Citino teamTwo sources of an answer
When an assistant names brands, the names come from one or both of these:
| Source | What it is | How fast it changes |
|---|---|---|
| Training data | What the model absorbed when it was trained | Slowly; only when a new model is trained |
| Retrieved pages | Web pages the assistant fetches while answering | Quickly; as soon as the pages change and are recrawled |
Answers built on retrieved pages are called grounded answers. Every major assistant now grounds at least some answers in live search: Google AI Overviews and AI Mode use Google Search, the Gemini app grounds answers in the Google Search index, Perplexity runs its own index, ChatGPT search uses OpenAI’s crawler and search partners, and Claude has had web search since 2025.
Grounded answers are where marketing work pays off fastest, because changing what the assistant reads changes what it says.
What happens when an assistant searches
The steps vary by engine, but the shape is consistent:
- It rewrites your question. A buyer types “best WMS for a 3PL”. The assistant turns that into one or more search queries of its own.
- It fans out. Google describes AI Overviews and AI Mode using query fan-out: “issuing multiple related searches across subtopics and data sources” to build one response. One question becomes several searches.
- It reads the results. The assistant pulls passages from the pages it found: review sites, comparison articles, community threads, vendor pages.
- It writes the answer. It names the brands those passages support, in an order that reflects how strongly and how often they are supported, and often links the pages it relied on.
Two consequences follow. First, you are competing to be on the pages that get retrieved, not only to rank your own page. Second, brands that many independent sources mention consistently tend to be named; a brand mentioned once, on its own site, tends not to be.
Why the answer changes from run to run
Ask the same question ten times and you will rarely get ten identical answers. There are three ordinary reasons:
- Sampling. Language models choose words with some randomness, so wording and order shift.
- Different retrieval. Fan-out queries and search results vary slightly between runs, so the set of pages read varies too.
- Context. Location, language and, in some products, signed-in history can change what is retrieved.
This is why one screenshot of one answer proves nothing. A brand named in 3 of 10 runs and a brand named in 5 of 10 runs may be equally visible; the difference can be chance. To compare, you need many runs and a margin of error.
What tends to get a brand named
Across engines, the brands that get named have a few things in common:
- They are on the pages that get read. Review sites, category directories, comparison articles and active community threads for the question.
- Their claims are stated plainly. If an answer recommends a competitor “because it bills each client natively”, the competitor’s pages say exactly that, in text. Yours should too, if it is true. (See content gaps.)
- Sources agree. The same positioning appears on the brand’s site, its review profiles and in what customers and writers say about it.
- The pages are readable. Content is in the HTML, not locked in images, PDFs or scripts, and the right crawlers are allowed. (See the technical checklist.)
What this means for your plan
Start from the questions your buyers ask, not from keywords. For each question you lose, look at the pages the winning answers read and ask two things: are we on those pages, and do those pages, or ours, state the facts that would make us the answer? Those two questions turn into most of the useful work. The rest is measuring carefully enough to know whether it worked.