Webinar! JavaScript SEO for AI Search: Training with Gray Dot Co

August's webinar was an extra-special super-duper exciting BONUS training module in our JavaScript SEO training course.

Yes, that's right! SEO legends, Sam Torres and Tory Gray from Gray Dot Co, came back to deliver a live training session on JavaScript SEO in the era of AI Search - 'cos:

• Do you actually know which LLM bots render JavaScript and which don't?

• Do you know what the potential impact is of having a JavaScript-reliant website in 2025?

• Do you know how to audit JavaScript from a search optimization perspective?

Well, these are some of the burning questions that Sam and Tory cover in this training session, complete with Q&A.  

This course, including this new training module, is in partnership with Women in Tech SEO but is open to all.

Webinar recording

Subscribe to the Sitebulb YouTube channel! You can also sign up here to be the first to find out about upcoming webinars. 

Webinar transcript

Erin Simmons:

Hello everybody and welcome to another portion of JavaScript SEO Training course. So, thank you all for coming. This training set up by Sitebulb and the Gray Dot Company came directly from last year's Sitebulb survey where folks went in and published what don't you feel comfortable with in terms of SEO and JavaScript? So, taking the results from that survey and seeing where folks needed more support, this team put together these wonderful trainings, this one, and of course, the free JavaScript SEO training course that's on the website to help folks find the skills that they need to be able to succeed in 2025 with SEO and JavaScript. So, knowing things like LLMs don't render JavaScript the stakes are going to be even higher as 2025 turns to 2026. So, Sitebulb is also running this survey again and we can drop that in the link.

If you go and fill that out, you'll be helping create more trainings like this that really speak to the skills that you want to learn. So, of course, Sitebulb brought in the wonderful Sam and Tory of the Gray Dot Company to help with this training, experts on this topic and brilliant insanely active WTS members so we are lucky to have them here and talking about this topic today. We're also going to drop into the chat the WTS way, which is our code of conduct. And if you could all just take a read through that, we will always strive to make sure that our spaces maintain a safe environment so that we can all learn, grow and be kind to each other. So, without further ado, I will throw it over to Tory and Sam to get it started. Thank you all for being here.

Tory Gray:

Thank you so much for having us. I'm very excited, I'd honestly prefer if it was JavaScript SEO, I'd prefer that would be more fun, I think.

Sam Torres:

All right, we're just going to say that the whole time, it'll be fun.

Tory Gray:

Good, we will.

Sam Torres:

But yeah, super excited to be here today, of course, JavaScript SEO is something we just love to talk about at Gray Dot Company so we are definitely in the right jobs. We probably enjoy this maybe a little more than we should, but super excited to be able to continue what we had before with the other courses that we taught and really just bring it to LLMs and what does it mean with this AI-powered search experiences. So yes, I'm Sam Torres, I'm the Chief Digital Officer at Gray Dot, obviously I don't say that enough 'cause I just doubled over that. But yes, I'm one of the chief people at Gray Dot Company. We have a lot of fun and I am of course joined by the wonderful Tory. I'm going to let her introduce herself.

Tory Gray:

Hello, I'm Tory Gray, I'm founder and CEO of Gray Dot Company and we love strategy, we love tech, we love data. So, at the intersection you'll find JavaScript SEO we really love this and now we're in a new AI-driven world apparently, everyone is making it happen whether we all like it or not so we're here for the ride and we're here to learn and grow from each other so thank you so much for joining us. We also just want to say thank you so much to our lovely sponsors, Sitebulb and Women in Tech SEO we love you both all the time, the best, the most. Thank you.

Sam Torres:

We do, we do.

Tory Gray:

All right, so what this series covers, if you've already joined us before, you've seen that in the previous three sessions we've covered things about what the heck is JavaScript SEO? How do you audit JavaScript for SEO purposes? What's the process and the tools? The next session we went through prioritising and explaining those issues. So, once you understand what the problems are, how do you communicate with your product and eng teams in order to deploy those fixes effectively? So, if you want to get a little more coverage, a little more reminders of what that was like, I believe they already put the link in the chat. But today we're talking specifically about how JavaScript impacts LLMs in again this AI world. So, without further ado, let's talk about rendering and how that matters for search and LLMs.

Sam Torres:

Yeah, which maybe you're thinking, "Oh shoot, I have to worry about rendering for LLMs too." And unfortunately the answer is yes, probably, right? It's a new frontier for the search experience. You can throw around all the three-letter acronyms you want, whatever you're calling it, it is obviously a channel that is getting adopted, it's things that users are actively using for lack of a better term. So yeah, we need to start talking about this and understanding how that works too. But there might be various reasons why you're here today so let's talk about this.

Tory Gray:

Yep. So, it might be a lot of reasons. It might be visibility, it might be influence, it might be money, you might be on either end of the LLM spectrum. You could want to appear and you're doing what you can to understand how to make sure you appear more frequently or as best as possible. You could be on the other end of the spectrum and try not to appear because you don't want to be a part of those results for whatever reason. There's not really one way to think about LLM visibility, it's not good or bad you just have to make the right decision for your business and your situation. But that said, I think a lot of us frankly are here because maybe your boss said you had to be.

So, even if that's the case, we welcome you, we will get this done, we will help you onto the other side of helping to understand what this all means. Okay. So, first what we're going to do is kind of cover our basis obviously and we're just going to go through kind of the facts, some groundwork for LLMs just to make sure we're all on the same page before we start digging in.

Sam Torres:

Yep. We want to make sure we're all working from the same base. So, this is a study provided by Hostinger, I hope I'm saying brand correctly, of just the use and the adoption of these different LLMs or AI platforms. Really what we want to highlight here is that not all of these are created equal, they're working with different technology platforms. Obviously Gemini has the benefit of being in the Google ecosystem so there's going to be some special things about them. But generally we are going to be talking about them as a whole, we're not really going to get too much into the specifics of one LLM versus another. And frankly that's because there's just a lot of these that we don't have that data, a lot of them are sitting in black boxes. So again, we're going to be speaking mostly in generalities with few exceptions.

