Brand Recognition vs. Discovery. A key framework throughout LLM Optimizer is the distinction between brand recognition — how well AI represents your brand when people search for it by name — and inbound discovery — how often AI surfaces your brand when people search your category without prior knowledge of you. Both matter, but they require different strategies. Brand recognition improves through authority signals, earned media, and training data presence. Discovery requires appearing in category-level content, answering the questions your audience asks before they know you exist, and being present in the YouTube videos, Reddit threads, and web pages that LLMs cite for category queries.

The emerging science of LLM visibility reveals a fundamental shift in how information gains authority online. The most significant recent finding comes from NanoKnow (2026), which demonstrates that content appearing frequently in training data more than doubles a model's accuracy on related questions — and that the advantage compounds when content is both memorized during training and retrievable at inference time. This means the traditional SEO playbook of optimizing for a single ranking algorithm is being replaced by a dual imperative: getting into training corpora through widespread, high-quality publication, while simultaneously remaining citable through structured, authoritative web presence.

Across the research, a consistent pattern emerges: AI search engines overwhelmingly favor earned media over brand-owned content, citing third-party sources 72-92% of the time. Content that includes quotations from authoritative sources gains +41% visibility — the single most effective optimization technique identified. Meanwhile, YouTube has rapidly become the dominant social citation source for LLMs, with its share doubling to 39% between August and December 2024. Critically, video LLMs process content through transcripts, not visual analysis — a 7B model trained on YouTube transcripts outperformed 72B models, proving that transcript quality matters far more than production value.

Reddit has emerged as the #2 social citation source for LLMs, with unique authority dynamics. Reddit was foundational in LLM training through datasets like WebText and the Common Crawl, and continues through $60M (Google) and $70M (OpenAI) annual licensing deals. Unlike YouTube's channel-centric authority, Reddit's influence comes from multi-user validation — upvoted comment consensus, especially in "best X for Y" recommendation threads, creates credibility signals that LLMs weight heavily. The Toronto GEO paper classifies Reddit as "Social" — a category AI search engines suppress in direct citations — yet Reddit's pervasive presence in training data means it heavily shapes baseline model knowledge even when not explicitly cited.

A critical "two-world" split has emerged between Google AI Overviews and standalone LLMs. 76% of AI Overview citations pull from top-10 organic pages — making traditional search rankings the primary signal for AIO inclusion. But for standalone LLMs like ChatGPT, only 12% of cited URLs rank in Google's top 10. The strongest predictor of AI citation across platforms is YouTube mentions (0.737 correlation), followed by web mentions (0.664) — not backlinks. Meanwhile, content freshness has become a significant signal: AI assistants cite content that is 25.7% newer than traditional search results, and 65% of AI bot crawl hits target content less than a year old. The explosive growth of AI crawlers (GPTBot up 305% YoY) makes robots.txt policy a direct lever for AI visibility.

However, this new landscape comes with important caveats. Citation accuracy across AI answer engines remains surprisingly poor (49-68%), with nearly a third of claims lacking any source backing. Citation concentration follows power-law dynamics, where the top 20 sources capture 28-67% of all citations. And LLMs exhibit strong positional bias, reliably attending to content at the beginning and end of context while ignoring the middle.

Compounding these challenges, model updates can sharply reduce citation volume. When GPT-5.3 replaced GPT-4o as ChatGPT's default, unique domains cited per response dropped 20.5% overnight — meaning brands that had achieved dynamic visibility through real-time retrieval lost it without any change on their end. This volatility reinforces the importance of parametric visibility (being embedded in training data) alongside dynamic visibility (being citable at inference time). Research into LLM parametric memory reveals that network centrality — being densely associated with high-authority brands in a model's knowledge graph — outweighs raw mention frequency. A brand that appears alongside category leaders in training data gains disproportionate visibility, even if it is mentioned less often overall. Together, these findings inform LLM Optimizer's scoring frameworks across answer optimization, video authority, Reddit authority, and search visibility analysis.