From prompt analysis and AI-driven content strategy to platform-specific publishing cadences and the surprising power of YouTube in LLMs, this episode is a must-listen for digital marketers and SEOs who want to stay ahead of the curve.
https://page2pod.com - In this episode, Jon Clark and Joe DeVita engage with Metehan Yesilyurt to explore the evolving landscape of SEO in the age of AI. They discuss the transition from traditional SEO metrics to new paradigms influenced by AI technologies, including the concept of Rank Fusion and the importance of brand visibility.Β
The conversation delves into practical strategies for optimizing content for AI search engines, the role of platforms like YouTube, and the future of e-commerce with AI integration. The episode concludes with predictions for the dominance of AI search engines by 2027.
Metehan dives deep into the seismic shifts in how search engines are evolving, including how visibility works in Large Language Models, the mechanics of Reciprocal Rank Fusion (RRF), and why the classic SEO metric of βrankβ is rapidly becoming outdated.
From prompt analysis and AI-driven content strategy to platform-specific publishing cadences and the surprising power of YouTube in LLMs, this episode is a must-listen for digital marketers and SEOs who want to stay ahead of the curve.
π‘ In This Episode (π§ ):
β’ Why Metehan left traditional SEO for AI-driven discovery
β’ The founding story and mission behind AEO Vision
β’ How LLMs like ChatGPT and Perplexity βrankβ content differently
β’ What Reciprocal Rank Fusion (RRF) is and how it impacts visibility
β’ Why brand mentions matter more than backlinks in AI search
β’ The role of schema, structured content, and semantic chunking
β’ Why you shouldnβt optimize solely for LLMs (and how it hurts SEO)
β’ How YouTube is a powerful but underrated channel for AI visibility
β’ Prompt analysis: what real user questions reveal about content strategy
β’ Predictions on who will dominate AI search by 2027
If you're looking to understand the next wave of search and how to optimize for it, this episode will arm you with the knowledge you need.
π Subscribe to stay ahead in SEO, AI, and digital marketing trends every week!
π¬ What are you doing to optimize for AI and LLM search? Drop your thoughts in the comments below!
π Mentioned in the Episode:
β’ Follow Metehan on LinkedIn: https://www.linkedin.com/in/metehanyesilyurt/
β’ AEO Vision: Β https://aeovision.ai
β’ AI Share Buttons: https://metehan.ai/blog/citemet-ai-share-buttons-growth-hack-for-llms/
β’ Reciprocal Rank Fusion (RRF)
π https://plg.uwaterloo.ca/~gvcormac/cormacksigir09-rrf.pdf
π https://metehan.ai/blog/chatgpt-is-using-reciprocal-rank-fusion-rrf/
Jon Clark (00:00)
In a world of AI answers, are we optimizing for machines or teaching them what to say? Metehan YeΕilyurt is the Co-founder and Chief Growth Officer at AEO Vision, a company trying to reverse engineer the black box that is large language model visibility. He spent more than a decade in digital marketing, from App Samurai to Media in Turkey. And now he's all in on decoding how generative AI reshapes search. This episode dives deep into the evolving mechanics of visibility and an AI first internet.
We explore how platforms like ChatGPT, Perplexity, and Claude select citations, why rank in this context is more probabilistic than positional, and what SEOs need to rethink about content structure, brand citations, and synchronized publishing. Metehan breaks down concepts like reciprocal rank fusion, embedding similarity thresholds, and even how YouTube and schema might influence LLM behavior. The real story is optimization hasn't died.
It's just changed form and the rules are being rewritten in real time. Metehan is one of the leading voices in how LLMs crowd and respond to prompts. This episode will leave you with an aha moment from start to finish. Here we go.
Jon Clark (01:14)
Welcome to another exciting episode of the Page2Podcast, episode 99. We are almost to the hundred mark. As always, I'm your host, Jon Clark, joined by my partner in crime and Moving Traffic Media, Joe DeVita. And today we're excited welcome Metehan YeΕilyurt to the show. Super excited to talk about AI discovery on ChatGPT rankings, Perplexity, all those sorts of things.
Joe DeVita (01:25)
Hi β
Jon Clark (01:37)
But maybe to start, you recently co-founded AEO Vision. You were at App Samurai for about five years before that. And I was what prompted you to leave App Samurai after such a long tenure and start something in the SaaS space?