And then also just want to highlight that expect this to change over time, that's going to be a consistent theme that we talk about today because this is actively changing and moving technology. Think about the .com bust if you're old enough to have remembered that like Tory and I are to call ourselves out, it was changing all the time. And even think about how much just SEO has changed even just from the lens of Google in the last 10 years. We're going to see the same kind of case I would guess. But like I said, we're going to be talking on the whole. So, from these LLMs, now let's move into what are the data sources? What are they using to actually be able to generate the responses that you're seeing? And the first one is definitely their training data.

So, this is preset data, it could be anything from books, newspapers, magazines, movie scripts, all kinds of websites. All of that is going into this training models. Now, the things that I think are really important here are that many of the LLMs there's not transparency on what that training data is. Maybe you've heard about a whole of these cases where certain LLMs are getting sued because they used copyright material and didn't give correct attribution. But what's really important here is that this training data, it's already ingested, it's already there, there's not a whole lot that we can do to change it. And then, also it's just kind of the web as a whole and depending on which LLM depends on how up to date it is.

Tory Gray:

Yep. How up to date it is, which sources they use, to what degree do they use. I mean, this all varies from LLM to LLM and as Sam said, they don't necessarily share that data. There have been leaks, we do know that about some sources, but that training material can vary. So, LLMs can also use the live web, I will say that the training materials is the default. That's what LLMs want to use, it requires less resources for them so I would call that compute in terms of the computer time. What do they want? What's the natural state and how do they use as minimal resources and therefore money as possible? And so, the default, yes, will be training materials. They can access the live web, they do that classically speaking through either an MCP or RAG. MCP is model context protocol, RAG is retrieval augmented generation.

They are slightly different tools that do, they accomplish the same goal. So, both of them are there to help an LLM interact with outside data but they do it in different ways and have different strengths. So, I think of it if you picture a Venn diagram, the RAG model is much more about text retrieval. So, it can say, "Hey, go to this web page and tell me what's on it," right? It can go live and fetch that and it can look at that compared to maybe what it knows in its training model if your page is already a part of that training data and it can know that and it can update it. The MCP is a much more interactive model. Think of it as more of that Agentic. It can take action, it can help you log in, it can build things, it can do things. And so, where those two things overlap in that Venn diagram is that they're both able to access APIs.

So, in theory both are able to query Google. So, that brings me to query fan-outs. We all love and know queries because we've been in search for maybe too long for some of us. Yeah, so I want to talk a little bit about a query fan-out, don't worry about the other slides we're coming back to that. But what the heck is a query fan-out? So, essentially we are taking this one succinct query, which is you have a goal. If you are a photographer and you need to schedule time to take an engagement photo shoot, you need to figure out what the best time is for that. So, you might go to an LLM in order to query this. So, you have again a specific query, but there are many different sub-intent queries based on this. This is an example from Google's one of their initial presentations, so this is three of frankly more than three sub-queries that this relates to.

So, if you think about Google along our path of searching, historically in order to get to the big picture of what we wanted to accomplish, we broke our search journey up into multiple steps. So, this is what light do the photographers prefer? When's the right time of day? What's the weather like? If it's an e-commerce and you're buying a dress for a special occasion, it might be what's the fit, what's the flow, what's the colour? Is it appropriately dressy? All those kinds of things. So, you can have a lot of different needs that bubble up to your one primary need. So, in other words, LLM results are in fact three Google searches in a trench coat. Thank you Myriam for bringing this lovely data to us. But yeah, it's Googling, it's let me Google that for you. It's true.

Sam Torres:

[inaudible 00:11:50] that doesn't exist anymore. I used to love handing out those stickers. So, the last piece where LLMs can get their data is actually from feedback as users are using it. So, this is both the LLMs sometimes have quality testing teams that actually are giving feedback on the results. And then also your own feedback when you're using the tool, that can all go into training and moulding the model. Definitely has led to some pretty embarrassing circumstances for some large brands. So, this is definitely how sometimes information gets leaked to other users because your own conversations might get used to help train that AI model, which can then be served to someone else. So again, with these different sources, training materials, live web and live feedback, I just want to stress, we don't know how much they use the one versus the other depending on which LLM you might be able to tell when it's accessing the live web versus when it's not. But there's really no just standard way to understand across all of these.

Tory Gray:

We are in the Wild West times y'all, things are changing all the time every day so ...

Sam Torres:

Exactly.

Tory Gray:

Good times. All right, so let's talk a little bit more about what LLMs see. So, like pictured here, this is Russell Crowe from A Beautiful Mind, and this is how I picture it. It's a bit of a personification of LLMs to help me visualise what this means. That can be loaded sometimes, sometimes it makes us misunderstand how the LLMs work but I do find this analogy helpful. So, if you want to picture Russell Crowe here and he's looking at all his formulas and he's looking at the information, he's looking at the web because he's an LLM in this scenario, and he's looking specifically at say, code, machine code and trying to understand and get nuance out of it.

So, if you get code, how do you get that to words? I picture that like here where he's visualising where kind of the words kind of pop up out at him and can see. They can see through the code to see the text, the copy, the contents underneath. So, with that in mind, let's talk about what the heck do LLMs see? But first I'm going to tell you what they can't see and sadly that for the most part with a big asterisk that we'll talk about later is rendered HTML, on the whole LLMs do not see rendered HTML.

Sam Torres:

So, what does that mean they see? They see the response HTML. So, everything that we were talking about in earlier courses and sort of what's in that initial server response. If you do a view page source in Chrome, what does that look like? All of that is what an LLM typically sees. So, let's go through this and let's talk more about the details 'cause there is some nuance here and there's some things that are a little bit more interesting or different than what we've been used to from the Google ecosystem and how they've kind of trained or moulded the SEO community.

