How AI assistants decide which brands to recommend: what is known and what isn't

AI assistants name a brand when two things line up: the model has a picture of it from training, and, for assistants that search, the pages they retrieve for the question describe it clearly. The exact ranking logic is not published by any major vendor, so anyone claiming to know the formula is guessing. What is documented is how the retrieval side works. What is measured, mostly by vendors with something to sell, points the same way.

This post separates three things: what the vendors say, what studies measured, and what is open. For the practical routine see our GEO guide for small teams and the shorter explainer.

Do assistants answer from memory or from live search?

Both, depending on the product and the question. A model can recall brands it saw in training data. Many assistants can also run web searches and write the answer from what they find. When search is used, the pages retrieved are the raw material.

Google describes this for its own features: AI Overviews and AI Mode may use "query fan-out", issuing multiple related searches across subtopics and data sources, and then show a wider set of supporting links than classic search (Google Search Central). Google's announcement of AI Mode describes the same idea: the question is broken into subtopics and many queries run at once (Google blog). Our glossary has an entry on query fan-out.

The consequence: your brand is not only competing for the question the buyer typed. It competes for the sub-questions the system generates behind the scenes, such as pricing, alternatives, reviews or "for small teams".

What do the vendors say about being included?

Not much about ranking, more about access:

• Google: a page must be indexed and eligible to show with a snippet. There are "no additional requirements" for AI Overviews or AI Mode, and no need for new machine-readable files, AI text files or special markup (Google).
• OpenAI: OAI-SearchBot is used to surface sites in ChatGPT's search features, and sites that opt out will not appear in ChatGPT search answers. OpenAI recommends allowing it (OpenAI bots).
• Perplexity: PerplexityBot is designed to surface and link websites in Perplexity search results (Perplexity).
• Anthropic: Claude-SearchBot crawls to improve search result quality, and blocking it may reduce visibility in search results (Anthropic).

None of these pages explains how a retrieved page becomes a recommendation. That part is not documented.

What do independent studies measure?

Three findings are reasonably solid as direction, with caveats.

1. Answers vary a lot. SparkToro and Gumshoe.ai ran 12 recommendation prompts nearly 3,000 times across ChatGPT, Claude and Google's AI features. The same list of brands appeared in fewer than 1 percent of repeated runs. Yet in the headphone prompts, four well-known brands (Bose, Sony, Sennheiser, Apple) each showed up in 55 to 77 percent of the 994 responses (Search Engine Journal). Not peer reviewed; one author's company sells tracking.

2. Cited pages overlap with top search results, but not completely. Ahrefs analyzed 1.9 million citations from 1 million Google AI Overviews in 2025 and found 76.1 percent of cited pages ranked in Google's top 10, 9.5 percent at positions 11 to 100 and 14.4 percent below position 100 (Ahrefs). That study looked at the top three citations per overview and only at AI Overviews. Treat the percentages as a snapshot of one method and one period.

3. Third-party pages carry a lot of weight for "best tool" questions. In DerivateX's May 2026 study of ChatGPT with web search across 40 B2B SaaS categories, 11.6 percent of recommendations cited the recommended tool's own site, and review aggregators such as G2 and Capterra got no citations in that sample (DerivateX). The sample is modest, one prompt per category, and the publisher sells GEO services. It does not cover other assistants or other question types.

What factors are plausible but unproven?

These are hypotheses to test on your own questions, not facts:

The GEO paper is worth reading in full: on its benchmark, adding citations, quotations and statistics improved visibility by roughly 30 to 40 percent on one metric, while keyword stuffing did little (arXiv). It does not show what happens on live commercial assistants today.

What is genuinely unknown?

• How much weight any signal gets. No vendor publishes weights.
• How much is training memory versus retrieval for a given answer.
• How stable the behavior is over time. Assistants and their search layers change. The studies above are snapshots from different months.
• Whether results for your market match studies from the US. Most published data is English-language and US-centered.
• What a personalized or logged-in answer looks like. Studies mostly use clean sessions.

This is why we recommend measuring your own questions instead of trusting any universal recipe.

What does that mean for a small team?

Three consequences you can act on without knowing the formula:

1. Be reachable. Allow the search crawlers, make pages render without JavaScript, keep them indexed.
2. Be explicit. State what you are, who it is for, what it costs and how you compare, in plain sentences. See comparison pages that AI assistants cite.
3. Be present where the answers come from. If the cited sources for your category are roundups and community threads, that is where the gap is.

What do we measure, and how can you check?

For each question and assistant, Nodpick records whether your brand was named, its position, and which domains were cited. Paid plans repeat this every week, because a single run tells you little. The report lists the domains cited in answers, so you can see which sources appear where competitors are named and you are not, the closest practical stand-in for "why them and not me". Details are on the method page. The free check (5 questions, once each, on ChatGPT, Perplexity, Gemini and Claude, no sign-up) opens with the launch; join the waitlist and we email you once, the day Nodpick opens. We have no published measurements of our own yet and will share them when we do.

FAQ

Does ranking first on Google guarantee a mention in ChatGPT? No. Overlap between search rank and citations exists but is partial, and ChatGPT's search layer is its own system.

Is there a ranking factor list? Not from any major assistant vendor.

Do Reddit and YouTube matter? Several vendor studies find both cited often, but the shares disagree between studies and months. Check which sources appear for your questions rather than chasing a platform.

What should I do first? Make sure the search crawlers can reach you, then publish honest comparison pages and measure with the tracking guide.

Why do two people get different answers? Wording, sampling, location, language and personalization all play a role, and so does plain randomness in generation.

Sources

• Google Search Central: AI features and your website
• Google: AI in Search, AI Mode update
• OpenAI crawler documentation
• Perplexity bots documentation
• Anthropic crawler documentation
• SparkToro / Gumshoe.ai study via Search Engine Journal
• Ahrefs: AI Overview citations vs. rankings
• DerivateX B2B SaaS citation study
• Aggarwal et al., GEO (arXiv)
• SE Ranking llms.txt study via Search Engine Journal
• Search Engine Journal on John Mueller and llms.txt