To Get Cited by ChatGPT and Other AI Search Engines, Brands Need to Rethink How They Create Content: Insights from Joshua Squires of Amsive
More than a year ago, I interviewed Joshua Squires and we talked about how AI would change the way people search online.
Back then, it was a prediction.
Today, it’s a reality: millions of people ask ChatGPT or Google’s AI Overviews who to choose, what to buy, and who to trust.
And your brand? Is it being cited, or is it off the radar?
Every now and then, as part of SEO Confidential, I like to revisit experts I’ve interviewed in the past to see how things have changed and evolved (or gone backwards!) in the world of SEO and AI Search.
This time, I’m catching up with Joshua Squires, Director of AI Search at Amsive. Josh has been working in SEO for 17 years and studying machine learning since 2016. He explored transformers from the moment Google introduced BERT. He’s worked with startups and Fortune 50 companies, from e-commerce to insurance, and today he leads a team of ten SEO specialists. In short, he knows what he’s talking about.
This interview covers topics that matter to your day-to-day work.
Why focusing on a single channel is a mistake.
How to create content that LLMs choose to cite.
Why AI monitoring tools are like studies on mice in pharmacology.
And how to build test accounts based on your personas to see what your customers actually see.
If you want to become the answer AI recommends in your industry, dive in. You’ll find plenty of useful tips.
“AI monitoring tools are like studies on mice in pharmacology: useful, but sometimes far removed from reality”
Josh, about a year ago, we talked about how artificial intelligence would change the way people search for information online. Today, AI Search has become a reality: millions of people use AI Overviews, ChatGPT, and other answer engines. In your opinion, what is the most important change that professionals and businesses need to understand today, but that many still underestimate?
I think many people are still stuck in a single-channel mindset. I see a lot of recommendations for tactics and strategies to apply within individual channels. But with large language models (LLMs) now influencing a significant portion of the audience, we should all be thinking in terms of multichannel strategies.
No single channel can cover all the touchpoints an LLM uses to formulate an answer. Brands should identify the most important questions and figure out how to answer each one across multiple channels, at different stages of the buying journey, and through different communication mediums.
This shift requires a fundamental rethink of how most marketing teams operate. It involves changes to organizational structures and internal processes, which are often particularly difficult to implement.
Many SEO professionals still create content assuming that AI retrieves the best page for a given query. In reality, RAG systems seem to select individual passages of text to build the final answer. How important is it to understand this difference, and how should it change the way we structure content?
I think this is very important. I would add that LLMs are increasingly drawing on parametric memory as well. When creating content, SEO, AEO, and GEO professionals need to take these factors into account and structure their texts to address each of these needs.
To optimize content so that LLMs can retrieve it, we need to consider both people’s needs and the needs of machines. LLMs already have an idea of what an answer should look like, and then they look for sources that confirm it, add nuance, and help establish a consensus. People, on the other hand, still need readable content that genuinely solves their problems and is easy to understand.
The point, then, is not so much to create the best page overall, but to provide the most suitable source of information for each individual user’s needs.
To achieve this, content creators need to understand their audience in depth and include relevant nuances in their texts, making the information more relevant to the groups of people they want to reach. They then need to organize the content into well-structured documents that LLMs can analyze quickly and effectively.
Understanding how different types of attention mechanisms work can help bring all these elements together.
In your opinion, what is the biggest misconception among SEO professionals about Query Fan-Out, and what approach should they take to create content that RAG systems can genuinely retrieve?
I think too much importance is placed on the topics covered by the questions generated through query fan-out, while too little attention is paid to users. It’s not necessary, and often not even possible, to answer every related question. Answering all these questions can help create comprehensive content, which may be useful for some audiences. Others, however, may be industry experts who find answers to basic questions unhelpful.
Look at the queries generated by fan-out and ask yourself which ones are genuinely relevant to your audience. Consider which questions might require slightly different answers for your users than for a broader audience or people with different characteristics. Also consider which questions need to be addressed specifically to reach and engage the audience you’re targeting.
You suggested starting with Google Search Console data and “People Also Ask” questions to build prompt clusters that more closely reflect real user behavior. Why do you believe this method is more reliable than simply tracking a list of keywords?
PAA (People Also Ask) are the related questions Google shows users after they perform an initial search. Essentially, Google shows you a set of questions connected to the original query. If you use Google Search Console (GSC) to identify a primary query, PAA gives you an excellent indication of the information your audience needs.
PAA often reveal the most common ways people phrase their questions, which is particularly important for LLMs because changing the wording or phrasing a prompt differently can produce different results. Knowing which formulations are most common among your audience helps you understand which answers users are more likely to see in AI visibility tools and allows you to reach the broadest relevant audience possible.
