23So I'd like to understand from your point of view why video is important when we talk about AO.
00:10:31I think that UGC and diverse opinions are really important.
00:10:36And video has a lot of that.
00:10:38So if you look at Google, one of the biggest complaints is that the results are all derivatives of each other.
00:10:44Everyone is using a content scoring tool and looking at the top ten results and then rewriting them.
00:10:52And all the results are just rewritten versions of each other.
00:10:55And so you have a lack of diversity.
00:10:58And then Google then wants to rank Reddit and Twitter and video to increase the amount of diversity.
00:11:04Like if you have one opinion on everything, that's bad.
00:11:06You want many opinions.
00:11:07You want the wisdom of the crowds.
00:11:08The wisdom of the crowds says that the more opinions that you have, the sum of the group of people will be better than the best expert.
00:11:17So if you said, how much does this person weigh?
00:11:23the if you pulled a thousand people the average Often would would be more accurate than the best guess of the best person within the thousand people.
00:11:33and so the more diversity the better.
00:11:35so if Google is all Rewrites of each other there's no diversity of opinion which which we already have that problem of.
00:11:40so then you want UGC.
00:11:41UGC is a great way to have you know way much larger wisdom of the crowd.
00:11:47so Then what do LMS want?
00:11:49So LMS want the exact same thing.
00:11:50And LMS are especially good at summarization.
00:11:53So the larger the data set, the more useful the LMS is because it can summarize the opinions of many people.
00:11:59And this is related to why answers vary.
00:12:02It's not because it's unpredictable.
00:12:03It's because it's a probability distribution.
00:12:06And so the bigger the distribution, the better the output.
00:12:10So then coming to video, where do we have the most unique information?
00:12:15We have it in... text and we have it in videos images.
00:12:18So in text we have it with reddit, Quora, X, LinkedIn, which is why those are huge inputs to LMS.
00:12:25And then we have video.
00:12:26And video is already a large input to LMS, but the majority of the information in videos is locked in the video.
00:12:33Most of what's used is the title in the description, maybe comments.
00:12:36But the majority of the context in the video is not accessible by the LMS.
00:12:41I think it will be at some point.
00:12:43And video is already a large data source.
00:12:46Once you can extract more information from the video, I think that video will overtake Reddit, overtake X, overtake written content, because there's so much context in videos.
00:12:55Vimeo, who we work with, just launched something where you can do a search, like a semantic search thing.
00:13:02I want you to find this type of a scene within a video.
00:13:06So you have... You have the title in the description, that's mostly what's used in the album now.
00:13:10Then you have the transcript, which you could make accessible.
00:13:14But there's so much information that the transcript doesn't really give you, like the semantic understanding, the one layer of abstraction about just the words that were said.
00:13:24And the thing that Vimeo just launched, they're not like feeding that into an LLM, but that type of thing when an LLM can understand what's happening in a video.
00:13:33Like for example, we're having this conversation, Right now.
00:13:37if you just looked at the words that I said there would there would be a decent amount of information.
00:13:40But if you could have a layer of abstraction about the types of things that I'm you know the concepts that I'm saying.
00:13:45like I talked about prioritization I talked about.
00:13:49I talked about types of people you could hire When you add a layer of abstraction above that then there's so much more unique Content that could be fed into the LMS and then you have this massive corpus of unique content and rather than having the ten rewritten articles that are all the same thing.
00:14:06You have this huge amount of UGC.
00:14:09So that's why I think the video is so important today.
00:14:11It will be even more important, maybe the most important thing over time.
00:14:15Let's quickly tap into the rewritten content problem.
00:14:19So I think this is something that is still happening, unfortunately, a lot.
00:14:26You guys are also creating on-site content.
00:14:30So not talking about UGC and the off-site initiatives here, but on-site content, for example, for Webflow or also other companies.
00:14:39How do you ensure that you're not part of the problem?
00:14:42So creating like rewritten content, but actually bringing in unique point of view, et cetera.
00:14:50How do you handle it operationally?
00:14:52There's two kinds of content.
00:14:55There's human content, then there's AI content.
00:14:57Most of what we do is human content.
00:14:59And so part of it's diversity and quality.
00:15:03So quality is easy.
00:15:05Quality, you just hire domain experts.
00:15:08It can be expensive.
00:15:09But we have.
00:15:11we have thirteen editors.
00:15:12We have multiple rounds of editing for every single piece of content that we create.
00:15:16We hire dedicated domain expert writers for every project that we work on.
00:15:21I think we accept five percent of all.
00:15:24applicants, so we're very rigorous about sourcing the writers, and then multiple rounds of editing.
00:15:30So that's on the quality side, and then on the uniqueness side, we try to have uniqueness and information gain.
00:15:37So how we could do.
00:15:38that would be, there's various ways to do that.
00:15:40So you could have expert quotes, you could have unique metadata, you can have stuff about how to use the product to do the thing, like how do I use Rippling to pay people in Argentina, so not just like.
00:15:51This is how theoretically to do it, but this is how to use the product to do that.
00:15:54So we try to find unique hooks about the company to add some amount of uniqueness or information gain.
00:16:01Let's draw from there to go to a content workflow a little bit deeper, because obviously we're talking about AO and we're talking about AI as a marketing channel, but AI can also be a very powerful tool for a lot of people.
00:16:15I think from the how people use chat GPT study that OpenAI did with Harvard for professional users, actually the biggest category of usage was writing.
00:16:26So I think forty percent of people like the latest data point was forty percent of the usage can be categorized as writing.
00:16:34know that you have done, I think, multiple studies, but the latest one was with Axios about AI-generated content.
00:16:44So my question is, why does AI-generated content not work?
00:16:50Yeah, so it does work, depending on how you do it.
00:16:55So usually when people think of, I'm going to... There is this eight, I forget what it was, the AI heist, the SEO heist, where I'm gonna take all your content and rewrite it with AI.
00:17:06Or I'm just gonna say ChatGPT, write a million landing pages, launch all of them.
00:17:10Here are all the keywords, make pages for it, done, free.
00:17:14That doesn't work.
00:17:16And that doesn't work for a few different reasons.
00:17:19I mean, reason one is I think that Google has an AI detector and they just generally are sensitive to that.
00:17:23Reason two is, well, I mean, That's the reason.
00:17:28Now, why shouldn't it work?
00:17:30I'll reframe the question.
00:17:31Why should it not work?
00:17:33And so if you just say ChatGPT write an article about this thing, you're essentially deriving that from information that everyone has access to.
00:17:41So you're adding nothing on top of what already exists.