Tory Gray:

Yep. Okay, so what sorts of things can they see in the response HTML? The first thing is the text that you can see on the page, most of the time big asterisk. And we're going to walk through some details because rendering matters and sometimes things are rendered. But on the whole they can see text on the page, on the whole they can see the links on the page, again based on rendering. Same thing for media references. Media might be less important to LLMs. They care more about text but they might care about specific media so logos, videos, things that would be ancillary and helpful to the search response of whatever people are looking up on your site.

Sam Torres:

Yep. And then, the one that I think gets more interesting is that it can also see the code that's in the response. So, that could be you have text snippets. Things that we're used to would be like your title tag, your metadata, open graph, social tags, those are things that as SEOs we're already used to, but as a user you don't inherently see it when you're on the webpage. When was the last time as a user you actually looked at a meta description on a webpage? You don't, right? And Google's rewriting them most of the time anyways. So, that's what we're used to. However, what we have seen is that LLMs are also seeing the content of comments in your code. Are there huge JSON snippets? Those are things that can all be understood that have traditionally been ignored by Google.

So, that may also mean if you have a JavaScript-based site, maybe a lot of the content that's populating that is coming through in JSON snippets or you have web components. LLMs are actually able to see those things. This is also why some of the crappy old-school black hat tactics of SEO still work. So, you can put white text on a white page, you can have texts that is too small for the user to actually read. So, we are seeing unfortunately those kind of black hat tactics are working with LLMs. I'm going to beg and play with you, please, this is not me saying to do that. That's a terrible idea, you're going to hire your SEO and ideally LLMs are going to start to really understand and be able to account for that because obviously that was something that Google had to learn and the goal is that LLMs will learn that as well. So, those tactics are working but I think if you try to engage in any of those, it will damage you in the long run.

Tory Gray:

Yep. Next we'll talk about what they use so we know what they can see, let's talk about a little bit about what do we know that they use. We know that they use text, right? They care a lot about text that's what they base their training models on. So check, we know they care about text. We know they care about links as well, whether that's for grounding and citations or whether that's so that they can find more data to crawl more pages. They care about links to the extent that they can access that. Do they care about media? Sure, in some cases they do, especially a brand logo, those kinds of things that I mentioned before. They can also care about again, the code. So, maybe sometimes depends on the context. I want to specify here that code is a language, LLMs understand code as well.

They can read HTML, they can craft HTML, they can break it apart, they can put it back together. Code is a language that LLMs understand, which is why there's say tools like Claude that can help you vibe code and all that jazz. We know that that's out there. So, how and when they use this for text-based data is certainly more of a contentious topic in SEO and for this world of LLMs. But we know that they can see it so it's important to pay attention to it. Okay, so let's talk a little bit about examples. So, we know that text that you can see on the page, again, for the most part they can totally access this data. So, this is me going to ChatGPT and saying, "Hey, what do you see?" This is the text from our homepage on our own website. Again, caveat sometimes 'cause sometimes that content is in the response HTML.

So, if you're using JavaScript to load in critical content on your page, that can mean that the LLMs aren't seeing that context. Depending on the page, depending on the text, you may or may not care. But again, here's an example from a big e-commerce website, Crate & Barrel where some text is missing from this page. So, how do you identify those things? You can look at our lovely favourite tool here, Sitebulb, here's three errors that we see a lot that you're going to want to pay attention to specifically in this context. If the h1 is only in the rendered HTML, that means it's not in the response. If the h1 is modified by the JavaScript depending on how much it changes, we might care because maybe they're missing critical information. And then, certainly if it contains JavaScript content so page copy, whether it's an h1 or not, we might care about that.

But much like we talked about in our last session of JavaScript training, again everything is contextualised in, do I care about this page or do I not care about this page? And do I care about the content that's missing or not, not all content is created equal. Sometimes it's the meat of the page and it matters to the context of the page. And maybe that page is your homepage or maybe that page is a critical landing page that drives a lot of business and revenue for you or it might be ancillary unhelpful content. So, context matters.

Sam Torres:

And so, it's very much the same with links. So, here we have an example of our website using our newly launched and released rendering difference engine Chrome extension. So, this is a free tool that you can use to see what does the site look like in response versus rendered, what's kind of the difference and it's serving what those differences are. So, for example, what we're seeing here is that all these links are being served in the response HTML and then the JavaScript actually destroys them and rebuilds them.

Tory Gray:

Which is very common. Don't worry, don't worry about that.

Sam Torres:

Super common, super common.

Tory Gray:

The fact that it's coming back is the part that you care about. If it goes away and you care about it, that's a problem.

Sam Torres:

So, these are links that LLMs would be able to see. So, this is where LLMs can see links but sometimes they can't. And again, it's because these links are showing up only in the rendered HTML and they're not in the response. And LLM is probably not finding that fall 2025 article unless maybe there's other ways that it's surfacing, but just based on what they can normally see, that's-

Tory Gray:

Internally [inaudible 00:21:54] potentially.

Sam Torres:

Yeah.

Tory Gray:

Yep. And it's the same deal for media, we were not going to go through all the examples 'cause it's the same thing. They can see it if it's in the response and they can't see it if it's in the rendered version.

Sam Torres:

So, when you are doing this what to see with links, so the site will report when you're wanting to do this at scale and not just looking at single pages. You will see they have a lovely alert for you contain JavaScript links. So, definitely consider that LLMS might not be able to see those links.

Tory Gray:

Yep. I'll say look for this contains JavaScript links specifically in terms of the in-page content that you're linking to specifically. Also in terms of your navigation are the two kind of common areas that we see that sometimes things like JavaScript loads in your nav, maybe then they don't see any of your sub nav and maybe that will impact what they access and can access because they know it exists on your site.

