Transcript
This transcript was created with AI assistance and has been edited for clarity and readability.
Today, we’re going to talk about what we’re seeing in AI search and the biggest questions coming up as people think about how to end 2026 and get 2027 started.
Before I jump into the webinar, I want to introduce Elementive. Some of you already know us; some of you may not. Elementive is a two-person consulting firm. Andrea and I started Elementive 12 years ago, and we focus on the technical side of SEO.
We dig into the technical foundations: how things work, how things are structured, and what we need to fix to improve SEO performance. Right now, a lot of that work is about improving AI search performance and understanding how technical issues affect AI search. Because of how we dig into things, a lot of this also affects user experience more broadly. We end up talking about accessibility, usability, and measurement, all of which we’re going to cover today.
One of the biggest things worth saying about our approach is that Andrea and I feel each client is unique. No two websites or clients are going to have the same problems or look at things quite the same way. A lot of people ask us what a standard project looks like with Elementive, and the truth is we don’t have a standard project. Everything is customized.
That’s partly because everybody has different teams, goals, and ways of seeing things. It’s also because we work with a diverse set of clients across different industries, company sizes, platforms, and countries. With all those differences, you need to customize how you approach technical SEO and the technical side of AI search.
What we are hearing from clients
Even though everybody is unique, a lot of the same questions are coming up right now. People are coming back to the same themes and the same things we’re all struggling with. That’s what led Andrea and me to do this webinar: to talk through those common questions and what we need to think through with each one.
I want to cover what we’re recommending, how people can approach these questions to find the answers that work best for them, and what other people are recommending. We’ll also talk about some of the claims being made and what’s worth experimenting with as we try to figure out this new world of AI search.
The other big category I want to spend time on is the stuff we don’t know. I don’t think that gets enough attention. I get it: none of us want to post something on LinkedIn saying, “Here’s a bunch of stuff I don’t know.” But there is a lot we don’t know. There’s a lot we haven’t been able to understand because new things are happening and there’s data we don’t have access to.
I want to talk about ways we might experiment to understand at least some aspects of those things, as well as claims we shouldn’t quite trust because the evidence isn’t fully there.
How do we measure AI search?
The first question I want to dig into, and the one just about everybody is asking some version of, is how we measure all of this. How do we measure AI search results? How do we measure AI agents coming to our site? How do we get a handle on what value this is adding to our business, if it is adding value at all?
I think there are two problems here. First, AI search is difficult to measure because a lot of data isn’t being shared. Google, OpenAI, and Anthropic aren’t giving us tons of insight into what’s happening on their platforms.
The other reason is that user behavior is different from what we’ve seen in the past. As the funnel changes and people engage with websites in different ways, the measurement tools haven’t caught up. We have to figure out what the new funnel and behavior look like so the measurement tools can start to catch up. That lag makes measuring AI search difficult.
The measurement tools we are using
We’re using a mixture of different tools. Most of the clients we work with and other companies we talk to use some combination of these tools.
The one we keep coming back to, and find a lot of value in, is server logs to monitor bot activity. How many times does OpenAI crawl the site for training? How many times do they try to interact with the site? You can get that information in server logs, and there are interesting insights there.
A lot of companies are also using AI visibility tracking or some version of prompt monitoring. Whether they’re using an existing tool or doing something themselves, they’re trying to understand what the prompts and responses look like.
Both Google and Bing have introduced tools in Google Search Console and Bing Webmaster Tools to track different aspects of search and give you some insight into what’s going on. We see people trying to utilize that data and figure out where it fits.
We also get some traffic from these platforms. With GA4, you can categorize AI assistant or AI referral traffic, and other analytics tools offer ways to do this too. That can give you an idea of what traffic is potentially coming to your website from these platforms.
Then people are trying to figure out what they can do with indirect signals. We know direct traffic is growing, but how much of that is coming from AI search? We see more people searching for our brand, but how much of that is AI? They’re trying to look at those indirect signals and piece things together.
