Advanced GEO in the AI Era — Your Own Site Is Just the Foundation. Off-site Is a 5.7× Bigger Battlefield

A counter-intuitive fact for marketing teams

For the past decade, marketing teams spent most of their time on “what should our website say”: landing pages, product pages, blogs, case studies. That was correct in the traditional SEO era — Google ranked your keywords based on your own content.

But the AI-search era reveals a distribution that’s worth brand owners pausing on:

About 85% of brand mentions in AI answers come from third-party pages; only 15% from the brand’s own site — off-site mentions are about 5.7× the volume of owned mentions.

This doesn’t mean your own site is unimportant — it remains the foundation of your GEO body (52% of Gemini’s citations still come from brand-owned domains; sites strong in SEO start ahead in GEO; well-written citable paragraphs get cited more stably each time).

But the 5.7× signal is: finishing your own site is just the ticket of entry, not the whole game. The brand story AI sees is distributed across many sources, and its judgement reads roughly as “multiple sources say this, so the conclusion is credible” (we call this Consensus is King). If 100% of your GEO budget is on your own site, you’re participating in only 15% of the conversation — the 5.7× larger battlefield is where someone else is shaping the narrative (yours or your competitor’s).

See Differences in how three engines pick citations — Gemini’s 52% brand-owned vs ChatGPT’s heavier reliance on Wikipedia / Reddit / Forbes is exactly this logic playing out.

<g>
  <circle cx="440" cy="180" r="40" fill="#1e293b" stroke="#60a5fa" stroke-width="2"/>
  <text x="440" y="178" text-anchor="middle" fill="#60a5fa" font-size="13" font-weight="700">Owned brand</text>
  <text x="440" y="195" text-anchor="middle" fill="#94a3b8" font-size="10">15% mentions</text>
</g>

<g>
  <circle cx="200" cy="100" r="58" fill="#1e293b" stroke="#fb923c" stroke-width="1.5"/>
  <text x="200" y="92"  text-anchor="middle" fill="#fb923c" font-size="13" font-weight="700">Local forums</text>
  <text x="200" y="110" text-anchor="middle" fill="#e2e8f0" font-size="11">Reddit / PTT</text>
  <text x="200" y="124" text-anchor="middle" fill="#e2e8f0" font-size="11">Dcard</text>
</g>

<g>
  <circle cx="680" cy="100" r="58" fill="#1e293b" stroke="#fb923c" stroke-width="1.5"/>
  <text x="680" y="92"  text-anchor="middle" fill="#fb923c" font-size="13" font-weight="700">Video platforms</text>
  <text x="680" y="110" text-anchor="middle" fill="#e2e8f0" font-size="11">YouTube</text>
  <text x="680" y="124" text-anchor="middle" fill="#e2e8f0" font-size="11">Transcript citations</text>
</g>

<g>
  <circle cx="200" cy="260" r="58" fill="#1e293b" stroke="#fb923c" stroke-width="1.5"/>
  <text x="200" y="252" text-anchor="middle" fill="#fb923c" font-size="13" font-weight="700">Knowledge bases</text>
  <text x="200" y="270" text-anchor="middle" fill="#e2e8f0" font-size="11">Wikipedia</text>
  <text x="200" y="284" text-anchor="middle" fill="#e2e8f0" font-size="11">Business databases</text>
</g>

<g>
  <circle cx="680" cy="260" r="58" fill="#1e293b" stroke="#fb923c" stroke-width="1.5"/>
  <text x="680" y="252" text-anchor="middle" fill="#fb923c" font-size="13" font-weight="700">Third-party reviews</text>
  <text x="680" y="270" text-anchor="middle" fill="#e2e8f0" font-size="11">Trustpilot / G2</text>
  <text x="680" y="284" text-anchor="middle" fill="#e2e8f0" font-size="11">Google Reviews</text>
</g>