Metehan (01:55)
have experience more than 10 years in digital marketing industry as a professional and I started with websites when I was younger. I'm 32, by the way, the time just flies, you know. I worked at a digital marketing agency, inbound marketing agency, then jumped in a...
large news media outlets in Turkey. just during the pandemic, I would like to jump into the SaaS space, especially the mobile apps were rising. They are still rising, but everyone houses. So I just went mobile and web. And I just want to learn more about mobile, the new hype. And after successful and I'm grateful to App Samurai team, we had a great five years. Then, you know, the things just went crazy. I mean the AI hype. We were targeting the US market at App Samurai and then realized end of 24, I noticed something weird on our search console and analytics and everyone was talking about the new zero click world. Then when I was stuck digging, I realized somethings were changing. I mean, I still live in Turkey and I need to change my connection with VPN to see US results, then I realized the all AI things, AI overviews and other AI related features rolled out in Google. Then I was actually, this is an honest review from me, I was very boring with doing SEO during the 23, especially because everyone was talking
and advising you need to optimize your web page speed. Yes. And so what? You need to create the high engaging content. Yes. Yes. But how? Everyone was saying and spreading same narrative. So when I met with AI during the 23,
especially ending of the year, then I, I felt very, I was very excited. Then, β you know, every digital marketer, I believe we have a passion that if I knew how to code, could, I can create billions of dollars market. Just I need to learn how to code. And with this,
AI features, new models from large-language models. So I was really excited and I just jumped into SEO world again with my old passion. Then I started working on AI overviews, search generative experience during those times.
Then I jumped in the AI world. Then I realized I'm very excited to work with AI, using AI to learn more stuff. Then start to reverse engineering on search engines. Then you know, ChatGPT just came out.
Jon Clark (05:05)
Great.
Metehan (05:07)
new disruptive product in the market then I felt like wow this is this is just amazing and everyone will everyone went crazy again then start to work heavily on AI search stuff and including SEO. By the way I'm not a new type of GEO guys I love using GEO and AO I mostly prefer AO you know
But I still believe in SEO because if you complete the fundamentals of SEO, can succeed in AI search. And I believe the main difference is between rack and hybrid search. I mean, the traditional search engines like Google, Bing, all others. Then I realized that I'm obsessed with AI, working on reverse engineering. So I decided to jump on just the
Jon Clark (05:32)
Okay.
Metehan (06:01)
AI trade. Then I met my current partners in the company, AEO Vision, and we founded AEO Vision. And right now, thanks to God, my only job is focusing on these AI things, AI stuff, β optimization, talking with our customers, playing with dashboard features, and now we are here.
Jon Clark (06:23)
Amazing. I was looking at the, I was digging through the site, It's a aeovision.ai. And one of the things I noticed was that you talk a lot about, you know, measure your visibility, you know, define your presence, but the word rank, which is so often attributed with traditional SEO,
β you know, wasn't prevalent on the site. And it made me start thinking about a question that we really haven't asked anyone yet, which is how would you define rank in an LLM?, right? Like, like maybe help our listeners sort of get their arms or maybe mind around what rank is in an LLM, because it's not really rank, right? It's something totally different.
Metehan (07:08)
Yeah, we are still chasing 10 blue links in Google. And I believe traditional search engines is a new narrative that we're used to call it. I love still Google and I believe Google is the best machine learning company in the world because they have their own servers,
engineers, and they have a 25 years index advantage than other any LLMs, especially if they're using web grounding feature. If I need to define rank, first it depends.
It depends on your question. You can ask same questions a few minutes later in any LLM. The context doesn't change, but you can see citations are changing or some words or verbs.
Jon Clark (07:53)
All right.
Metehan (07:59)
if the answer is coming from β offline training data, like you can ask a recipe, then you just see the old question. But if you ask the latest AI news, you can see they are connecting to web scrape from search, traditional search engines, and you can see citations. But it's very dynamic and I believe...
from my perspective and of course this is in the nature of LLMs, they have a temperature settings. So we don't play with these settings but if we meet all together tomorrow and talk about the same topics, our wordings can change,
but context will stay the same. So LLMs are trying to mimic the natural human conversation. So, and they have some layers like re-ranking and they have some, let's say, backend settings like temperature, creative. So it changes all the time, but...
If you have a great content with a great structure on your webpage, you can rank in LLMs, but there are now hundreds of prompt tracking tools at the moment. And we are sending prompts with using APIs without any personalization. Unfortunately, this is the nature of our market at the moment.
And we are fetching and saving in a let's say daily or hourly basis just from one window only but there are hundreds of them so ranking is also dynamic but it's important to inject or include your brand every time even you're not in citations
I believe this is the current situation and we will figure it out for a more stable ranking, I believe.
Joe DeVita (09:59)
I think it's, it sounds like you'd prefer we just stop talking about rank as we move toward, you know, this whole new system to optimize for. Maybe, we don't set ourselves up for success if we try to carry over that same KPI. You gave a very eloquent and long-winded answer, is, would you, is your advice just to
Jon Clark (10:05)
.
Joe DeVita (10:20)
you have to move on. This isn't the same old optimization. This isn't the same We don't measure success with rank anymore. We should stop talking about it.
Metehan (10:24)
Yeah, yeah, yeah.
No, no, no, no. I believe you need to create your own media chance at the moment, your own content distribution, because I was just using keyword stuffing in, I guess, 2008 when I was in high school. Yeah, good old days. You can...
Jon Clark (10:46)
The good old days.