Sam Torres:

And then just like the text, think about are the links that are appearing only in the rendered HTML, do they actually matter to you? So, in this example it's to accessibility and disclaimers, things like that. So, probably not the biggest deal and it's going to be okay, so you're always going to need to prioritise based on what are your overall goals.

Tory Gray:

Absolutely. All right, so now we're going to talk a little bit about these code examples. So, what can we know that they see? They will see stuff like meta tags. Again, big asterisk, we don't know how they use this and how much they use this but they can see this data. And metadata can be nice because it's nice kind of chunked snippets of nicely defined things that LLMs tend to like. They can also see stuff like open graph data. They can also see stuff like schema, which is a whole bag of worms, whether or not they use this, how it gets used or not, how it gets destroyed.

What I can tell you is that we know, at least for the big LLMs, it is used on live web fetches and there's been some really interesting case studies and I can share a link after, wherein they basically found during a live fetch that the data pulled from the pages is much more detailed, more correct. It helps with grounding in the data on those live fetches. It's less clear exactly how much that might be used or not or how and how effectively that's used in the training data. But they can see it so, maybe it's worth paying attention to and experimenting with it.

Sam Torres:

Yes. And we know that this is a spicy take and a pretty-

Tory Gray:

We're okay with that, we like spicy.

Sam Torres:

Lively conversations happening right now.

Tory Gray:

Yes. Okay. So, other things they can potentially see. So, we talked a lot in our previous JavaScript sessions about what counts as the link so this is a screenshot from that. What counts as the link and what counts as on-page copy. In terms of an LLM, LLMs aren't rendering so they don't know what's rendered and they don't know what's showing up on the page, they see text. Is it text or is it not text, right? So, things like links that aren't linked as a hrefs, maybe they're linked as on-clicks or some of these span hrefs, these common other ways that apps do links that work for users to the extent that the LLM can detect that that is a URL-like pattern, they might use that to crawl. Same thing with this example below is a web component.

So, this is a particular site that we worked on and we did a lot of work to make sure that the h1 wasn't headline type h1 equals this and instead it was actually classified within semantic HTML as an h1. That matters a lot for Google to understand, "Hey, this is special text, this is an h1, this is not just page text." Or, "Yes, this is actually a link and Google you should crawl, you should go to the next page and you should value this link further." Whether or not Google can detect that it's a link and maybe it decides to crawl it, unless it's an a href, they're not going to value that link, it's not going to spread link equity, it won't matter for SEO. But in terms of LLMs, potentially that content is an interesting and fair game.

Sam Torres:

Now, where it gets a little bit more interesting with code and kind of fun. So, things that have been typically ignored by Google would be comments in the page. But LLMs actually see this and we know that they do because of some of the embarrassing PR stories that have popped up about major brands. Like there was a major technology company who had all of the internal project names leaked and in a massive way because they weren't actually protecting it, it was showing up in comments, which technically anybody would've been able to find that at any time, it just hadn't gotten the publicity for it yet. So, for example, we have this one on the left, someone's calling Richard an idiot, which seems really unkind and maybe that's not what you want your brand to be associated with. Or this one on the right, "Yeah, we all just got rickrolled."

So, really think about those are things that can actually pop up, it happens. Developers think they're being funny, they put in all kinds of things. So, think about there might be rants, there might be comments about, "Hey, this doesn't work," I've written comments like that. Like, "Come back to this, you need to fix it. Are there Easter eggs," April Fool's jokes, all of those things can come together. And then, also think about what about any kind of confidential information? As a user is on the page are you pulling in any of their PII to help render the page? Things like that.

Tory Gray:

Yeah, these are all the things we want you to look out for. To be clear, they were always problematic in your source HTML because anyone could look at it at any time. But I think the risk is greater now because now there are tools specifically looking at that and paying attention to that and using that and maybe accidentally misconstruing things. So, if you think about the LLM example, I think everyone's heard the Reddit, how to make the cheese stick to the pizza, you put glue on it, LLMs don't understand jokes per se. So, they might not understand that your April Fool's joke is an April Fool's joke unless you say April Fool's at the bottom. And even then with tokenization, it might not come together, they might not get it.

So, what do you look out for? What could be in your view source that could be problematic, period, end of story. And also in the context of LLMs, is it embarrassing? Is it inaccurate? Is it confidential or proprietary? You definitely don't want that in view source because it's there. Obviously PII and privacy-related concerns. And then overall, obviously we still care about what is JavaScript changing, adding, actually it doesn't matter what they remove, that only matters for SEO because they won't know it's removed. [inaudible 00:28:51].

Sam Torres:

Well, unless it's getting removed because you didn't want people to see it, right?

Tory Gray:

Too late.

Sam Torres:

That I would stress don't serve it ever. So yeah. So, really the biggest thing to know when it comes to LLMs is basically that if it's rendered you can't really depend on LLMs being able to see that. Or I'm sorry, yeah, if it requires rendering, I think I said that right. Now, there are other ways we know that they can do live web fetches, maybe they interact with Google searches. So, that's not to say that that's just like a blanket, it's never going to happen, it's just more unlikely.

Tory Gray:

Significantly more unlikely, especially for any non Gemini crawler specifically.

Sam Torres:

Yeah. So, you're always going to want to be able to understand on your site what's in response, what's waiting for rendering to show up, do you care about those differences? Do you care about what's showing up only in rendering? But it's LLMs and Google so you have to, and other search engines, I guess not just Google, but you really need to consider both audiences but still always it's about the user at the end of the day. So, don't forget the number one reason.

