Table of Contents
Most GEO advice still reads like recycled SEO advice with “AI” swapped in for “Google”: publish more, publish owned listicles, hope for citations. On September 14, 2026, Search Engine Land published a detailed breakdown of two independent, data-logged experiments that put that advice to the test across ChatGPT, Claude, Gemini, Perplexity, Google AI Mode, and Grok. The results, drawn from 1,073 total logged citation events, don’t just tweak the conventional playbook. They invert a chunk of it.
Executive Summary
Here’s what you need to know before diving in:
- Across a 30-day “cold start” test with zero baseline AI presence, earned placements such as guest mentions and third-party listicles generated 85.8% of AI citations, versus 14.0% from the brand’s own owned content.
- In a separate, longer-running “consultant test” tracking 15 keywords across four platforms, listicle placements alone produced 72.4% of 775 logged citation events, with PR contributing another 24.1%.
- Roughly half of all citing sources stopped being cited within 30 days in both experiments, meaning AI visibility decays quickly without ongoing maintenance.
- Adding named competitors to content lifted citation rates by 12.25 times compared with purely promotional copy, and citation volume did not reliably correlate with actual referral traffic.
Two experiments, two different starting points
The first experiment, which the article calls the “consultant test,” tracked 15 commercial-intent keywords across ChatGPT, Claude, Gemini, and Perplexity over several months, running queries manually with and without a VPN to account for location-based answer differences. The brand’s visibility peaked at 37.01% keyword presence on April 29, and the researcher logged 775 total citation events across the four platforms: 148 on ChatGPT, 96 on Claude, 87 on Gemini, and 64 on Perplexity, according to Search Engine Land’s September 14, 2026 report.
The second experiment, the “cold start test,” began from zero. A SaaS link-building agency with no existing AI presence was tracked across six platforms, adding Google AI Mode and Grok to the original four, over a 30-day window from May 30 to June 28. From that zero baseline, the brand generated 298 total appearances: 104 on Gemini, 95 on Google AI Mode, 59 on Claude, 32 on ChatGPT, 4 on Grok, and 4 on Perplexity. Crucially, the researcher also tracked business outcomes, not just citation counts, finding that 18.5% of new users arrived via referral traffic tied to AI answers, with another 3.25% attributed through Google Analytics 4’s AI Assistant channel.
The finding that should worry anyone betting on owned content alone
The headline number from the cold-start test is stark: earned placements, meaning mentions the brand did not publish itself, accounted for 85.8% of all AI citations, while the brand’s own owned listicle content produced only 14.0%. Owned listicles were also the slowest content type to get picked up by AI platforms, with time-to-citation across all content ranging from 1 to 18 days. Three sources, specifically Indie Hackers, one guest post on a site called Bruce Jones SEO, and the brand’s single owned listicle, generated 78% of all citations in that test combined.
That data point runs directly against a common piece of GEO advice: publish a big, comprehensive owned resource and wait for AI models to cite it. Per this research, an owned listicle functions more as a credibility anchor than as a growth engine. The growth came from earning mentions elsewhere, particularly comparison-style content; adding named competitors to a piece lifted its citation rate by 12.25 times relative to purely promotional versions of the same content.
The consultant test reinforced a related but distinct point: source authority outweighs content quality alone. Government and education domains ranked highest for citation reliability, and being mentioned alongside recognized experts measurably improved a brand’s own citation odds, an effect the report attributes to entity association. At the same time, exact-match keyword phrasing remained important despite the advances in semantic search that AI models are supposed to represent, and placing a direct answer within the first 100 words of a page improved citation likelihood.
Both experiments converge on a less comfortable finding for anyone tracking AI visibility as a vanity metric: citation volume did not correlate reliably with referral traffic. A source could be cited frequently without sending meaningful visits, while a less-cited source sometimes drove outsized referral activity. And in both tests, roughly half of all citing sources stopped appearing within 30 days, meaning a brand cannot treat a citation as a permanent asset the way it might treat a stable page-one ranking.
ARC Marketing’s Take
“The instinct in this industry is always to publish our way to visibility, but this data says the opposite: earn your way in first, and let owned content do the supporting work, not the heavy lifting,” ARC Marketing. “If 86% of your AI citations are coming from someone else’s site, your content calendar should be spending less time on your blog and more time pitching the sites AI models already trust.”
We’ve started reallocating a meaningful share of client content budgets, in the range of 25 to 30% depending on the account, away from net-new owned blog posts and toward earned placements: guest contributions, comparison mentions on third-party listicles, and structured outreach to the small number of sources that already show up in AI answers for a client’s category. Given that half of citing sources drop out within 30 days per this research, we’ve also moved AI citation checks into the same monthly cadence as rank tracking, rather than treating them as a quarterly audit item.
FAQ
What were the two GEO experiments about?
One tracked 15 commercial-intent keywords across four AI platforms over several months for a brand with existing visibility. The other tracked a SaaS company with zero AI presence across six platforms over a 30-day window, logging every citation from scratch.
Does this mean owned content doesn’t matter for GEO?
No. Owned content still generated 14.0% of citations in the cold-start test and functioned as a credibility anchor other sources pointed to. The finding is that owned content alone is a weak growth lever compared with earned placements, not that it’s worthless.
Which AI platforms were included in the research?
ChatGPT, Claude, Gemini, and Perplexity were tracked in the consultant test. The cold-start test added Google AI Mode and Grok, covering six platforms total.
Why did roughly half of citing sources disappear within 30 days?
The research doesn’t give a single cause, but it indicates AI models continuously re-evaluate which sources best answer a query, meaning citations behave more like a rotating shortlist than a fixed ranking.
Does more AI citation volume mean more traffic?
Not reliably, according to both experiments. Citation volume and referral traffic did not consistently correlate; some frequently cited sources sent little traffic, while some less-cited sources performed better on referrals.
What’s the single most actionable takeaway from this research?
Prioritize earning mentions on third-party sites, especially comparison-style content that names competitors, over publishing additional owned resources, and revisit AI citation tracking monthly rather than quarterly given how quickly source lists change.
Sources
- Search Engine Land, “Two GEO experiments challenge conventional AI visibility advice,” September 14, 2026: https://searchengineland.com/geo-experiments-challenge-conventional-ai-visibility-advice-488342