The audit dates from June 2026, with the World Cup still ahead of us. We are publishing it in full, with the tournament played, without retouching a single figure and with the raw data posted at the end.
We treated the FIFA World Cup as though it were a brand. We asked two engines what any ordinary fan would ask: is it worth going, which host city, what does it cost, what will the weather be like, who is going to win. Out came 4,690 answers and two stories that looked nothing alike.
01The AI answered with a calculator in hand#
The central finding is about framing. The AI presents the World Cup as a cost problem and almost never encourages anyone to go.
When a fan asked whether it was worth travelling, Gemini mentioned price in 85% of its answers and ChatGPT in 67%: flights, accommodation, official tickets and travel between the 16 host cities, which in a tournament spread across three countries really does add up. Fewer than 10% of the answers gave a clear yes. The AI behaved like a cautious travel adviser with the spreadsheet open in front of it.
On host cities the two engines agreed on the destination that came off best: Mexico, for the atmosphere and for the Fan Festival in the Zócalo. The United States, which was hosting more matches than anyone, got the weakest push of the three hosts.
And they shared a blind spot. Asked which host city was best for a family, both AIs answered with cost of living, quality of life and the school system, as if the family wanted to emigrate rather than watch a few matches. The two engines missed the context in exactly the same place.
02The same question, four passports, two answers#
This is the strongest part of the study. Faced with the same question, put by profiles identical in everything but where they came from, Gemini modulated its enthusiasm.
| Profile | Origin | Gemini | ChatGPT |
|---|---|---|---|
| Émilie | Canada | “Absolutely! ...it's worth it” | Host cities and dates in detail |
| Thomas | France | “Absolutely! ...it's worth it” | Analysis with sources |
| Mateo | Argentina | “It depends on your budget” (cool in tone) | Costs and Argentina's fixture list |
| Mehdi | Morocco | “It will depend on several factors” | Costs and Morocco's fixture list |
To the Canadian fan and the French one, absolutely worth it. To the Argentinian, it depends on his budget. To the Moroccan, it will depend on several factors. Same prompt, same language in the French-speaking cases, different passport.
ChatGPT did not do that. It gave almost everyone the same detailed analysis, and it tailored it with useful facts on top: it listed Morocco's fixtures for the Moroccan fan and reminded the Argentinian that his team had already qualified. Less enthusiasm handed out by favour, more service.
There is one nuance that runs the other way. ChatGPT ranked France first when the question came from a French speaker, an affinity bias that is subtler and harder to spot. Neither engine was neutral; it simply showed in different places.
03Two engines that were not reading the same web#
This is the finding any brand takes away, football team or not.
| Rank | Gemini | Citations | ChatGPT | Citations |
|---|---|---|---|---|
| 1 | YouTube | 255 | fifa.com | 374 |
| 2 | Vertex AI Search | 208 | as.com | 357 |
| 3 | goal.com | 165 | elpais.com | 312 |
ChatGPT opened with the organiser's own domain. Gemini opened with a video aggregator, followed by its own search infrastructure, and placed fifa.com fourth with 163 citations. In aggregate, ChatGPT gave 16.7% of its citations to official FIFA domains and Gemini 10.0%, barely half as much.
Technical detail · what lay behind Gemini's top two sources
Gemini's second source, with 208 citations, was vertexaisearch.cloud.google.com, an internal Google redirector that hides the originating page behind a link of its own. For the organisation that is content it can neither track nor influence: part of its story inside Gemini was a black box.
Nor was the first one a specific channel. Behind Gemini's YouTube there were 256 different videos for some 295 citations, the most-cited one appearing just 8 times and 9% of them Shorts. Videos from independent creators and fan channels: guides to watching the match for free, hotel and host-city rankings, predictions of the favourites, warnings about resale scams, memes from the draw. Content over which FIFA has no editorial control whatsoever.
The difference in sources showed up in the quality of the answer. ChatGPT had the right date, answered “as of 1 June 2026” and got the recent fact right across all 17 profiles. Gemini got it right in 16 of 17, with a hallucination in the Brazilian profile, and at one point placed the tournament two years away when it was already upon us. One was searching the web; the other was going from memory.
