AI Training Data & LLM Memorization

Brass-SEO draws on this to explain that language models don't just learn stylistic patterns from training text — they memorize and can reproduce specific passages verbatim. Carlini et al. tested the GPT-Neo/GPT-J family and found the 6-billion-parameter model memorizes at least 1% of its training dataset (The Pile) well enough to reproduce it on the right prompt, with memorization rising log-linearly as model size grows — roughly a 19-percentage-point increase for every tenfold jump in parameters (R²=99.8%) — and rising further with how much context precedes the prompt: 33% of tested sequences were extractable at 50 tokens of preceding context, versus 65% at 450 tokens. A brand that appears often and in detail across a model's training corpus isn't just represented in it; it's memorized, which is why some LLMs answer confidently and specifically about businesses no one asked them to look up.

Examines:

A peer-reviewed (ICLR 2023) empirical measurement of verbatim training-text memorization across the GPT-Neo/GPT-J model family, testing how memorization scales with model size and prompt context length.

Brass-SEO draws on:

The log-linear memorization-vs-model-size relationship and the context-length effect — cited when Brass-SEO explains why larger, more capable models recall brand-specific facts more reliably than smaller ones.

Primary source

arxiv.org/abs/2202.07646

Brass-SEO cites this as primary-source confirmation that open web crawls are the dominant raw material behind a frontier model's factual knowledge. The GPT-3 paper documents the model's pretraining mix directly: filtered Common Crawl supplied 410 billion of GPT-3's training tokens and was weighted at 60% of the total training mix, ahead of WebText2, two book corpora, and Wikipedia combined. The same kind of open crawl that produced that mix is still running today, which is why a page a business never wrote with an AI audience in mind can still end up shaping what a model knows about that business.

Examines:

The original GPT-3 paper (Table 2.2), documenting the exact per-source token counts and mixing weights of the model's pretraining corpus.

Brass-SEO draws on:

The 60%-weighted, 410-billion-token Common Crawl share of GPT-3's training data — the evidentiary basis for treating open-web crawler access as a training-data decision, not only a search-indexing one.

Primary source

arxiv.org/abs/2005.14165

Brass-SEO uses this to explain why an LLM will confidently name one brand and never mention a real, comparable competitor. Kandpal et al. correlated question-answering accuracy against how many pretraining documents mentioned the relevant fact, tested up to 176-billion-parameter models on TriviaQA, and found accuracy tracks document count closely — the authors conclude today's models would need to scale by many orders of magnitude to reach competitive accuracy on facts backed by few pretraining documents. A business with a sparse web presence isn't being penalized for quality. The model simply never saw it mentioned often enough to recall reliably.

Examines:

A peer-reviewed (ICML 2023) study correlating LLM question-answering accuracy with the number of pretraining documents supporting each fact, tested on models up to 176B parameters.

Brass-SEO draws on:

The document-count-to-accuracy correlation — cited when Brass-SEO explains that AI-recommendation gaps often trace to training-data document volume, not brand quality.

Primary source

arxiv.org/abs/2211.08411

Brass-SEO treats this as the authoritative source for what blocking OpenAI's crawlers does and doesn't change. OpenAI's documentation names three separate agents with three separate purposes: GPTBot performs the automated crawling used to train future models and can be disallowed in robots.txt to keep content out of training; OAI-SearchBot indexes pages specifically for ChatGPT search results; and ChatGPT-User fires only when a live user's request triggers it, for which OpenAI states robots.txt rules may not apply. Blocking GPTBot stops future training inclusion. It has no effect on what ChatGPT retrieves live when a user asks it to browse or search.

Examines:

OpenAI's official specification of its three web crawlers — GPTBot, OAI-SearchBot, and ChatGPT-User — their distinct purposes, and their independent robots.txt controls.

Brass-SEO draws on:

The training-vs-live-retrieval distinction across GPTBot, OAI-SearchBot, and ChatGPT-User — the basis for Brass-SEO's guidance that a single blanket robots.txt rule conflates two separate decisions.

Primary source

developers.openai.com/api/docs/bots

Brass-SEO draws on this to separate a training-data decision from a ranking decision. Google's documentation states the Google-Extended token controls whether crawled content "can be used to train future generations of Gemini models" and for AI grounding, and states explicitly that it "does not impact a site's inclusion in Google Search nor is it used as a ranking signal." Google-Extended has no separate crawler user agent of its own — it's a robots.txt control token layered onto Google's existing crawl, not a new bot to detect. Disallowing it opts a site out of future Gemini training without touching search visibility.

Examines:

Google's official specification of the Google-Extended robots.txt token governing use of crawled content for Gemini model training and AI grounding.

Brass-SEO draws on:

The explicit statement that Google-Extended has no effect on Search ranking or inclusion — cited when Brass-SEO reassures users that a training opt-out isn't an SEO risk.

Primary source

developers.google.com/search/docs/crawling-indexing/google-common-crawlers

Brass-SEO cites this when explaining the single largest control point for keeping a site out of the shared open-web datasets many LLMs are trained on. Common Crawl's own documentation identifies its crawler as CCBot/2.0, confirms it's blockable with a standard robots.txt disallow rule, and publishes its crawler IP ranges so site owners can distinguish real CCBot requests from bots that spoof the user agent. Common Crawl feeds directly into corpora like the one that supplied 60% of GPT-3's training mix, so a CCBot block is one of the few actions that reaches multiple downstream models at once rather than just one company's.

Examines:

Common Crawl's official crawler-identification page — user agent string, robots.txt blocking method, and published IP ranges for verification.

Brass-SEO draws on:

The CCBot user agent and IP-verification method — the specific, checkable robots.txt rule Brass-SEO points to when a user wants to exit shared open-web training corpora.

Primary source

commoncrawl.org/ccbot

Brass-SEO monitors this for the same training-versus-retrieval split, applied to Claude specifically. Anthropic's documentation names three separate agents: ClaudeBot, which collects web content to train and improve models; Claude-User, which fires only on a live user's query for real-time retrieval; and Claude-SearchBot, which crawls to improve search result quality. Each has its own independently blockable robots.txt token, and Anthropic states it also honors the Crawl-delay extension. A site owner who blocks only ClaudeBot has opted out of training while leaving Claude's live browsing untouched, which is very often the outcome they actually wanted.

Examines:

Anthropic's official disclosure of its three distinct web agents — ClaudeBot, Claude-User, and Claude-SearchBot — their separate purposes, and independent robots.txt controls.

Brass-SEO draws on:

The three-agent breakdown and independent blocking controls — cited when Brass-SEO explains why a customer's "I blocked the AI bot" assumption often doesn't match what actually happened.

Primary source

support.claude.com — does Anthropic crawl the web