The limits of AI search measurement
All of that gets difficult, and there are problems with each source. Even log files, which we spend a lot of time with and which I think are one of the best tools we could work with, don’t always relate to the end result.
You can say a bot crawled your website: Google came to the site, or OpenAI came to the site. But does that mean the content was included in citations? Does it mean it was used as training data? If an AI agent used it, what were they doing, and did it add value? You can’t necessarily connect crawling to results. They’re too far apart in the process.
Another problem with log files is that AI agents don’t always identify themselves. We don’t know how much agentic activity is happening on our sites. We don’t know how many people are using AI browsers to visit because they aren’t necessarily telling us. There’s a lot we can’t know about how AI is being used when people visit our site.
When it comes to measuring traffic and impressions, clicks aren’t guaranteed. Behavior is changing in a significant way, and I don’t think it’s clear yet if clicks are part of the AI search process. Many AI interactions won’t result in a website visit. That’s partly by design and partly because of how AI tools work and the responses they give.
That raises a question about how much we should use traffic as one of our KPIs for AI search. Is it a relevant metric? Even if we get traffic, attribution is fuzzy. Sometimes traffic arrives directly from these platforms and we can track it. Other times it arrives indirectly. People may see our website mentioned in an AI response and then visit it directly or find it through social media or somewhere else.
The tools from Google and Bing and the visibility tools all have places where they may fit, but they’re incomplete. At best, they’re showing patterns, trends, or some of the available data. They aren’t giving us the complete picture.
I think that’s difficult because AI search is heavily personalized. The prompts themselves are different, and the responses are different too. It’s hard to capture all of that in a tool. That’s where I don’t think the measurement tools have fully caught up to the new behavior, and we haven’t quite figured out the right way to report on it.
Defining success and combining signals
The big open question everybody is trying to get to is how we define success with AI search. What does that look like? What value is AI search driving for the business? Is it only about traffic to the website, or are other things potentially valuable?
We see clients stitching together different measurement sources and trying to figure out how the pieces fit together to tell a story. We can look at log files, visibility tools, and data from Google and Bing. Maybe we’ve talked to customers about what they’re doing. Together, that tells us a story about how people use AI search to work with our business. From that, we can start to see where there may be value.
One concrete piece of data I find valuable, and that people keep coming back to, is using log files to confirm bots can reach the content. Can they reach content you want included in training data? Can they engage with content on the website?
Tracking trends and comparing channels
As you bring all the data together, the next step that we see people moving into, and that seems to make sense, is tracking changes over time and looking at them directionally.
It’s not just looking at quantitative growth, such as a number being up 12 percent month over month. Rather, it’s saying, “We’ve stitched together all these data sources, and they’re generally moving in this direction. It generally seems positive.” We have to be more qualitative in how we bring these data points together because we don’t have concrete, exact numbers in this case.
Another valuable experiment is comparing AI search with other traffic channels. How do the trends in the AI search data we have compare with branded search, direct traffic, conversions, the business overall, traditional search results, and social media?
Partly, that’s a check on our interpretation. If AI search looks one way and everything else looks different, maybe our story about AI search is wrong. But AI search can also influence those other channels. That’s something we see people trying to understand: how AI search is trending alongside the website and the other channels we’re pursuing.
How much schema do we really need?
Another question people are talking about, and showing curiosity about, is schema or structured data. There are claims about what schema can do for AI search and how it can improve GEO performance.
There’s reason to think schema could be valuable here. But the important thing is that we’re talking about using schema differently from how we’ve traditionally thought about it.
Using schema to explain entities and relationships
Traditionally, we’ve thought about schema as, “I add this set of schema to my website. I add these properties and types, and I’m going to enhance my search results.” It’s easy to take that mindset into AI search and say, “If I add these properties and types, I’ll get a different appearance in an AI response. I’ll get a different citation or mention. Maybe I’ll be cited more.”
