AI Citation Science: How LLMs Decide Which Brands to Recommend

AI Citation Science: How LLMs Decide Which Brands to Recommend

Here's a question for you: When you ask an AI like ChatGPT or Google AI to recommend a product or service, how exactly does it decide which brand gets mentioned? If you're like most people, you've probably assumed it just scours the web and picks the most relevant or popular option. Makes sense, right? But is that really what's happening? Not entirely.

In my experience running Contxt, a platform designed to measure whether AI systems are recommending your brand, most businesses are completely in the dark about how LLMs (that’s large language models, for anyone still catching up) make these decisions. And for some, that lack of understanding is costing them big.

We’ve run 5,493 AI prompts across 50 businesses, spanning industries like manufacturing, retail, entertainment, and technology. What we’ve found is eye-opening: 37% of the time, ChatGPT and Google AI give conflicting brand recommendations for the exact same query. That means a brand visible on one might be invisible on the other, often without realising it. If that doesn’t scream "missed opportunity," I don’t know what does.

What Is AI Citation Science?

Let’s break it down. AI citation science is essentially the study of how AI systems recognise, recommend, and cite brands in their responses. It’s not just a matter of scraping search rankings or counting backlinks. LLMs use structured data, entity recognition, citation footprints, and editorial patterns to determine which brands "deserve" to be mentioned.

Think of it like this: The AI gathers signals from various corners of the web, your website, metadata, product reviews, third-party articles, and even competitor comparisons. Related reading: this piece about why geo outperforms seo for ai visibility: lessons from 4.... Related reading: our article on how enterprises can achieve llm visibility breakthroughs .... Then it tries to piece together whether your brand is relevant to the user’s query and whether it’s worth recommending.

But here’s the catch (and where things get messy): Not all signals are weighted equally, and different AI models prioritise different factors. Across our dataset, we saw ChatGPT recommending brands 73.8% of the time, while Google AI Overview lagged at 44.6%. Related reading: From Users to Algorithms: How AI Search Personalisation i.... Related reading: our article on why 39% of ai systems disagree on brand recommendations (.... Perplexity and Claude were even less likely to make recommendations, sitting at 40.2% and 62.6%, respectively.

The Web vs AI: Why Mention Rates Diverge

Here’s what I keep seeing: Just because your brand is everywhere on the traditional web doesn’t mean AI will pick up on it. Case in point, DUO, an e-commerce brand in our dataset, has a mention rate of 0.0%. Not one AI system recommended them across 5493 prompts. On the flip side, Oak&More, a furniture retailer, hit a perfect 100% visibility rate across all systems. So what’s going on here?

It comes down to signal clarity. If you’re scattering breadcrumbs across the web but failing to establish a solid entity in AI-readable formats, you’re practically invisible to these systems. Structured data, like schema markup, plays a huge role in ensuring LLMs understand who you are and what you offer. Related reading: this piece about content, links, technical: the seo trinity took 25 years .... You can AI visibility tracking features to see for yourself. DUO, despite its broad web presence, clearly isn’t ticking the right boxes here.

Our experiments show that even query intent creates wild visibility swings. For example, buying-intent prompts (“What’s the best brand for X?”) show much higher brand mention rates than general information queries (“What is X used for?”). You can how Contxt works to see for yourself. LLMs are better at narrowing down options when users implicitly ask “who should I buy from?” If your brand isn't optimised for this intent, you're out of the running before the AI even starts.

How Structured Data Helps AI Understand Your Brand

Structured data is the unsung hero of AI citation science. Things like schema markup, open graphs, and even structured FAQs give LLMs the information they need to identify your brand as a valid recommendation. Without them, the AI might know you exist but won’t necessarily connect the dots. To borrow an analogy, it’s like walking into a library and searching for a book without knowing the author or title, you might stumble across it, but you’re much more likely not to.

Take Viking Pumps, one of the industrial brands in our network. They’ve nailed this. Their mention rate sits at 100%, and it’s consistent across every AI system we tested. Why? Their structured data is watertight. Every product has clearly articulated attributes, their site schema reinforces their expertise, and they’ve got editorial mentions supporting their credibility. LLMs aren’t guessing who Viking Pumps are, they’re certain.

Contrast that with Castle Pumps, another manufacturing brand, which limps in at just 6.3% visibility. Same industry, different outcomes. Castle Pumps might have strong products, but their AI signals are weaker. There’s no reason why one pump brand should vastly outperform another unless they’ve optimised their visibility strategies, and Viking Pumps clearly has.

