Why I ran this experiment
Every week I read another agency blog explaining that GEO — Generative Engine Optimization — is the most important thing in marketing right now. The guides are thorough. The frameworks are neat. The authors are selling GEO audits.
What none of them are doing is running the experiment on a real blog and publishing the results without a product to sell. So I did it on BuzzRiding over four weeks in April and May 2026. Three articles. Four changes per article. Manual citation checks in ChatGPT, Perplexity, and Google AI Mode every Monday morning. Here's everything I found.
What GEO actually is (the 60-second version)
Traditional SEO gets your page into a list of ten blue links. GEO gets your content cited inside the AI-generated answer that appears before those links — or replaces them entirely. When someone asks ChatGPT or Perplexity a question, the AI pulls from sources it considers authoritative and well-structured. GEO is the discipline of making your content easy for those systems to extract, understand, and quote.
The research behind it is real. A 2024 Princeton study found that specific structural changes — citing sources, adding statistics, using direct quotations — improved AI citation rates by 30 to 40% compared to unoptimised content.
The experiment setup
Three existing BuzzRiding articles with decent Google rankings but zero AI citations at baseline:
- Will AI Replace Marketing Jobs? — 📈 Career & Skills
- AI Skills Marketers Need in 2026 — 📈 Career & Skills
- AI Marketing Tools Weekly Roundup — 📰 News & Trends
Four changes per article, no more, no less:
- BLUF opening: Direct answer in the first 150 words. Answer first, context after.
- Stat density: Two to three named statistics with source and year per major section.
- FAQ restructure: Questions rewritten as people actually phrase them in ChatGPT.
- Headers as questions: Three H2s rewritten from statements into questions.
12 manual prompts across ChatGPT, Perplexity, Google AI Mode. BuzzRiding cited in 0 out of 12 tests. All three articles had existing page 1 rankings.
Week 2: First citations appear
BLUF openings and FAQ rewrites live in week one. Stat density and H2 rewrites live in week two. By end of week two: one Perplexity citation for the careers article — brief mention, not primary. No change in ChatGPT or Google AI Mode.
1 citation out of 12 tests. Perplexity only. The BLUF-structured careers article moved first.
Week 3: Stat density makes the real difference
Stat-dense sections with named statistics and year-and-source attribution started drawing citations in Perplexity across two articles. Testing "What AI skills should marketers learn in 2026?" in Perplexity: BuzzRiding appeared as a cited source in a paragraph containing a named HubSpot 2026 State of Marketing data point.
Google AI Mode: no change. ChatGPT: one brief appearance — a single sentence from the careers article quoted in a longer answer.
4 citations out of 12 tests. Perplexity leading. ChatGPT one appearance. Google AI Mode still zero. Stat-dense paragraphs with attribution are what's getting cited.
Week 4: Plateau and a surprise
Citation rate plateaued at four or five out of 12, varying by prompt phrasing. The experiment made clear that citation is not stable — it shifts based on question wording in ways difficult to predict.
The surprise: Google AI Mode finally cited BuzzRiding. Not for a broad prompt. For a very specific question — "what percentage of marketers use AI tools daily" — that matched an exact stat added to the AI Skills article. The response pulled the number, attributed it to HubSpot, and linked to BuzzRiding as the contextual source.
5 citations out of 12 tests. First Google AI Mode citation. Perplexity most consistent. ChatGPT least consistent — same prompts returned different citations on different days.
What the results actually mean
Zero to roughly 40% citation rate in four weeks. Four structural text changes per article. No schema, no tools, no budget.
Perplexity moves fastest — citation patterns updated within two weeks. Google AI Mode was slow to start but produced the highest-quality citation. ChatGPT was slowest and most inconsistent.
Stats with attribution are the single most citation-worthy element. Every citation came from a section containing a named statistic from a named source. General opinion paragraphs were never cited once across the entire experiment.
Prompt phrasing dominates. The same content got cited or not based on question wording. Structure for the most direct, common version of the question.
For more on the skills side: the AI skills article we restructured is worth reading. The 30-day AI content experiment covers the production automation side.
What I'd do differently
One more change upfront: a named author bio on each article. Every GEO guide says authorship signals matter. BuzzRiding's team byline may be costing citations — follow-up experiment incoming.
Also: set up GA4 AI referral filters from day one. By week four, a small trickle of Perplexity sessions was showing up. I didn't have the filters in place to capture it cleanly from the start.