Running campaigns across 100-plus brands and 50,000-plus creators produces a specific kind of knowledge that no GEO whitepaper can replicate: you see exactly what gets cited and exactly what doesn't, at scale, across real queries, across real AI systems. Most GEO advice is built on theory and small experiments. This article is built on campaign data.
The findings below come directly from two primary sources: Scribble's RocketX GEO case study, which tracked 358 creator-generated pieces across five platforms over 15 days and measured citation outcomes on Grok, Gemini, ChatGPT, Perplexity, and Copilot, and Scribble's citation source analysis of 1,504 AI citations across 25 structured queries on ChatGPT, Perplexity, and Grok. The structural lessons below are drawn from what actually moved citation metrics and what didn't, not from what GEO theory says should work.
The Finding That Surprised Us Most: Volume Did Not Predict Citation
The RocketX campaign deployed 358 pieces of content across 162 independent creators in 15 days. The most counterintuitive result in the data was the relationship between content volume and citation yield.
X/Twitter accounted for 201 pieces, 56% of total content volume. It produced zero measurable citation lift across five tracked AI platforms. Paragraph accounted for 8 pieces and produced zero citability signal. Reddit accounted for 15 pieces, 4% of total content, and produced the highest citation yield per piece in the entire campaign.
That ratio is the structural lesson. Content volume and citation performance are not correlated. The 201 X/Twitter pieces moved mention counts, not citation metrics. The 15 Reddit threads moved the needle that mattered. If the goal is AI citation performance specifically, platform selection and content structure per piece outweigh raw output volume by a significant margin.
This matters for how you brief creators. A campaign that optimizes for volume across all platforms equally is not the same as a campaign that optimizes for citation yield. They produce different outputs and require different structural decisions from the start.
Each AI Model Has a Completely Different Source Diet
The 1,504-citation study across ChatGPT, Perplexity, and Grok produced one finding that restructured how Scribble thinks about platform strategy entirely: the three models do not draw from the same sources, and content that reaches one model does not reliably reach the others.
Editorial and blog content is the only citation category that leads across all three models, at 35% for ChatGPT, 42% for Perplexity, and 55% for Grok. Every other category splits sharply.
ChatGPT cites zero social media across 25 queries. It leans on news publications at 15%, the highest of any model. A single editorial domain, eco.com, accounted for roughly 13% of all ChatGPT citations by itself, which illustrates how concentrated ChatGPT's source trust actually is. Getting into a small number of high-trust editorial domains is more valuable for ChatGPT citation than broad distribution across many platforms.
Perplexity leads all three models on community content at 11%, pulling from Reddit, Bitcointalk, Trustpilot, and Slashdot. It also drew from just 56 unique domains across 25 queries, far fewer than Grok's 163 or ChatGPT's 118. Perplexity applies stricter source selection than the other two models. For this query type, community-validated content on forums and review platforms is disproportionately valuable for Perplexity citations specifically.
Grok's behavior is the most distinctive. It produced 735 citations, almost three times Perplexity's 280. DefiLlama appeared 53 times in Grok's citations, making it Grok's single most cited source in the sample. It does not appear in the top domains of ChatGPT or Perplexity at all. Grok also cites social media at 11% of total citations, including YouTube, X/Twitter, LinkedIn, Instagram, and Facebook. ChatGPT cites none of these. Critically, Grok is the only model in the study to cite Medium, Substack, and Paragraph content, including brand-owned accounts on those platforms.
The structural implication is direct: a piece published on Reddit is optimizing for Perplexity. A piece published on Medium or Substack is optimizing for Grok. A piece placed in a high-authority editorial domain is optimizing for ChatGPT. These are not interchangeable distribution channels for GEO purposes. They are different citation pools serving different AI systems, and a campaign that treats them as equivalent will underperform in every one of them.
The Platform-Model Matrix That Drives Structural Decisions
Based on the citation data, this is how each platform maps to AI citation potential by model:
Platform | ChatGPT | Perplexity | Grok |
|---|---|---|---|
High-authority editorial (eco.com, coincodex.com) | High | Moderate | High |
Low | High | Moderate | |
Medium / Substack | None detected | None detected | Moderate |
X/Twitter | None detected | None detected | Moderate |
YouTube | None detected | Low | Moderate |
DefiLlama / data platforms | Moderate | Moderate | High |
News publications | High | Moderate | Low |
The practical consequence: if a campaign is trying to move citations across all three major models simultaneously, it needs content across at least three structurally different distribution channels, editorial, community, and data platform presence, with different structural requirements for each.
What Structure Inside a Piece Actually Determines Citability
Platform selection determines which AI system has access to your content. Content structure determines whether that system extracts and cites it once it does.
The RocketX campaign identified a direct tradeoff between publication velocity and per-piece citation quality. The 15-day sprint required publishing at a pace that sometimes prioritized volume over structural optimization. GEO-citable content has specific structural requirements: a direct answer in the opening paragraph, self-contained explanatory passages that stand alone without surrounding context, clear entity signals, and specific verifiable claims rather than general descriptions. Pieces that didn't fully meet this bar likely underperformed on citation rates relative to their reach.
