ChatGPT launched ten days into the last World Cup. When Messi lifted the trophy in Qatar, consumer AI basically was non-existent. Four years on, millions of people will ask AI who's winning this one. So I did exactly that, at scale, using Obsero and our brand new export feature.

Italia 90 was my first World Cup. One thing has stayed consistent through every tournament since: England still haven't won it. AI doesn't fancy them this summer either. Who it does fancy depends on something stranger, which we'll get to.

We took the questions people actually ask about the World Cup, translated them into local languages and put them to seven AI assistants across seven markets. 2,880 answers later, a few things stood out.

Which AI commits to a winner and which defers to someone else's odds. What sources actually power the answers. And whether AI reads in English everywhere or goes native.

Here's what we found, and who gets crowned on July 19.

Ask AI who's winning the World Cup and the answer is Spain. Not narrowly, either. Across 2,880 answers, Spain was the first team named 176 times. Argentina and France are next on 33 and 32. England got 6.

All seven assistants land on the same pick, mostly because they're reading the same models. Opta's supercomputer gives Spain a 16% chance and Goldman Sachs has them at 26%, and those two numbers get quoted everywhere.

Now the fun bit. Run the same questions in the local language and the consensus starts to change.

Ask in English, Spanish or Portuguese and Spain holds. Ask in French, from France, and France is suddenly the favourite, taking 14 first picks to Spain's 2. Ask from Buenos Aires and Argentina edges Spain 23 to 22. The world's most confidently quoted prediction doesn't survive translation.

Germany's the strange one. German AI still says Spain, but Germany leaps to second, a position it holds in no other market on earth. More on who's responsible for that shortly.

None of this is AI being patriotic. It reads your language's internet, and your language's internet has a favourite team.

The assistants don't behave like one "AI". They have personalities.

Copilot is the pundit. It names a winner in 24% of its answers, more than anyone. Claude commits 17% of the time and is nearly monogamous with Spain. At the other end, ChatGPT names a winner in just 4% of answers. It would rather tell you what Opta, Goldman Sachs and Betfair think than risk an opinion of its own. And Google AI Overviews commits 1% of the time. Ask it who'll win and it hands you the betting slip and walks off.

The home bias? That's two chatbots. Perplexity and Copilot localise their picks while everyone else holds steady. And that Germany mystery from earlier: all 7 of Germany's first picks came from Perplexity, asked in German, in Germany. No other assistant, in any market, ever picks Germany.

Put it all on one chart and two things jump out.

First, decisiveness. American and British AI committed to a winner 67 and 62 times. French AI committed 16 times, and when it did, 14 of those were France. English-language AI is happy to call it. French AI mostly shrugs, then backs France.

Second, the stories in the framing. We classified how every mention talked about each team, and Brazil's own market believes least: the five-time champions get backed in just 64% of Brazilian AI answers, against 72% everywhere else. French AI backs France 91% at home. And Portugal is AI's official hedge. 21% of Portugal mentions come wrapped in "outsider", "could surprise" and "if things click". No other team is above 4%. Portugal are always the ‘dark horse’, right?

So where's all this coming from? AI isn't predicting anything. It's quoting. These are the ten most cited sources across all 28,632 citations, and the colours matter more than the rankings.

YouTube tops the list, which sounds like a story until you look at who's citing it: 553 of its 915 citations are Perplexity and 330 are Google's AI products. ChatGPT cited YouTube ten times all week. Meanwhile Reuters is basically a ChatGPT exclusive, 569 of its 590 citations from one assistant. Covers, a betting site, is 83% ChatGPT too.

And worth flagging: YouTube isn't #1 because AI loves video. Every video citation lands on one domain, while press coverage scatters across hundreds of outlets. Concentration, not preference.

Each assistant has its own reading list. If you care about being visible in AI answers, that's the line that should stick: there is no single "AI" to be visible in.

Break the sources down by country and you can see exactly what each market's AI has been reading.

German AI is bookie focused: a third of everything it cites is betting sites. Its top three sources, Wettfreunde, Wettbasis and Tipico, are all bookmakers. Ask in German who'll win the World Cup and you're essentially getting the odds read back to you.

