What isTraining data
Training data is the body of text, and sometimes images, code and other content, that an AI model learns from before it's released.
It shapes the model's general knowledge, including what it "remembers" about brands and categories. It's separate from the pages a product retrieves at answer time, and many AI companies offer separate crawler controls for each.
Connected terms
The terms next to Training data in AI search, and where to read more.
LLM→An LLM (large language model) is an AI model trained on very large amounts of text so it can predict the next words in a sequence, and through that, answer questions, summarise and write.Knowledge cutoff→A knowledge cutoff is the point in time after which an AI model has no information from its training data.AI crawler→An AI crawler is an automated bot that fetches web pages on behalf of an AI company, for training models, building an AI search index, or retrieving a page when a user asks a question.All Fundamentals guides→Checked on 31 October 2026. "Our view" marks MentionX opinion. Ledgerline is a fictional company used for illustration.
01Training data vs retrieved content
| Training data | Retrieved content | |
|---|---|---|
| When it's used | Before the model is released | At the moment of a question |
| Can change quickly? | No, only with new model versions | Yes, as pages are recrawled |
| Typical crawler control | GPTBot, ClaudeBot, Google-Extended | OAI-SearchBot, PerplexityBot, Claude-SearchBot, Googlebot |
| Shows citations? | No | Often |
02What opt-out controls exist?
- OpenAI: disallowing GPTBot indicates content shouldn't be used for training; OAI-SearchBot is separate (OpenAI).
- Anthropic: restricting ClaudeBot signals future content should be excluded from training (Anthropic).
- Google: Google-Extended manages use for Gemini training and grounding, without affecting Search (Google).
03Example
Illustrative. A media company blocks GPTBot and ClaudeBot to keep articles out of training, but allows OAI-SearchBot and Claude-SearchBot so its reporting can still be cited in AI search answers.
04Why it matters for GEO
Training-data decisions are business and legal choices, not only marketing ones. Make them deliberately, and don't confuse them with search access. For visibility in cited answers, retrieval controls are what matter most in the short term.
05Common misconceptions
- "Blocking GPTBot removes me from ChatGPT." Search visibility is controlled separately, through OAI-SearchBot.
- "Opting out deletes past training." Controls apply to future crawling.
- "More training data about me means more mentions." No company documents a direct link between the two.
06From MentionX data: answers with and without sourcesOriginal data
Method. Read-only aggregate queries over MentionX scans from 22 August to 29 October 2026 (UTC): 2,217 engine runs (2,154 completed) across 18 anonymised projects and six engines.
| Engine | Brand mentioned | Answers with no citations |
|---|---|---|
| Gemini | 47.3% | 50.7% |
| OpenAI | 42.9% | 30.2% |
| Perplexity | 31.5% | 4.1% |
| AI Mode | 31.3% | 12.6% |
| Copilot | 26.0% | 47.0% |
| DeepSeek | 24.0% | 99.0% |
When an answer has no citations, it's likely drawing mostly on training data. The share varied from 4.1% to 99.0% by engine.
Unsourced unbranded answers also named more brands on average (1.44 against 1.11).
Our view: for engines that rarely cite, the practical lever is consistent, accurate descriptions of your brand on widely read public pages, which is what future training data will contain.
Limits. A convenience sample of mostly B2B software projects, run through engine APIs, which can behave differently from the consumer apps. Perplexity, Copilot and DeepSeek figures come from a single day (26 October 2026), so treat them as a snapshot, not a trend.
07Worked example: Ledgerline's old description
Ledgerline is a fictional accounts-payable tool. Figures are illustrative.
DeepSeek describes Ledgerline as "an invoice scanning tool", which was accurate in 2023 before it added payments and approvals. No sources are shown.
Ledgerline can't edit the model. It updates its own homepage and About page with a one-sentence description, then asks the three review and directory sites that still use the 2023 wording to update theirs. It records the current answer and checks again after the next model release.
08What opt-out controls exist?
OpenAI documents GPTBot for training (OpenAI); Google offers Google-Extended for Gemini model training and grounding, separate from Search (Google); Anthropic documents ClaudeBot (Anthropic). These affect future collection only. Weigh opting out against the chance that models know less about you.
09Common misconceptions
- "If I'm not in the training data, AI can't mention me." Engines that search can find you at answer time.
- "Opting out removes me from AI." It stops future collection by that crawler; it doesn't touch search or existing models.
Questions people ask
3 answersCan I remove my site from AI training data?
+
You can opt future content out with crawler controls such as GPTBot, ClaudeBot or Google-Extended. Past training isn't undone.
Does opting out of training hurt AI search visibility?
+
Companies document separate controls for training and search. OpenAI, for example, separates GPTBot from OAI-SearchBot.
Is training data the same as the search index?
+
No. Training data shapes the model; a search index is what retrieval systems query at answer time.