00:17:45And the similarity will probably be pretty high.
00:17:47And there'll probably be, I'm guessing Google has some sort of similarity plus AI footprint.
00:17:51And if it's AI, but it's unique, it's fine.
00:17:54AI is not unique, it's not fine.
00:17:56That's my guess about how they're algorithmically doing that.
00:17:58Because we have multiple examples where AI content is working, but it works when you have a unique input.
00:18:06That's not just a, you know, and it's adding, it's additive.
00:18:10So a couple of specific examples.
00:18:12Example one is we work with Ceremo, which is a community for doctors where they can talk about.
00:18:18they gave their patients a particular drug and this is what happened.
00:18:21Like I gave my patients this particular SSRI and they had these conditions.
00:18:26and this is what happened.
00:18:27And this is all proprietary.
00:18:29It's closed.
00:18:30So it's not in the public.
00:18:31And then we have landing pages saying summarize what the doctors are saying are working and not working.
00:18:37And that actually works quite well.
00:18:39They rank right underneath WebMD and drugs.com.
00:18:43And it's AI generated.
00:18:44But it's an AI generated summary.
00:18:47of unique information, really high quality, unique information.
00:18:51And it's adding.
00:18:52You can't find that anywhere else.
00:18:54So that works.
00:18:55We also have examples with Webflow where we have FAQs or product related content.
00:19:00And that also works.
00:19:01But that's because, again, it's stuff that's unique to Webflow.
00:19:04So when you're adding, like AI is great at summarizing things.
00:19:08And if you're summarizing unique information that is not accessible by the public, then you're adding.
00:19:14And so, again, algorithmically, my guess to what Google's doing is they're saying AI detector, yes, no, and then is unique, yes, no, if unique in AI, good, if not unique in AI, bad.
00:19:26Okay, makes sense.
00:19:27Can you give us... A little bit more of an understanding of how you use, for example, Aerobs with Webflow because Aerobs from my understanding is also strongly pushing towards AI content generation and like this content engineering movement.
00:19:42So how do you use it and what are also your thoughts on the whole content engineering movement?
00:19:48I think it's great.
00:19:49I also think.
00:19:51AI content is the future.
00:19:53Clearly, it's better if you have AI as a co-pilot than if you don't have it at all.
00:19:58And it's also bad if you're just rewriting each other's content with no value add.
00:20:03So I think clearly, you want an AI co-pilot.
00:20:06Just like with mobile and the internet, there's a bunch of creator apps.
00:20:09There's Instagram and TikTok, and it's way easier to create videos now than it was in nineteen eighty.
00:20:15You would need to spend hundreds of thousands of dollars and now it's free.
00:20:18So I think that AI will make it.
00:20:20There'll be a whole suite of creator apps.
00:20:22AirOps is more on the professional side.
00:20:24There'll probably be consumer side ones like Captions and others.
00:20:30But AirOps is focused on the enterprise professional content workloads.
00:20:36I think it's very interesting.
00:20:37So the kinds of things we've explored there are FAQs.
00:20:41I think also... things like category pages where you could have something like an apartments in Santa Monica page and you have unique metadata about the prices and the inventory and, you know, you want a list of apartments and you want to know average price inventory kinds of features.
00:20:58You don't need, that doesn't need to be, you don't need to hire someone from the New York Times to write what the average price of an apartment in Santa Monica is.
00:21:06But what you do want is you want unique metadata.
00:21:08So then you could use something like an AirOps to summarize this metadata.
00:21:11So feed on the unique metadata and get a summarization of that.
00:21:14FAQs, I think that product content is also very interesting.
00:21:17A lot of what people ask in LMS is, I want a product with these features and these attributes.
00:21:21Again, that doesn't need to be someone from the New York Times writing a beautiful story about whether or not you have an integration with Rippling.
00:21:29You just need the information.
00:21:30so then an AirOps would be good at that as well.
00:21:33So I think that this space is quite interesting.
00:21:35But it shouldn't just be some form on top of chat GBT where you say write a resume and then it just sort of suddenly gets this thing.
00:21:43You wanted to be hooked into workflows such that you have that unique metadata and you're leveraging that as much as possible.
00:21:50Got it.
00:21:50We already touched on tooling now with ERABS.
00:21:53I'd like to go a little bit deeper into the whole AEO tool landscape because I know that you also published a couple of things about that.
00:22:03So first of all, maybe If people are thinking about getting an AEO tool or I think like marketers always tend to think about, hey, I need a tool for that if there's a new channel.
00:22:16What would you say?
00:22:16What should we look for?
00:22:18Or what should people look for?
00:22:20I think you should look for.
00:22:23what do you want to do with it?
00:22:24And then what's the cost?
00:22:27And if there's a new feature, is it as sold?
00:22:30So what do you need to do with it?
00:22:32You know, SEO tools are a good example.
00:22:35There's Screaming Frog, one of my favorite tools.
00:22:39The majority of the columns are not useful.
00:22:42There's a few columns that are useful.
00:22:43So when I use Screaming Frog, I'll do a site crawl, and typically I'm just looking to see the index pages, how's the internal link coverage, and so I look at unique inlinks.
00:22:54The rest of the columns are not useful to me.
00:22:56That doesn't mean that they're never useful, but they're not useful to me.
00:22:59So if I had some other Screaming Frog competitor, and it was one-tenth the cost, and it didn't have the meta data, the meta description character count.
00:23:08Like, I don't care about that.
00:23:09I don't need that.
00:23:10So, what do you actually need?
00:23:13Don't just look for a feature list in a series of checkmarks and like, well, I need that feature.
00:23:18Like, what do you actually need?
00:23:20Number one.
00:23:21Number two is what's the cost?
00:23:23And I think that the tool space generally is going to... rapidly commoditized, and we've already seen that, but just like SEO tools, there's no moat around SEO tools.
00:23:35So SEO tools are generally a commodity, which is why everything costs between eighty dollars and one hundred thirty dollars a month.
00:23:41The exact same thing will happen with eight EO tools.
00:23:44I think that an air ops and workflows and stuff like that is different, but I'm just talking about an SEM rush, a trust competitor.
00:23:52It's going to cost eighty dollars to one hundred fifty dollars eventually because of commodity.
00:23:57And then there's new features.
00:23:59So there's four features that I think are interesting, and probably long term, there's only four features.
00:24:07It's visibility tracking, prompt volume, workflows, and content scoring.
00:24:14So visibility tracking, I think everyone's generally familiar with this concept, but for keyword tracking, you would use an AHRFS SCM rush and say, what do I rank?