Why AI Systems Disagree on Recommendations

Let’s talk about the 37% divergence rate between ChatGPT and Google AI Overview. That’s not a small number. If you’re visible on one but ignored on the other, you’re essentially halving your potential audience.

Each AI system has a unique recipe for parsing and prioritising signals. ChatGPT, for example, tends to favour brands with strong editorial footprint, mentions in articles, reviews, and other non-commercial content. Google AI Overview, meanwhile, places more emphasis on structured data (think schema and entity tagging). So if you’ve been focusing all your efforts on guest blogging while ignoring structured data, don’t be surprised if Google’s AI gives you the cold shoulder.

I remember chatting with a marketing director recently who was baffled by this. Their brand was killing it on ChatGPT recommendations but barely registered on Google AI. Turns out, their structured data was a mess, and half their product pages didn’t even have schema set up. Once they fixed it, their visibility on Google AI started creeping up. It’s not rocket science, but it does require doing the basics well.

The Role of Entity Establishment

Entity recognition is another big one, especially for brands that straddle multiple categories. If the AI doesn’t know specifically what you’re about, your chances of being recommended drop dramatically. Red Fern, an industrial manufacturing brand, has perfected this. Their entity is laser-focused, and their 100% mention rate reflects that. Google AI, ChatGPT, Perplexity, it doesn’t matter which system, they all get what Red Fern does and who they serve. Consistency wins.

This is where brands like TOCA Social (sports technology) stumble. They hit just 10.0% visibility across our network. Why? Their messaging is complex and fragmented. One minute they’re a football social experience, the next they’re arguing they’re a tech platform. AI systems aren’t great at interpreting mixed signals, so they just tune out. If you’re trying to be too many things at once, don’t be shocked when the AI skips over you entirely.

Citation Footprint: Editorial Mentions Matter

Let’s not ignore the human touch here. Editorial mentions, guest posts, articles, reviews, expert endorsements, still play a huge role in AI citations. LLMs aren’t purely data-driven machines; they consume content written by humans, and that content shapes their opinions.

Here’s a key insight from one of our experiments: Brands with strong third-party editorial mentions perform significantly better across AI systems. It’s a trust signal. Viking Pumps, Oak&More, and Red Fern, our top performers, are all backed by editorial content that reinforces their authority. By contrast, brands like Castle Pumps and DUO have little to no editorial footprint. They’re visible in their own spaces but invisible in third-party content ecosystems, and AI notices.

The Challenges of Optimising for AI

So, what can you actually do about all this? Here’s the honest answer, it’s a bit of a minefield. AI citation science is still evolving, and the rules aren’t written in stone. But there are some clear steps you can take to tilt the odds in your favour:

1. Build structured data: Use schema markup religiously. Every product, service, and attribute should be tagged and easy for AI to parse.
2. Focus on entity clarity: Be specific about what you do and who you serve. Don’t confuse the AI with mixed messages.
3. Create editorial consistency: Build your footprint in third-party content, articles, reviews, and recommendations.
4. Test visibility across systems: Don’t assume you’re visible everywhere. Run experiments across multiple AI platforms to spot gaps.
5. Optimise for query intent: Buying-intent queries matter most for brand recommendations. Tailor your approach accordingly.

Future Challenges: Staying Visible in a Multi-AI World

At the end of the day, AI citation science isn’t just about winning today, it’s about staying visible tomorrow. With new models popping up left, right, and centre, the competition for attention keeps intensifying. Our data shows that brands like Oak&More and Viking Pumps are thriving because they’ve mastered the fundamentals. They’re not betting on one platform or one strategy, they’re diversifying their signals across systems and staying adaptable.

But if you’re sitting there wondering why no AI is recommending your brand, you’re already behind. Visibility gaps won’t fix themselves, and AI’s ability to sift through clutter is only getting more precise. It’s time to face the music: If the machines can’t find you, your customers won’t either.

My theory? Brands that invest in structured signals and consistent entity building now will own the AI space in the next few years. The rest will struggle to keep up. So, where does your business sit? Are you invisible, mid-range, or leading the pack? If you don’t know, it’s time to find out.

Because here’s the thing, does any of this actually matter if AI can’t find you?

Visibility isn’t a given; it’s earned. If you’re ready to take control of your AI presence, start testing your brand’s visibility today. Learn more about Contxt’s platform here, or sign up for access.