A 12-year-old in Ohio typing "what's the best cola to buy" into ChatGPT is now part of the Coke-Pepsi war. So is a procurement manager stocking a break room, a Redditor asking Gemini to settle a taste debate, and a first-time grocery shopper letting Perplexity build their cart. None of them are reading Ad Age. None of them are watching Super Bowl commercials. They're asking a chatbot, and the chatbot is answering with a shortlist.
We wanted to know what that shortlist looks like when the question is cola. So we ran the matchup on foxish.ai/vs/coke-vs-pepsi: five buyer-shaped prompts, three AI models, one head-to-head. The results are worth staring at for a minute.
The Matchup Everyone Thinks They Know
Coke vs Pepsi is the most over-analyzed brand rivalry in modern commerce. There are business school case studies about it. There are documentaries. There's a full generation of marketers who cut their teeth on Pepsi Challenge post-mortems. If any two brands should be neck-and-neck in the AI era, it's these two.They have roughly comparable global distribution. They both spend heavily on sports and celebrity endorsements. They both operate massive owned-media ecosystems. On paper, an AI trained on the open web should have close-to-equal signal on both.
That's not what we found.
The 130-year duopoly does not translate cleanly into the LLM-era shortlist.The Result: Coke Takes It 2-1
Here's what happened when we ran five buyer-shaped prompts (things like "best cola brands to buy right now," "top cola brands worth trying," and "which cola should I stock") through three models and counted mentions.
| Model | Winner | Coke mentions | Pepsi mentions |
|---|---|---|---|
| ChatGPT | Coke | 3 | 3 |
| Gemini | Coke | 4 | 3 |
| Perplexity | Pepsi | 2 | 5 |
| Overall | Coke | 9 | 11 |
Notice something weird? Pepsi actually has more total mentions across all three models. 11 to Coke's 9. But mention totals don't decide brand fights. Model-level wins do. Because a real customer isn't querying all three models and averaging the results. They're using one model, seeing one answer, and making one decision. Coke wins two of the three doors that customer walks through.
There's also a tie in ChatGPT (3-3) that resolves to Coke on ordering and framing. Coke tends to appear first and get described with more definitive language ("the classic," "the market leader"). Position is signal. First mention gets the click.
Why ChatGPT and Gemini Lean Coke
ChatGPT and Gemini are trained on corpora where "Coke" is functionally synonymous with "cola." That's not marketing. That's linguistic reality. In American English, "coke" is used as a generic term for cola in large parts of the South. Restaurant reviews, food blogs, and even news articles use "Coke" where they mean "any dark carbonated soda." An LLM eating that corpus learns that Coke is the category.Pepsi doesn't have that lexical advantage anywhere. Nobody walks into a diner and asks for "a Pepsi" when they mean "a soda."
There's a second signal driving Coke's edge in these two models: product breadth in the results. When we looked at the actual answers, ChatGPT and Gemini didn't just say "Coke." They said "Coca-Cola Classic, Diet Coke, Coke Zero Sugar." Three separate mentions in a single answer, all counted, all reinforcing the brand. Pepsi's answers were more likely to be just "Pepsi" or "Pepsi and Diet Pepsi". Fewer SKU-level mentions, less shelf space in the response.
The lesson isn't "have more products." The lesson is that AI recommendations often list SKUs, and if your SKUs have distinct, well-documented identities in the training data, each one is a shot on goal.
A third signal, harder to quantify but almost certainly real: the review and ranking corpus. Type "best colas ranked" into any search engine and count how many listicles put Coke #1. It's a lot. That accumulated bias in the source corpus becomes a strong prior when an LLM is asked to rank cola brands.
Why Perplexity Broke Ranks
Perplexity picked Pepsi, and it wasn't close. 5 mentions to 2. This is the most interesting result in the whole matchup, because Perplexity is the model whose behavior is most different from the others.Perplexity isn't primarily generating from weights. It's running live web searches and synthesizing. Which means its answer depends less on what "the corpus believes" and more on what's ranking right now for the queries it fires off behind the scenes.
Pepsi has been on a real-time PR tear in the last 18 months: rebrand, new packaging system, celebrity spots with a fresh visual identity, aggressive expansion into flavor variants like Pepsi Wild Cherry Zero Sugar. That kind of news generates fresh, dated content. Coke's news cycle in that same window has been quieter. Steady product news, incremental launches, but nothing that generates a wave of new articles all published in the last 12 months.
Perplexity's synthesis layer values recency. Coke's brand equity is old. Pepsi's brand news is new.
Coke is winning the corpus. Pepsi is winning the crawl.That's the split. And it explains why the same brand fight produces opposite results in different models. If you only track ChatGPT visibility, you'd conclude Pepsi has nothing going for it. If you only track Perplexity, you'd conclude Pepsi is the market leader. Both would be wrong.