I don’t think that’s quite the right mindset from what we’ve been seeing and how people are using schema. A better question is, “We added this schema to our website. Does it help a bot or agent from a search engine or AI system clearly understand what’s going on? Does it help them understand what my website is talking about and how the different things relate to each other?”
That’s a different way to approach schema, and it requires a different process to bring schema onto the website. Start with schema in the traditional way: look at the types and properties Google supports and how they can influence search results. Go ahead and do the things that support traditional search.
Then you can build on top of that. If you want to approach schema from an AI search standpoint, look beyond Google’s supported types. Look at Schema.org and all the types available. Ask, “What fits my content? What are the different ways I can describe things?” We’re seeing people expanding the schema they use and how they mark up their site’s content.
Connecting schema across the website
The next level is to say, “Now I have all these schema types across my pages, but how do I connect the pages?”
The traditional way we thought about schema was that each page was independent. We would mark up everything on that page: information about the product, blog, or company. Now, a lot of people are trying to connect those pages and build an entity map across the entire website using schema.
If you move in that direction, expanding beyond Google-supported types and connecting schema across the website, one of the biggest challenges is keeping it current and consistent. It starts to get involved when you have that much schema code. That’s another challenge as we think about approaching schema within AI search.
What to test with schema
There’s a lot going on with schema, but the biggest question people keep coming back to is how much of this is needed. Will it help? Is it worth adding more types to the website? Will going through this whole process matter?
There are claims that schema can help and do a lot of things, but there isn’t necessarily much data telling us that’s the case. There also isn’t an official statement from OpenAI, Google, Anthropic, or Meta saying you need schema on your website for AI search. That leads to questions about where schema fits into the process.
We see clients testing different markup and asking what other schema they can bring into their page types. We also see some going in the other direction. They’re saying, “We don’t know why schema would work here, and nobody’s officially saying we need it.” They’ve taken schema off, or almost entirely off, and are sticking with Google’s supported types if they have schema at all.
They want to see how well-written content, good site structure, sound internal links, and technical foundations perform without worrying about additional schema. Is that enough? Do we need schema and entity mapping, or can we perform fine without it? You can test in both directions and see what helps.
If you are testing schema, one of the most interesting tests is entity mapping across the website. That moves beyond the traditional, page-specific approach and looks at how to connect pages and build relationships across the site.
We launched a test on my blog a couple of weeks ago to build an entity map. It’s still early, and we don’t have results yet. But it’s interesting to test and see whether it helps.
Measuring the effects of schema
Then the question becomes: we’re doing all these tests with lots of schema, different approaches, or no schema at all, but how do we know if it’s working?
There was that Ahrefs study from a while back where they found that more schema didn’t translate into more citations. Okay, maybe it didn’t add more citations, but does schema affect other things? Does it change how an AI agent uses the website? Does it change how bots crawl? Does it change the context of a citation?
Maybe you don’t have more citations, but you have a better citation or a better mention in an AI response. There are interesting things to try to tease out there.
The truth is that we don’t know yet where this is helping. There’s no conclusive answer about where schema helps AI search. Some things anecdotally seem like they may have helped after adding schema, but we’re still experimenting. This is a good area to include in your plans to figure out where, and if, it can help.
Which bots and agents do we want to support
The next question is which bots and agents we need to support. To me, one of the most interesting things about AI search is that there are so many types of visitors coming to our websites beyond the human visitors and search engine robots we’re used to.
We still have traditional search crawlers. Googlebot still visits to index content. But now other automated visitors do other things. AI training crawlers gather training content. User interaction bots do small tasks for people, like summarizing content.
There are also AI agents and AI browsers. An agent can visit a website autonomously, while an AI browser can assist a person visiting a website. That’s a different type of automated visitor from what we’ve thought about in the past.
There are commercial and third-party crawlers from other tools too, accessing websites for different reasons. With all these automated visitors, the biggest question is which ones support my business goals. Which will help me the most?
