Swedish is 0.61% of the training data. Then Swedish media blocked the retrieval bots too.

Nought point six one percent

CulturaX is one of the cleaned multilingual text mountains used to train language models: 6.3 trillion tokens across 167 languages. English accounts for 2,846,970,578,793 tokens, or 45.13% of the whole thing. Swedish accounts for 38,486,181,494. That is 0.61%.

The ratio is roughly 74 to 1. For every Swedish sentence the model has seen, it has seen about seventy-four English ones. That does not make Swedish a low-resource language in the way Faroese is, but it decides what the model has read most about when it comes to answer a question on mortgages, insurance, or which robot vacuum tested best.

"But Swedish Wikipedia is enormous"

True, and it is the best argument against the whole thesis. Swedish Wikipedia has 2,626,776 articles and is the fifth largest edition in the world, ahead of French and far beyond what ten million people should reasonably produce.

The reason is a bot. Lsjbot, built by Sverker Johansson, mass-produced articles about species and geographic features. When Swedish Wikipedia passed one million articles in 2013, the bot had written 454,000 of them, close to half the entire corpus. The same bot is why Cebuano, a language of the Philippines, now has the second largest Wikipedia at 6.1 million articles.

Volume is not the same thing as depth. A stub about a beetle is an article in the statistics but not much to ground an answer on. Sweden has quantity without necessarily having the kind of content a model leans on when someone asks about an industry.

14 of 23 Swedish news sites block OpenAI's retrieval bot

A thin evidence base could be offset by Swedish media showing up in the answers. So we read the robots.txt of 24 Swedish news and business sites on 5 August 2026. One could not be fetched. Of the remaining 23, fourteen block OAI-SearchBot outright.

That is the number that matters, because OAI-SearchBot is not the training crawler. It is the bot that fetches pages when ChatGPT is answering a question right now. Block it and you do not exist as a source in the answer, however good your content is.

One detail worth noting: Aftonbladet and SvD explicitly block Google-Extended, CCBot, ClaudeBot and PerplexityBot, but have no rule for GPTBot at all. It gets in by default, through the * rule. That looks less like a policy and more like a list somebody filled in once.

The pattern follows ownership. Bonnier's titles close everything: DN, Expressen, Di, Sydsvenskan and HD block both training and retrieval bots, and Ny Teknik does the same outside that group. Schibsted's titles let OpenAI in while shutting Perplexity out. And SVT makes the distinction that is actually considered: training bots refused, answer bots welcomed. SVT was the only site in the sample that explicitly allowed all three retrieval bots.

Blocking training and disappearing from answers are two different things

This is the most expensive misunderstanding in the industry right now, and an easy one to make. GPTBot, CCBot and Google-Extended collect text for training. Closing them is a reasonable copyright negotiating position and costs you nothing in visibility.

OAI-SearchBot, ChatGPT-User and PerplexityBot do something else entirely. They fetch the page the moment a user asks a question, in order to cite it. Block those and you have not protected your material, you have only ensured somebody else answers in your place. We have written more about why "block all AI bots" now costs you visibility.

Most of the fourteen sites above have in all likelihood made a deliberate decision about training data and taken the retrieval bots along with it. The effect is the same either way: when somebody asks ChatGPT about Swedish business, Dagens industri is not in the room.

We built our own model, and the size tells the story

Sweden has not ignored the problem. AI Sweden, with RISE and WASP, built GPT-SW3 in six sizes from 126 million to 40 billion parameters, explicitly to generate text in Swedish and the Nordic languages.

The interesting part is the scale. The entire training set is 320 billion tokens, and it has to cover Swedish, Norwegian, Danish, Icelandic, English and program code together. The English portion of CulturaX alone is 2.85 trillion tokens, close to nine times as much text for a single language. The number says less about GPT-SW3 than about what a small language has to work with.

What it means if you run a Swedish company

The conclusion is not to give up. It is that the Swedish gap is an opening, because few people are filling it.

1. Check your own robots.txt first. Open it and look for OAI-SearchBot, ChatGPT-User and PerplexityBot. If any of them carries Disallow: /, you have made yourself invisible in AI answers without necessarily meaning to. It takes five minutes to correct.
2. Write the Swedish material that is missing. When 0.61% of the text is in your language, every substantial Swedish page about your industry is disproportionately valuable. Concrete numbers, prices, comparisons and dates, not marketing prose.
3. Expect English sources to compete for Swedish questions. Your competitor in the AI answer is not always another Swedish company. It can just as easily be an international site that happened to write about the subject in more detail.
4. Start with ChatGPT. As we have shown before, ChatGPT dominates Swedish AI search by a wide margin. That is where the effect shows up first.

The first step is knowing how you are mentioned today, and what language the sources behind that answer are written in. Most Swedish companies have never checked.

Questions we get about the Swedish source picture