Short answer: ChatGPT visibility should not be audited as one static result set. New Semrush research reported by Search Engine Land found that ChatGPT’s low-reasoning and high-reasoning modes cited overlapping sources only 25.6% of the time across tested buyer-journey prompts. For SGO teams, that means prompt testing needs to separate fast-answer behavior from deeper “thinking” behavior.
What changed
Search Engine Land summarized a Semrush and Kevin Indig study showing that ChatGPT’s reasoning depth can change which brands and sources appear in answers. Semrush’s own write-up, Only 25% of cited sources overlap between ChatGPT’s different reasoning modes, frames the same issue as an AI visibility measurement problem: the same prompt can produce different citation sets when the model uses a different answer path.
Why this matters for SGO
- One prompt result is not enough. A brand may appear in a quick answer but disappear when ChatGPT reasons longer, or vice versa.
- Source diversity matters. Deeper answers may draw from comparison pages, reviews, definitions, documentation, or third-party authority signals.
- Reporting needs labels. Record the mode, prompt wording, context where possible, and whether citations or mentions were counted.
What SEO teams should do this week
- Split AI visibility tests into quick answer prompts and deep research / reasoning prompts.
- Audit whether core pages answer the direct query and supporting pages cover comparisons, alternatives, use cases, pricing, implementation, and objections.
- Track citation overlap by prompt cluster instead of averaging all ChatGPT outputs into one score.
- Strengthen internal links from pillar guides to pages that answer buyer-journey questions.
Use this alongside SGOinsights’ guides to how AI search engines choose sources, query fan-out, and the AI Search Optimization Checklist.
Editorial takeaway
The practical lesson is not “optimize for Thinking Mode.” It is to stop treating AI answer engines as a single rankings page. SGO measurement should test multiple answer paths, then map which content types and authority signals are present when a brand is cited.