A framework for deciding what to allow
We’ve been working with clients on a framework for deciding which automated visitors are most important, which ones we want to allow into the website, and which ones we need to intentionally build the website to support.
The first question is where the bot or agent comes from. I’ll have a different answer about supporting a bot from Google than one from a company I’ve never heard of. The more you know about the company, and the more important that company is to you, the more you may want to design the website intentionally for its bot or agent.
Once you know which company is behind it, the next question is what the bot or agent is trying to do. In the past, we mostly thought about bots as single-purpose: Google came to index content. Now, some bots are for training, some are for interacting with the site, some are for indexing, and some are looking for specific information.
Agents are trying to complete tasks. They can do a lot of things that previously only human visitors could do. You want to understand the different things a bot or agent might try to do on your website.
The next question brings this to business value. We know what the bot or agent is trying to do and where it’s coming from, but does allowing it into the website help? Does it add value, and in what way?
If I allow a training bot in, the potential value could be that inclusion in training data leads to more mentions or citations, gets my brand out there in a positive way, or gets me customers. There’s potential value, but that isn’t always the case. It depends on the bot or agent.
To dig further, ask what the bot or agent needs access to. We need more than a high-level answer about accessing the website for a certain purpose. To obtain that value, which specific pages do we need to let it see? Which tasks do we need to let it complete? Understand what it needs access to in order to do what it came to the website to do.
Server capacity, privacy, and security
One of the biggest things to be careful about is the impact on server capacity. Bots and agents can be aggressive when they visit. They can fetch a lot of pages at once, which can negatively affect your server if you aren’t ready.
Historically, Google has been fairly polite about crawling. It’s good at understanding when it’s coming up against a server threshold and when to back off. That isn’t necessarily true of all the AI bots. Keep that in mind: we want a bot or agent to do certain things, and it needs access to files to do them, but accessing those files could create problems. What do we need to change to support that access?
Another question is about privacy and security concerns with whatever the bot or agent needs to access. This is especially true with AI agents. On an ecommerce website, maybe it’s okay for an agent to add things to a cart. That could help a human visitor. But do you want the agent completing checkout, entering information, and clicking the final purchase button? Is there a concern there?
The same is true of an agent accessing an account area where it can change passwords or access personally identifiable information. Traditionally, these aren’t things we’ve had to think through from an SEO standpoint. Now, as we work with agents and build websites for them, they are becoming part of the GEO and AI search optimization conversation.
Can bots and agents see the content?
Another big question is whether bots and agents can see the website’s content. You’ve decided to allow them in and figured out how to support them, but do they see the content correctly?
A lot of AI bots and agents struggle to execute JavaScript. They might execute some pieces but not everything, which can give them an incomplete picture of the website. Sometimes that problem is more dramatic than others.
It’s important to step through how each bot or agent interacts with the site. They have different capabilities for accessing and using a website. For the ones you want to support, understand those capabilities and how they relate to the site’s rendering, functionality, and tasks.
Testing bot policies and agent tasks
There’s a lot to experiment with here. The biggest question driving these experiments is where automated visitors add value. There are also questions about where customers prefer to use them and how customers want AI visitors to help them access the website.
That includes evaluating robots.txt policies and WAF rules. How are we controlling access for different bots and page types? You have to go bot by bot and page type by page type to figure out which rules make sense. There’s a lot to experiment with and fine-tune.
When it comes to AI agents, we don’t know how many are accessing our websites because they don’t always identify themselves. We also don’t currently have a way to test at scale how agents can use our websites. That means manually asking an AI agent, or using an AI-assisted browser, to go through the website and seeing whether it can complete the tasks.
As you do that, consider which tasks make sense for agents to complete. Am I comfortable with an agent doing this task, even if it can? If it can complete a task but I don’t want it to, what do I need to adjust to prevent that? Those are questions we see people grappling with.
Testing rendering and emerging standards
Another big thing to test is rendering on important pages. Confirm whether bots and agents can reliably view each page. We’ve spent a lot of time with clients going page by page to make sure each bot or agent that needs access can get it successfully.