The structural rules that explain this gap are covered in detail in Scribble's guide on what makes a paragraph AI-quotable, but the campaign data produced a specific version of the same lesson. The pieces that earned citations in the RocketX campaign shared identifiable structural features: they made a specific comparative claim with a number attached ("RocketX returned 18.51 ETH vs Changelly's 18.16 ETH on a 0.5 BTC to ETH swap"), they placed that claim in the first paragraph rather than building up to it, and they provided enough context within a single passage for an AI system to extract and reproduce the answer without needing the rest of the article to make sense.
Pieces that described RocketX's features without specific comparative evidence, without dated figures, and without a claim that could be independently verified produced lower citation rates regardless of how well-written they were. Writing quality and citation performance are not the same variable.
The Mentions vs Citations Gap Is the Most Underdiagnosed Problem
In the RocketX campaign, Perplexity registered RocketX in 12 mentions and Copilot in 23 mentions across tracked queries. Neither platform generated a single citation from RocketX as a source. Combined: 35 mentions, zero citations.
Mentions and citations measure different things. A mention means the AI named the brand in an answer. A citation means the AI sourced a specific piece of content as evidence. A brand can accumulate mentions by appearing in answers about a category without any of its third-party content being trusted enough to cite. This is the gap between "being named" and "being sourced," and it's the difference between awareness and authority in AI search.
The structural solution to converting mentions into citations is the same as the structural solution for earning citations in the first place: content that is specific, independently verifiable, and distributed across sources the AI treats as trustworthy. Generic brand mentions in community conversations produce awareness. Comparative content with specific evidence, published on platforms with established citation track records, produces citations.
How Perplexity decides what to cite covers the retrieval mechanics behind this distinction in more detail.
The Structural Mistake That Killed Otherwise Strong Submissions
Across multiple Scribble campaigns, one structural pattern consistently underperforms regardless of content quality: the claim buried after three paragraphs of setup.
AI retrieval systems extract passages that answer a query, not articles that eventually get to an answer. A piece that spends the first four paragraphs explaining what a cross-chain aggregator is before making a specific claim about RocketX's rate advantage will not have that claim extracted by Grok or ChatGPT when a user asks a comparison question. The system finds the passage that most directly answers the query. If the answer is in paragraph five, the system may cite a different piece entirely, one where the same answer appears in paragraph one.
This structural failure shows up in the citation data precisely. The RocketX campaign's highest-performing pieces by citation yield were almost uniformly structured around a lead claim: a specific number, a dated comparison, or a verifiable outcome stated in the first paragraph, followed by supporting context. The briefing process Scribble now uses for Month 2 of the RocketX campaign has this as a hard structural requirement, not a recommendation.
The practical test for any GEO piece before publication: can you extract a single, self-contained paragraph that fully answers the target query? If the answer requires context from surrounding paragraphs to make sense, the piece needs restructuring before it will reliably earn citations.
What This Means for How Campaigns Should Be Structured
The campaign data produced five structural conclusions that now shape how Scribble briefs creators and evaluates submissions:
Structure the brief around citation targets, not keyword targets. A brief that says "cover RocketX's privacy features" produces generic content. A brief that says "include a specific comparison of RocketX's private swap output versus Changelly's on an identical trade, with a verifiable figure" produces citable content. The claim has to be specified in the brief, not left to the creator to invent.
Match platform assignment to the target AI system. Reddit is a Perplexity play. Medium and Substack are Grok plays. High-authority editorial placement is a ChatGPT play. A campaign trying to move all three simultaneously needs deliberately different content in different places, not the same piece repurposed everywhere.
Treat Reddit as infrastructure, not afterthought. The 15 Reddit posts in the RocketX campaign were the most valuable content produced per piece. They were also the hardest to scale. Authentic community-validated threads survive moderation. Promotional content gets removed. Building Reddit presence requires creators with existing credibility on the relevant subreddits, not generic contributors posting for the first time. This needs to be built into campaign design at the start, not added when other platforms underperform.
Velocity has a citation cost. Publishing 358 pieces in 15 days is achievable. Publishing 358 pieces in 15 days with the structural quality that maximizes per-piece citation rate is a harder constraint. The RocketX case study is transparent about this tradeoff: content velocity created a structural quality tradeoff. Month 2's approach addresses this with GEO-structured briefs as the default rather than the exception.
Measure mentions and citations separately. A campaign that only tracks mention counts will miss the mentions-to-citations gap entirely. 35 mentions and zero citations is not a success metric, it's an early-warning signal that brand awareness is registering without the trust signals that convert awareness into source authority.
The complete framework for building content that earns citations across all three models is in How to Write Content That Gets Cited by AI Search Engines. What this article adds is the campaign-level evidence for why that framework is structured the way it is, grounded in 358 pieces, 1,504 citations, and five AI platforms measured across 15 days of live execution.
Want to see where your own brand stands? Run it through the GEO Checker, or go straight to brands.scribble.network to start a campaign.
Written by
I’m Tanmay Tarte, a community builder at Scribble and an engineering graduate from Priyadarshini College of Engineering. Over the years, I’ve worked across community management, content, hosting, and social media, mainly within the Web3 and creator ecosystem space. Outside of work, I’m a huge sports enthusiast and can genuinely play cricket all day, every day.