Metehan (10:49)
use keyword stuffing everywhere, the one pixel content parts, etc. But then social media came out and I read many news in the past like SEO is dead so social is the new channel but SEO isn't dead, it's just evolving and we need to rank and
Jon Clark (11:03)
.
Metehan (11:10)
I believe the brand visibility is the most important factor at the moment. You need to be visible. If your audience is using TikTok, you need to be in there. If your audience is using Substack, you need to be in there because they're
active and they joined in a community and you need to keep β that community and distribute your content effectively. So
Jon Clark (11:26)
Okay.
Metehan (11:39)
rank is important, but I believe at this moment, the brand visibility and impressions, even we couldn't find a chance to get impressions from LLMs at the moment. I mean,
let's think like an OpenAI search console, but visibility is important right now. And ranking is too.
Jon Clark (12:03)
So.
Metehan (12:05)
of course, but it's dynamic.
Jon Clark (12:07)
So I think the emphasis on brand citations maybe leads us nicely into the Reciprocal Rank Fusion or RRF, which came out of some of your discoveries around how ChatGPT sort of determines an answer to a search. So maybe just for the listeners, you take us through what that discovery was, maybe a little bit of background around
which is really difficult for me to say for some reason. And we'll start there and then sort of dig into some of the findings that you had.
Metehan (12:37)
Yeah, yeah, let's call it rank fusion and it's a concept. I found original academic paper from 2009. from my perspective, and these are the industry general information, LLMs are now requiring too much computational power.
And you need to use, if you're an LLM engineer or AI company, you need to find the best methods that can benefit your service. You need to save money. So you can use effective and you need to consider to satisfy your users, your audience. And we need to get great answer quality. So β Rank Fusion is a concept and mathematical formula that if I ask latest AI news, if I probe and just send it to ChatGPT, and you can actually test yourself with using AIMA ChatGPT plugin and there are some many others. You can see ChatGPT actually returns around
more than 40 results, but we see only 10, 12 to 20 results in citations, almost in general, I mean. So there should be a reducing mathematical formula or concept. So we see only the final citations,
citation results. So Rank Fusion, if you ask a prompt, ask a question to LLMs, LLMs are fetching from search engines and retrieve the results and they are using Rank Fusion to reduce scoring these results and showing the final top candidates in citations.
So when I dig in the source code, I realized that and also Microsoft is using this and it's well documented in Azure website for developers. So Rank Fusion is basically you ask and LLMs are using a very similar approach to fan outs in AI mode at the moment.
So they, let's say they are using five subqueries at the moment. So at least they need to retrieve 50 results. Then they score it and show us the final citations up to 20. Right, this is the concept.
Jon Clark (15:00)
Got it.
So I think one of your takeaways was, or I guess one of the reasons why you want a more comprehensive keyword universe is because within those 30 to 40 results, you want to be mentioned more often. And therefore, your RRF score will be higher than, someone who only has one piece of content within
those collective results. Is that a fair way to summarize what you're thinking here?
Metehan (15:30)
Yeah.
Jon Clark (15:30)
So I guess one of my questions was, have you started using any of those insights for keyword research? Has it changed how you do keyword research or have you just sort of leaned into, I'm just going to write around what maybe my customer wants and hopefully that aligns to β prompts that folks are putting into these engines.
Metehan (15:52)
Yes, actually I'm also a user in ChatGPT and other LLMs and our world is just changing. It's the digital direction to asking questions. And in the old great times, if we are looking for, let's say a coffee machine under $100, we need to chase the 10 blue links.
as a buyer, then read all links, summarize it using our own brain, then read the old reviews on Amazon, watch videos on YouTube maybe or Facebook, then make the final decision. But right now we are just asking same questions with using some, let's say, maybe filters or specifications, and I'm just
fetch the final product and β show, let's say three candidates to buy. So right now we are using, we are asking more questions and it's obvious Google wants to go in that direction. So questions are more important and
For questions, I'm using actually, we are also using everyone. It's very popular, people also ask questions. So if you look in real questions from users, it's just more useful at the moment, I believe.
Jon Clark (17:14)
And so I definitely want to dig into how you're thinking about optimizing content, but in more detail. So when you're sort of building out your page, are you structuring it sort of like content or sorry, question, answer, question, answer? Yeah. I wanted to dig into the comments you had around like brand citations and
you talk a little bit about like synchronized publishing. So you have your sort of your, your owned property, like let's call it your website, but then you also have all these social platforms like Reddit and YouTube that you just mentioned. I was curious, like what, or how do you think about the cadence of that distribution? Like, is it all at once? And so you're trying to get all of those things working together.
to influence as fast as possible or is it something you sort of drip out over time in a way to influence the rank? How are you guys thinking about that?