<g stroke="#475569" stroke-width="1" stroke-dasharray="3,3" fill="none">
  <line x1="258" y1="100" x2="400" y2="170"/>
  <line x1="622" y1="100" x2="480" y2="170"/>
  <line x1="258" y1="260" x2="400" y2="195"/>
  <line x1="622" y1="260" x2="480" y2="195"/>
</g>

<text x="440" y="312" text-anchor="middle" fill="#94a3b8" font-size="11">The owned brand is just one node; AI citation actually relies on the "consensus signals" across all four quadrants</text>

1. The landscape: Reddit, Wikipedia, YouTube are the three gateways

Semrush analysed 150,000 citations across major LLMs in June 2025, and three sources lead:

These three add up to nearly 90%. The remaining 10% is split across news media, forums, review sites, and government domains.

But this landscape isn’t static — it’s dynamically reshuffling:

• Between August and September 2025, Reddit’s share within ChatGPT dropped from 60% to 10% (OpenAI changed retrieval parameters), with the released share absorbed by PR Newswire, Forbes, and Medium
• Between August and December 2025, YouTube’s share in AI citations rose from 18.9% to 39.2% (OtterlyAI YouTube Citation Study 2026)

In other words: betting on a single platform is highly risky. Below we unpack each quadrant.

2. Knowledge bases: Wikipedia is still GEO’s single strongest signal

Wikipedia holds a special place for LLMs — it enters both the training corpus and the real-time citation pool. Even in pure-training, non-retrieval LLM modes, Wikipedia content is already “baked into” the model weights.

What it means for brands:

• Being in Wikipedia ≈ your brand entity has an anchor in the LLM’s training-corpus layer, making the AI’s “implicit understanding” of you more stable
• Not in Wikipedia ≈ the AI knows you only via real-time search, with both citation rate and accuracy taking a discount
• Entries need to be credible and supported by secondary sources — Wikipedia deletes “self-PR” articles directly
• Even if you can’t make Wikipedia, having entries in Wikidata, Crunchbase, Bloomberg, etc. acts as a knowledge-base substitute

We’ve covered why Wikipedia is the strongest signal and the legitimate path to a listing in a separate post: Why Wikipedia inclusion is one of GEO’s strongest signals

3. Video: YouTube is the fastest-rising citation source post-2025

YouTube has been the fastest-growing single platform for AI citations in the past year. Key numbers from OtterlyAI’s 2026 study:

• Share rose from 18.9% to 39.2% in 5 months (Reddit fell from 44.2% to 20.3% in the same period)
• Long-form videos generated 574,420 AI citations in 2025 — 51× the citation volume of Shorts
• Of videos cited by AI, 40.83% had fewer than 1,000 subscribers — channel size is essentially uncorrelated with citation rate (Pearson ≈ 0)

What it means for brands:

1. Transcript matters more than title — the AI actually reads the transcript, not the SEO tags. Videos without complete transcripts can’t even be “consumed” by the AI
2. Topic matters more than subscriber count — good news for smaller brands: you don’t need to hit a million subscribers before Perplexity will cite you
3. Long-form > Shorts — AI citation values content depth, not format trendiness
4. Perplexity is the main YouTube citation engine — it displays transcript snippets directly in answers

In other words, video content is already GEO’s second battlefield, on par with your own site and Wikipedia — a configuration that simply didn’t exist in the SEO era.

4. Local forums: Reddit is #1 in English; Taiwan’s equivalents are PTT and Dcard

Reddit’s 40.1% share in AI citations (Semrush 2025-06) leads across all major AI engines.

Why is Reddit so strong? Because it’s the largest pool from which AI engines infer “what real users are saying about this brand” — AI increasingly discounts “marketing copy,” while giving high weight to “actual user experience writing.”

The problem for Taiwanese brands: mainstream LLM training corpora contain very little Traditional Chinese from Taiwan (see Why AI models don’t know Taiwanese brands), so whether PTT and Dcard truly enter the AI citation pool doesn’t have publicly-reported share data the way Reddit does.

But the functional role is the same: PTT and Dcard are publicly-crawlable real-user discussion spaces in Taiwan that Common Crawl picks up. Their role to AI engines is analogous to Reddit’s role in the English-speaking world.

