AI Visibility Isn’t Enough: How AI Decides What to Recommend

Imagine you’re the VP of marketing at a software company. Your marketing team is struggling to hit its growth targets, and you’re trying to figure out what to do about it. You could hire an advertising agency. You could bring in a consultant. You could hire another full-time employee, or you could try to solve the problem with the team you already have.

So you go to ChatGPT and explain your situation. You tell it about your company, your marketing team, the problem you’re facing, and what you’re trying to accomplish. Then you ask what you should do. ChatGPT recommends hiring an agency.

Now imagine we change a few details. Instead of a team that knows what needs to be done but doesn’t have the capacity to execute, you’re a team that’s struggling to figure out why its marketing isn’t working in the first place. You have the capacity to take on more work, but you don’t know what work you should be doing. This time, ChatGPT recommends hiring a consultant.

Both are perfectly reasonable recommendations. But something interesting just happened.

We didn’t ask ChatGPT to rank agencies against consultants. We asked it to help solve a business problem. It interpreted the situation, made some inferences about what the company needed, and recommended a particular type of solution. And depending on which recommendation it made, an advertising agency might have gained or lost a potential customer before that customer ever started researching agencies.

This is already playing out in real purchasing decisions.

In its 2026 AI Search Insight Report, G2 surveyed more than 1,000 B2B software buyers and found that 69% reported choosing a different software vendor than they initially planned because of guidance from an AI chatbot. A third purchased from a vendor they hadn’t previously heard of. Those numbers matter because we’re no longer just talking about AI helping people find information. We’re talking about AI influencing how people evaluate their options and decide what to buy.

Which raises a question that I think marketers need to spend a lot more time thinking about: What makes an AI recommend one company over another?

The difference between being found and being recommended

For most of the internet’s history, marketing has been concerned with getting information in front of people. That’s the basic idea behind search engine optimization. You publish content, Google indexes that content, and you try to understand the signals that determine where your website appears in the search results.

Of course, Google is doing a great deal of evaluation itself. It’s looking at relevance, authority, quality, and a whole collection of signals that determine which pages deserve to rank. But there’s an important distinction between a search engine giving you a list of resources and somebody taking those resources, evaluating them, and coming back with a recommendation.

Think about what happens when you ask a colleague to research a software platform for your company. They don’t just return with ten links and tell you to figure it out yourself. Ideally, they read the information, compare the options, think about your company’s particular needs, and come back with a recommendation. They might say, “I looked at these five companies. These three aren’t really built for what we’re trying to do. This one is probably more than we need. I think we should go with this one.” That’s a fundamentally different service from simply finding information. And it’s much closer to what large language models are doing when people ask them for advice.

In fact, this is increasingly how AI companies themselves describe their products. Anthropic’s research agent can conduct multiple searches, use earlier findings to decide what to investigate next, and synthesize the results into a response.

AI is less like a search engine and more like a researcher with a point of view. And that introduces a whole new problem for marketers.

Let’s say you’re a software company and you’ve invested heavily in AEO, or Answer Engine Optimization. You’ve made sure your website is accessible to AI systems. You’ve structured your content properly, implemented the right metadata, and published articles addressing the questions your customers are asking. That’s all useful work. But what happens when ChatGPT finds your content, draws the wrong conclusions about what your company does or who it’s for, and recommends a competitor to a customer you would have been better suited to serve? Your visibility strategy worked. The model found you. It just misunderstood you. And if all you’re measuring is whether AI knows about your company, you’re missing the most important part of that interaction.

What did the model conclude about your company? Who does it think your product is for? What does it consider your strengths and weaknesses? What assumptions is it making about the customer? Those aren’t really search optimization questions anymore. They’re psychology questions.

What does AI think about your company?

For decades, researchers have studied why people recommend certain products to one another. Robert Cialdini studied the psychology of persuasion. Jonah Berger studied why people talk about and recommend particular products. Marketers have spent enormous amounts of time and money trying to understand these behaviors because recommendations are incredibly valuable. Now we have machines making recommendations to people, and we’re largely approaching them using the same playbook we developed for search engines.

I think that’s missing an enormous opportunity.

What people tell a model about themselves changes what it recommends. Researchers at Carnegie Mellon University studied whether the information users disclose about their identities affects the recommendations they receive from language models. They found that it does, and that models can produce stereotyped recommendations based on identity cues, including information the user didn’t necessarily intend to make relevant.

