When an AI assistant recommends three brands for a question, it is fair to ask why those three. The honest answer is that no provider publishes a formula. But a good part of the mechanics is documented, and the rest can be observed by measuring answers over time.
This article separates what is known from what is inferred, and turns both into practical levers.
Two ways an assistant “knows” about a brand
An assistant can mention your brand for two broad reasons.
1. Model knowledge
Language models are trained on large amounts of text. If your brand was discussed widely and consistently in that material, the model may associate it with your category. This knowledge is fixed at training time: it does not update when you publish a new page today.
Providers document separate crawlers for training. OpenAI states that GPTBot is used to make its generative AI foundation models more useful and safe, and Anthropic describes ClaudeBot as the crawler that collects web content that may contribute to model training.
2. Retrieval at answer time
Many assistants also search the web while answering, then cite the pages they used. This is where your current content and third-party coverage come in.
The documentation is fairly explicit:
- OpenAI says OAI-SearchBot is used to surface websites in ChatGPT’s search features, and that sites opted out of it will not be shown in ChatGPT search answers.
- Anthropic documents Claude-SearchBot for improving search results and Claude-User for fetching pages when a user asks a question.
- Perplexity says PerplexityBot is designed to surface and link websites in its search results, and that Perplexity-User fetches pages in response to user questions.
- Google states that AI Overviews and AI Mode may use a “query fan-out” technique, issuing multiple related searches across subtopics and sources to develop a response.
In practice, most answers about products and services are a blend: the model’s prior understanding shapes the framing, and retrieved pages supply the specifics and the citations.
What this implies about brand selection
From those two mechanisms, some reasonable inferences follow. They are inferences, not documented rules, so treat them as hypotheses to test with measurement.
Consistency across sources matters. If many independent pages describe your brand the same way (“a payroll tool for small agencies”), both training data and retrieved pages point in the same direction. Mixed or outdated descriptions produce mixed answers.
The question shapes the candidate set. “Best CRM” and “best CRM for a 5-person real estate agency” trigger different searches and different sources. Niche brands often appear on specific questions long before they appear on generic ones.
Cited pages are often comparisons, reviews and discussions. For recommendation questions, assistants frequently draw on pages that already compare options. If you are absent from those pages, you are less likely to be in the answer. Your measurement data will show which page types dominate in your market.
Accessibility is a precondition, not a guarantee. If a search crawler cannot access your site, your pages cannot be retrieved for that assistant. Being accessible does not mean you will be cited.
Answers vary. The same question can produce different brands on different runs. A brand that appears some of the time is in a different position from one that appears nearly always, and you can only see that difference with repeated measurement.
What Google says about its own AI features
For Google specifically, there is useful clarity. Google states there are no additional requirements to appear in AI Overviews or AI Mode, no special schema.org markup needed, and no need for new machine-readable or “AI text” files. Pages must be indexed and eligible to show a snippet. The usual guidance applies, including Google’s advice on helpful, reliable, people-first content.
That is a useful reality check: for Google’s AI features, solid SEO fundamentals are the foundation, not a separate discipline.
Practical levers
Here is what you can act on, ordered from most to least within your control.
| Lever | What to do | How to check it worked |
|---|---|---|
| Crawler access | Confirm search and user-fetch crawlers are not blocked in robots.txt | Retrieval possible; your pages start appearing as cited sources |
| Clear positioning on your site | State plainly who you serve, what you do, where, and at what price level | Assistants describe you accurately |
| Answer-ready pages | Pages that directly answer the questions buyers ask, including comparisons and use cases | Your pages cited on those questions |
| Presence on cited third-party sources | Be listed and accurately described on the comparison, review and directory pages assistants cite | Mentions rise on the relevant question group |
| Reputation signals | Respond to reviews, correct inaccurate public information | Sentiment of mentions improves |
| Consistency over time | Keep names, descriptions and facts aligned everywhere | Fewer contradictory descriptions across assistants |
An example
A boutique hotel group finds that assistants rarely mention it for “family-friendly hotels in Porto”, while two competitors appear often. The cited sources for that question are mostly travel guides and review pages. The hotel’s own site mentions family rooms only in a booking widget, not in readable text.
The actions are concrete: a clear page on family stays with facts (room sizes, cots, nearby activities), reaching out to the guides that are cited, and working on the review profiles that assistants link to. Then the question group is measured again over the following periods to see whether the change holds.
What not to do
- Do not chase a secret formula. None is published. Work from documented behavior and your own measurements.
- Do not block search crawlers by accident. Blocking training crawlers and blocking search crawlers are separate decisions, documented separately by each provider.
- Do not judge from one answer. Variation between runs is normal.
- Do not ignore third-party sources. Your own site is only part of what assistants read.
Checklist
- Search and user-fetch crawlers allowed for the assistants that matter to you
- One clear, consistent description of your brand used everywhere
- Pages that directly answer your buyers’ main questions
- A list of the sources cited for your key questions
- Presence and accuracy checked on those sources
- Repeated measurement before and after each change
How SearchFirefly helps
SearchFirefly asks your tracked questions across ChatGPT, Gemini, Perplexity, Google AI Overviews and Google AI Mode, and shows for each answer which brands are named and which sources are cited, along with your share of voice against competitors. Reputation views (Google reviews, web mentions, Reddit discussions, sentiment and themes) show more of what shapes those answers, and prioritized recommendations spell out the exact fix to make.