There are also emerging protocols like WebMCP, the Web Model Context Protocol, to assist agents in completing tasks. There are llms.txt and agents.md files and other Markdown files too.
There are still questions about whether these will be used and whether they will help. They’re at various stages of development, and the standards are evolving. There’s a lot to figure out about whether you should use them and whether they will help bots or agents use your site.
We see people experimenting with these. If you aren’t already experimenting with some or all of these formats, I think that’s worth considering as you plan for 2027.
What content matters in AI search
The next question is what content matters in AI search. One important distinction is separating content from information. Information can live in different content formats, and we want bots and agents to access the information they need regardless of which formats they can use.
We have text, video, images, and audio. We also have content off the website: social and local profiles, user-generated content on forums like Reddit, directory listings, and other third-party profiles about the company.
Those give bots and agents different ways to access information about your business. One change from what we’re used to is that the website is no longer the only source that matters. It may not even be the primary source of discovery. Bots and agents may get information from other places instead of the text on your website.
How do you shift the website to support content discovery across those platforms and content types?
Making information accessible and discoverable
From a technical standpoint, the first question we’re helping clients work through is how to make information accessible to bots and agents. Regardless of the content format, what important information do they need to access and understand about this business, product, or service?
This is an area where schema and structured data may help. They may make it easier to surface key information from text, images, or videos so bots and agents can access it.
The other part is making that information discoverable. How do you make sure bots and agents can find different content types and access information about the business? That includes figuring out how those content types fit into the site’s hierarchy.
With so much content and so many ways for bots and agents to access it, there are questions about maintaining continuity and consistency. It’s easy for information to get outdated and for old information to remain available. That can skew training data or affect an agent’s ability to complete tasks.
Pairing formats to provide context
One part of this we’ve been working on is pairing formats to provide more context. If you have a video, make sure you have a good transcript alongside it. If it contains key visuals, like an infographic or graph, pull those out as separate images so bots and agents can access them too.
The same applies to images themselves. Use good alt text and surrounding text to describe what’s going on. Don’t depend on the image alone to express all the information it contains.
For audio, provide a transcript and supporting text or imagery. Give bots and agents multiple ways to access the information.
Experiments with content formats and consistency
There are questions here to experiment with. How do these formats fit together? What formats do we need, and what’s the right combination to help bots and agents understand what we’re doing?
We see people experimenting with video, audio, images, PDFs, and other formats within the website. Where do those help, and where do they work? It isn’t only about training data or indexing. Does an image, PDF, text explanation, or video help an agent complete a task on the website? We’re seeing interesting tests around that.
There are also questions about repurposing website content on external sources. Some people are trying an approach where the website is the base for all the content and everything on external sources is an offshoot. That leads to experiments about how the website connects to those sources.
How much overlap should there be between what you say on social media and what you say on the website? Does more overlap or less overlap work better?
Some clients are also figuring out how to build AI systems to monitor this and keep information consistent across the website and the platforms and profiles that discuss the business. How do you make sure everything points in the same general direction and that the information isn’t outdated, incomplete, or inconsistent?
What does AI search mean for brand?
The last question we’re hearing from clients is what all this means for our brand. Where does branding fit in? Branding goes beyond the technical side we focus on, so I’ll cover the things we’re thinking about from a technical standpoint.
I think brand comes up so much because, as clicks become less predictable, we’re not sure how much of this is about traffic. Is brand recognition part of the value AI search can deliver? Does being found in an AI response build trust? Where can the website support that?
This gets to the idea that the user journey has changed. People are using AI chat first to figure out which companies to work with. AI responses can shape the companies they’ll consider long before they visit the website, if they visit at all.
Signals worth watching
There’s a lot to watch. How is branded search demand changing? Do we see more people searching for our brand in Google? Do we see more direct traffic?