In another study published at EMNLP in 2025, researchers experimented with different persuasion techniques in product descriptions. They found that social proof could improve a product’s recommendation ranking, while other familiar tactics, including scarcity and exclusivity, could actually make the product less likely to be recommended.

Think about what that means.

LLMs learned from human writing, but they don’t always respond to stimuli in human-like ways. You can’t just take a playbook from consumer psychology, apply it to LLMs, and assume you’re going to get the same outcome. LLMs have their own unique patterns of behavior to be studied. The same principle applies to how LLMs conceptualize entire markets.

Imagine asking ChatGPT to describe the average customer of Nike versus the average customer of Adidas. The model may have different associations with those two brands. It might associate one with serious athletes and the other with casual runners. Those associations may or may not accurately describe the products themselves. But if those associations influence which company the model recommends, they have commercial consequences.

Now imagine we systematically vary the information a customer discloses about themselves. They’re a beginner runner. They’re an experienced marathon runner. They’re looking for the best possible performance. They’re looking for something affordable. Under which circumstances does ChatGPT recommend Nike? When does it recommend Adidas? Which details about the customer actually change the recommendation? And once we’ve identified those patterns, can we understand what produces them?

That’s the kind of question I’m interested in answering. Not just what AI says, but under what conditions does it say one thing instead of another?

Turning AI behavior into something measurable

Instead of sitting down with ChatGPT, asking a few questions, and making assumptions about how it behaves, we can construct experiments. We create large numbers of hypothetical situations, systematically change particular details, and send those situations to different language models. Then we collect the responses, classify them, and analyze the results.

It’s a little bit like conducting survey research, except instead of surveying a thousand people about what they would recommend, we’re surveying language models about how they would handle a thousand different situations. By varying those circumstances systematically, we can begin to map the conditions under which models favor particular solutions. We can investigate how they conceptualize categories, what variables they use to differentiate companies, what they infer about customer needs, and which circumstances change their recommendations.

And that gives us something much more useful than a collection of anecdotes about what ChatGPT happened to say. It gives us a way to investigate the behavior itself.

Understanding comes before optimization

For a business, the important question isn’t simply whether AI mentions your company. It’s whether AI understands your company well enough to recommend it to the right customers.

Suppose an LLM consistently recommends your competitors to customers you would be better suited to serve. There are a number of possible explanations. Maybe it has outdated information about your product. Maybe it’s associating your company with a customer segment you no longer serve. Maybe it understands your capabilities but is making an incorrect inference about what the customer actually needs. Those are different problems, and they may require different solutions. But until you understand what’s causing the recommendation, you’re effectively guessing about what to optimize.

This is what I think is missing from a lot of the current conversation around AI marketing.

We’re spending enormous amounts of effort measuring whether brands show up in AI answers, without spending nearly enough time investigating the judgments that produce those answers in the first place.

The point isn’t that visibility doesn’t matter. It does. The point is that being visible to AI and being understood correctly by AI are two different things. And understanding that difference changes where we need to begin.

Why Construal Research exists

The reason I started Construal Research is that I think we’re entering a world where businesses need to understand not only what their customers think about them, but what the AI systems advising those customers think about them.

Traditional SEO and AEO still have important roles to play. Making your content discoverable, accessible, and understandable to machines is a necessary part of the equation. But there’s another layer beyond visibility. How does AI conceptualize your category? How does it distinguish between your company and its competitors? What does it infer about a customer’s needs? Under which circumstances does it recommend your product, somebody else’s product, or a completely different way of solving the problem? Those are the questions we’re building Construal Research to investigate.

Machine psychology is the study of how AI systems interpret information, evaluate alternatives, and make decisions. I tend to think of it as behavioral science applied to machines. We’re especially interested in how these behaviors shape recommendations that influence commercial decisions.

The amount of progress we’ve made in the past three years is astonishing, but there’s still an enormous amount of work to be done here. We’re still learning how to investigate these systems, and I think the right response to that uncertainty is curiosity.

AI is changing how people make decisions and how companies compete for customers. And there’s still a great deal we don’t understand about what happens between the moment somebody describes their problem to an AI and the moment that AI recommends a solution.

Those are questions we can investigate.

That’s the purpose of my research.