00:24:22And you rank position three for content workflows, let's say.
00:24:26And so then for LLMs, you would say, well, when I ask what's the best payroll management software, I want to know if my brand shows up.
00:24:33So this is how often do you show up and where do you rank?
00:24:36I think everyone's familiar with this.
00:24:38Things to consider with this are where the data come from.
00:24:43So there's the API, there's logged out scraping, and then there's logged in scraping, and then there's paid logged in scraping.
00:24:52These are four different things.
00:24:54They all have different types of data.
00:24:56So let's say that I'm calling the API, it's going to give you citations that are very different from logged in citations.
00:25:05And what you really care about is probably logged in citations, like nobody's using the API.
00:25:09And so the answers will actually probably not vary that much, but the citations will vary a lot.
00:25:15So if you're doing offsite, for the seed station company that I mentioned.
00:25:19If you're using API data, it's gonna be way off.
00:25:21You're gonna be optimizing citations that are not used in logged in.
00:25:25So you wanna consider where's the information coming from.
00:25:27And again, for any new feature or data source, I suggest getting multiple sources of truth and comparing them.
00:25:35So rather than just picking one thing and assuming that it's correct, get a few different sources of truth.
00:25:39So I would get, I would probably use two tools and then I would also compare with search data.
00:25:44And this brings me to.
00:25:46So for visibility tracking, I would manually take questions, log in and manually go into a spreadsheet and put in the citations that you see and compare it.
00:25:55You don't need to do that for thousands of things, but I would at least sanity check how close is this tool to my actual experience.
00:26:01So that's visibility tracking.
00:26:03So there's prompt volume.
00:26:04So prompt volume is interesting.
00:26:06This would be how many people are looking for this prompt?
00:26:08What are the most popular prompts?
00:26:09It's kind of like search volume.
00:26:11How many people are searching for this particular keyword?
00:26:13And so for search, This comes from Google, essentially.
00:26:18Google Ads API gives you search volume.
00:26:20Google Search Console gives you impression counts.
00:26:22Bing gives you impression counts.
00:26:24Bing API gives you search volume.
00:26:26So we have actual first-party data from four different sources.
00:26:29Google Bing Search Console and Google Bing Ads.
00:26:33So we have a lot of good data there.
00:26:36We don't have that for any of the prompts.
00:26:38None of the LMs give us prompt data.
00:26:40So then we have to get data that's as close as possible to that.
00:26:47There's panel data.
00:26:48So what panel data would be would be I have some subset of the total population and that would be coming through browser extensions.
00:26:57You bought it from someone like.
00:26:59you bought it from a browser extension.
00:27:01You bought it from an LLM.
00:27:02There's a bunch of LLM apps.
00:27:03I could just pay them and say, please give me your prompt volume.
00:27:07There's ISP data.
00:27:09So there's all these different sources and it's a subset of the total population.
00:27:13So then the question is, how representative is the subset.
00:27:17And the way that I think about this is it's not just about the volume, but it's about how representative it is.
00:27:23So if I did a presidential poll and I said, who are you going to vote for?
00:27:27Most presidential, so the voting population I think is two hundred million plus, and most presidential polls are a thousand people.
00:27:36And they're, I mean, you know, they're not perfect, but they're pretty close.
00:27:39You're not going to be off by twenty percent for presidential poll.
00:27:42So it's a thousand people out two hundred million people, which is a tiny fraction of a percent.
00:27:46So the relative sample is extremely low.
00:27:48However, if it's representative, then it can get you pretty close.
00:27:52Now, let's say instead I poll all of California and say, who are you going to vote for?
00:27:58It's going to be a much larger sample, but it won't be representative.
00:28:01So I think it's like.
00:28:03thirty million people in California or something for voting.
00:28:05So if I pulled thirty million non-representative versus a thousand representative, it's a way better answer.
00:28:10So the way to think about panel data is not the size, but the level of representativeness.
00:28:14And so that's what people are generally using for prompt volume.
00:28:18Now there's discussion about how close particular prompt volumes are, and I come back to you want multiple sources of truth.
00:28:25So I would compare it with search data.
00:28:27And how to think about this is that Search engines get roughly twenty-five times more page views than LLMs.
00:28:35So therefore you would expect that the prompt volume would be one-twenty-fifth of the search volume, generally.
00:28:41It's not exactly the same, but that's generally what you would expect.
00:28:44So the closer prompt volume is to search volume, the more off it is.
00:28:48And so if you just want to generally sanity check, is this prompt volume reasonable?
00:28:52I would just say the expected relative size is take the search volume.
00:28:56divide by twenty-five.
00:28:57The further away it is from that, the more inaccurate it is.
00:29:01Let's step into the the visibility aspect a little bit deeper because we got a question also beforehand from the community and I'm trying to weave in these questions now a little bit.
00:29:13If you have a favorite tool to measure visibility.
00:29:17Tools that I use and like we have our own internal tool but we decided to non-commercialize that so that's just for internal usage.
00:29:25and then other tools that we use and like our Aerops has visibility tracking, PKI, Surfer SEO and EverTune.
00:29:37And I haven't tried the rest.
00:29:40So the exclusion of the rest is not because I tried them and I didn't like them.
00:29:43It's because I have not tried them.
00:29:45So I would say, I think I haven't had a bad experience so far that they're all good.
00:29:52But again, I would compare the data and the citations with your actual personal experience.
00:30:00What we also do internally is we just have actual people logging in and copying and pasting things.
00:30:08So like if we really care, we're literally just having humans copy and paste things into spreadsheets.
00:30:13and that's how we do tracking.
00:30:15Refreshing.
00:30:17Another question we got beforehand was basically It's a couple of questions and I feel like they fit in very well at the moment.
00:30:24So the first one is how do you decide which prompts are worth tracking?
00:30:29There's a few different ways to think about this.
00:30:32So there's prompt volume from panels, which you talked about.
00:30:35There's search data and then there's other conversation data and I mostly use the second two.
00:30:42So search data I think is great because The way that the relative volume of search is probably gonna be pretty similar to the relative volume of prompting.
00:30:52The format will probably be different for prompting prompts are longer than searches But I you know the the relative number of people looking for payroll management software and search is probably gonna be similar to prompting.
00:31:04and if people related to payroll management software are looking for integrations and or you know What's the best ETF with these features?
00:31:11It's probably gonna be roughly similar.
00:31:13So for search data what I do is I either take my existing search data, or I take my competitor's search data and paid data.
00:31:22So if I'm a brand new credit card company and I want to know which prompts to care about, I'm going to go find another credit card company.
