Question
The Role of Wikipedia in Training LLMs to Recognize Your Brand
LLM answers 4 min read
In GPT-3’s 2020 training mix, Wikipedia was a small part by volume and a larger part by influence. OpenAI’s GPT-3 paper, which describes its training mix in detail, gave English Wikipedia 3% of the training weight, about 3 billion tokens. Because the authors sampled sources they considered higher quality more often, the model saw that Wikipedia text about 3.4 times during training, while it saw most of the much larger web crawl less than once. If current models are trained in a similar way, which their makers no longer document in that detail, an accurate Wikipedia article about your brand may be among the most repeated descriptions of you a model encounters. If it does not, you cannot simply write one, and you should not try.
How Wikipedia figures in training data
The GPT-3 paper lists the weighting of its training mixture:
| Dataset | Share of the training mix |
|---|---|
| Common Crawl (filtered) | 60% |
| WebText2 | 22% |
| Books1 | 8% |
| Books2 | 8% |
| Wikipedia | 3% |
The weighting was deliberate. In the authors’ words, datasets were “not sampled in proportion to their size, but rather datasets we view as higher-quality are sampled more frequently,” which is how a 3% share turned into repeated exposure.
An earlier OpenAI model shows the same source being treated as special in a different way. For GPT-2, OpenAI built a dataset called WebText from pages linked from Reddit, and the GPT-2 paper says it “removed all Wikipedia documents from WebText since it is a common data source for other datasets.” Wikipedia was common enough in other training sets that the researchers excluded it to keep their analysis clean.
These papers describe models from 2019 and 2020. The crawler documentation companies publish today names their training bots but does not list training datasets, so the exact role of Wikipedia in current models is not documented in the same way. In both papers Wikipedia was handled deliberately, weighted up in one and excluded from the other.
Wikipedia also shows up at answer time
Wikipedia’s influence is not limited to training. In Pew Research Center’s look at Google AI summaries from March 2025, the three most cited sources were Wikipedia, YouTube and Reddit, which between them made up 15% of the sources listed.
So a Wikipedia article about your company can matter twice: as text a model may have learned from, and as a page a system may retrieve when someone asks about you.
Who gets a Wikipedia article
Wikipedia decides, not the company. Its notability guideline says a topic is presumed suitable for an article “when it has received significant coverage in reliable sources that are independent of the subject.” It adds that “presumed” means significant coverage creates an assumption, not a guarantee.
The key word for a brand is independent. Press releases, sponsored content and the company’s own website do not establish notability. Coverage by reputable publications that chose to write about the company does.
Wikidata, the structured knowledge base connected to Wikipedia, has its own notability policy. An item is acceptable if it meets at least one of three criteria: it has a valid link to a page on a Wikimedia project such as Wikipedia; it “refers to an instance of a clearly identifiable conceptual or material entity that can be described using serious and publicly available references”; or it fulfills a structural need in Wikidata.
What a company may and may not do
Wikipedia’s conflict of interest guideline is direct: “COI editors are strongly discouraged from editing affected articles directly, and can propose changes on article talk pages instead.” Editors with a conflict of interest, including paid editors, are expected to disclose it whenever they seek to change an affected article.
In practice:
- Do not write or edit your own article. Propose corrections on the article’s talk page, disclose your connection, and cite independent sources for each requested change.
- Do not pay someone to edit undisclosed. Paid editing must be disclosed under Wikipedia’s rules.
- Fix errors with sources. If an article contains outdated or wrong information, a talk page request backed by reliable, independent sources is the legitimate route.
Why this matters for AI recognition
If models learned about your company partly from Wikipedia, errors in your article can become errors in AI answers, and those can persist until a new model is trained. Correcting the article through the proper process can help future training and immediate retrieval at the same time.
If you have no article, AI systems describe you from whatever else they find: your own site, news coverage, directories and discussions. That makes the rest of the public record more important, not less.
What to do
- Check whether an article exists about your company or its founders, and read it for accuracy.
- Correct errors properly, through talk page requests with disclosure and independent sources.
- If there is no article, focus on independent coverage. The coverage that would justify a Wikipedia article is the same coverage that helps AI systems recognize you.
- Keep your own information consistent. Your website, business profiles and press materials should describe the company the same way, so that whichever source a model learned from, the facts match.
- Check Wikidata. If an item exists for your company, make sure its basic facts are accurate, following Wikidata’s rules for sourcing.
Start by reading what Wikipedia already says about you, if anything. An accurate article earned through independent coverage can improve how models describe your brand; an article you write yourself runs against Wikipedia’s conflict of interest guideline and puts at risk the credibility you hoped it would add.