GEO Experiments Reveal What Actually Drives AI Citations

TL;DR: Two back-to-back GEO experiments tracked 775 citation events across ChatGPT, Claude, Gemini, Perplexity, Google AI Mode, and Grok. Listicle placement on sources models already cite, reinforced by PR and guest posts, drove the majority of AI visibility gains. Source authority consistently outperformed content quality, and a cold-start brand reached 298 AI appearances in 30 days using observed citation data to guide outreach.

What These Experiments Actually Measured

Most GEO advice circulating right now is theoretical. Blog posts citing screenshots, case studies with no controlled variables, or single-campaign anecdotes that could have moved for any number of reasons. These two experiments were different: structured, manually tracked, and run sequentially so findings from the first could be pressure-tested against the second.

Experiment one ran over several months for an established consulting brand. It tracked 15 commercial-intent keywords across ChatGPT, Claude, Gemini, and Perplexity, with every query run manually and repeated through a VPN to control for location variance. The brand reached 10 to 12 of the 15 tracked keywords at peak, with 37% keyword presence on the best single day. Total citation events: 477.

Experiment two was a 30-day cold start for a SaaS link-building agency with zero measurable AI presence at baseline. Six platforms were tracked, adding Google AI Mode and Grok. From May 30 to June 28, the brand accumulated 298 appearances. Combined, the two experiments produced 775 logged citation events — enough to identify patterns that hold across verticals and across platforms.

Listicles, PR, and Guest Posts Are One System

The single most durable finding across both experiments: the three main content formats work as a reinforcing loop, not as independent tactics. Listicles introduce the claim. Press releases amplify the listicle. Guest posts reference both. Each layer performs measurably better when the others are active.

In experiment one, a single comprehensive listicle placed on Indeed SEO — a source ChatGPT was already citing for the tracked keywords — generated 190 citation events. That one placement outperformed every other source combined. The listicle had been noticed by ChatGPT roughly 10 times before outreach began. After placement, it became the dominant citation source for the duration of the experiment.

Listicles accounted for 72.4% of all citations in experiment one. PR added 24.1%. Guest posts, the owned site, and LinkedIn split the remaining 3.5%. But those percentages understate the interdependence: Claude cited guest posts that referenced Yahoo and Business Insider features, even when it didn’t cite those publications directly. PR appeared capable of amplifying a strong placement, but not of rescuing a weak source. Better-written listicles on low-authority general sites produced near-zero visibility in experiment one’s phase two.

Operators running paid media alongside GEO efforts should note that this compound effect mirrors what happens in well-structured ad ecosystems: touchpoints reinforce each other, and removing one layer degrades the others.

Source Authority Is the Constraint, Not Content Quality

One of the cleaner findings from experiment one: roughly half of all sources stopped generating citations within 30 days of initial appearance. A single MEXC placement dropped from 29 mentions to 11 in one week. A Triple Review listicle that had been cited consistently declined without any change to the content.

The practical hierarchy that emerged from citation data: government and education sources at the top, followed by news publications, then industry-relevant sites, then general sites. A placement on a strong-authority industry publication supported by PR outperformed a better-written piece on a general site by a significant margin. Depth within a single authoritative piece also outperformed frequency across multiple weaker ones. The first five listicles in experiment one covered topics only briefly and generated minimal visibility. The comprehensive piece kept earning citations.

There was also a peer-set effect. One listicle placed the brand alongside recognized industry experts. When those names were removed, performance declined within days. The models appear to evaluate surrounding entities, not just the individual mention — a meaningful finding for anyone planning co-citation strategies or contributed content.

If your brand is operating in a high-CAC vertical and you haven’t mapped which sources the models currently cite for your commercial keywords, a structured marketing audit should start there, not with content production.

Cold Start Results and the Citation-First Outreach Method

Experiment two validated the most operationally useful tactic from experiment one: build the outreach list from observed citations, not from domain rating alone. Before any outreach began, all 15 target keywords were run through every tracked platform. Sources that appeared were logged and ranked by citation frequency. That ranked list became the outreach priority list.

Five of eight outreach targets appeared in the final citation mix. Indie Hackers increased from 44 observed mentions during prospecting to 146 after placement — a 232% improvement. Bruce Jones SEO went from 26 to 69 (up 165%). TechBullion rose from 15 to 37 (up 147%). Not every placement moved the needle: RankTracker and HR.com showed fewer citations after placement than during prospecting, and two shortlisted sites hadn’t appeared at all by the end of the measurement window. The method improves hit rate; it doesn’t guarantee citations.

Platform distribution in experiment two differed sharply from experiment one. Gemini led with 104 appearances. Google AI Mode followed with 95. Claude produced 59, ChatGPT 32, Grok 4, and Perplexity 4. Gemini and Google AI Mode together accounted for roughly two-thirds of total appearances — the inverse of experiment one, where ChatGPT led. The comparison is imperfect since AI Mode wasn’t tracked in experiment one and the two niches weren’t directly comparable, but it signals that platform weighting varies by vertical and query type.

Business impact was real despite the small window: 18.5% of new users arrived via referral traffic, and another 3.25% came through GA4’s AI Assistant channel. One Perplexity referral — from the lowest-volume platform in the experiment — converted to a paying customer.

What This Means for High-CAC Vertical Operators

Forex brokers, iGaming operators, crypto exchanges, and law firms all share the same structural problem: high cost-per-acquisition, heavily regulated content environments, and buyers who run extensive research before converting. AI search is becoming a meaningful part of that research process, and the citation patterns from these experiments map directly onto how those buyers evaluate options.

For forex acquisition teams, the peer-set effect is particularly relevant. Being mentioned alongside recognized analysts or trading educators in a listicle may carry more citation weight than a standalone brand feature on a stronger domain. That changes how co-citation strategies should be structured.

For iGaming operators, the query-type finding matters. Software and tool queries favored high-authority review sites. Service queries more often returned listicles. Matching placement type to query type improved citation likelihood. If your brand is targeting “best online casino” variants, the placement format should reflect that query structure — not just the domain authority of the target site.

For law firm marketing teams and crypto lead generation operations, the freshness finding is actionable: recent data, visible publication dates, and current references correlated with stronger citation performance. Evergreen content that hasn’t been updated in 18 months is likely losing ground to fresher placements on lower-authority sources.

On the structural content side: FAQ content performed better when visible by default rather than hidden behind expandable sections. Question-based headings outperformed comparison-style headings. Answers appearing within the first 100 words of a page correlated with stronger citation performance. A key-takeaway block near the top produced a larger improvement than any other single on-page change tested. These are findings that precision-targeted content strategies should be incorporating now, not after another round of tests.

Capitalization of queries also appeared to affect which sources were retrieved. Capitalized and lowercase versions of the same query returned different citations in three repeated tests. This has implications for keyword mapping: exact-match assets for high-value commercial terms still matter. The first experiment ranked for “Best LLM SEO Consultant” but barely appeared for “Best AI SEO Consultant” despite similar intent, because only the first phrase had a dedicated listicle. Broader semantic coverage didn’t bridge the gap. High-value commercial keywords likely require dedicated assets, not umbrella content that tries to cover multiple variants.

For operators evaluating AI-driven lead qualification alongside GEO investment, the referral-to-conversion path from experiment two is worth noting. Low-volume platforms can still produce high-value conversions. Attribution models that deprioritize AI referral channels based on volume alone will misallocate budget.

Originally reported by Search Engine Land, September 2026.