Tory Gray:

Yep. So, in summary, it's a bit of a spectrum. In JavaScript SEO we care about what gets rendered, that it gets rendered and that Google can effectively reach it. And do we want to pay that sort of performance cost of rendering? So, it's a different ball game for LLMs where we care about is it only in rendered because chances are an LLM can't access that? Or alternatively things you don't have to worry about at all on the SEO side, what can they see that you don't want them to see? So, how do we go about that?

Sam Torres:

And so, let's talk about what some of the exceptions are because of course they exist because SEO, right?

Tory Gray:

It depends.

Sam Torres:

So, it depends. So, the first being that we are in a consistent state of change. So, like I was talking about before, it's going to be constantly changing. What we're talking about today, it could change next week, it could change tomorrow for goodness' sake, so really just be prepared. I know the AI burnout is real, we're all feeling it, but just know you're not alone we're all in it together. Other things about it. Gemini actually has access to the rendered training data because they have Google powering them. So, that's one of the few that definitely has that. And of course, it might not end up being the only one, it's just really the major one right now.

Tory Gray:

That is today, yeah. So, the other thing is obviously LLMs can query the live web, including using API calls to Google on Google. And if Google's already rendered that data and can answer that question within Google, they can get access to that rendered data. That doesn't mean they can access it directly on your page, that doesn't mean as they're crawling your site, they're doing the compute work that would require rendering because they're not. I'd also say can LLMs render? Sure, if you hook it up to a GitHub puppeteer plugin that's telling it to go to, it can code, it can do things, you can make it do things because you're telling it to do things. Does that mean it's doing it? It does not mean it's doing it.

Google can understand the contents of a video, for example, but that doesn't mean you don't want a transcript, that doesn't mean you don't want to set context and add video schema because it makes it easier for them to understand the contents of video without going to that level of work. So, LLMs are a similar thing. But again, LLMs, again this changes every minute. Obviously we're exhausted, but LLMs could develop the tech, they could partner with others, they could enter partnerships, they could licence that data. So, watch this space. I mean, Perplexity last week, right? Makes a bid for Chrome. Who renders? Chrome does. So, Perplexity could get access to that data, were that a serious bid, which it was not, but whatever.

Sam Torres:

Yeah, it was not a serious bid. And then also I don't know, we got an invite to Comet, which is their own browser so I think honestly the bid was probably just a publicity play. But yeah, so Perplexity is coming out with a browser so it's obviously something that, or it seems like it's something that LLMs do care about. Yeah, it's definitely important to understand that even as they develop these tools, they may not render the same way. If you've heard of the site Can I Use, the reason why that's so important is because as designers or UXers, we have to understand, is the experience going to be the same for my user who's on Microsoft Edge versus Chrome versus Firefox versus Safari? None of those tools work exactly the same when it comes to how things get rendered, how web pages actually appear. Now obviously we try to go for standards, but that is something to keep in mind that as they develop these tools, it might be another way that we need to understand how things are rendered.

Tory Gray:

Exactly. Just like Googlebot renders in ways that Googlebot renders, which is different frankly than how Chrome renders. So, if you use your Inspect and you're using your web developer tools in Chrome to understand how Google is crawling, that'll be right 85, 90% of the time but there are small use cases where if you put the same page into a Google Search Console and you look at the Googlebot rendered code, it's somewhat different. And I have seen cases where that matters a lot to the bottom line of a business because something that worked in the browser didn't work for Google. So, today we have specific rules and guidelines that we're talking about that work today, but these are things you need to pay attention to because if and when inevitably more LLMs start to render, we're going to need to start auditing in the context of that LLM and hopefully they'll be tooling at that point to figure out how do they render, what issues do they run into, what things are they able to overcome that Google is or isn't and all that jazz.

So, those sorts of normal JavaScript SEO auditing rules we believe will apply in the future, even if they don't today and they do today, apply in theory to Gemini because if Googlebot can access it, then it's not part of the training data and then Gemini can't use it. So, here's another good piece to read. We're all learning, this is all evolving and changing every day. The tech is changing literally every day so some of these things we expect to be inaccurate soon, but it's what it is. Okay. So, if that was exhausting, you're right it was, there's a lot happening. But you might be saying, "What the heck do I do now?" And what is our advice for you? So, what do we want you to do now? Yeah, first you're going to audit to understand what is in the response HTML, including what shouldn't be there, right? We're also going to audit to understand what's in the rendered HTML. Then just like in a JavaScript audit, we're going to determine what if anything is missing and how much you care about that thing or not.

If you want a refresher on how to do that, you should go check out our JavaScript training, we have that for you. But it'll be different in that you'll be looking at, oh, it's only in response, I'm sorry, it's only in rendered. Okay, then how do we get it into the response? Or if you don't want LLMs accessing your content, maybe that's an interesting barrier you could use to keep LLMs out your way and off your content with a grabby hand. Obviously we love SiteBulb, it's how we do JavaScript audits, especially for enterprise brands across all sites. We also would love for you to check out the Rendering Difference engine, which is a Chrome extension that you can use, and it's specifically for page-level audits, as Sam said previously.

Sam Torres:

Yep. So, with all these changes, now you know how to run the audit, but obviously what you're looking for might change, so how are you going to keep up? So, I will say subscribe to the Women in Tech SEO newsletter. Yes, I'm biassed, doesn't make me wrong though, it's phenomenal content. So good. And then also Women in Tech SEO, SiteBulb, our own brand, we're going to continue talking about these things and trying to surface what actually matters and how it impacts what you should do or how to think about certain scenarios so that you can make the best decision for your business. Because I do want to say that while we're talking about mostly with the lens of you probably want to increase your LLM visibility, that is not true across all brands and that is okay, you need to figure out what works for you.

And then of course, really start building out and becoming comfortable with the tools at your disposal to do the audits, to run the test so that you can make good decisions, start building out your roadmap and really start looking at where do you want to take it, what kind of visibility do you want to have in LLMs and AI search? Start looking at what rendering changes are really going to be most important to you based on how you want users to perceive your users and LLMs to perceive your brand, your products, your services, whatever it is you're trying to sell.

