AI model behavior

Why assistants mention or skip brands, and how training crawlers differ from retrieval crawlers.

Updated 2026-08-06

AI assistants do not behave like a single blue-link ranking algorithm. Mentions emerge from a mix of training memory, live retrieval, prompt context, and safety/policy filters. VisibleMax helps you improve the signals those systems can use — it cannot rewrite vendor models.

What assistants tend to cite

In practice, assistants are more likely to name a brand when public content makes the entity easy to identify and compare:

  • Clear entities — consistent brand, product, and company names (not only metaphors or taglines).
  • Comparisons — explicit category fit, alternatives, and differentiation.
  • FAQs — direct answers to buyer questions in plain language.
  • Structured facts — pricing posture, who it is for, integrations, and proof points.
  • Machine-readable aids — schema, sitemaps, and llms.txt where relevant (supporting signals, not magic).

Training crawlers vs retrieval crawlers

Vendors use different bots for training vs live retrieval — policies are not interchangeable.
KindTypical purposeVisibility implication
Training / foundation crawlersCollect content that may inform future model trainingAffects long-horizon knowledge; blocking them does not remove you from live search-style answers overnight
Retrieval / search crawlersIndex or fetch pages for grounded, up-to-date answersOften the more direct path to appearing in AI search and cited answers

Examples of the split include OpenAI’s GPTBot vs OAI-SearchBot, Anthropic’s ClaudeBot vs Claude-SearchBot, and Perplexity’s PerplexityBot / Perplexity-User fetchers. Allowing retrieval bots you care about is usually the higher-leverage robots.txt decision for AI answers.

Why brands get skipped

  • Ambiguous positioning (“AI platform for everything”) with no concrete offer.
  • Thin public HTML or SPA shells that crawlers and retrievers cannot read.
  • Stronger competitors with clearer comparison and review coverage.
  • Prompt mismatch — the buyer question does not match how you describe yourself.
  • Index lag or provider-side retrieval gaps unrelated to your on-page quality.

FAQ

If I allow GPTBot, will ChatGPT recommend me?

Not automatically. Training access and live recommendation are different. Prefer also allowing retrieval/search bots and publishing clear, citable pages.

Do probes equal what every user will see?

Probes are controlled API/assistant checks with fixed prompts. End-user chats vary by personalization, tools, region, and model version.

Related

Still stuck? Contact support with your project URL and what you already tried.