00:31:28I'm going to look at what keywords they're bidding on, which are probably going to be the valuable ones, and then transform them into questions.
00:31:34And so you just go into Ahrefs and put in your competitor, get their top paid keywords, export CSV, and then go into chat GPT and say, make a table with these keywords and then make a question version of that.
00:31:45That's actually pretty good.
00:31:47It's not going to be exactly what people are prompting, but it's going to be pretty good.
00:31:50It's going to be directionally good.
00:31:52So that's a shortcut for finding prompt volume, and that's what I do.
00:31:58So that's option one.
00:32:00Option two is panel data.
00:32:02And I talked a bit about this.
00:32:04I don't actually know how accurate it is because I don't know what the panel is.
00:32:08So back to the presidential poll, I don't know anything about the poll.
00:32:11So if I don't know anything about the poll, it might be really accurate.
00:32:13It might not be accurate.
00:32:14I just don't know.
00:32:15So if I don't know, I'll wait until I have some more confidence before I trust panel data.
00:32:20But I know that it eventually could work.
00:32:22I just don't know if it works really well today.
00:32:24Then the third would be other conversation data.
00:32:27And for prompts, the tail is larger and longer.
00:32:32There's really specific prompts that have never been searched for.
00:32:35So how do we know what that looks like?
00:32:38With search data, it necessarily won't have it.
00:32:41So how do we get information about the tail?
00:32:43The way to get information about the tail is try to find conversation data and other data sources that you do have access to.
00:32:49So you have conversation access to things like your sales calls, your customer support, Reddit, what are people asking about your brand on Reddit?
00:33:00What are they asking on YouTube?
00:33:02What are they asking on G to?
00:33:03these are all actual conversations and this can help us inform what the tail looks like.
00:33:08So what I would do potentially would be going to chat GPT and say summarize what people on redditor asking about my product or reddit actually has reddit answers where you can Say hey, what are people asking about my product?
00:33:18I'm gonna get a summarization of the kinds of questions people have.
00:33:22So that's how it fill in the tail.
00:33:23So, you know, just to summarize Search data.
00:33:26start with that and that'll tell you most of the information and then reddit answers G to tail, and then the more we know about how representative the panels are, then we can shift to panel data whenever we have confidence around that.
00:33:40And how many prompts do you typically monitor at once, or what would you recommend people to monitor?
00:33:47We work with later stage companies, and they'll typically do one to five thousand.
00:33:54But you could also ask, how many keywords should you track?
00:33:58If you have no information, how many keywords should you track?
00:34:01I mean, ideally you'd probably be tracking tens of thousands.
00:34:05If you really want a full picture, like how many keywords should Webflow track if they want to know how they're doing?
00:34:11Probably tens of thousands.
00:34:13So how many prompts?
00:34:15Probably ideally tens of thousands, but I think it's too expensive to actually do that.
00:34:19So I would do at least one to five thousand.
00:34:21This is again for enterprise companies, which is typically who we work with.
00:34:24So probably like one to five thousand.
00:34:26And then what I would do is I would take all of my search topics and then make question versions of that.
00:34:32So for Webflow, they want to target product managers and designers and solopreneurs.
00:34:38And then there's, you know, there are various features like creative feature persona matrix.
00:34:42And then you have your keywords.
00:34:44Are you ranking in search?
00:34:45And you have your prompts.
00:34:46Are you ranking for these prompts?
00:34:48And yeah, probably one to five thousand.
00:34:51And then you sort of do like ten prompts per page and then do this like visibility matrix.
00:34:58Okay, right.
00:34:58I think this will surprise a lot of people if they hear a thousand prompts because most start with like ten to fifteen.
00:35:05So what would you tell them?
00:35:07Yeah,
00:35:08I think it's.
00:35:08I think it's good to start with ten.
00:35:09ten to fifteen and ten to fifteen for a Brand new company or you know, let's say graphite.
00:35:16do we need to track five thousand prompts for ourselves?
00:35:19Probably not.
00:35:20We probably all need to be looking at ten to fifteen So that's totally fine.
00:35:24It's also a good way to get started.
00:35:26I'm more speaking to.
00:35:27if you're a webflow and you're a multi-billion dollar company, you probably want something more substantial.
00:35:33So it just depends on how big your company is and what the footprint of what you want to rank for is.
00:35:39Makes sense.
00:35:40Another question we got, and I found this very interesting, is given that LLM outputs can vary significantly because of each user's personalized layer, do you think tracking prompts is still meaningful?
00:35:53And if so, why?
00:35:56I disagree with the premise.
00:35:58So I have not seen a significant difference in answers based on the person.
00:36:04I think that that will happen, but I don't think that that's happening now.
00:36:08The majority of the variability of the answers is back to what I described earlier, which is that the answers are probabilistic.
00:36:16So if you said something like, what's the best website builder?
00:36:19There's a distribution of potential answers.
00:36:21Webflow will show up some percent of time and framer and lovable and and ghost and like You know, they'll have a probability distribution.
00:36:30Maybe a better example is what's the best flavor of ice cream?
00:36:34chocolate vanilla the most popular or there's hundreds of kinds of flavors.
00:36:38and So there's a distribution.
00:36:39depending on when you ask.
00:36:40you might get cinnamon and you might not get cinnamon like.
00:36:43maybe you get cinnamon ten percent of the time but you get chocolate ninety percent of the time and so depending on It's like a coin toss.
00:36:51So depending on when you ask it, you'll get a different answer.
00:36:55But that's not because it's because I asked it and it's not because the model is changing constantly or it's a black box and unpredictable.
00:37:02It's because it's a probability distribution and you're getting a different answer based on that.
00:37:06I think that someday you will get personalization, but I don't think that's the case today.
00:37:10So then how do you think about tracking?
00:37:13You ask the same question multiple times.
00:37:15So the more you ask the question, the more of the view of the distribution you get.
00:37:19We did a study that saw that if you ask about seven to ten times, you'll get a decent distribution, depending on the answer.
00:37:26For ice cream, you probably need to ask it more times to get the distribution.
00:37:29For website builders, seven to ten times is probably fine.
00:37:32So short answer is I would ask it multiple times and that for most prompts, that'll give you a pretty good sense of the distribution.
00:37:40So I just asked trajectory was the base flavor of ice cream?
00:37:42and it said the objectively correct answer is a vanilla.
00:37:47and ask it the second time and also said vanilla.
00:37:49so My anecdotal evidence says a vanilla is objectively the best flavor of ice cream.
00:37:55Do you agree?