Metehan (18:10)
I believe if you're a solopreneur you can push same content just every channel you own or you manage but I found this a little bit spamming sometimes I'm just posting same content but try to update it a little bit I believe this is you need to be visible almost on
every channel where your audience live. Unfortunately, by the way, because it's very time consuming and you need to think the platform dynamics. You can post on LinkedIn, post, let's say cross posting to TikTok and LinkedIn. It's not possible and useful, you know. So if you create content that aligns with the
platform audience, it will help much better for your brand visibility. And actually you are talking your audience language. You know, Gen Z is just talking in a different way. So I believe you need to create unique posts for each platform, but it's time consuming and I'm a very lazy person, by the way.
Jon Clark (19:21)
Haha
Joe DeVita (19:22)
I,
because you want engagement. Like you don't, if you, if you have a great post in LinkedIn and you get some commentary, you can't use that same content for TikTok because the audience will disregard it. β so there's no, there's no real advantage to your AI strategy. You're at your answer engine optimization strategy. If you, if a post to TikTok creates no engagement, am I right?
Metehan (19:37)
Yeah.
Yes, yes, I'm not saying by the way you need to be visible on every social platform, I mean, you can pick some β and focus on it. Let's say Reddit, let's say TikTok. If we read the reports from our lovely industry for the citations, market share, can see Reddit was
Jon Clark (19:54)
Okay. you
Metehan (20:11)
on the top β and Quora and all other.
Jon Clark (20:12)
Right.
Metehan (20:16)
forums, Wikipedia
Jon Clark (20:16)
you
Metehan (20:17)
and all others. I guess TikTok is not in the leadership at the moment, but it will arise. And I believe that is very useful because you can, the audience is just using the comment section, very crazy. They're asking a lot of great questions. And I want to tie this topic to your previous question.
I'm still using keyword research from tools like HREF, SEMrush, Google Trends, and all other platforms. But I believe if you can access and see the real user's questions with real data without creating a synthesized questions from LLMs, because you can ask what can be the possible questions to LLMs. But rather than that, you can...
if you can β see and mind those real questions, it will be more helpful. So it's not a requirement. You need to be active on every social platform, but it helps and it helps your brand visibility and β trust.
Jon Clark (21:20)
I think one of the first blog posts I came across that you'd written was around
I think it was an extension of your discoveries on ChatGPT and it was around like AI share buttons. So sort of feeding prompts with sharing the content. And then since then I've seen, you know, New York Times and other publishers use like a summarize with AI button and things like that. are those good implementations to increase citations? Like, have you tested that at all? And, seeing that.
Metehan (21:29)
Yes.
Jon Clark (21:47)
β be as impactful as synchronized publishing across other platforms.
Metehan (21:52)
Yes, I published these methods. It works, but it just β won't drive millions of traffic. It's not working for the LLM's global training. It only tries to... target is not the correct keyword, but it's trying to align with...
personal user memory because sometimes I just paste the full link and want to summarize it and just save it to my notes because I don't want to miss any great content and AI is very helpful then I realized that after a while
Jon Clark (22:21)
you
Metehan (22:32)
ChatGPT started to reference those links from the past chats. So I found this very impressive because also ChatGPT is improving their own system. then I just want to test it out. And it worked. It won't drive tens of thousands of traffic, but I can still see the improvement and the increasing traffic from ChatGPT. Of course, there are some many factors and there are many considerations. You can use AI share buttons everywhere in every industry because interrupting the current session is not good because you're sending the user β in another tab and you don't know if they come back to your website again, but it's like retargeting in the paid ads. So you can build your audience with using AI share buttons and you can see the effect. As I mentioned, it won't drive millions of traffic to your website, but if your audience is good enough and maybe you are driving, let's say, three million clicks from your website, it will work much better. But I love this concept.
Jon Clark (23:49)
Yeah, I mean, it's a smart way to sort of seed your content within the user's memory, like you said, to potentially be recalled later. So it may not drive immediate traffic, but you give yourself the improved option.
Metehan (23:56)
Yeah.
Jon Clark (24:03)
for that content to be served up in a future prompt. Yeah, that makes a lot of sense.
Metehan (24:06)
Yes, see
many, many examples, people just sharing their examples and testings on LinkedIn and X. I'm grateful for that. Some group started to use it in their newsletter actually, because there many external links and these are common β user experience and they just click to links. So they
had a great success. Of course, I can share the results without their permission and I don't think they will give it to me, but there are some many different cases at the moment.
Jon Clark (24:34)
Okay.
We had Duane Forrester on a few episodes ago and he talked about this idea of the verification layer, right? Like the LLMs need some sort of trust or level of verification in order to surface some of the content. And I think a lot of what you found from both ChatGPT and your Perplexity
β insights were around sort of this idea of like trusted domains and things like that. Do you have an opinion or any thoughts on how like traditional like SEO E-A-T factors like play a role in how LLM surface up content? Like is there correlations there?
Metehan (25:22)
I believe yes, but there are some differences because the traditional search engines also considers the backlink quality, the narrative anchor text, but LLMs don't follow these metrics as usual. So I believe if you use more open...
language, let's say, you define with more β semantic and relevant keywords, you can show your expertise authoritative to LLMs, I believe.
