Two GEO Experiments, 775 AI Citations: What Really Drives AI Visibility

Introduction

For the past two years, almost every article about generative engine optimization (GEO) has repeated the same advice: publish comprehensive content, add schema, get mentioned in listicles, and the AI models will cite you.

Zeeshan Yaseen, an SEO and AI search consultant, decided to test that advice instead of repeating it. Published by Search Engine Land on September 14, 2026, his write-up documents two structured experiments that logged 775 citation events across ChatGPT, Claude, Gemini, Perplexity, Google AI Mode and Grok — and it ends with him retracting one of his own conclusions.

That last detail is what makes this study worth your attention. Most GEO advice online is built on a single screenshot or one campaign. This one ran a playbook twice, on two different brands in two different verticals, and only kept the findings that survived both tests.

The headline result: AI visibility is shaped less by the content you control than by who else mentions you, who you appear next to, and how recently those associations were reinforced.

The News / Trend

Experiment one: four platforms, several months, 775 events

The first experiment worked with an existing brand over several months. It tracked 15 commercial-intent keywords across four AI platforms — ChatGPT, Claude, Gemini and Perplexity — with every query run manually, both with and without a VPN, to catch location-based differences.

The strategy was to place listicles on sources that the models already surfaced for those keywords, supported by PR, guest posts and organic LinkedIn activity. By the end, the brand appeared for roughly 10 to 12 of the 15 keywords, peaking at 37.01% keyword presence on April 29.

Citation volume broke down unevenly across platforms: ChatGPT 148, Claude 96, Gemini 87, Perplexity 64. Listicles accounted for 72.4% of all citations and PR for 24.1% — guest posts, the owned site and LinkedIn split what was left.

One placement dominated everything. A comprehensive listicle on Indeed SEO produced 190 mentions — more than all other sources combined.

Experiment two: a 30-day cold start

The second experiment was a far harder test. It took a SaaS link building agency with no measurable AI presence at all and ran a 30-day program from zero, this time across six platforms (adding Google AI Mode and Grok) and 15 buying-intent keywords.

The result: 298 appearances in 30 days from a standing start. Gemini led with 104, Google AI Mode followed with 95, then Claude 59, ChatGPT 32, Grok 4 and Perplexity 4. Gemini and AI Mode together produced roughly two-thirds of all appearances — the opposite distribution from experiment one, where ChatGPT led.

That mattered commercially. During the 30-day window, 18.5% of new users arrived through referral traffic and another 3.25% through GA4's AI Assistant channel — together just over one-fifth of all new users. A single Perplexity referral, from the lowest-volume platform in the test, became a paying customer.

Why the timing matters

This research lands in the same week Google confirmed it is running an "AI contribution pilot" that pays publishers when their content meaningfully contributes to AI Overviews, AI Mode and Gemini responses. The economics of being cited by AI are being formalised right now. Which makes the question "what actually earns citations?" suddenly a revenue question, not an academic one.

Why This Matters

Your keyword list is not your target list anymore

The most useful operational finding is that you should build your outreach list from observed citations, not from domain rating or traditional prospecting data. Before any outreach, the second experiment ran all 15 keywords through every platform, logged which sources appeared, and ranked them by citation frequency. That ranked list became the placement target list.

Five of eight targets went on to appear in the final citation mix. Indie Hackers climbed from 44 mentions during prospecting to 146 after placement — a 232% improvement. Bruce Jones SEO rose from 26 to 69 (up 165%), and TechBullion from 15 to 37 (up 147%).

Citation volume is not business outcomes

The most-cited source in experiment two, Indie Hackers, had flat referral traffic. TechBullion produced fewer citations but lifted sessions from 1 to 64. Claude.ai referrals tripled and ChatGPT referrals rose 166%.

In other words, ranking sources by raw citation count can point you at the wrong investment.

Owned content is a foundation, not a growth lever

This is where the study's author revised his own position. After experiment one, he recommended publishing listicles on your own site. Experiment two contradicted that: the owned listicle was the slowest source to be cited — 18 days, longer than every third-party placement — and generated only 14% of mentions, against 85.8% for earned placements (PR contributed 0.2%).

But there is a twist worth copying. When the owned listicle was rewritten to be less promotional and more comparative — adding leading competitors instead of presenting the brand alone — mentions jumped from 4 to 49, a 12.25-fold increase, and the daily visibility curve rose from 22 to a peak of 95 within four days. Same domain, same author, same keyword target. Only the scope of the content changed.

Exact-match still matters — more than you'd think

The brand ranked for "Best LLM SEO Consultant" but barely appeared for the near-identical "Best AI SEO Consultant". The first phrase had a dedicated listicle; the second didn't. Broader semantic coverage did not close the gap.

