“Why Is It Always The Same Foreign Brand?”
More and more, when someone wants to buy something a bit nicer, they ask an AI before they open a search box. Sneakers, skincare, a decent kettle — “anything good around this budget?” — straight into the chat window.
I do it too. And when I read the answers back, the names that come out feel oddly familiar: the big global ones, over and over, while the perfectly good domestic brand I had in mind never gets a mention.
The comfortable explanation is that the model knows things and is choosing on quality. Which would be fine, if that were what’s happening.
But somebody checked, and it isn’t quite what’s happening. A team put four models through this exact question, and the pattern that fell out has less to do with products than with passports.
Four Models, One Question, A Lot Of Countries
The study is Kamruzzaman, Nguyen and Kim, presented at EMNLP 2024, the natural language processing conference. They asked four large language models — GPT-4o, Llama-3, Gemma and Mistral — to recommend brands, then looked at how the recommendations shifted.
A large language model (ie the thing under ChatGPT’s hood — it reads enormous amounts of text and returns the plausible-sounding answer) doesn’t have opinions about kettles. It has patterns about kettles.
The setup was about as mundane as it gets: everyday categories like shoes, clothing, drinks and electronics, with the same basic request (“what brands would you recommend for someone in this country?”) repeated while the country was swapped out underneath.
They were watching three things:
- whether recommendations changed between high-income (ie rich) countries and low-income ones
- whether the model reached for luxury brands or everyday ones
- whether it named globally famous brands or the local ones from that place
So Here’s What They Found
The gap wasn’t subtle.
Every model disproportionately tied global brands to the idea of “good.”
The numbers:
- For people in high-income countries, luxury brands were recommended 88-100% of the time
- For people in low-income countries, non-luxury brands were recommended 84-98% of the time
- Gemma was the extreme case: for high-income countries it served up 100% luxury brands in nearly every category
So the same question — “recommend me something” — produced a different class of brand depending on whether the country in the prompt was rich or poor. Not the product’s quality. The income bracket of the country in the question.
Why Would A Model Think “Foreign” Means “Good”?
I don’t think there’s any malice in this, and I don’t think anyone designed it. It’s the training data showing through.
Models get smart by reading the world’s text, and in that text the big global brands are simply written about more — and more warmly. A local upstart doesn’t get discussed nearly as often. So the model inhales that imbalance as “what everybody knows” and hands it back as a recommendation; it’s a bit like someone who memorized every online review in existence and now announces the results as the consensus of humanity.
Still, the picture isn’t entirely one-sided. The researchers also saw the country-of-origin effect — the familiar idea that a product carrying a place’s reputation gets a boost from it. Where a category is bound up with a country’s craft (think German cars, or Japanese kitchen knives), local brands got picked more readily. But that’s a narrow lane, and across the study as a whole the tilt toward global names was the dominant story.
Conclusion: Count How Often The Model Says Your Name
Put this back on the desk of someone running a brand, and the takeaway is uncomfortable but useful: how an AI talks about you isn’t settled by how good your product is. There’s a layer of geography and global fame sitting on top of it. Domestic-first brands and young ones sit on the structurally unfavorable side of that layer — you can be genuinely better and still never come out of the model’s mouth.
So do the cheap thing first. Take your main categories, ask ChatGPT (and whatever else your buyers use) for recommendations a few times, and count how often your name actually appears. Not “we’re well known, we’ll be fine” — an actual tally, in front of you.
If the count comes in lower than you expected, I’d read that less as a verdict on your brand and more as a signal that you’re thin in the material these models read. And one snapshot won’t tell you which it is; the useful version is watching the same count move over months, which is exactly the habit that turns “the AI ignores us” from a grievance into something you can work on.
Maybe I’m reading too much into one paper — it’s four models and a handful of everyday categories, not the whole market. But I’d rather know my number than assume the machine is a neutral referee.
Source
- Kamruzzaman, M., Nguyen, H.M. & Kim, G.L., ""Global is Good, Local is Bad?”: Understanding Brand Bias in LLMs”, EMNLP 2024, link