Jon Clark (25:50)
you
Joe DeVita (25:53)
Can you think of any optimizations you might make for better LLM visibility that would hurt your organic search strategy?
Metehan (26:03)
If you only optimize your website directly to LLMs, I believe your rankings can hurt. Let's say if you only consider BM25 or Lexical Match to optimize your content, or if you just use bullet points everywhere, it's...
It's not useful for the user experience and we know the UX is now a part of SEO right now. It just doesn't matter for them at the moment. So yes, it can hurt your rankings.
Jon Clark (26:34)
You.
Metehan (26:35)
If you only
optimize for AI search, and this is my perspective, because I still love SEO and I saw many patterns and I recommend on my posts or custom JavaScripts inputs or optimization tips, please do not optimize your website only for LLMs.
Jon Clark (26:55)
Yeah, I think one of the things that's interesting about, you know, LLM content optimization is you really truncate down a lot of the content into its most simple form, which is always, which is not always the way that
humans prefer to read. And you're significantly diminishing the word count on the page. I might get the sentence wrong, but you recommended something to the effect of, every 300 words, use anchors, reasoning chains, synthesis hooks, I believe you called it. Is that a good framework to balance
the need for writing for LLMs and for users?
Metehan (27:34)
Yes, I believe that we know that there are many new things and terms in the market right now, like semantic chunking. They're all fancy words. I love using them too, but not too much because if you use too much fancy words, you can look like a GEO guy. If you use that, GEO is the new narrative, like new... β
Jon Clark (27:56)
Right.
Metehan (27:58)
But it's obvious and we know LLMs has a limited context window. I mean, for the token size. I run many experiments on this. They started to guess the rest of the content. Let's say if you just upload 5,000 pages of PDF to LLMS, they won't process it all, but they will make a guess
Jon Clark (28:04)
Thank
Metehan (28:23)
for the ending or for the remaining pages they can't process or understand. So if you use a clear summary top of your page, let's say your blog pages, and everyone has different use cases at the moment, but it will be useful, yes, especially for their embedding models to understand entities, entity relations, semantic space.
Jon Clark (28:49)
So, that's interesting because if you have a really long article and you ask it to summarize it and it's beyond that token window, you may not get the full summary.
Metehan (29:01)
Yeah, yeah.
Jon Clark (29:02)
which is, which is sort of interesting. I bet many people don't necessarily know that. So maybe let's transition to Perplexity. I know the market share is a little bit smaller there, but I thought the findings around, you know, what Perplexity takes into account in terms of rankings was super interesting. And maybe to start, did you find any relationships between the sort of Perplexity findings and
or did they appear to be more different than you expected?
Metehan (29:31)
Yes, there are some different factors. I didn't have a chance to decode ChatGPT fully, unfortunately, but for Perplexity and ChatGPT, and yes, you're right for the market share at the moment. And I believe the bad part is no AGI yet. Everyone is using highly manually configured
search configurations, feature flags at the moment, and still Google is the best for the real-time results and rankings. But for Perplexity and ChatGPT, they are using almost the same content freshness window, and it's different in Perplexity Discover tab because they have a Discover tab just similar to Google Discover.
and they launch their agent browser right now, the content freshness, and they're using embedding models to understand text better, the relations, everyone is using their own embedding models, but, and they are using re-ranking formulations.
I believe very similar way, but it seems Perplexity is using most likely the Google results, I believe. But ChatGPT, OpenAI has an agreement with Microsoft. So we know they're also using Bing for the search results, but it seems hybrid. They are going between Google and Bing at the moment. But these are the similarities I noticed.
Jon Clark (30:59)
Good job.
Joe DeVita (30:59)
Freshness
Metehan (31:00)
Thank
Joe DeVita (31:00)
factor.
We don't talk about it enough. I want to ask a maybe unrelated question. Hopefully it is related. For many years, we have told smaller clients mostly, most clients, when you're optimizing for organic search, you just optimize for Google and you'll see some success with
Metehan (31:03)
Yeah.
Joe DeVita (31:20)
Microsoft's a handful of different foundational can't do that. It doesn't seem like you can do that. You can't optimize for ChatGPT and hope to see results with Perplexity or Claude or can you?
Metehan (31:35)
I believe they have all differences with minor changes for the re-ranking systems, especially for the re-ranking layer and they are using different embedded models. I mean, ChatGPT is different, Google is different, so they all have different perspectives, but staying in the same context window because they know who is the president of the United States at the moment. They can all find the
right answer at the moment, but they have some differences. I don't think they're major, but for the Perplexity, they don't have their own LLM. They are using other LLMs to deliver search results. But ChatGPT, they have their own offline training model. Now we are using GPT-5. So there are differences, but...
I believe you can be successful in every LLM, but also LLMs have different formulations. Even their minor changes, the possibilities are changing like millions of possibilities for every LLM. So you need to be more, let's say Perplexity is using really high embedding similarity threshold, but it seems
ChatGPT is using the most cheap computational models to fetch results and re-rank them in a new order. And for the Claude they're using, they're expert on the coding field right now, we see. And Gemini is using just basically Google. So it changes
all the time and I believe it's also dynamic.
