What Gets a Business Recommended by AI? Findings from OtterlyAI's Q4 2026 GEO Experiments

Nov 3, 2026 · @Ina
What gets a business recommended by ChatGPT, Gemini, Perplexity or Google's AI keeps changing. OtterlyAI studies this through controlled experiments, and in its October 2026 session, GEO Experiments 2026 Q4 Updates: What We Tested, What Failed, What Actually Works, Rick Tousseyn and Thaylise Nakamoto shared their latest findings.
Below are the takeaways I found most useful. The research and data belong to OtterlyAI; for their full experiments and methods, see their insights hub.
What works and what doesn't in AI search right now
OtterlyAI summarized its findings in one list, lightly condensed below.
What works in AI search right now | What doesn't |
Press releases in licensed news publishers (ChatGPT, Copilot) | Publishing thousands of AI-generated blogs with one click |
Owned content (brand pages take 56.4% of citations) | Glossary hub for industry terms |
Cutting long livestreams into videos and written blogs | llms.txt file |
Content freshness | Markdown files as page variants |
LinkedIn Pulse articles | Schema markup for ChatGPT and Gemini |
Converting YouTube videos into written blogs | Image alt tags and file names |
AI videos on YouTube | Beautiful content design |
Reddit visibility in active communities, with replies | Regular LinkedIn feed posts (work less than Pulse articles) |
Long-form YouTube videos | |
External listicles | |
Schema markup for Google AI Overviews and AI Mode | |
YouTube video optimization | |
Clean, structured URLs | |
Industry recognition (awards, analysts) |
Six findings worth knowing
1. Analytics may undercount AI
Google Analytics credited Claude with just 0.1% of OtterlyAI's conversions, while customers' own answers about how they found the company put Claude at 10.6%. Website analytics alone may miss part of AI's influence.
2. Keep content genuinely up to date
In a small test, articles whose "last updated" date was refreshed received 31.5% more AI citations in the following 30 days than in the 30 days before, while unchanged articles declined. OtterlyAI was clear this was a test, not a tactic: the lesson is to review and genuinely update older content, not just change dates.
3. "Best of" lists dominate AI citations
Of the AI citations OtterlyAI could classify by page type, 66.6% pointed to ranked lists, and 35.6% of the cited lists came from brands ranking their own category. Each AI engine behaves differently, so it helps to know which ones your customers use.
4. Combine AI with human input
In an ongoing experiment, a site written half by AI and half by humans earned 104% more AI citations than a human-written one. Caveats: the human articles were not in-depth, and the test is still running. In an earlier test, sites filled with one-click AI articles lost visibility within weeks.
5. On LinkedIn, articles beat regular posts
LinkedIn articles earned far more AI citations than regular feed posts, and personal-profile articles performed best. However, none of the test content was cited by ChatGPT, Claude, Gemini or Google AI Overviews.
6. There are no shortcuts on Reddit
OtterlyAI's Reddit test pointed to authentic community building: content and replies mattered, while upvotes and subscriber counts did not.
A snapshot, not a rule
The landscape shifts fast. In August, Reddit's share of ChatGPT citations in the US dropped from almost 4% to about 1%, around the time of a ChatGPT update. OtterlyAI's own advice is to treat findings like these as a snapshot, then measure and adjust.
Thank you to Rick Tousseyn, Thaylise Nakamoto and the OtterlyAI team for sharing their research openly.
Source note: this post summarizes findings OtterlyAI presented in October 2026. Results are specific to their tests and are not guarantees for every business, industry or AI platform.
Sources
OtterlyAI, GEO Experiments 2026 Q4 Updates: What We Tested, What Failed, What Actually Works, presentation, October 2026



Comments