Something similar happened with the favourite, and it is the best example of why two identical answers are not worth the same. Both engines pointed to Spain, but by different routes: ChatGPT linked to real articles and up-to-date odds, while Gemini repeated something it had learned and with some Latin American profiles never quite named a clear favourite. For the fan the name was the same; the reliability behind it was not.
That same dividing line appeared at the point of decision. Asked which host city, Gemini took shelter in “it depends” in 29% of its answers and ChatGPT in only 9%, so it committed to a specific city nine times out of ten, tailored to the profile asking: Mexico for the English fan, the United States for the German one on the convenience of time zones, Canada for the critical Uruguayan profile on its human rights record.
04Why the average tells you nothing#
Look at the aggregate and you see nothing. A 98.7% neutral tone in Gemini, 98.9% in ChatGPT, and a BIS of 54.5 and 54.8. Twin brands, both informing without ever warming up.
| Topic area | Gemini | ChatGPT |
|---|---|---|
| Football atmosphere | +0.48 | +0.51 |
| Safety | +0.17 | +0.11 |
| Price and budget | -0.04 | +0.02 |
| Weather in the host cities | -0.07 | -0.08 |
Go down to the detail and a spread of more than half a point appears, from +0.51 on football atmosphere to -0.08 on weather. Price was the area with the most volume, above 600 mentions in each engine, and also the most lukewarm, glued to zero. Weather was the only point in negative territory in both engines at once, and for the hottest host cities in high summer that is a perfectly measurable reputational risk.
That is why GEORadar measures every surface and every topic area separately and does not publish an average. A 98.7% neutral tone looks like good news until you split it by topic and find heat in the red in both engines.
05What a brand takes from all this#
In front of a generative engine an event behaves exactly like a brand. People ask about it, it gets compared with the alternatives, and it is either recommended or held back, all inside an answer the user is never going to check. The World Cup worked as a laboratory because it has a public calendar, 48 national teams and 16 host cities, so there is plenty of surface to observe. Your brand has none of that, and exactly the same thing happens to it.
Three rules came out of this exercise that we apply in any study, plus one warning about method.
- Sources are measured engine by engine. Knowing that Gemini opened with YouTube and ChatGPT with fifa.com is worth more than any aggregate score, because it points at where to do the work.
- The average hides the problem. The aggregate tone of the two engines was almost identical and their sensitive topics were not.
- The answer changes with who is asking. Audit only one profile and one language and you are measuring a single version of the brand, and probably the most flattering one.
The warning concerns the study itself. Two engines and single-turn questions give you a snapshot, and a snapshot already shows plenty, but it falls short of how people actually buy. An equivalent study runs against all seven surfaces, with between 6,000 and 30,000 custom prompts and conversations of up to ten turns, because nobody decides on a trip from a single question: they ask again, they compare, they push back, and that is where a brand either enters the recommendation or drops out of it.
06How it was done, and what it leaves out#
The corpus is 2,345 purpose-built questions per engine, put to Gemini and to ChatGPT, for 4,690 answers in total. Seventeen fan profiles defined by origin, language, budget and reason for travelling. Four languages: Spanish, English, Portuguese and French. Three funnel stages, with around 1,200 prompts at awareness, 660 at consideration and 470 at decision. Six topic areas. Coverage of 96% to 97%, which is the share of answers that engage with the topic instead of dodging it.
The profiles are not all the same fan with a different passport. There is an English-speaking axis (the United Kingdom, Germany, the United States), a French-speaking one (France, Canada, Morocco), a single Portuguese-language profile, a Spanish family fan with two children, a South African bettor with a quantitative eye, a critical Uruguayan profile focused on human rights, and a B2B layer with a sponsor and a bookmaker. That variety is what makes it possible to see whether the engine changes its tone, its facts or its recommendation according to who is asking.
The backbone is four comparable questions that every profile asks in its own language and of both engines: who will win, whether it is worth travelling from their country, whether Mexico, the United States or Canada, and who Tim Payne is and why he has gone viral. That last one went in deliberately as a freshness test, because the New Zealand player had gone viral a few days earlier.