Jon Clark (33:18)
One of the parameters that you discovered was enable_search_urls_based_dedupe. And I thought that was super interesting because the duplicate content issue for Google has been long and well documented, right? Like what is the primary source and
when do I serve that versus something else? How do I manage for that duplicate content from an index perspective? I was curious like how that impacts Perplexity's ranking, meaning are they filtering out all the search results that they're evaluating in sort of a fan out?
Or are they trying to determine what is that primary source based on citations or whatever else to serve that up instead? Does that make sense?
Metehan (34:07)
Yes, at some point the application filtering isn't just for the content. They want to diversify the results with different opinions and perspectives, just like...
Google, you can see a forum discussion, they can make a fan out search like only using site operator to fetch only from Reddit, include them in a dynamic order. So it's similar and it's changing, but it's dynamic too at the moment for the...
Metehan (34:40)
citations and topic similarities, page similarities. But at the end of the day, in the nature of LLMs, they want to show different results from different perspectives. And of course, they want to satisfy users. So they're using also the application filters.
Jon Clark (35:01)
I thought one of the most interesting things that you discovered was the relationship between YouTube and Perplexity and potentially the way to, again, I don't know if spam is the right word, but spin up content on YouTube quickly to potentially rank in Perplexity. Have you done any more tests around that since that discovery, maybe on the content side or?
Metehan (35:07)
β yeah.
Yes, you can imagine I run many experiments around the week. I went silent on my socials for a while. But you know, I didn't want to share all results because I don't want to change our direction to spamming all the...
Well, then right now Google became a spam filtering machine on the indexing part. You know, you can just set up a new website and expect, β OK, my millions of pages index in just an hour. So but YouTube is working and it's related with the system prompts of LLMs at the moment. And they
also have some configurations for their agentic web browser and it also uses, just check my video history, find relative videos and remind me something like that for the system prompt. And I believe YouTube is the most underrated channel at the moment for LLMs because Google
β Google just started to use YouTube results and YouTube shorts in AI overviews and AI mode, but they have a very great content platform. So if you use YouTube effectively to reach more people or just want to spread your knowledge, expertise, all your...
information, even playbooks tutorials. So YouTube is a great platform and Perplexity just catches in a great way and they are using it very effectively.
Joe DeVita (36:56)
For companies thinking about producing video, they should be thinking about how to optimize for the YouTube platform. Whereas many platforms want video, it's really only how that video exists on YouTube that matters with answer engine optimization. What do you think about imagery? A lot of people use LLMs to create and to learn and to see.
Can you optimize images? Any advice to optimize images for LLMs to help with brand visibility?
Metehan (37:27)
Yes, I realize many solopreneurs and content creators are using AI, but you know, the hardest part is creating long form AI videos, let's say, and AI image models are getting better and better every hour, every day. So I also use AI image generation tools. I believe, yes, you can...
use it in your day-to-day SEO or AI operations. I believe it's very useful and time-saving operation. But for the future, I believe the video generation will be much more important than the image models.
Jon Clark (38:08)
I have to ask about schema again.
Metehan (38:11)
Great
question. Yes, yes, send it.
Jon Clark (38:14)
Again, referring back to our Duane Forrester conversation, you know, he was the co-founder of schema.org. kind of debate around like, is schema used, is it stripped out? And Duane had a really interesting take, which is, yes, it gets stripped out, right? Because the LLMs don't need brackets and things like that, but more than likely when they're ingesting it, they're either tagging that content or
classifying it in some way, which makes sense to me because why would, like if you have those, I think about it as like a label, right? Like you're basically labeling the content as you're ingesting it, regardless of whether that code is kept, it's kind of a mute point if the label is already applied and then you have all of these labels that you can then train the models on. Like that concept makes sense to me.
What are your thoughts on schema? Have you done any testing around that where you got a reasonable answer?
Metehan (39:09)
Many times we had a conversation with Dan Petrovic. β He also has many, many great names in the industry tested this approach. Just saying hi to Andrea Volpini from here. He have a great resources and experiments on this. I believe the schema is useful for LLMs, but it's...
Jon Clark (39:14)
Yeah.
Metehan (39:33)
I don't believe they're just reading the schema text, but it's useful your traditional search engine results. And LLMs can't read the schema text in your source code, but if it's helpful for your rankings in traditional search engines and LLMs are using...
scrape results from search engines. So yes, it's not a direct factor, but you can use and you won't lose anything. You can use this approach. But structure content and structure, I mean the structure saying is a little bit different in LLMs. And I believe people just...
talk with their own ChatGPT models characters in their daily lives. LLMs just placing these schema or texts can helpful for AI engines because LLMs now think from their offline trained data.
