Getting your brand into LLM training data — the hard truth, and what to do while you wait

The one-line version: Approximately 60% of ChatGPT conversations are answered using the model’s innate knowledge — patterns built up during pre-training and post-training — not from live SearchGPT retrieval (Seer Interactive). If your brand is not represented as a coherent entity in that training data, live-retrieval optimization can only fix part of the problem. Brands seeking inclusion in an LLM’s innate training data should expect to wait months or even years.

The 40/60 split, and why it matters

According to Seer Interactive’s testing:

• ~40% of the time, SearchGPT is triggered to supplement ChatGPT’s innate knowledge. It formulates a search query and pulls relevant webpages from Bing’s index.
• ~60% of the time, ChatGPT answers directly from its innate knowledge — pre-training data plus post-training fine-tuning.

That 60% is the part of the game that traditional SEO doesn’t touch. And it’s the part where brand mentions elsewhere on the web are the currency — because the model has learned about your brand from every page that mentioned you before the training-data cutoff, not from your own website in isolation.

The evidence: 48% of citations are earned media

Omniscient Digital’s analysis of 23,387 unique citation sources across 240 branded queries on ChatGPT, Perplexity, Gemini, AI Mode and AI Overviews found (Outpace SEO summary):

Only 23% of citations come from the brand’s own website. The other 77% comes from what other people say about the brand. For customer-review intent specifically, earned media rises to 82% of citations (Outpace SEO). See the standalone earned-media article for the wider dataset.

What LLM training data actually looks like

• LLM training data is a “black box” — OpenAI has not published the weighting or inclusions for ChatGPT 4. Stanford’s Foundational Model Transparency Index rated ChatGPT 4 at 0/10 for transparency (Seer Interactive).
• Some providers are more open — Meta’s Llama has been unusually transparent about training inclusions.
• Reverse-engineering the training corpus is described as “comparable to reverse-engineering Google’s algorithm” — you infer from what the model produces, not from a published spec.
• Training data updates happen when new frontier models are released, not continuously. Being absent from the last snapshot means being absent for the current generation.

What the model needs to know to represent your brand well

The Outpace/Omniscient framing is that a brand either:

1. Exists in training data as a coherent, well-defined entity — associated with specific topics, product categories and value propositions the model can reproduce; or
2. Has insufficient presence to be recognized or confidently associated with relevant topics — the model either doesn’t know you, or worse, misattributes you.

Getting into category 1 requires third-party corroboration of your brand’s identity across the pages the training crawl saw. That’s why owned content alone is not enough: the model looks for pattern consistency across many sources.

The playbook — three parallel workstreams

Workstream A: Earn media that makes your brand’s identity obvious

• Independent editorial coverage in outlets your buyers read. Not press releases, not link-building schemes — reported pieces that describe what you do and who uses you.
• Named appearances in industry roundups and review sites. Being included in “Best CRMs for small teams” articles from 30 different sources is what teaches the model that you belong in the category.
• Consistent brand description across every appearance. If half the coverage calls you a “CRM” and the other half a “sales enablement tool,” the model may split the entity into two weakly-associated concepts.

Workstream B: Own the pages that describe your entity in machine-readable form

• A clean Wikipedia article if you meet notability thresholds. Wikipedia is one of the most heavily-weighted single sources in every major LLM training pipeline.
• A well-formed Wikidata item with subclass of, industry, founded, country, official website and named after filled out. LLMs draw on Wikidata for entity relationships.
• A schema.org Organization JSON-LD block on your homepage with sameAs links to your Crunchbase, LinkedIn, X, GitHub, Wikipedia and Wikidata pages. This lets any crawler resolve your brand across surfaces.

Workstream C: Optimize live retrieval for the 40%

• Rank well in Bing — SearchGPT retrieves from Bing’s index. Bing indexation is a hard prerequisite for showing up in ChatGPT Search’s live-retrieval half.
• Publish content that answers specific sub-questions cleanly. Live retrieval favors passages, not pages. See answer-first writing for the mechanics.
• Update canonical pages quarterly. Freshness matters much more on the live-retrieval path than on the training-data path.

What NOT to do

• Don’t buy links or planted mentions. Modern retrieval systems apply the same trust filters that Google does; low-authority sources are discounted or ignored. Worse, when discovered, negative content follows and does get cited.
• Don’t try to trick the training crawl. Prompt-injection attempts in page markup, hidden entity-stuffing, or fake structured data don’t survive current filtering, and they get flagged in model evaluation.
• Don’t expect fast results on the training-data path. Months or years is the honest timeline. Live retrieval is where you’ll see 30–90 day movement; training data is a multi-year positioning exercise.

The one-sentence strategic conclusion

Treat GEO like a two-track discipline: retrieval optimization to win the 40% of queries answered from live search, and entity development to win the 60% answered from training-data patterns. Neither track alone is sufficient, and the second one takes an order of magnitude longer than the first.

Related reading: Earned media is 89% of citations. Entity-first SEO covers the Wikidata / knowledge-graph mechanics in depth.