Most AI visibility advice conflates two fundamentally different citation mechanisms. Retrieval-augmented engines (Perplexity, Bing Copilot) respond to on-site changes in days. Base models (GPT, Claude without browsing) require training cycles that run 6-18 months behind. The optimization approach differs entirely.
The most common mistake in AI visibility strategy is treating all LLM citations as equivalent. They are not. There are two fundamentally different citation mechanisms, and they require different optimization approaches, different timelines, and different measurement methods.
Mechanism 1: Retrieval-augmented generation (RAG)
Perplexity, Bing Copilot, and ChatGPT with web browsing enabled retrieve pages at query time. They run a web search, pull the top results, extract relevant passages, and cite the source. For these engines, the citation mechanics are similar to search engine optimization — with the addition that the model evaluates passage quality, freshness, and authority before including a source.
Changes to your on-site content, structured data, and sitemap can appear in retrieval-augmented citations within 24-72 hours of indexing. FAQPage schema, Article schema, and freshness signals (sitemap lastmod, dateModified) directly influence which pages get retrieved and cited. These are the engines where short-term tactical improvements produce measurable results.
Mechanism 2: Base model knowledge
ChatGPT (without browsing), Claude, and Gemini Base draw from pre-trained knowledge with a hard cutoff date. These models do not retrieve content at query time — they recall from training data. Content published after the training cutoff is invisible to base model citations, no matter how well-optimized it is.
The training cycle gap is 6-18 months for most major models. OpenAI typically trains on data through a cutoff 6-12 months before release. Anthropic and Google have similar patterns. This means on-site changes you make today will not influence base model citations until the next training cycle includes your domain — which could be a year from now.
For base model citations, the relevant optimizations are entity infrastructure: Wikidata entities, Wikipedia articles, Organization schema with sameAs links, and Common Crawl coverage. These are re-evaluated at each training cycle and produce stable, permanent citation improvements across training runs.
What this means for how you allocate effort
If you need citation results within 4-8 weeks, invest in retrieval-augmented optimization: FAQPage schema, Article schema with fresh dateModified, complete sitemap with current lastmod, and fast page load times. Measure results in Perplexity and Bing Copilot.
If you are building for sustained citation presence over 12-24 months, invest in entity infrastructure. Wikidata completeness, Wikipedia notability, and sameAs graph depth produce stable citation rates that survive training cycles without requiring ongoing maintenance.
Most teams should do both. The mistake is investing only in entity infrastructure and then wondering why Perplexity still isn't citing them — or investing only in on-site optimization and being surprised that base model citations haven't changed.
This week's citation data
Across the tracked domain set, retrieval-augmented citation rates respond to on-site changes with a median lag of 6 days. Base model citation rates for the same domains show no correlation with on-site changes in the current tracking window — consistent with the 6-18 month training cycle gap. Entity coverage scores (Wikidata completeness, Wikipedia presence, sameAs depth) show strong correlation with base model citation rates across the entire domain cohort.
The separation is clear in the data: entity infrastructure predicts base model citations; on-site signals predict retrieval-augmented citations. Treating them as the same thing produces strategies that are half-optimized for each and fully optimized for neither.
Scan your domain to see your current scores across both dimensions — entity coverage for base model citation potential, and on-site signals for retrieval-augmented citation readiness.
Put this into practice
See how your domain scores on the signals covered in this edition. Veezow runs a free AI visibility scan — robots, sitemap, structured data, bot access, and off-site presence.
Run a free scan →