Tory Gray:

Yep. And a thing to throw out there is if LLMs are very exciting to your executive team and you've been fighting tooth and nail to fix your rendering issues for SEO or to get more things in the response, maybe this is the underlying data you need in order to say, "Hey boss, I need you to get this in the response HTML, I need to prioritise this work or you're not going to show up in LLMs," so use it is the point.

Sam Torres:

30 points in the next sprint, give it to me. Right? This is this how you do that.

Tory Gray:

Yep.

Sam Torres:

Okay, so what questions? I know that was a lot and I think there's actually quite a bit of Q&A for us, which is great. Maybe it looks like Chad is busy so that's good. So yes, let's get started on some questions.

Erin Simmons:

Cool. So, Jojo will throw some questions up for us and we can take them one by one. One second, lost my screen, there we go. Okay, so first question. I've seen a lot out there about structured data being great for LLMs but recently there are people saying that they don't look at structured data as structured data and just as words. What are your thoughts on that? From Emma Russell.

Tory Gray:

Thank you, Emma. That's a good question. And again, this is yes, it's contentious and it's contentious because we don't know. So, we can make, there's been explorations, there have been interesting tests lately that we can speak to. I can say Google and Microsoft have confirmed publicly in various presentations that they use structured data. Structured data is something that machines enjoy.

Sam Torres:

In the context of LLMs, it wasn't just in the context of search, they said that in the context of LLMs.

Tory Gray:

They did. And so, at various different conferences, different Googlers. Google has also come out and said they're not investing in schema anymore and then they contributed to it a month later. So, what you do in Google, I don't know, and I can't predict the future. And Google is just separate, right? Google isn't Google, Google is various people with various priorities pushing for various things.

Sam Torres:

Yeah, that's sounds like one team did not talk to the other.

Tory Gray:

Yeah, exactly, yeah. So, which one will win the war? I can't tell you, I can't. I can tell you, I'm going to share, I'm going to find this link for you. And that's the case study that I was referencing earlier I'm putting it in chat. It's a really cool test that they did where they added content in schema versus not and then those pages are still live. So, you can do your own tests in ChatGPT or Claude or Copilot or Perplexity, well, whatever your LLM of choice is, you can go see what it does. What the test found is that during a live test, again that the details of the page more detailed, better information, significantly more accurate results. So, during a live test period, end of story it matters. So, then the question is, how often is Google or I'm sorry, any LLM, how often are they using live data versus how often are they using training data?

I have been digging into this and I have not found any really good answers. ChatGPT on the whole doesn't like to. I asked ChatGPT how often other ones did it and they assured me that Microsoft said they're doing it all the time, but then you click the citations and that's not actually what they said. They were talking about something completely irrelevant. So again, we don't know. I would say again, a lot of the arguments against it are that as an LLM is doing its thing, it's chunking content and it's breaking it into bits and the semantic integrity in terms of its content. And that may well be true. I'd say a counterpoint to that is that LLMs understand language, they understand code. They can take a block of code, HTML, JavaScript, CSS, whatever it is, they can tokenize it, they can break it apart, they can put it back together, they can know what it means, they can restructure it, they can do things with it.

So, to me to say they don't understand how to handle it, that doesn't make sense. Are they using it in that way actively in their training data? I can also say historically, no, but more and more of them are experimenting with them. They are experimenting with what they're collecting and why they want to grow their training data. This might be helpful information. So, if nothing else, I personally believe this is something they're actively experimenting in to see if they can use it certainly in live batches and perhaps also in training data. But watch this space is what I'd say.

Sam Torres:

Yeah, and I would say in particular we have a client who uses web components that actually it's been really interesting journey with them of trying to get more things into the response HTML to make the overall Google understanding them easier. But what we're finding is actually their visibility in LLMs and that content is really strong. So, that tells me that LLMs are seeing the JSON in those web components and they can understand it just fine. And I'm sorry, it's a combination of web components and JSON so it's a little bit tricky. But overall, yeah, I think at Gray Dot we've seen anecdotally too many instances where content of structured data is being surfaced that it's being used, how much they use it, how much they rely on it.

Tory Gray:

Is it special?

Sam Torres:

Basically, we don't know.

Tory Gray:

Yes,

Sam Torres:

It's the same, we don't know how much an H1 weighs against a P tag. We kind of have an understanding of Google prioritises that or thinks it's more important. We don't know any of that with LLMs. But to just say that it's not being used I think is a complete misdirection.

Tory Gray:

At the end of the day, it is text and it is nice chunked text that LLMs love. It is completely succinct and it tells a complete story. So, as we're hearing advice about LLMs like keep your sentences whole and have the context set within them. The nice part about schema is it is a sentence that does that. And so, also it'll help you in some cases in SEO so like why not experiment with that? Should you go for broke and put everything you can think of on every single page ever? No, I'd never recommend that for any [inaudible 00:44:29] anywhere. You do what makes sense for your business and experiment and test and see and let this evolve over time and keep watching data. But they say they do, are they lying? I don't know, but they're probably experimenting with something if nothing else.

Erin Simmons:

Yep, great one. It sounds like watch this space and hey, why not? It does more good than any harm that's being seen. All right. Next question we have from Maria, how would you recommend to display content that currently lives in accordions or tabs so it can be read by LLMs?

Sam Torres:

Ooh, great question. And if you attended the other ones you know how I feel about accordions. But the good news is with accordions there's actually really great ways to design and code those without having to hide them in JavaScript, it can be done purely in CSS, which means that all of the content is showing up in that initial response HTML, that is something that LLMs can see. No, Patrick you're totally wrong, I hate accordions so much, so much. But yes, there are ways to do it and of course I say this, and actually we're probably building out an accordion module for our own site so I'm going to have to eat my own words. But all of that being said, there are ways to do it without JavaScript so just ensure you are doing those. Yeah, just go for the pure CSS method and you should be fine.