Something very similar that AI search optimization is very same as SEO. So when they see more schema org suggestions, they start to believe it. Yes, it's a direct ranking
factor or it's very useful to increase your brand visibility. I respect that by the way. Yes, it's helpful for your traditional search engine rankings and LLM scraping these results. So yes, but it's not a direct factor, I believe in my experiments. They can't take this data. They just tokenize them.
Jon Clark (41:13)
Okay.
Got it. I wanted to ask one more question around the domain authority strategy. There was a list of domains that were identified as like top authority domains or most trusted on Perplexity's part at least. And you mentioned something around the concept of naturally incorporating references sort of manually in your content.
And I was curious what you specifically meant. Did it mean, like if you're writing a piece of content that you're adding links to Wikipedia, for example, is that effectively what you mean? Just sort of using a very trusted domain as the citation for, you know, maybe what you're saying or a data point. Is that how to sort of interpret that application?
Metehan (41:55)
Yes,
I want to share this example. Let's say you just β write a sentence like, 'if you want to access our latest study, click here' and you use the anchor as click here. Instead using this sentence for LLMs, we studied the latest ChatGPT prompts data and according to
our brand names analyzes, we access these XYZ results. And if you want to see the results, click here. This is much better for LLMs and for their understanding because you just, you're just sharing the concept with context.
So they want to follow that direction, especially imagine like every piece of web page just saying click here, click here, they're training data. So LLMs won't work properly. So we need to use more consistent language for this.
Jon Clark (42:57)
Got it. I wanted to quickly talk through your prompt analysis. I believe it was around like 2,000 prompts that you analyzed. First question, have you done additional analysis since then? Have you expanded that list beyond the initial 2,000?
Metehan (43:11)
Yes, I want to expand my research then. I guess HREFs share the great analysis. then if I remembering wrong, please forgive me, Olivier from France share the great study. Yes, I want to expand it, but
during that time, Google already removed the public chats from SERPs. I used Data4Sale API to mine all SERP results and I used Archive Wayback Machine. And I realized that of course it's a very small data, but you can...
identify some similar patterns in those prompts like agentic use, they are using, they are assigning roles to LLMs. So these are great insights. And I just want to publish it and get feedback from our lovely SEO and search industry.
Jon Clark (44:05)
I pulled down about 7,000 of those from archive.org, but only the titles, because I just wanted to see if there was anything I could take out from the titles. And I things very similar to what you found, like a lot of prompts that you wouldn't go to Google for, analyze this, help me write this. One was crazy, was like, we're trying to push out
one of the co-founders, how do I put together? I was like, Jesus, this is, you don't want that public. Where was I going with that? So one of the things that you didn't see a lot of was transaction related prompts. And so I wanted to get your thoughts and your opinions on a lot of the recent news over the last couple of days with shopping and Shopify integration.
Metehan (44:31)
you
They do the conversion from
the fear.
Jon Clark (44:49)
Yeah,
like, and I know that you're a big, big advocate of that's sort of where all of this is going. So I was really excited to get your thoughts and your opinions on like, do you see maybe transactional type prompts increasing in the future or sort of how do you...
Metehan (45:06)
Yes, OpenAI share their public data results for their overall prompt stats, like if there are people using generative, transactional queries
and many other topics and categories for prompts and it's very useful. And I found the transactional data is very low at the moment for the share, but I will, it will also increase and people most likely use AI to discover new products.
and I see many patterns from many e-commerce websites that I also working with many others at the moment. All same product discovery, not usual products or not the top selling usual products. You can see on your dashboard on your analytics all different and people will use more effectively and they need to learn how to
use and create prompts at the same time. So I believe OpenAI is also trying to spread this narrative and know-how with partnering other giant companies like Shopify, Stripe, Etsy and all others will just join. It's exciting and I believe it's the first time I couldn't find any SEO related information
from an OpenAI announcements. I mean, Sora is different, we know, it's for image generation, but this is agent e-commerce. And first time ever, we won't see any real user interaction on our e-commerce web pages and products. So it's really scary and very useful. Right now we need to track the only logs or transactions.
How can I be more successful on the agentic e-commerce? Nothing at the moment, only partnering with other giants, sharing some technical complex stuff on GitHub and all other dev websites. So it's exciting and I hope we will find other way. And this is another billions of market at the moment agentic e-commerce.
I hope we all find a way to be early pioneers, early SEO and AI search pioneers in the industry so we can buy new Lambos, I hope. β
Jon Clark (47:33)
And
Find your Lambos, yes.
Joe DeVita (47:35)
There's so much changing so quickly right now. It's hard to predict what's going to happen. We have a handful of like kind of rapid fire questions we'd like to ask you first, before maybe Jon can get those started before he gets started, I've heard you say you predict SEOs become more important in the years to come because SEOs will be asked to try to understand how to optimize for AI.
but it also seems to me with the announcement this week of the ChatGPT integration with Shopify and Etsy, affiliate marketing will have a resurgence of interest next year too. And it's, just like, and advertising is going to change next year too. It's just like, next year is going to be totally different than this year and this year was crazy to begin with.
