TL;DR: Getting into LLM training data means being present, consistently and credibly, in the corpora labs actually train on: Common Crawl’s web archive (the March 2026 crawl alone spans 345 TiB across roughly 2 billion pages), Wikipedia, news archives, forums like Reddit, and licensed publisher content. The play is durable third party coverage published before model training cutoffs, plus a consistent entity story engines can absorb. Training mentions persist for years after retrieval citations churn away.
When an AI names your brand, it is drawing from one of two layers: live retrieval (the engine searched the web just now) or training data (the model absorbed your brand during pretraining, months or years ago). Most GEO work targets retrieval, and rightly so, it moves faster. But the training layer is the one that makes a model volunteer your brand unprompted, with no search involved, in every conversation where your category comes up. This post covers how brands actually get into that layer.
What is actually in LLM training data?
Mostly the open web, filtered. Every frontier lab starts with Common Crawl, the free monthly archive of the public web: the March 2026 crawl measures 345 TiB, and the archive contains on the order of 250 billion pages collected over years, with each monthly snapshot covering roughly 2 billion pages. Labs filter this raw crawl into cleaned datasets (C4 is the best known derivative), then layer in higher trust sources: Wikipedia, books, academic papers, code, news archives, and licensed corpora from publishers that signed data deals, plus high signal community text from forums like Reddit, which has licensing agreements with both Google and OpenAI.
The filtering is the part that matters for brands. Labs aggressively deduplicate and quality filter, which means thin, repetitive, low engagement pages get cut. What survives filtering is exactly what you would expect: established publications, reference sites, active community discussion, and pages other sites link to and quote. Presence in the raw crawl is cheap. Surviving into the training mix is earned.
Why do training data mentions outlast retrieval citations?
Because retrieval is rented and training is owned. A retrieval citation exists only for the moment an engine happens to fetch your page for one answer; the next model update, ranking shift, or competitor page can take it. A brand absorbed during pretraining is part of how the model understands your category. Ask an assistant to name brands in a space with search disabled and it still produces names, the ones woven into its training data. Those mentions persist for the operating life of the model, typically a year or more, and often carry into successor models trained on overlapping corpora.
The two layers also compound. Models that already associate your brand with a category from training treat retrieved mentions of you as confirmation rather than new information, which shows up as better placement and framing in blended answers. We covered the retrieval side in how to get your brand mentioned by AI; this is the slower layer underneath it.
How does timing against training cutoffs work?
Models have public training cutoff dates, and content published after a cutoff does not exist for that model, no matter how good it is. The practical rule from practitioners working this layer: if a major model family trains in Q3, content published by Q1 has had time to be crawled, replicated across sources, and picked up in the datasets that feed the run. Coverage earned the week before a cutoff probably missed the train.
You cannot schedule around every lab’s runs, and you should not try. The durable strategy is continuous presence: a steady output of coverage and citable content means whenever any lab snapshots the web, your brand is there. Brands that publish in bursts around product launches leave gaps; whichever months the crawl and cutoff land in decide whether the burst counted.
What actually gets a brand into the training mix?
Third party coverage in sources that survive filtering. Ranked by durability:
Earned media in established publications. News and industry press are core training corpus material: high authority, well linked, archived for years. One feature story naming your brand, what you do, and who you serve is a permanent training data asset. This is the same mechanism behind digital PR for AI visibility, operating on a longer clock.
Wikipedia and Wikidata, if you clear notability. Wikipedia is among the most heavily weighted sources in every corpus. It is also the most policed; the honest assessment of when it is worth pursuing is in our Wikipedia for AI visibility guide. Wikidata entries have lower thresholds and feed entity understanding.
Authentic community presence. Reddit’s licensing deals put its threads directly into training pipelines. Organic discussion of your brand in category subreddits, real users comparing and recommending, reads as ground truth to a model. Manufactured astroturf gets deleted by moderators before it is ever crawled, and the Reddit playbook covers the line.
Data and research other sites cite. Original statistics get quoted and republished across dozens of domains, and each republication is another training data instance associating your brand with your category. A single annual industry report can seed hundreds of crawlable mentions.
Your own site, structured and stable. Owned content is the weakest single signal but the substrate everything checks against. Consistent entity data (same name, description, category everywhere) across your site and profiles teaches models who you are, and stable URLs that persist across years of crawls beat pages that churn.
What keeps brands out of training data?
Three own goals. First, blocking crawlers: robots.txt disallows for CCBot (Common Crawl), GPTBot, and ClaudeBot, or CDN defaults doing it silently, remove you from the corpora at the source. Opt out signals like robots.txt and llms.txt are respected by major labs, which is exactly why accidental opt outs are so costly; audit yours with should you block AI crawlers. Second, thin content at scale: quality filters cut it, so a hundred generated pages contribute nothing while one dense page survives. Third, inconsistency: if your brand name, positioning, and category description vary across sources, the model absorbs a blurred entity, and blurred entities do not get volunteered in answers.
How do you measure training data presence?
Ask the models with retrieval off. Query ChatGPT, Claude, and Gemini in modes that do not search (“without searching the web, what brands do you know in X category?”) and log whether you appear, how you are described, and who else is named. Do this quarterly and on every major model release. New model versions are the scoreboard: if coverage you earned in the past year shows up in a model trained after it, the pipeline is working. Pair this with the retrieval side metrics in what is AI visibility for the full picture, and expect the training layer to move on a 6 to 18 month lag, not weeks.
What does a 12 month training data campaign look like?
Quarterly waves of durable coverage, each aimed at a different corpus source. Quarter one: publish one original data asset, a survey, a pricing study, an industry benchmark, and pitch it to trade publications, because data earns coverage that names you and gets requoted for years. Quarter two: entity infrastructure, Wikidata entry, consistent profiles across the platforms your category uses, structured data on your own site, so every future crawl absorbs a clean record. Quarter three: community layer, sustained authentic participation where your buyers discuss the category, plus expert commentary placements that put your people’s names next to your brand in crawlable text. Quarter four: refresh the data asset with new numbers and pitch the update, which converts a one time story into an annual citation habit for the journalists who covered it.
None of these steps knows when any lab’s next training cutoff falls, and none needs to. By the end of the cycle your brand exists in news archives, reference databases, community threads, and structured data, the four corpus sources every filtering pipeline keeps. Whenever the snapshot happens, you are in it.
FAQ
Can I pay to be included in training data? Not directly, as a brand. Labs sign licensing deals with large publishers and platforms, not individual companies. Your indirect path is coverage in the publications and platforms that hold those deals.
Does retrieval based GEO still matter if I focus on training data? More, not less. Retrieval moves in weeks and compounds into the training layer over time, since the coverage you earn for retrieval is the same material future crawls absorb. It is one pipeline with two payout schedules.
How do I know if my site is in Common Crawl? Search your domain in the Common Crawl index (index.commoncrawl.org). If recent crawls show your pages, you are in the raw corpus; surviving filtering is then about quality and links.
Can bad press get trained into models too? Yes, the corpus is neutral. Negative coverage that dominates your entity’s footprint becomes part of how models describe you, which is another argument for steady positive coverage volume.
Do smaller brands have any realistic shot? Yes, within their category and geography. Models absorb long tail entities fine; a med spa consistently covered in local press and active community discussion becomes the entity models know for that service in that city. The bar is consistency, not fame.
Building the coverage layer that feeds both retrieval citations and the next training run is exactly what we do. Get in touch to map your current entity footprint, or start with the ROI calculator.
Tagged