What is Training Data (for LLMs)?
Training data is the corpus of text a language model learns from, largely web content gathered by crawlers, plus licensed and curated datasets. A model's default beliefs about your brand are a compression of what that corpus says, which is why what the web wrote about you two years ago can still shape AI answers today.
How brands enter training data
Models learn about brands from crawlable web pages, review sites, forums like Reddit, news coverage, and reference sites. Content published today is a candidate for the next training run. That lag means GEO is partly an investment: the third-party coverage you earn this year becomes model priors next year.
What you can influence
- Allow or block training crawlers per bot in robots.txt
- Keep authoritative, factual pages about your brand crawlable
- Earn mentions on high-signal third-party surfaces
- Fix wrong public facts at their source; models learn the errors too