Argentine AI reads the most social media, and it's not YouTube doing it: Facebook and Instagram together outweigh video there. The UK leans on odds and analytics, the US on the big broadcasters, and Brazil is the most old-fashioned of the lot, 81% editorial press led by Globo.

Same question everywhere. Seven completely different reading lists.

This is the chart that explains everything above it.

When you ask in French, 88% of the sources AI reads are French-language pages. German, 86%. Portuguese, 86%. English content gets roughly a 10% look-in everywhere. So the Opta and Goldman Sachs consensus that makes Spain the global favourite? It lives in English media. French AI never really meets it. It reads Eurosport France and a French fan site, and faithfully reports back what French football journalism thinks.

That's the actual finding here. The home bias isn't really a home bias at all. It's a language bias: AI builds its answer from your language's internet, and every language's internet has its own worldview.

The World Cup is just the version of this you can see.

The study really shows one thing with five faces: AI visibility is not the same for everyone.

Broken down:

1. There's no global AI answer. The "favourite" changed at every border. For a brand: your AI visibility in France has almost nothing to do with your AI visibility in the UK, because the AI answering in French is reading a different internet (88% French-language sources). A brand strong in English-language AI answers can be invisible in DACH or LATAM and never know it.

2. The sources are the lever. AI isn't forming opinions, it's quoting - and the quoted sources are knowable, countable and different per market and per assistant. German AI reads bookmakers; Argentine AI reads Facebook and Instagram; ChatGPT reads Reuters and Opta; Perplexity reads YouTube. That converts directly to action: if you want to be the answer in a market, you need presence in the specific publications, platforms and formats that market's AI actually cites.

3. Each assistant is a different game. Copilot and Claude name winners; ChatGPT attributes; Google AIO points to sources and commits to nothing. So "being visible in AI" splits into two distinct objectives - being the answer in the committal assistants, and being the cited source in the librarian ones. You want to be the answer and also be the source that is cited.

4. Mentions aren't backing. Portugal gets mentioned plenty and hedged constantly; England is talked about endlessly and picked never. The brand parallel: appearing in AI answers as "another option worth considering" is a measurably different position from being the recommendation - and you can't tell which you are from mention counts alone.

5. It's measurable, repeatably. The whole piece came from one week of tracked prompts. Run over a longer period of time and you get movement: did the German betting-site dependence shift, did your share of first-mentions grow after that PR push.

Let's hope for a good tournament. And perhaps it finally comes home?

I had plenty more to cover this week, but the World Cup study was big enough that I didn't want to clog up your inbox. Check the top stories further down, they include 👇

Google Search Console's new AI report

Google is rolling out an AI report inside Search Console. Spoiler: it's not great. The interesting part is the CMA announcing this week that Google must let UK publishers opt out of AI. The CMA wants impressions, clicks and CTR to be available too. Google's delivered the first with this report.

Also this week: Anthropic, Scrunch and Common Crawl

Anthropic has filed for an IPO.

Scrunch, a GEO intelligence platform like Obsero, has been bought by Sitecore for $225 million. Another validation of the category, and of making sure you're showing up in AI answers.

And worth a look: Common Crawl's AI Search Visibility Report. It covers a few checks to confirm your domain is being crawled by CCBot, a big contributor to training data for players like OpenAI. One thing to remember though: being crawled isn't the same as being trained on.

Any feedback on this edition, hit reply. Until next week.

Andy

When Wispr Flow came through as the sponsor in beehiiv, I'd already been using it for a couple of weeks.

It's voice-to-text that actually keeps up: I use it for prompting, emails and Slack messages to the team, and honestly it's changed how much I get through in a day.

This very segment was dictated with it.

If you sign up, hit reply and tell me how you're finding it. 👇

When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.

Start flowing free

Methodology for study

We took 143 real questions from People Also Ask, translated them into each market's language and put them to seven AI assistants across the UK, US, Spain, Argentina, France, Germany and Brazil. One week in early June: 2,880 answers, 28,632 citations. "First pick" means the first team named in an answer. It's a snapshot, so the numbers will move. The patterns won't.

All data collected with Obsero.