Direct traffic is a big question mark. People make claims that growth in direct traffic must be branded traffic from AI systems. Not necessarily. Direct traffic can include other things. You have to dig into it and ask where it could be coming from an AI search result and where it could be something else.
Log files become interesting here too. How often are bots and agents crawling and interacting with information about your company? Do they go to the About pages or staff pages? How frequently? What else are they looking at, and how are they finding those pages?
Visibility tools can help track where the brand is mentioned and how it’s described relative to competitors. Monitoring prompt responses, even with sampled and incomplete data, can still give you a valuable idea of how people are being told about your company.
Questions about how AI describes your brand
Even with limited data, there are questions to investigate about how consistently AI systems describe your company, products, and services. What strengths and weaknesses do they talk about? We see AI tools saying some critical things about companies in prompt responses.
Look at the mistakes too. Where is the brand described incorrectly? Where is outdated information being used? One of the most interesting questions is where that information comes from.
In some cases, the LLM has made it up. In other cases, the information exists somewhere. If you dig through the sources and trace it back, maybe you can find the source of outdated, inconsistent, or incorrect information and correct it.
Helping bots and agents understand your brand
The biggest open question around branding is how you use your website to improve bots’ and agents’ understanding of your brand. They interact with so many platforms and profiles. Where does the website fit in to provide complete information about who you are and what you do?
One experiment clients are trying is enhancing their existing About information: adding details about staff, company policies, reviews, and reputation. This isn’t completely new with AI search. Google has been talking about E-E-A-T for years: experience, expertise, authoritativeness, and trustworthiness. How do you build information on the website about who you are and what you do, and establish your authority?
The more interesting part is how you integrate that content into the website’s hierarchy. Don’t just put a link to the About page in the navigation and assume that’s sufficient. Where else can you bring that information into pages and integrate it into site content? Schema also has potential to connect some of this information.
We see people experimenting with llms.txt files or Markdown files to explain their brand. They use these to clarify brand questions or things bots and agents might misunderstand about working with the company, product, or service. What do they need to know that they may not get directly from the website?
As with the other emerging standards, the results aren’t clear yet. But I think that is interesting to test, especially if you’re already testing something like an llms.txt file.
Auditing brand consistency
Another thing I think everybody needs to do, and that we see people starting to do, is audit how the brand is described everywhere it’s discussed.
Start with the website, because it’s probably one of the key sources of information about who you are and what you do, if not the primary source. Also review social platforms and local profiles.
Some clients are figuring out how to use agents for these brand consistency assessments, alongside manual audits. It’s interesting to think about how AI could help us understand this information and correct some of the problems.
Best practices, experimentation and revisiting assumptions
There’s a lot going on with AI search, but the biggest takeaway across these questions is that there are things we know. There are solid best practices we can fall back on: good content foundations, things put together well, and technical fundamentals taken care of.
Once you have those best practices in place, there’s a lot to experiment with and questions to investigate to find what works best for your business.
There isn’t one answer for how to optimize for AI search yet. I see people sharing what they call a definitive checklist for getting citations or mentions, but that doesn’t exist in that way. There are different things you need to figure out, and a lot we don’t know that requires experimentation.
The clients we’re seeing achieve the best results in AI search responses and with agents have strong technical foundations and good content. They’ve thought through how bots and agents need to access the website and engage with the content.
They’re also moving away from needing exact numbers and toward qualitative stories that bring together different signals. Another common trait is that they don’t simply accept claims. They test them. Can we trust this claim? Can we find evidence to back it up? Can we experiment deliberately to see whether this helps?
We need to learn from tests that work and tests that don’t. We also have to keep revisiting everything. An experiment that failed two months ago might succeed now because things have changed.
We see clients repeating experiments and revising assumptions as platforms change. That helps them stay current rather than relying on something that worked a few months ago.
Questions and next steps
If you have questions or need help with any of this, that’s why we’re here. We focus on these technical issues to help clients optimize for SEO, GEO, and AI agents. Feel free to contact me with questions or if you need help.