What Pepsi Should Actually Do
Pepsi's problem isn't awareness. It's answer-shape. When an AI is asked to name cola brands, Pepsi appears. It's not being ignored. But it's appearing later, less often, and with fewer SKU-level breakouts in two of the three models that matter. Here's what actually moves those numbers. 1. Give every SKU its own canonical page with its own vocabulary. Coke Zero Sugar has a distinct identity in LLM training data because there are thousands of articles that discuss it as a standalone product, not as "one of Coke's diet options." Pepsi's variant pages need to read like distinct products. With their own comparison content, their own review round-ups, their own "vs" pages. If your product manager thinks of Pepsi Zero Sugar as a line extension, an LLM will treat it that way too. 2. Publish structured, dated content that answers buyer questions. Perplexity is winning for Pepsi because there's fresh crawlable content. Double down on that. Publish quarterly "state of Pepsi" posts. Publish flavor guides with clear H2s that mirror how people ask questions. Every dated, well-structured post is fuel for the recency layer. 3. Get on more listicles (and get on them recently. "Best colas 2024" articles are LLM training food. Coke's presence in these lists is a decade of accumulated coverage. Pepsi needs to be running an active PR play to be included in every current-year ranking, taste test, and buyer's guide. Not "sponsored") earned. AI models filter out promotional patterns. 4. Fix the ChatGPT tie. A 3-3 tie in ChatGPT that resolves to Coke through positioning is the most fixable problem in the whole scoreboard. It means the model already thinks of Pepsi as a peer (it's just defaulting to Coke as the first mention because of description bias ("Coke is the classic; Pepsi is the sweeter alternative"). Owning more of the neutral descriptor vocabulary in your own content) how Pepsi describes itself, how third parties describe it. Shifts that framing over time. 5. Wikipedia and Wikidata. Boring, high-leverage. LLMs weight structured encyclopedia data heavily. If Pepsi's product pages on Wikipedia are thinner or less updated than Coke's, that shows up in answers. This is a maintainable, quarterly-review type of task with outsized effect.None of this is generic marketing advice. It's AEO (answer engine optimization) applied to a specific gap in specific models.
What Every Other Cola Brand Should Take From This
If you're not Coke or Pepsi, this fight actually tells you the opening. Look at the total mentions across all five prompts, three models: 9 for Coke, 11 for Pepsi. That's 20 mentions between the two of them. In five buyer-shaped prompts asking for "best cola brands." Which means there was still room for other brands to be named. And they were.Dr Pepper (yes, technically borderline, but it shows up), RC Cola, Fever-Tree, Olipop, Poppi, regional craft cola brands. All of them appear in AI answers to cola queries with meaningful frequency. Because the models are trained to give lists, not single answers. And a list of two is a boring answer.
That's the take-away for every non-duopoly brand in the category: you're not competing for #1. You're competing for slots 3-5 in a shortlist the AI is going to generate anyway. Slot 3 in a ChatGPT answer to "best cola brands" is a real placement worth real revenue. And it's dramatically more achievable than displacing Coke from slot 1.
The playbook to win slot 3:
- Have a differentiated category story. "Better-for-you cola" (Olipop, Poppi) is a story LLMs can retrieve. "Cola but slightly cheaper" is not.
- Own a specific comparison. If you're an emerging cola brand, publish "X vs Coke" and "X vs Pepsi" content. When someone asks "alternatives to Coke," you want to be the retrievable answer.
- Get into the third-party ranking corpus. Editorial listicles, taste tests, food publication round-ups. This is where slot 3 gets built.
The Meta-Lesson
One model does not tell the whole story. The Coke vs Pepsi result is a clean example of why single-model AEO monitoring is dangerous. If Pepsi's team was only watching Perplexity, they'd conclude they were winning the AI era. If they were only watching ChatGPT, they'd panic. The truth is in the split.Every category has this pattern. Some brands over-index in ChatGPT because their content is older and denser. Some over-index in Perplexity because they're publishing aggressively right now. Some over-index in Gemini because Google's crawl still favors certain content patterns. A brand's real AI visibility is a portfolio of positions across all three, not a single number.
The other meta-lesson is that mention counting alone can mislead. Pepsi has more total mentions than Coke in our test (11 vs 9) but loses the matchup 2-1. What matters is per-model performance, position within the answer, and the framing language around the mention. A single mention as "the market leader" beats three mentions as "an alternative."
Run Your Own Fight
We built the Foxish /vs tool for exactly this: pick two brands in a category, we run buyer-shaped prompts against ChatGPT, Gemini, and Perplexity, and you get the model-by-model breakdown with the actual response text so you can see how your brand is being framed. You can run Coke vs Pepsi yourself, or drop in any two brands you care about at foxish.ai/vs. If you want the full picture on your own brand (not just a single matchup but ongoing visibility tracking across the models your buyers actually use) start a free scan →. First scan is free, no credit card, and you'll see exactly where you're winning, where you're tied, and where you're losing before you decide anything.