PAA can be combined with queries generated through query fan-out to create a richer, more comprehensive list of relevant questions to answer.
Many tools for measuring visibility in AI search engines are now emerging. How much can we trust the data they provide? And what features should a good monitoring system have to accurately represent how people actually use ChatGPT, Gemini, or Perplexity?
I would say these tools are the equivalent of studies on mice in pharmaceutical research. We can gain some useful insights from them, but the data and interactions we observe probably differ from what our actual users experience.
The reason is that AI visibility tools cannot adequately account for personalization factors, such as geographical location, chat and search history, installed apps, and so on. Including these variables in a prompt does not reproduce the way they influence behavior within the application.
Both Google and Bing now offer some features for monitoring visibility in AI-generated search results. We also have web analytics data at our disposal. Interviewing your audience and conducting studies and surveys is probably still the best approach, although these activities can be difficult for many businesses to carry out.
This is precisely why AI visibility tools still have a role to play.
I recommend that brands create dedicated accounts for different personas, using virtual devices and free accounts signed in to the relevant major LLMs. These accounts should have browsing histories and conversation histories consistent with the target audience. This makes it possible to check and evaluate the results returned by AI visibility tools and compare them with results obtained in more realistic contexts.
Some publishers, as reported by the Wall Street Journal, are beginning to see AI crawlers more as a cost than an opportunity. Do you think more websites will start restricting access to their content, or will new models of collaboration with AI platforms emerge in the coming years?
This is already happening, but mainly in sectors where LLMs’ acquisition of content undermines the profitability of websites. Online publishers are a particularly clear example.
Payment models based on the number of content crawls (pay-per-crawl) already exist. Cloudflare also seems to be investing heavily in developing technical solutions to the problem.
I think the situation will eventually find its own resolution as the technology evolves. However, significant innovation in the publishing industry may be necessary, and we have yet to see that materialize.
If models increasingly feed on content generated by other AI systems, do we risk creating a self-sustaining web that becomes increasingly uniform and less original? Is model collapse a genuine risk, or is it being exaggerated today?
It’s a fairly broad topic. Some researchers argue that model collapse is already happening. Others believe that, rather than producing negative effects such as hallucinations, we may primarily see the spread of mediocre, standardized content in which nuance is lost. Other research suggests that biases in responses are increasing.
One of my favorite recent pieces of research was this article, which hypothesizes a gradual decline in the effectiveness of the RAG (Retrieval-Augmented Generation) process, meaning the generation of answers based on retrieving information from external sources: here is the study published on alphaXiv.
The interesting thing is that even experts don’t seem to have any definitive answers yet. To return to the example of studies on mice, the way LLMs behave in the real world seems to differ somewhat from what we observe in the laboratory.
Apparently, only time will tell.
Several studies show that some AI models can alter the meaning of a text when rewriting it, particularly on politically sensitive topics. Is this a technical problem that will eventually be solved, or do you think complete neutrality is effectively impossible?
I’ve often wondered how much of this problem stems from content and comments generated by online bots, and how much is due to the current climate of “relativity of facts,” in which different groups believe different things to be true, sometimes regardless of what demonstrable facts clearly show.
Perhaps LLMs are struggling to deal with a reality that we ourselves have distorted. I think this reflection connects to the problem of model collapse, but I also believe this question is likely to remain unanswered for now.
Research in this field has yet to provide definitive answers. Perhaps this is a good time to remind readers that AEO and SEO professionals are not scientists or engineers specializing in artificial intelligence and machine learning. Answering this question would require a much deeper understanding of the technology than our industry can currently provide, in my view.
Want to Win Over LLMs? Your Content Needs to Be Useful to People and Understandable to Machines
Well, I told you Josh Squires would have some valuable insights to share.
First: stop thinking in terms of a single channel. No single channel covers all the touchpoints an LLM uses to generate an answer.
Ask your SEO agency for a strategy that goes beyond your website: which questions matter to your business, and where should you answer each one at every stage of the buying journey? This also means getting teams that currently work in silos to collaborate. It’s not easy, but there’s no alternative.
Second: don’t try to answer everything. Query fan-out generates dozens of questions. Find out which ones are genuinely relevant to your customers and why. Create content for them, not for some imaginary average reader.
Third, and this is important: personas.
AI search visibility tools use prompts without history, location, or habits. That’s not your customer.
Josh recommends creating accounts for two or three real personas, each with a consistent history, and comparing the results with the reports from your tools. If they match, great. If they don’t, the tool is showing you a version of reality your customers may never see.
That’s all for today. A huge thank-you to Josh for his honesty and willingness to share his insights, and thanks to you for reading this far. I’m sure you found this interview useful!
See you next week, right here on SEO Confidential.