Source decay is fast

Roughly half of all sources stopped being cited within 30 days. One placement fell from 29 mentions to 11 week over week; another declined with no intervention at all. One-time publication does not sustain AI visibility.

How to Adapt Your Strategy

1. Run your keywords through the platforms first. Before buying a single placement, query your target keywords in ChatGPT, Gemini, Claude, Perplexity and AI Mode, and log which domains come back. That list — not your competitor backlink export — is your outreach plan.

2. Prioritise sources the model already cites. A placement on an already-cited source compounds fast. The Indeed SEO listicle was surfaced around 10 times before outreach; afterwards it became the largest citation source for months.

3. Choose depth over frequency. Five brief listicles produced little visibility; one comprehensive piece kept earning citations.

4. Structure pages answer-first. Pages performed better when the answer appeared within the first 100 words. A key-takeaway block near the top beat every other on-page change tested. FAQs visible by default outperformed FAQs hidden in accordions, and question-based headings ("How is AI SEO different from traditional SEO?") beat noun-phrase headings.

5. Keep sections self-contained. "What to look for when hiring an LLM SEO expert" and "Where to hire one" worked better as separate sections than merged into one.

6. Stop writing only about yourself. The 12.25x lift came from adding competitors and recognised names. Meanwhile, when a listicle that had placed the brand alongside Lily Ray and Aleyda Solis removed those names, performance declined within days. Models appear to evaluate the surrounding entities, not just the individual mention.

7. Reserve dedicated assets for high-value terms. The two strongest keywords across the whole study were "Best SaaS Link Building Agency in USA" and the same phrase with "2026" added — 18 appearances each. High-value commercial keywords seem to require their own page, exact match included.

8. Match placement type to query type. Software and tool queries favoured high-authority review sites; service queries returned listicles more often.

9. Budget for maintenance. With half of sources decaying inside 30 days, treat AI visibility as a subscription, not a one-off purchase.

10. Measure repeatedly and track referrals. Time to citation ranged from 1 to 18 days, and four live placements had not been cited by the end of the window. Checking once, one week after publication, is not a measurement — it's a coin flip.

Key Takeaways

• Two GEO experiments logged 775 citation events; the author retracted one of his own conclusions after round two.
• Earned placements generated 85.8% of mentions, owned content only 14% — and owned pages were the slowest to be cited (18 days).
• Build your placement list from observed AI citations, not domain rating or traditional prospecting.
• Answer within the first 100 words; keep FAQs visible, headings question-based and sections self-contained.
• Exact-match keywords still matter — near-identical phrases produced very different visibility.
• Roughly half of cited sources decayed within 30 days. Maintenance, not publication, is the ongoing work.
• The best-cited source was not the best traffic source; citation volume alone is a poor investment signal.

FAQ

Q: What is GEO (generative engine optimization)?

GEO is the practice of earning visibility and citations inside AI-generated answers — AI Overviews, Google AI Mode, ChatGPT, Gemini, Claude, Perplexity and similar systems — rather than only ranking in a classic list of ten blue links.

Q: Do keywords still matter if AI answers replace clicks?

Yes, and more precisely than before. In this study, exact-match targeting decided whether a brand appeared at all: it ranked for "Best LLM SEO Consultant" but barely for "Best AI SEO Consultant". High-value commercial phrases still justified dedicated, exact-match assets.

Q: Which AI platform should I prioritise?

Whichever one your data points to. In experiment one, ChatGPT led with 148 citations. In experiment two, Gemini and Google AI Mode produced roughly two-thirds of all appearances. Platform priority is niche-specific — run your own keywords through each platform and measure.

Q: How long does it take to earn AI citations?

Between 1 and 18 days after publication in this research, with substantial variation. Some placements were cited overnight; the slowest was an owned listicle at 18 days, and four placements never converted within the window.

Q: Is publishing on my own site worth it?

As a foundation, yes — but not as your primary growth engine. Owned content took the longest to be cited and produced 14% of mentions. Its performance improved dramatically (4 to 49 mentions) when it stopped being promotional and started comparing the brand with competitors.

Conclusion

The most valuable thing about these two experiments is not a tactic. It is the method: define your keywords, run them through the AI platforms, record what the models already cite, and build your strategy from observed results instead of inherited advice.

The pattern that held across both brands and both verticals is association. Which sources mention you, who appears alongside you, and how recently those relationships were strengthened — that is what AI visibility is made of in 2026.

The fastest way to start is to stop guessing which keywords and topics are worth targeting. Generate a data-backed keyword set, test it across AI platforms, and let the citation data tell you where to invest next at SocialKeywordGenerator.com.