What it means for brands:

• Don’t treat PTT / Dcard as “PR crisis management” targets — they’re part of your brand narrative
• Being discussed ≠ being recommended: AI can distinguish negative consensus from positive consensus
• Absolutely no “astroturfing” — AI gets more sensitive to marketing-toned content over time; astroturfing reduces trust scoring; only naturally-accumulated genuine discussion works

5. Third-party reviews: 3× citation rate, but a double-edged sword

Third-party review platforms (Trustpilot, G2, Capterra, Yelp in English; in Taiwan: Google Reviews, business listings, Pchome product reviews, Mobile01 reviews) have a measurable amplifying effect on AI citation:

• Sites with profiles on these platforms see their ChatGPT citation odds rise 3× (industry research consensus)
• The correlation coefficient between brand mention count and citation rate is 0.664 — being mentioned more places is nearly linearly tied to AI citation rate
• Earned + owned media accounts for 90% of AI citation traffic; paid media only 10%

But there’s a crucial caveat: review quality determines direction. The 3× number is conditional on “positive reviews.”

If your Google Reviews average below 3.5 stars, the AI carries that fact into the citation — which is worse than not being cited at all.

6. Hard evidence from Princeton at KDD 2024

To anchor the observations above to an academic data point — Princeton’s GEO paper at KDD 2024 tested 9 optimisation techniques. The three most effective:

The third technique’s improvement is far larger than the first two — academic-level evidence for the “off-site authority” logic: AI doesn’t just care what your site says, it cares about what your network of links to the outside world looks like.

7. Concrete recommendations for Taiwanese brands

Premise first: your own site is always the foundation (GEO body score, citable paragraphs, schema, SEO technical basics all live here; 52% of Gemini’s citations are still brand-owned domains). The order below isn’t a “ranking of importance” — it’s the direction in which most brands are currently severely imbalanced and need to rebalance. Most clients today spend 80% of their budget on their own site, leaving only crumbs for off-site work. The list below is about rebalancing that ratio.

1. Google Business Profile, Wikidata / Wikipedia (if eligible) — the knowledge-base layer; first priority for AI’s “implicit understanding” of your brand entity
2. Third-party reviews (Google Reviews ≥ 4.0 average) + organic accumulation on local forums (naturally occurring, no astroturf) — the consensus layer
3. Long-form video content with complete transcripts — the newly-risen GEO second battlefield
4. Media PR (earned media) — industry reports, expert commentary, list-style mentions
5. Continued SEO / GEO maintenance on your own site — to be done regardless; don’t cut this in order to do the other 4

Read this correctly: item 5 is “don’t cut” — not “do this last.” Items 1–4 are the “most brands’ missing pieces.” Plug the gaps to a non-fatal level first, then come back and polish the foundation.

8. How to start

GEO’s off-site dimensions aren’t a “hire a consultant to fix the Wikipedia entry” one-off — it’s a long-term build that spans content, PR, community, and review management. But the first step is always to quantify the current state:

👉 Free GEO health check — 3 minutes to see your site’s scores across 12 dimensions, including “external authority signal” and “cross-source consensus” which map directly to the off-site quadrants in this post.

Once the health check reveals gaps, building Wikipedia / video / third-party reviews / media PR tracks systematically is the core scope of our consulting service: contact@geoweb.tw

Data sources: Semrush cross-LLM citation analysis (150,000 citations, 2025-06; summarised via Soar); OtterlyAI YouTube Citation Study 2026; Contently: Do YouTube Transcripts Influence AI Search Summaries; 5W AI Platform Citation Source Index 2026; Aggarwal et al., GEO: Generative Engine Optimization (KDD 2024); the 3× third-party-review citation rate and 0.664 brand-mention correlation are summary values from a 2025 cross-section of GEO industry research (including Frase, Averi, Similarweb analyses). PTT / Dcard AI citation share has no publicly-reported data; this post infers their functional role from their position in Common Crawl and as Taiwan’s real-user Chinese discussion spaces.