Tory Gray:

Yep.

Erin Simmons:

Can't wait to see that accordion live on the Gray website. All right, next we have a question from Simon. So, earlier versions of ChatGPT could not access live websites. When did this change or are they just scraping Google Search results and other AIs?

Tory Gray:

I mean, they're certainly using APIs to access Google on the live web. They're also directly able to access. When did they make that change? Is that 2.0? I don't remember.

Sam Torres:

I don't know, what is time? I don't know.

Tory Gray:

Sometime in the last year.

Sam Torres:

Great question. I remember LinkedIn being very excited about it but I couldn't tell you when that was so sorry Simon we're kind of failing you on that one. But I'd say as far as where are they getting the data, we've seen enough studies to show that all of the above is true. It was kind of funny and ironic when we saw that ChatGPT is accessing Google results when they have so much paying money, but then also they're serving Bing ads so it's a, all of those things could be continued. Oh, great Erin, yes, so I'm going to let you talk about that, what you just linked.

Erin Simmons:

Oh, I don't know anything about it I don't do SEO I'm just here for the entertainment. But yeah, remember Elayna did this wonderful study that yeah, it showed that ChatGPT is actually using Google SERP snippets for some of the answers.

Tory Gray:

Yep. So, that's the query fan-out.

Erin Simmons:

I just read the title.

Tory Gray:

Yep. So, when you search for something in ChatGPT or another LLM, they can turn that into a series of sort of sub-Googles and then they go and then they Google each of those things individually, they get those answers and they contextualise and pull it all together. So, they're doing that via APIs, maybe other methods too 'cause seeing lots of interesting things in all of our GSC accounts and even in Google Trends I'm seeing weird data that looks very machine oriented, very patterned, so something weird is happening there. Yes. I can also say definitively they are looking at live websites. So, for example, as I was experimenting with the schema stuff myself.

We have a test site, we set it up where there's an individual page and all it says on the page is you can't have my address. And then in schema, we could add local business markup and add an address in there and then ask ChatGPT. We put it on two minutes prior and then we can go to ChatGPT and say, "Look at this webpage, what's the address?" And it can pull the address. And so, there was no time for Google, Google hadn't crawled that, Google didn't have access to that domain, it hadn't indexed that content so definitively they can access sites and update that in real time in live fetches.

Erin Simmons:

Yep. All right, thanks for that one. Next one is from Chris. So, what are your thoughts at this point about the utility of LLMs.text files?

Sam Torres:

So, this one's fun. The number, I think probably every single one of our clients has come to us asking about this. It's not really being adopted by anyone so if this is like you have some extra time on your hands, go for it. But in our opinion, it should not take the effort away from a rendering fix or content improvement, content pruning. It's not being adopted. It's one of those where I don't have any evidence to show that it hurts, but you can't really expect any results from it either.

Tory Gray:

Yeah, it doesn't hurt. It may be a waste of resources but maybe those are minimal resources and it's worth it for you to experiment. Absolutely go nuts and have fun I guess. Google does tend to crawl it from what I'm seeing. Someone posted a study where they were seeing which LLMs were crawling the LLM.txt file and most of them just didn't even see it. I think if you link to it in say your robots.txt file with a sitemap protocol kind of reference, maybe they can find it in that way and maybe they'll look at it. But for the most part, they're not looking at it at all. Google might be the exception but from what I've seen, they're treating it like any other page on the site or a txt file. So, it's not special, it's just another page.

It is another page with links so that's maybe interesting to make sure, especially if you like internal linking issues on your site, it could be a way to be like, "Here's a bunch of links or a bunch of folders that matter on my site." So, there might be use cases where it could be helpful because we're pointing that out and fixing problems on crawlability. Again, watch this space at any time Google will be like, Sure, why not?" But also we have the robots.txt so why couldn't you give instructions for different bots in that place? And we already have the robots.txt and we already use that to control LLMs, and they already don't listen to it, right? Because LLMs, their incentives are not necessarily aligned, they want your content, they're willing to steal copywritten content.

Didn't Meta download all these copywritten books and use that to train their model? They're willing to break laws to get the content that they want, so why do we think they're going to listen to this volitional instructional file? I don't expect them to because right now they don't have a lot of incentive to. That might change very much in the future when there's market leaders like a Google that says, "I'll do this 'cause I'm a good player," and Bing as well. But today it's probably a waste of time but it's fun to experiment so [inaudible 00:52:12].

Sam Torres:

Yeah.

Erin Simmons:

Nice. So, a little lawless right now but do your own experiments, watch this space. All right. We have about seven minutes left. Next one is from Myriam. So, what are your hot takes on Perplexity being caught with their hand in the cookie crawling jar by Cloudflare?

Sam Torres:

I think all of them are probably guilty of it. Just like what Tory was talking about they don't care. They're going to try to get as much content as they can because that's who's going to end up really kind of winning is how much data can you get to continue growing the model. So, I think as a consumer I'm really interested in this idea that Cloudflare has of LLMs need to start paying to crawl that so as a brand you can start charging. I have no idea how that's going to be respected, how it's going to be enforced, but I guess I'm encouraged by seeing that people are trying to think of solutions. I don't think that's going to be the unicorn of solutions that Cloudflare thinks it's going to be, but it is nice to see that we're seeing some innovation there. But yeah, I think ultimately it being Perplexity, it was going to be one of them at any point.