00:37:56Did it mention any other flavors?
00:37:58Yeah, it mentioned that chocolate.
00:38:03So depending on your mood chocolate.
00:38:04and then I think there it got a little bit fancy.
00:38:07and it also says salted caramel.
00:38:09when you want to feel a bit fancy without committing to chaos
00:38:14Salted caramel.
00:38:14and if you asked it again right now without cashing, I will bet you it will not say salted caramel.
00:38:20I bet you it'll say some other.
00:38:22Yeah type of ice cream.
00:38:23Yeah, I got cookies and cream pistachio and then mint chocolate chip.
00:38:28Yeah, so I got chocolate mint chocolate chip pistachio Cookies and cream.
00:38:33and
00:38:34yeah, that's what I got.
00:38:35so it's mostly overlap Not salted caramel.
00:38:39Yeah, salty caramel was the fancy one if you feel a bit fancy.
00:38:43Yeah, maybe that shows up five percent of the time.
00:38:46So now this transition.
00:38:48I'm very proud of the transition that will come now.
00:38:50We talked about ice cream now.
00:38:51Let's talk about pies Because I want to talk about incrementality between SEO and AO.
00:38:59So I saw your case study from Webflow where you basically doubled the percentage of signups from LLMs, which is obviously something that a lot of people are very interested in.
00:39:13So because it's, I think one of the first like real publicly documented cases of people getting a lot of business value from LLMs.
00:39:23So, but my question is, isn't the gain you get from LLMs the pain, so to say, you lose from Google.
00:39:31So is the pie really getting bigger or is it just a shift from one to the other?
00:39:38The pie is getting bigger.
00:39:39So based on my data, my data, so there's a few ways to look at this.
00:39:45Ultimately, I think what you would want to look at is incremental conversions across tens of thousands of companies, like across every company in the United States or in the world, or the total number of conversions increasing.
00:39:58I don't actually know the answer to that.
00:39:59I don't have the data for that.
00:40:00I do have data for individual companies.
00:40:01One step before that would be is the usage incremental.
00:40:05So that I do know.
00:40:06If you look at similar web data, you'll see that visits for LLMs or page views are one-twenty-fifth the size of search, and they're generally incremental.
00:40:17Google is not going down, and maybe it's going down a tiny, tiny amount.
00:40:21But if you sum them, it's incremental.
00:40:23The pie is getting larger.
00:40:25Okay.
00:40:26And what about professional users?
00:40:28Because if we think about maybe you and me also, I can see us working in our day-to-day a lot with AI tools.
00:40:35And I think the probability of us then also turning to claw Chatchabity.
00:40:41perplexity for research is higher.
00:40:44But isn't it plausible that one prompt or maybe two prompts we do with Chatchabity is a substitute to, let's say, six, seven searches we would have done on Google earlier?
00:40:57Definitely.
00:40:58So my statement is about the macro effect.
00:41:01You're asking about the micro effect.
00:41:03What about different specific use cases?
00:41:05I definitely think that LMS are used more for specific use cases.
00:41:10LMS are very good at things like analysis, summarization of large data sets, research, and I would expect that People would use them instead of search because they're much better at that than searches.
00:41:22Just like I would expect for video, people would look for travel ideas in Instagram and TikTok and YouTube.
00:41:32More than they might look in Google.
00:41:34They might look in Reddit.
00:41:35More than they might look in Google.
00:41:37Same with beauty and things like that.
00:41:40Different services are better at fulfilling use case.
00:41:44Same with LMS, but I don't think that search is generally going down.
00:41:47I think that these new surfaces like YouTube, Instagram, TikTok, Reddit, LMS are additive to search.
00:41:56The pie is getting larger.
00:41:57The slice for search is the same size and you're just adding slices on top.
00:42:02And would you say that AEO as a marketing channel has the same potential across different verticals?
00:42:11So if we think about beauty, if we think about tech products, if we think about B to B SaaS, if we think about your example, the doctor platform, so medical topics.
00:42:25Would you say there are differences in how important or how impactful AEO can be across these verticals?
00:42:34Definitely.
00:42:34I mean, the relative size of these is where the impact is.
00:42:37You mentioned a few specific examples.
00:42:39Also, in looking at some Reddit data, it looks like the kinds of things people are discussing on Reddit are, LLMs are especially useful for things like analysis, coding.
00:42:50Interestingly, mental health is another area where people are using LLMs.
00:42:55I'm trying to dig into that more.
00:42:57Help with education and learning, help with research, ideas, brainstorming.
00:43:02I think that LLMs are very good.
00:43:03We talked about a sea of sameness where LLMs... our derivatives of existing information.
00:43:09It's actually great at suggesting ideas for information gain.
00:43:13Like if you said, what are some non-obvious questions about website builders?
00:43:17It would actually give you great, or like, what kinds of people aren't using website, no-code website builders, and they should be.
00:43:26It'll give you really good ideas for that.
00:43:28So it's actually great at brainstorming, coming up with new ideas.
00:43:32Has your own research behavior shifted like from Google to... chpd claude or some other tool?
00:43:42i use llm's for research for summarization actually for some of my research it's great for trying to see if they're academic journal articles about the thing that i'm exploring.
00:43:55it was quite painful to look for.
00:43:57there's millions of probably millions of academic journal articles.
00:44:00there's no way i can search through all these but it's actually great at saying has there been any prior research on uh ai content?
00:44:07not Not really.
00:44:08Okay, cool.
00:44:09If there are, here's what they said.
00:44:11I'm looking at evaluating the effectiveness of AI detectors.
00:44:18And there's actually very few studies on this, and the sample sizes are really low.
00:44:22But I created a table of all the different studies about the evaluation of the effectiveness of AI detectors and the sample sizes and where the samples came from.
00:44:35taken hours to do that in Google search and now I can do that in a few seconds in lm.
00:44:40so these are the some of the some of the things that I use for.
00:44:42also I'll say I went to New York and there are all these different things to do in New York.
00:44:46they're these different shows.
00:44:48and so then I asked reddit answers is this event good?
00:44:52and I went to masquerade which is this phantom of the opera immersive theater and the reddit answer said yes this is a great event.
00:44:58and then there was another one where it was like a light show.
00:45:01And I said, what do you think, what does Reddit think about this?
00:45:03And they said, well, people say it's overpriced and it's kind of lame.
00:45:06So whereas in Google, I wouldn't get that.
00:45:08So Reddit is a great source of information for things to do.
00:45:11Similar, I went to Hawaii and then I said, well, what do people say you should do in Kauai?