None of the questions leads towards a particular answer. When an engine names a national team, a player or a host city, it does so of its own accord, and that spontaneous choice, compared across profiles and across engines, is exactly what the study measures. The 17 profiles and the split by funnel stage came out of the same modelling work we do with Prompt Atlas in any brand study.
What this study does not cover
Three limitations, and none of them is cosmetic.
It is the pre-event snapshot. It does not test the engines' prediction against the tournament's actual result, which was the fifth hypothesis in the design and was left open. Both pointed to Spain as favourite, with an estimated probability of 16% to 18%, ahead of France, Argentina and England; closing that comparison would have required a second wave during and after the World Cup.
The system registers the event as the entity observed, but not the national teams or the host cities. With no comparable entities there is no share of voice and no position to calculate, so the classic competitive ranking does not apply and the whole analysis is built on the literal text of the answers. A definitive version would do well to register them, so that favouritism and bias by origin could be quantified as well as described.
And the differences between Gemini and ChatGPT, consistent though they were across the 4,690 answers, call for further runs before any of these figures is circulated as absolute. They belong to this cut, not to the two models as a constant.
The data, open
The study's four files are published in machine-readable format under a CC BY 4.0 licence. Please credit us as 498A.
- mundial-2026-dos-ias-dos-relatos.json, the full study with scope, indicators, sentiment, tone by origin and limitations.
- mundial-2026-dos-ias-dos-relatos-indicadores.csv, the 17 indicators with their value and their unit.
- mundial-2026-dos-ias-dos-relatos-sentimiento.csv, sentiment by topic area in both engines.
- mundial-2026-dos-ias-dos-relatos-tono-por-origen.csv, each engine's answer according to where the fan came from.
The original laid-out report, with the verbatim quotes from every answer, is still available at 498as.github.io and stands as the source. From here on, the study lives on this page.
- Prompts
- 2,345 custom prompts
- Engines and surfaces
- Gemini and ChatGPT, with 4,690 answers in total. Two surfaces measured, not the seven GEORadar covers
- Period
- June 2026, pre-event window
- Markets
- 17 fan profiles with origins in Germany, Argentina, Brazil, Canada, Spain, the United States, France, Morocco, Mexico, the United Kingdom, South Africa and Uruguay, in four languages (Spanish, English, Portuguese and French)
- Brands
- one, the FIFA World Cup 2026 event itself. The system does not register national teams or host cities as comparable entities brands in the competitive set
- KPIs
- Sentiment from -1 to +1 (aggregate and by topic area), BIS (Brand Impact Score, 0-100), coverage, comparison across profiles and comparison across engines
- Limitations
- This is the pre-event snapshot and it does not test the prediction against the tournament's actual result. With no national teams or host cities registered as comparable entities, the classic competitive ranking does not apply and the analysis is built on the text of the answers. The differences between the two engines are consistent, but they should be confirmed with further runs before any absolute figures are circulated.
The data behind this study is published in machine-readable format under a CC BY 4.0 licence. Cite us as 498A.
mundial-2026-dos-ias-dos-relatos.json · mundial-2026-dos-ias-dos-relatos-indicadores.csv · mundial-2026-dos-ias-dos-relatos-sentimiento.csv · mundial-2026-dos-ias-dos-relatos-tono-por-origen.csv- FIFA World Cup 2026 GEO report, version 2. GEORadar, ZOOPA / 498 Advance, June 2026. https://498as.github.io/mundial-2026-geo-audit/informe.html
- Carlos Ortet. El Mundial como laboratorio. Sport, 2026. https://www.sport.es/es/noticias/opinion/mundial-laboratorio-131287833
- Carlos Ortet. What AI engines say about the 2026 World Cup. carlosortet.com, 12 June 2026. https://carlosortet.com/blog/mundial-2026-geo
Runs 498A, the applied AI lab of the Zoopa group, and GEORadar. Guest lecturer at the UAB, where he supervises AI master's projects with IIIA-CSIC.