AI assistants generate answers from two sources: what the model learned during training, and what it retrieves from the web at the moment you ask. Training gives broad, general knowledge that's frozen at a cutoff date. Retrieval adds current, specific information with citations. Most answers about products and brands blend both. Understanding how they combine tells you what you can influence, and how fast.
Source 1: what the model learned in training
A large language model (LLM) is trained by reading an enormous amount of text: public web pages, books, code, forums, reference material and licensed data. During training it learns statistical patterns: which words, facts and ideas tend to go together.
The result is a model that can:
- Explain concepts ("what is a CRM?")
- Summarize common knowledge ("what are the main project management methods?")
- Recall associations ("Brand X is known for its free tier")
Limits of training knowledge
- It has a cutoff. The model knows nothing from after its training data was collected. A brand that launched, rebranded or changed pricing since then may be unknown or misdescribed.
- It's compressed. The model doesn't store web pages; it stores patterns. Niche facts mentioned rarely may be fuzzy or wrong.
- It reflects consensus. If most sources describe you one way, that's what the model learns, even if it's outdated.
- It changes slowly. It updates only when a new model version is trained and released, typically months apart.
What shapes training knowledge about your brand
How often and how consistently your brand appears across the web before the cutoff: your own site, news coverage, reviews, forum discussions, directories, documentation and partner pages. Many companies let some or all AI training crawlers access their content; see our AI crawlers and robots.txt guide for how to decide.
Source 2: what the assistant retrieves live
To answer questions about current or specific topics, assistants use retrieval: they search an index, read relevant pages and write an answer grounded in them. This is often called retrieval-augmented generation (RAG).
The typical flow:
- Decide whether to search. The assistant judges whether the question needs fresh or specific information.
- Write search queries. It turns your question into one or more queries, often adding more specific variants.
- Retrieve candidates. It pulls results from a search index: its own, a partner's like Bing or Google, or both.
- Read passages. It extracts the most relevant chunks from the top results.
- Generate and cite. It writes the answer from those passages and attaches source links.
Strengths and limits of retrieval
- Current. It can use a page published yesterday.
- Citable. Users can see and click the sources.
- Dependent on access. If crawlers can't fetch your page, or your content only appears after JavaScript runs, you're invisible here.
- Passage-driven. The best-matching chunk wins, not the best overall page.
When does an assistant search?
It depends on the platform and the question:
| Question type | Likely to trigger search? |
|---|---|
| "What is [concept]?" | Often not, unless it's new or niche |
| "How do I [task]?" | Sometimes |
| "Best [product] for [use case]" | Usually |
| "[Brand A] vs [Brand B]" | Usually |
| "[Brand] pricing / reviews" | Almost always |
| Recent news, events, local info | Almost always |
Perplexity searches for almost every question. ChatGPT, Claude and Gemini decide case by case, and users can also ask them to search. Google's AI Overviews and AI Mode are always grounded in Google Search. See how each assistant finds sources for the details.
How the two sources combine
Even when an assistant searches, the model's training still shapes the answer:
- Query writing. The model's existing understanding determines what it searches for. If it doesn't associate your brand with a category, it may not search for you by name.
- Interpretation. It reads retrieved pages through what it already "knows," which affects how it summarizes them.
- Gap-filling. Where retrieved sources are thin, the model fills in from training, which is where outdated or wrong descriptions often come from.
This is why a brand can rank well in search and still be under-mentioned in AI answers, or be described with an old positioning even when the current website is accurate.
What this means for AEO
Work on both sources, with different expectations for speed:
For retrieval (fast feedback)
- Allow AI search crawlers and user fetchers.
- Serve key content in server-rendered HTML.
- Publish pages that answer specific buyer questions directly. See how to write content AI assistants cite.
- Make sure the third-party pages assistants retrieve, such as reviews, roundups and directories, describe you accurately.
For training (slow, compounding)
- Describe your brand, category and positioning consistently everywhere.
- Earn coverage in widely read publications and communities.
- Retire or update outdated pages that describe old products or pricing.
- Use clear structured data to anchor your brand as an entity.
How to tell which source an answer used
- Citations present? The answer used retrieval. Check which pages were cited.
- No citations, general description? Likely model knowledge. Compare the description with your current positioning to spot outdated training data.
- Mix of both? Look for claims that don't appear in any cited source; those came from the model.
Tracking this across many prompts, as covered in AI visibility metrics, shows whether your problem is mostly retrieval (fixable quickly) or mostly model knowledge (a longer project).
The bottom line
AI answers come from a frozen-but-broad memory and a live-but-selective search. You can't edit the memory directly, but you can shape what future models learn by being consistent and widely covered. You can influence retrieval this week by being crawlable and by publishing the clearest answers. Strong AEO programs do both.