00:45:15Everyone says, do the helicopter and make sure you don't have the doors on.
00:45:19So I did that, had a great time.
00:45:21But yeah, these are some of the ways that I use LMS.
00:45:25Okay, awesome.
00:45:27We touched already a little bit on the results that you have gotten already from AEO, especially for example from Webflow.
00:45:35So I'd like to focus on the topic of attribution a little bit because I feel like people are still having a hard time seeing showing up in AI answers being connected to driving real business results in terms of signups, in terms of pipeline.
00:45:51contribution.
00:45:52So from your perspective and your experience, what's the issue with attribution in LLMs?
00:46:00The attribution issue is that most of the answers are no click.
00:46:05So if you say what's the best payroll management software, you probably aren't going to click on Rippling.
00:46:10You're probably going to open a new tab and you'll either search for Rippling in Google and then click on their homepage or their ad.
00:46:17Their their ad where they're bidding.
00:46:18Oh, they have to bid on their own brand term.
00:46:20or they'll open a new tab and type in rippling.com.
00:46:23So either you're attributing it to direct branded search or branded paid search.
00:46:30But it actually came because the answer said rippling is great and And this is really hard to track now.
00:46:35Sometimes there's something to click on but most of the time there's not.
00:46:37so the majority of cases where you actually got a conversion There's not something that's traceable and clickable.
00:46:43So that's that's the main issue.
00:46:44the second issue is is that's similar to SEO, where there's a user journey for rippling, you probably heard of them a hundred times.
00:46:54So there's so many touch points.
00:46:55So if you just looked at the last thing that they did before they converted, you're missing the previous ninety-nine things that happened.
00:47:02And even though you can trace the click, you still don't know what happened.
00:47:07You have a very skewed view of what happened right before that.
00:47:11And so that's why you need to do things like multi-touch, or mix media modeling, or just ask them, how did you hear about us?
00:47:18That's search.
00:47:19Now, answer engine optimization gets even more messy, because you don't actually know the volume of the prompt.
00:47:24So like you could track, I'm showing up for this prompt.
00:47:26You don't know how many people are looking for that.
00:47:29You have to ask multiple times to see where you appear.
00:47:33There's no click through rate.
00:47:34So who knows if somebody, if you were at the top or the bottom, like, did somebody see you?
00:47:39We don't know.
00:47:40So there's all these compounding pieces of compounding errors on top of the fact that it's not traceable.
00:47:47So that's why it's hard.
00:47:49So my suggestion is focus on the beginning and the end.
00:47:54The beginning would be, did I appear for these prompts?
00:47:57Yes, no.
00:47:58And was I ranked high?
00:47:59And just make a guess about what the volume might be, probably based on search data.
00:48:05And then for conversions, ask the person, where did you come from?
00:48:10and they'll tell you.
00:48:11Now, there's issues with self-reporting, clearly.
00:48:14Who knows what the weight of that... They said it was from this one source, but we know that it was from many different sources.
00:48:20What were the weight of all those sources?
00:48:21We don't know.
00:48:22But that's the closest you can get, I think, is, did I appear for prompts with assumed volume with large error rate?
00:48:30And then, what did people tell me after they converted?
00:48:34Very messy.
00:48:34Yeah,
00:48:36already sounded a little messy.
00:48:41How did you solve the attribution at Webflow or what you can maybe share about the case with Webflow where you were able to actually attribute a double in signups from LLMs?
00:48:54I don't have a full answer, but that is a good question.
00:48:58But I kind of know, but I don't have the full picture.
00:49:02Okay, no worries.
00:49:04But I know that their data science team spent time on it.
00:49:07Yeah, probably they're doing a good job there.
00:49:10And this is why I was generally curious to understand it in more detail.
00:49:14But I could also totally see why Webflow and you would not be comfortable sharing it publicly.
00:49:22But let's talk about another aspect of the whole attribution thing.
00:49:27You already mentioned that most of the searches or the prompt entries are zero click.
00:49:35Can you see and also then other LLMs or AI chat, but go to showing more links over time?
00:49:44Or do you think like we are now locked in with the amount of links and the probably very low click-through rates we are at currently?
00:49:55No, I think it'll become way more clickable.
00:49:57And I think that LLMs will start becoming more and more similar to search.
00:50:02And we are already seeing that.
00:50:04So if you ask something like where should I go in Hawaii?
00:50:08or what's the best TV?
00:50:09You will see clickable shopping cards just like you see clickable shopping cards in search and it's useful.
00:50:15because I want to be able to.
00:50:17I don't want to have to open a new tab and Go find rippling and click on their brand search ad on Google like.
00:50:23that's not the ideal experience.
00:50:24The ideal experience is just click on rippling.
00:50:27So I think that it's very likely that things will become way more clickable.
00:50:31and how Google Looked at this is in the early days, they had the Ten Blue Links and they added maps and shopping and travel and events and flight booking.
00:50:44So I think the LMS will just go one by one and start having vertical specific experiences.
00:50:50And as a user, I want to click on stuff.
00:50:52It's better if I can just click on something and go somewhere rather than having to open a new tab.
00:50:59I do also think that there will be autonomous Asians where maybe I don't need to click on something to actually buy or book.
00:51:07We're already starting to see that with, you know, I think ChatGPD wants to do native stuff within ChatGPD so you don't have to leave to buy a product or to book something.
00:51:17So I think that we might also see that.
00:51:18So it wouldn't be zero click, but it would be click within LM.
00:51:24And how closely do you watch AI mode?
00:51:27Because, I mean, now obviously you're in the US, we're in Europe.
00:51:32And AI mode is now also available, but it has been rolled out significantly with a significant delay also due to the EO regulation obviously.
00:51:41But how important do you see Google's AI mode already being alongside chat GPT or maybe also the more AI native interfaces like perplexity?
00:51:54I think that When Google enters a new space, they have multiple teams working on the exact same thing with multiple solutions that all do something kind of similar, and then they merge them.
00:52:06And we've seen that multiple times.
00:52:09So I think what will happen is AI overviews, AI mode, and Gemini will eventually all merge because it's essentially doing the same thing.
00:52:15You don't need three different products that all do the same thing.
00:52:17So I think that they have multiple teams trying, you know, multiple really good teams trying this out.
00:52:22and figuring things out, and then they'll merge them, and they'll be a single unified experience.
00:52:27Now, if you add AI Overviews, AI Mode, and Gemini, that's a really big footprint.
00:52:33And so I am generally looking at that, but I expect that it'll be a converged unison of all three of these.
