Where AI Gets Its SaaS Recommendations: An 80-Answer Source Study

When buyers ask an AI assistant for the best software in a category, the answer leans on the vendors' own websites. When they ask which options are most trusted, it switches to review platforms and analysts. Across 80 live answers from ChatGPT and Gemini in 10 SaaS categories, the wording of the question changed the source mix more than anything else we measured.

We used two neutral prompts per category: a discovery prompt asking for the best tools for a growing SaaS team, and a trust prompt asking which tools are most trusted and what public sources support that. Discovery answers cited independent sources only 9% of the time, and review platforms appeared in 7 of 40 answers. Trust answers cited review platforms in 40 of 40 answers.

The practical reading: the pages a vendor publishes about itself shape whether it makes the shortlist, while third-party reviews shape whether the assistant vouches for it.

The most widely cited independent domains were g2.com (40 answers), gartner.com (34 answers), trustradius.com (23 answers), capterra.com (18 answers). Reddit appeared in 19 answers across 8 categories, the only forum with meaningful reach. Analyst and review pages were cited across nearly every category, which makes them some of the few sources a vendor can influence once and benefit from in many prompts.

The engines differed in volume, not direction: ChatGPT averaged 25 citations per answer and Gemini 12.3. Both relied on review platforms for trust questions.

For discovery prompts, make your own site answer the buyer's question directly: category and use-case pages, honest comparison pages, and clear pricing that an assistant can extract. Those are the pages cited when you are named.

For trust prompts, invest in the review platforms that assistants already cite in your category: complete profiles, recent reviews from real customers and accurate category placement. Then measure both prompt types separately — a single blended visibility number hides which lever is failing.

Answers were collected on 2026-09-30 through provider APIs with web search enabled, sampled twice per prompt, in US English. All 80 registered samples completed. The design was reduced before collection to two engines and two source-focused prompts; the full registration, limitations and the aggregate CSV are on the research page.

"Named vendor sites" are unclassified domains whose brand appears in the answer text; review, forum, media, social and institutional domains count as independent; everything else stays unclassified. A citation is not an endorsement, and two engines in one market are a snapshot, not a census.