Why do AI answers change every time I ask?

Language models generate text with some randomness, engines with search may retrieve different pages, and products update their models often. That is why visibility is measured as a percentage across many questions and runs.

Do AI assistants favour big brands?

They favour brands they have seen described often and consistently, which tends to help established brands. Smaller brands win by being the clearest answer for specific, narrower questions.

How ChatGPT, Gemini, Claude and Perplexity choose which brands to recommend

By Get AI Vizible · Updated 24 October 2026

Short answer

AI assistants recommend brands they can describe confidently. That confidence comes from what the model learned in training (how often and how consistently a brand is described across the web) and, for engines that search, from the pages they retrieve at question time. Brands that are clearly positioned, consistently described, frequently mentioned by trusted third parties and easy to crawl are named more often. The wording of the question, the engine and even the run can change the answer.

Two sources of knowledge

Model memory. During training, a model reads a large share of the public web. Brands that appear often, in many independent sources, with consistent descriptions, become associations: "for X, people use A, B and C".

Live retrieval. Some engines search the web while answering and cite the pages they used. Perplexity does this on every answer; ChatGPT, Gemini and others do it for many questions. Retrieved pages can introduce brands the model did not "remember".

Clear category and audience. "A CRM for small real-estate teams" is easier to match than "the future of customer relationships".
Corroboration. The same facts on your site, review sites, directories and articles.
Recency. Current pricing and features; old or conflicting information lowers confidence.
Comparisons. Engines answering "best X" questions lean on comparison and list articles.
Accessible text. Facts in HTML text, not only in images, PDFs or client-side scripts.

How engines differ

Engine	How it usually answers	What helps most
ChatGPT	Model memory, with web search for many recommendation questions	Broad, consistent coverage; OAI-SearchBot access
Gemini	Model memory and Google's index	Strong Google presence and structured data
Claude	Model memory, with web search when enabled	Clear, factual descriptions on credible sources
Perplexity	Live web search with citations on every answer	Pages that answer the question directly and rank well
Grok	Model memory, plus X (Twitter) and web content when searching	Presence in public discussion
DeepSeek	Model memory, with search in some apps	Broad web coverage, including non-English sources for global brands

These patterns are general and change as the products evolve, which is why measuring each engine regularly matters more than any rule of thumb.

Question wording changes the answer

"Best project management tool" and "best project management tool for agencies under 20 people" produce different lists. Track the specific questions your buyers ask, including audience, location and budget.

What you can control

Your own pages, your listings on third-party sites, the accuracy of what is said about you, and your crawler settings. Measure, fix the gaps, and measure again: see how to get recommended by ChatGPT.

Frequently asked questions

Why do AI answers change every time I ask?: Language models generate text with some randomness, engines with search may retrieve different pages, and products update their models often. That is why visibility is measured as a percentage across many questions and runs.
Do AI assistants favour big brands?: They favour brands they have seen described often and consistently, which tends to help established brands. Smaller brands win by being the clearest answer for specific, narrower questions.