00:52:43And then if you look at the data, ChatGBT and Gemini, Gemini or sorry Google Gemini AI mode a ovaries is not that far off from chat GPT and then everyone else is a destined distance.
00:52:53second so those two open AI and and Google are by far the largest market share.
00:52:59and do you think like Open AI or chat to PT will be able to steal market share?
00:53:06like let's maybe think one two years in the future from Google?
00:53:09or do you think Google will be able to either maintain the status they still have or maybe even grow market share again because AI because of this whole Bringing together of AI mode AI of ruse in Germany.
00:53:26You just outlined
00:53:29So there's the surfaces and then there's the companies.
00:53:32So do I think that search will go down?
00:53:35I don't think that search will go down, but I think that search plus LLM usage will increase.
00:53:40so search will just stay flat and then LLM usage will be on top of that incremental.
00:53:46then The question is, the two companies, and will OpenAI take market share from Google?
00:53:54I could definitely see that happening because Google has such a large market share for search.
00:53:57I think it's ninety-five percent, depending on what you look at.
00:54:01So it's hard to maintain ninety-five percent market share forever.
00:54:04I would expect that that eventually would move over to someone that could be OpenAI, that could be a combination of others, that could be Bing.
00:54:12But I wouldn't expect that anyone in any category would ever have a ninety-five percent market share permanently.
00:54:18I would expect eventually something that high would go down.
00:54:22I think Google had a little bit of an awakening moment when I think it was Robbie Stein joined the company again.
00:54:31And they were basically already declared that because Chajapiti won market share and won like.
00:54:41What was it?
00:54:42seven hundred eight hundred million weekly active users and now Google somehow comes back with AI reviews and also Gemini the new models Suddenly being very popular also in in app download charts.
00:54:56So have you been surprised by this?
00:54:58Come back or would you even agree that it's this?
00:55:01it is a comeback by Google?
00:55:03It is a comeback by Google.
00:55:06I'm not surprised but I'm impressed.
00:55:09So why I say that is because There's the innovators dilemma where big companies will eventually be disrupted by and this is a perfect example of a category that would lend itself to the innovators dilemma where New start new entrance small market big company cruise liner can't turn their boat quickly because they're not adaptive.
00:55:29tiny boat comes in steals market share.
00:55:31This is exactly what what the innovators dilemma would lend itself to and that has not happened.
00:55:37You know chat GPD has blown up but Google has has adapted and is growing quite well.
00:55:44And there are other companies who are not.
00:55:46Google is the only one that's actually catching up to, or even getting even close to catching up with OpenAI.
00:55:53I'll add that it's very hard.
00:55:54OpenAI was working on this for, I don't even know, ten years plus.
00:55:58It's very hard to build an LLM and have the talent to do that.
00:56:03So the fact that Google has been able to do that is quite impressive.
00:56:07So I believe that this is true.
00:56:10I am not surprised and I am very impressed.
00:56:13Great.
00:56:15Let's go to some community questions again because we got some really good questions beforehand from people that I definitely want to get your answer or at least your ideas on.
00:56:29The first one we got was from Someone working at a review platform, so for full context, review platform in Germany, so basically like Captura or G-II, it's called OMR Reviews, so like a software review platform.
00:56:44And the person asked, what do platforms like review platforms need in terms of content and positioning so that users still actively come to their sites and not just consume the information indirectly via AI or LLM answers?
00:57:01Well, what I would want, would be, you know, I mentioned Reddit answers.
00:57:05So Reddit answers is an LM summarizing what happened on Reddit.
00:57:09What I would want would be a summary, I mentioned SIRMO.
00:57:12SIRMO is summarized what the doctors are saying on SIRMO.
00:57:14I would want the review sites to have their own summarization.
00:57:18So, and I don't want to read thousands of reviews.
00:57:21I can't, I don't have time for that.
00:57:23What I want to know is generally, what are people saying?
00:57:25So.
00:57:25what I would really like for a G-II or I think I forget the site that you mentioned is I would want an AI summarization of other reviews.
00:57:32That would be super useful.
00:57:33And then maybe also the review, the review flow can help guide me to add stuff to my answer to make it even more useful.
00:57:40Like Airbnb does a good job about asking about the cleanliness and the responsiveness of the host.
00:57:45So for a general review site about B to B products, have lots of reviews, encourage.
00:57:53part of the reviews to answer the questions that people also have and some diversity of opinion, and then have your own AI summarization within your platform, so I don't need to go somewhere else.
00:58:02Okay,
00:58:04great answer.
00:58:06Another question we got was, do you think that what you shared on Lenny's podcast has aged well, or would you change or add anything?
00:58:18I think it's aged pretty well.
00:58:20Yeah, I can't think of... I mean, would I add anything?
00:58:22I would add this stuff that I... that have been sharing posts, Lenny.
00:58:26I mean, I have more insights today than I did in September when I filmed it.
00:58:32I wouldn't take back anything.
00:58:36So I asked, after Lenny, I got a bunch of people adding me on LinkedIn and saying, you had a great podcast.
00:58:45And then I asked follow up questions to each of the people who said that the podcast was good.
00:58:52And I asked them, I would love your feedback.
00:58:55Is there anything that you disagree with?
00:58:58And what would you, what did you find most useful?
00:59:01And is there anything that you would like me to discuss more?
00:59:04And then I got about fifty answers.
00:59:07And then I did an AI summarization of the feedback that I got about the Lenny's podcast.
00:59:12The number one request was help center optimization content.
00:59:17It's like, tell me how to optimize my help center.
00:59:20And I'm still early in that.
00:59:22I haven't mastered that yet, but I think what I would want to add would be, I mean, what I would want to add would be actual real case studies, so the more case studies the better, which I didn't, you know, I shared what I had then and I have more now.
00:59:33And then the second would be some more specific tactical stuff around help center optimization and product content, content around describing how your product works and stuff like that.
00:59:43Can you share at least some of your early additional thoughts on help center optimization?
00:59:48because I also saw help center content actually show up in citations a couple of times now with clients we're working with when we talk about very specific also like integration related prompts and like.
01:00:03we're also trying to explore this.
01:00:05but yeah obviously very selfishly I'd like to get your insights from it.
01:00:11Yes, so step one is What are people asking about?
01:00:14And we talked about how to do that, but I think customer support, sales calls, Reddit answers, what are people asking about my product G-II?
01:00:24This will give you information about what product questions people have.
01:00:29And then step two is answering the question.
01:00:31I don't think that there's anything not obvious about this.
01:00:35If somebody says what meeting note transcription tool integrates with Zoom, have a page about that.