Tory Gray:

And it actively is one of them. I mean, I've also heard [inaudible 00:53:39] of maybe the reason we're seeing things in our GSC file is because LLMs aren't going to Google and using the API, maybe they're scraping results directly. Different tools are doing things in different ways. Or according to or against terms of [inaudible 00:53:59], yeah. And again, they're breaking laws. They don't care right now, they need more data and they're willing to break rules to get it and that sucks. But maybe if you put all your stuff in render, if you really care about that, then they can't access it. But that will also have SEO repercussions so proceed with caution.

Sam Torres:

Yeah.

Erin Simmons:

All right. Next up, we have Clarissa. For metadata and whether it is available to be seen by LLMs, what if it is rendered on client-side versus SSR?

Sam Torres:

If it's SSR, it'll be available in the response HTML. If it's requiring rendering, then that won't be part of the regular training data, it might get pulled into a live fetch at times. They might be pulling that from search results or other, Google being any of those. Yeah SSR, it should all be in the response HTML.

Tory Gray:

Yeah, those are typically the two ends of the rendering spectrum, where it's almost all in the response HTML or almost all in the rendered HTML. And so, on the whole, yes, can they access it or not? Is it in rendered or is it in response? SSR it's probably there and client-side rendering it's probably not. There are exceptions. I've seen client-side render site get H1s, for example, or a meta title specifically in the response. So, you have to audit your site directly but those are the two ends of the spectrum. So, it could be that but that's most likely what's happening.

Erin Simmons:

Awesome. Really great questions. Thanks y'all. All right, we have another one from Myriam. So, when are you launching a little video series helping us navigate this from a technical standpoint? This question is specifically for Sam.

Sam Torres:

Okay. I love you Myriam, so much for calling me out. So, I guess yeah, we'll use this as announcing now. So yes, I'm going to be coming out with a video series called Offscript SEO that's aimed at helping to teach or whittle down some of the complex technical perspectives and concepts. So, mostly targeted at dev questions but kind of the introduction, the fundamentals so that SEOs who want to do more of that type of work, you can have that. So, it is going to be answering questions like, when JairoX shares a GitHub repo, what do you do with it? So, how do you use GitHub? How do you do version control? What can you find in Google Chrome Inspect and the developer tools? How do you interact with Google APIs, whether it's YouTube or the Google Search Console. So yes, that is on my to-do list. I will say I'm really scared-

Tory Gray:

It's been for a bit but you just said it publicly, Sam so ...

Sam Torres:

I know. I will say be gentle, I'm a little scared, I am 100% self-taught developer so there are a lot of times that I don't have formal training, I don't have formal education so be kind. We're all learning together and my goal is actually to learn with all of you and hopefully help clear up some of those questions. But when you're vibe coding you can make better decisions and better products.

Tory Gray:

Yes.

Erin Simmons:

Super exciting. You heard it here first, Unscripted SEO, is that what we are?

Sam Torres:

Offscript SEO.

Erin Simmons:

Offscript SEO. Very exciting. Okay, this will be the last question for us coming from Catherine. So, you mentioned a concept of keeping LLMs hands off your content, could you offer more insight? How and why might you do this?

Tory Gray:

So, this is my content so our comment so I'll take this. Different brands have different feelings about LLMs, right? Certain brands might have issues with the ethics of AI, with what information they're using to train it. There's a lot of horror stories out there. They're taking illegal and unethical actions in order to launch this. So, for whatever business ethics reasons, certain companies don't want LLMs accessing their data. And we're working with some of these clients. So, that might be because you're a big brand and you want your content to be something that people pay for. And then if an LLM gives it away for free and answers all the questions without people having to pay for it, maybe that's a thing that cuts into your revenue for example. So, if you're a community, if you're a media site. There's a reason New York Times and all these big brands X are having a lawsuit against ChatGPT for the use of their content and they want to licence that content. They're not unwilling to share but they want them to pay for that.

So, answering the why, there's a lot of why for different brands, not everyone wants to chase down the visibility piece. Or you might have objections to the degree to which you don't get links. So, what's the point of influencing a result in LLMs if that answer doesn't reference your brand, doesn't drive traffic, doesn't drive brand awareness, doesn't drive value for your business? You just did work and you might not get benefit out of it, right? I think we've seen those click ratios where some of these things are like, "Oh, every 3,000 answers reference any link ever," so what's your customer funnel of what you're driving to your site? So again, why is a big question. The how is, if at least today we know that rendering is required for LLMs, certain brands are already on completely client-side rendered JavaScript stacks, though that content is not available without rendering.

So, if LLMs on the whole can't access that, that is perhaps a barrier to preventing them to do it. Would I change my brand today to be client-side rendered in order to stop LLMs? Probably not because well, that'd be a lot of work and it would hurt me in other channels like SEO and it'd be a lot of work and also again, that might change any second. So, it might be a lot of work for really no gain in the long-term. But maybe there's something secret, confidential, proprietary, like putting that in rendered is the level of security. I think Google also, didn't Google change their SERP results to be rendered and that's why all the tools broke, all the keyword rank tracking tools work all of a sudden? So, Google wanted to stop being scraped so much because that costs money. They're hosting that content that their LLMs are coming through and scraping so they made it rendered to increase the barrier to entry so that that would happen less often and they'd save resources. So Google, they did it and they're doing it today.

Erin Simmons:

Yep. All right, so that brings us to the end of this training course. Sam, Tory, thank you so much for putting this together, answering all these questions. Really excited to get this recording out to everybody and all the folks who couldn't make it. So Jojo, who is behind the scenes who makes all of this happen, does all of the work to get us all in here, thank you so much, Jojo. We'll be sending out an email with a recording, also that survey that directly resulted in these trainings so please fill that out because if you want to get more of these trainings that are going to help you, that's how you do it. And the past trainings for the JavaScript SEO courses will also be included in that. So, thank you Sam and Tory, thank you Jojo, thank you Patrick, and thank all of you for being here and all the great questions, we appreciate y'all. Have a great day.