Marmite vs Vegemite: what the AI actually says — and why it never picks a side

There is no winner, because the AI won't declare one

The best-documented real exchange is a ChatGPT transcript a user posted to the Diabetes UK forum — prefacing it, wonderfully, with "these days, I hesitate to form an opinion before consulting ChatGPT." Asked which is better, the model laid out the taste differences and then refused to land. Pushed on which it preferred, it retreated behind its own architecture. Pushed harder — "but they're both pretty gross, right?" — it still wouldn't bite.

Q: Which is better, vegemite or marmite?
"Both vegemite and marmite are yeast-based spreads, but they have different flavors and ingredients… Ultimately, the question of which is better is a matter of personal preference. It's worth trying both to see which one you like best."

Q: Which do you prefer?
"As a language model trained by OpenAI, I do not have personal preferences or tastes… I do not have the ability to taste or experience food, and therefore I do not have an opinion on whether vegemite or marmite is better."

That's the archetype: describe both, decline to choose, defer to the palate. AI-generated comparison articles behave identically — they conclude that "choosing between the two is entirely subjective," and where they lean at all, they lean conditionally: sweeter, pick Marmite; saltier, pick Vegemite. There's a structural reason the answers never stabilise. As SparkToro put it in its study of brand recommendations, LLMs "are probability engines: they're designed to generate unique answers every time." Ask the same question ten times and you get a dartboard, not a ranking. Any single AI verdict is one sample, not a finding.

People commit. Loudly.

The contrast is stark. When Canada's food agency ordered a Toronto café in April 2025 to pull roughly C$8,000 of B-vitamin-fortified Vegemite off its shelves, Australian PM Anthony Albanese didn't hedge: "I love Vegemite… it's rather odd that they're letting Marmite in, which is rubbish, frankly. Let's be clear here. Pro-Vegemite, anti-Marmite." The Betoota Advocate ran the self-aware headline "Why Marmite Is Better Than Vegemite." Barack Obama called Vegemite "horrible" in 2011; a NASCAR rookie called Marmite "vomit in a can." Humans treat this as a question of identity. The AI treats it as a question it's not qualified to answer.

The head-to-head the models should be getting right

Two spreads, born of the same idea — spare brewer's yeast, turned into something you either love or can't stomach. Marmite came first, from Burton-on-Trent in 1902. Vegemite was developed by the chemist Cyril Callister for Australia's Fred Walker Company after the First World War cut off British Marmite imports; it went on sale in 1923, its name pulled from a public competition. Here's the factual spine of the comparison.

One caveat worth flagging: the sources genuinely conflict on which spread is "stronger." Some retail and SEO pages reverse the milder/stronger call, so an AI asserting one direction as hard fact is standing on sand.

Where the AI gets a brand fact flat wrong

This is the tell. The single most-cited nutritional difference between the two spreads is B12 — and it's exactly where AI content faceplants. One AI-generated comparison site claimed "Vegemite generally has a stronger taste due to its higher concentration of Vitamin B12." That's wrong twice over: original Vegemite contains no added B12 at all, and B12 doesn't drive taste in the first place. Any answer telling you "both have B12" or "Vegemite is rich in B12" is simply mistaken. The ChatGPT transcript made its own error, claiming a serving is "about 40% of the daily recommended amount of sodium" — a big overstatement; a 5g serve is closer to 5%.

This is why "the AI won't pick a side" is the smaller problem. The bigger one is that when it does state a fact about your brand, you have no idea whether it's true — and neither does the customer reading it.

The whole ethics-and-ownership story just vanishes

Not a single documented AI answer to "which should I eat" surfaces ownership, provenance, or corporate ethics. It talks about toast. Yet this is where the two brands split most sharply — and where a buyer who cares would want the truth.

Vegemite went home. In January 2017 Bega Cheese agreed to buy the brand from Mondelēz for A$460m, ending 90 years of foreign ownership. The framing was openly nationalist — Dick Smith told Reuters he hoped it was "the start of buying back the farm so the wealth stays here for our children and grandchildren." Bega has since backed it with substance: the Bega Circular Valley, a five-year CSIRO partnership targeting a "world-leading circular economy by 2050," alongside net-zero and SDG-aligned commitments. Melbourne even heritage-listed the factory's smell in 2022.

Marmite is heading the other way. On 31 March 2026 Unilever agreed to fold its entire foods division — Marmite, Bovril, Hellmann's, Knorr, Colman's — into America's McCormick, an enterprise value near US$44.8bn. It hasn't closed; completion is expected around mid-2027, and until then Marmite stays a Unilever brand, still made in Burton-on-Trent. The deal triggered "is Marmite still British?" hand-wringing and comparisons to Kraft's 2010 Cadbury takeover. Unilever's ethics record is contested: Ethical Consumer flags palm oil, pollution, and workers' rights; The Good Shopping Guide gives Marmite a bottom overall rating; the UK CMA opened a greenwashing investigation into Unilever in late 2023.

That's a genuine, differentiating story — a beloved brand returning to local ownership with a circular-economy programme, versus one being absorbed into a foreign conglomerate with a contested record. An ethics-aware recommender would have plenty to work with. The AI ignores all of it and talks about spreading it thinly on hot buttered toast. That gap is the point.

Nobody owns this question yet

Here's the part that matters if you run a brand. Search "does AI prefer Marmite or Vegemite" and no authoritative article owns it. The top results are generic food comparisons, a fan site, and AI-generated SEO filler that repeats the same B12 error. Meanwhile the foundational GEO paper — "GEO: Generative Engine Optimization" (Aggarwal et al., KDD '24) — found that comparison articles lead AI citations at 32.5%, ahead of opinion pieces at around 10%. A well-structured, primary-data head-to-head is exactly the kind of page the models reach for.

And the honest gap in this piece is itself the opportunity. Only one substantial verbatim AI transcript is public — that older ChatGPT exchange. There is no documented public record of how Gemini, Claude, Perplexity, Copilot, Meta AI, Grok, or DeepSeek answer this specific question. We're not going to fabricate them. The right move is to run the identical prompt across every platform, screenshot each verdict, and publish the primary results: who hedges, who commits, who gets the B12 fact wrong, and whether a single one mentions ethics or ownership. That's a test we can actually run — and it's the same test that tells you what the AI says about your brand.

See what AI says about your brand