01:00:42and say that you integrate with Zoom and explain how you do that, that's really all you need to do.
01:00:45Otter shows up for that particular example and they show up because they have a page saying, we have an integration with Zoom, here's how it works.
01:00:52The non-obvious thing with that particular example is have multiple pages describing this.
01:00:58So if you ask that question, meaning out transcription product integrates with Zoom, there's a help center page on Otter, there's a feature page.
01:01:08like otter slash zoom or something like that.
01:01:11There's an article about it, like how to use it.
01:01:15Then there's landing pages on zoom about the otter integration.
01:01:20So there's all these different pages.
01:01:22And so have multiple takes at answering this particular question.
01:01:27And especially if I have a page on the integration partner site.
01:01:31So let's say that I'm a brand new meeting note transcription tool.
01:01:36And I integrate with Zoom, and no one's heard of me.
01:01:41Assuming your friends with Zoom, ask them to make a page about your integration.
01:01:46All your integration partners, have them have a page about your amazing integration, and then you'll show up twice.
01:01:53That's a good one.
01:01:54That makes a lot of sense.
01:01:56We have another question.
01:01:58We maybe covered a little bit of that, but I still want to be sure to ask it.
01:02:04It's where and why do you see most AI content workflows fail.
01:02:12It's because you're just prompting ChatGPT to derive information from the public and nothing more.
01:02:19It's just a wrapper around ChatGPT.
01:02:21I think that a better version of that is.
01:02:24you have something like the apartments in City Example.
01:02:29People want to know about the price and the features and the areas to get an apartment and the pros and cons.
01:02:34Okay, now I have unique metadata for each of those four different questions.
01:02:38Then I have a workflow and a prompt saying summarize the cost, you know, the information about the price and include this, this and this.
01:02:47That's what they're lacking.
01:02:49It's like somebody being thoughtful about configuring the prompt and feeding in unique information to get an output that is more useful than just a derivative of what's already available to everyone.
01:03:01Got it.
01:03:02This is a community question for myself.
01:03:05Do you think more people should get a super me?
01:03:09Yes.
01:03:10Well, if you have more people should use SuperMe, for sure.
01:03:17And I'll explain what SuperMe is.
01:03:19So SuperMe is my friend's company where you can feed in your thought leadership and LinkedIn content and stuff like that.
01:03:27And then you can have a SuperEthan where it takes all of my webinars and my articles and my LinkedIn posts and things like that.
01:03:33And then you can ask AIEthan.
01:03:36What tools do you like?
01:03:37How would you approach building on an AEO strategy?
01:03:40And the answer is a summarization of my thought leadership.
01:03:42So it's not a derivative of public information.
01:03:45It's Ethan's specific thought leadership.
01:03:48So it's quite useful.
01:03:49It's only as useful as the input.
01:03:52So if you have a bunch of novel ideas that you've documented, then you should have a super me.
01:03:57If you haven't done that.
01:03:59No judgment probably no reason for you to have a super me.
01:04:02So that's not thinking about it, but but every everyone should be using it.
01:04:05I I looked at.
01:04:07I have had over five hundred questions to super Ethan and The answers are pretty good.
01:04:14Yeah It looked very interesting.
01:04:17Probably you have to do interesting stuff and have interesting things to say so that you qualify to having a super me.
01:04:25I actually have two more questions.
01:04:28one is One actually came from your colleague Emily.
01:04:35So don't blame me please.
01:04:37She wanted to know what's your favorite Backstreet Boys song?
01:04:41Okay, the answer to this is my favorite, I have to say in.
01:04:45sync is my favorite over Backstreet Boys.
01:04:48And my favorite song is Tearing Up My Heart.
01:04:52And my favorite performance is the, I think it's the, I think it's the, I think it's the, with Britney Spears and in sync, tearing up my heart.
01:05:03That's one of the best performances of all time.
01:05:05Okay, so then, but the answer to Backstreet Boys is, I would say, Shape of My Heart is my favorite Backstreet Boys song.
01:05:12I very much want to see them at the Sphere in Las Vegas, so I hope that they continue their residency so that I can see them live.
01:05:20Okay, and in this case, you wouldn't risk asking Reddit answers if it's good and because if it says it's not good.
01:05:28you don't want to be spoiled because you definitely want to see it.
01:05:34I've heard good things.
01:05:35I think I checked Reddit answers and I heard good things.
01:05:38One of my biggest regrets is I didn't get to see Brittany at her residency in Las Vegas and I've never seen her perform, so I can't make that mistake again and I have to see the backseat boys.
01:05:49Okay, thanks.
01:05:50Then the last one, and I think actually just bluntly stole this from Lenny, but I also started asking it on my podcast and the answers were so great that I just wanted to also have it in here.
01:06:04What's something that we didn't talk about, but should have talked about?
01:06:12Where to get information on, like who to read, where to get information about the subject.
01:06:20What I would say, because there's not that much information.
01:06:25There's ideas, there's this conversation, but there's a lack of first-party experiment data.
01:06:35What would you recommend people to go to, or where do you yourself get your information from?
01:06:43I like Kevin and Dick does good stuff.
01:06:46Lily Ray does good stuff.
01:06:49Both Ahrefs and S.C.M.
01:06:50Rush and Surfer all do pretty good.
01:06:53thought leadership and studies.
01:06:55And there's probably others.
01:07:00I have a full-time job, so I'm not spending all day long reading other people's stuff.
01:07:03So there's probably many people that I have not read that are doing interesting things, but those are people that I personally read that I find useful.
01:07:12Awesome.
01:07:14Ethan, this has been a very insightful conversation.
01:07:18I feel like it's a very decent follow-up to the Lenny's episode.
01:07:23Thanks so much for sharing.
01:07:26So much, also such in-depth thoughts, maybe also some raw thoughts.
01:07:31I really appreciate it.
01:07:34I hope also that everybody listening and viewing had a great time.
01:07:38I think it's obvious if you want to learn more about what Ethan is doing, either follow him on LinkedIn or go to SuperMe.
01:07:45I think it's superme.ai.
01:07:51So like Ethan, but just with the E-Smith and ask questions to SuperMe, Ethan, before you ask Ethan, because he has a full-time job, so please respect it, people.
01:08:02But yeah, Ethan, thanks so much for coming on.
01:08:05It was a real pleasure.
01:08:09Thank you for having me.

The Podcast

The Conference

The Index

The ultimate guide to AEO (Director's Cut) | Ethan Smith, CEO @ Graphite

Show notes

Full conversation