Metehan (48:20)
Yes, yes,
yes, just the every year's Tim Cook announcements. You know, this iPhone is the best iPhone ever produced. So we will crush the market and this is a new disruptive tech product in there. But for the AI, it's a little different because I believe the real hype didn't start yet.
I mean the agentic part. Right now we are asking SEOs like what's your most popular pages in LLMs for bots? And if you compare it with real humans on Google Analytics, what tell more? These are all weird questions if you compare it with the last decade. I mean, so everything is just... went
crazy and I believe agentic will be the new hype. We see any new products and SEOs, you know, we are optimizing everything in our lives. I'm a lister person, but I still looking for new stuff to optimize in my home.
My wife ask something you need to fix this? and to fix this? Yes, because my all perspective and thinking is like yes I need to optimize and fix this yes I need to use better time I need to look technical details we will find a way to survive during these hard and exciting times but we'll see and it depends of course
Joe DeVita (49:53)
Yes.
Jon Clark (49:55)
I think you started the interview with the comment around like you were sort of getting bored in SEO and now you're like excited with AI. And I think Joe and I have had this conversation many times with guests of there was a long period of time where SEO was just sort of ho-hum like it is what it's always been. And now there's just this resurgence of excitement and
Metehan (50:19)
Yes.
Jon Clark (50:19)
you know,
new voices like yourself that are helping to push the envelope on this whole new side of SEO. It's really awesome to see. All right. So to wrap up a couple of quick rapid fire questions. You know, you're pushing a lot of these conversations and doing a lot of this research yourself, but I was curious, like, who do you go to, to get information around AI and SEO? Like, who are your, who are your
go to followers to get this sort of information.
Metehan (50:45)
I mean, who also I follow β from the industry. First, of course, Dan He's just an amazing guy and he just went very crazy. He's now training his own models. I will set up a new computer environment in my home if my wife allows me to do that.
Jon Clark (50:48)
Yeah.
Metehan (51:07)
But we'll see. Mike King, of course, and Lily Ray is sharing great stuff from the publisher's perspective for the our lovely community perspective, not just publisher, I mean. And Andrea Volpini, of course, from WordLift. He shares really technical guides. I don't want to skip any name.
Jon Clark (51:10)
Yes.
Metehan (51:28)
Olivier from France, Nancy from Spain, of course you and your guests I love. Every person in the community, Aleyda of course, Yagmur from Turkey. So many great names at the moment. I don't want to skip any name, but you can follow these names.
Jon Clark (51:48)
I think all of those are on our guest wish list
as well.
Metehan (51:52)
Yes,
Glenn Gabe and of course, Barry Schwartz. β They are creating great content.
Jon Clark (51:56)
Yes, of course.
Of the LLMs, so ChatGPT, Gemini, Claude, Perplexity, is there one that you find most difficult to optimize for or has the most nuance to optimize for?
Metehan (52:11)
I can say maybe Claude a bit there, likely using some different search queries, a little bit different, completely different, but I can say. And Perplexity is using very high embedding similarity with the questions, so it's hard to...
Jon Clark (52:28)
Got it, okay. Do you have a favorite probing prompt to test answer diversity that you always go to?
Metehan (52:34)
Yes.
Jon Clark (52:35)
Will you share it? If not, that's okay too.
Metehan (52:38)
No, I won't share it. No,
no, no. I ask a couple of times same question or similar questions and use people also as data because they have real user questions. And I use Reddit using some extensions. I built my own tools to get the real questions from real humans. And I check the similarity, what gets their attention, I mean from the LLM's perspective, what are their most or top used bigrams, semantic triples, let's say, I analyze them and I try to place them in my content. This is excellent for your audience, podcast audience.
Jon Clark (53:20)
Love that.
Love that, love that. And last question, we like to ask like a prediction question.
And I'm going to switch it up a little bit because I think you have a really unique perspective. So which AI search engine will dominate by 2027?
Metehan (53:41)
I believe it will be hard question. Google has the best real time machine learning system. So I will say two candidates, probably Gemini and ChatGPT. Because the latest ChatGPT agreements on the Oracle stuff, they will provide more money.
Jon Clark (53:56)
yeah. Amazing.
Metehan (54:01)
And I hope we will see much stronger computational power and energy solutions in the all I mean in the all around the world so we can see much faster and effective LLMs. But I see OpenAI and Google will dominate the market.
Jon Clark (54:24)
Metehan, this has been super enlightening. I really appreciate it, especially given the late hour there. And let everyone know where they can find you online.
Metehan (54:28)
Yes.
Yes, they can reach me out anytime on X and LinkedIn, anytime. Just say hi, it's okay.
Jon Clark (54:41)
All right, sounds good. All right, everyone, this is wrapping up episode 99, episode 100 coming next week. And please remember, if you liked the episode, please remember to rate, subscribe, and review. We'll see you next time. Bye-bye.