2. Each AI Cites Different Sources

GEO Optimization Guide — 전체 시리즈

1. 1. What Is GEO - AI Citation Strategy Beyond SEO
2. 2. Each AI Cites Different Sources ← 현재 글
3. 3. On-Site GEO Technical Architecture - From Product DB to JSON-LD
4. 4. Off-Site GEO - How to Win Over AI That Ignores Your Official Site
5. 5. AEO - Why Coding Agents Read Documentation Differently

Do Not Take “We Show Up in AI” at Face Value

After deciding to implement GEO, the first thing most people do is ask ChatGPT about their brand. “Our company shows up – great, it’s working.” And that is where it ends.

But ask the same question on Perplexity, and you get a different answer. Google AI Overview gives yet another result. On some platforms, the official site gets cited; on others, blog reviews become the source. Even within ChatGPT itself, citation sources change depending on whether web search mode is on or off.

You cannot treat AI search as a single channel. Each platform has its own citation logic and preferred source types.

Platform Citation Sources, by the Data

Combining analysis data from three providers reveals stark differences across platforms. While Yext, Qwairy, and GrackerAI differ in scale, they all point in the same direction.

Gemini inherits Google’s search logic and trusts structured content from official sites. ChatGPT relies on external search layers, making it heavily influenced by directories and listing sites. Perplexity pulls answers from community threads where people actually exchanged opinions.

What is striking is how little overlap exists between platforms. Only 11% of domains are cited by both ChatGPT and Perplexity. Visibility on one does not mean visibility on the other.

Source and Citation Are Different Concepts

When you look closely at AI responses, you will notice two types of attribution.

One is the source – reference links listed at the bottom of the answer. A URL gets included here if it is deemed trustworthy. The other is the citation – a hyperlink embedded mid-answer that supports a specific sentence as evidence.

Getting cited requires passing two gates: domain-level trust of the URL, and information trust of the page content. If the domain is reliable but the content structure is hard to parse, it shows up as a source but not as a citation. Conversely, if the content is well-structured but the domain itself is weak, it stays in the source list only.

Connecting this back to the three GEO principles from the previous article :

• Identity – The foundation of domain trust. GTIN, Organization schema contribute here
• Context – Information quality of the content. Categories, use cases, and variant relationships must be structured for parsing
• Citability – The gateway from source to citation. JSON-LD and FAQ Schema determine this stage

Results Differ Based on Web Search Mode

If you ask ChatGPT the same question twice, you may get different answers. The difference lies in whether web search is on or off.

With web search off, answers are based on pre-trained data. Content that existed at the time of training becomes the citation candidate. When web search is on, it switches to real-time crawling. At that point, the current state of structured data, robots.txt permissions, and content freshness drive citations.

Failing to distinguish this when monitoring GEO leads to misjudgment. “Our brand shows up fine on ChatGPT” – if you tested with web search off, that result came from pre-trained data. You need to test with real-time crawling enabled to see your current GEO status accurately.

YouTube Is Surging in Google AIO

The most notable change in Google AI Overview’s citation patterns is YouTube.

YouTube is the number one cited domain in AI Overview (Ahrefs Brand Radar). Its share grew 34% in six months. Among social platforms, YouTube ranks second after Reddit in citations (OtterlyAI).

What is interesting is that it is not high-view videos getting cited. Nearly half of YouTube videos cited by AI had fewer than 1,000 views. Plenty had only a few dozen likes. AI looks at information structure, not popularity. Timestamps, chapter divisions, clear titles – these are what determine citation.

In contrast, YouTube citations are nearly nonexistent on ChatGPT and Perplexity. The same video content has entirely different value depending on the platform.

Check Competitors’ robots.txt to See Their AI Strategy

The fastest way to understand how competitors approach GEO: type https://competitor-domain/robots.txt into your browser.

E-commerce companies often block AI crawlers to protect product catalogs – preventing prices, inventory, and product details from being exposed to competitor AI systems. B2B SaaS companies do the opposite and open things up, since visibility in AI search benefits lead generation.

For large conglomerates, robots.txt policies often vary wildly across subsidiaries. Some block GPTBot entirely while others allow everything. The result of each subsidiary setting its own policy without a group-level standard.

In fact, when surveying the subsidiaries of a major retail group, zero had applied FAQPage Schema, and only one – a hotel subsidiary – had adequate AI citations. Even with robots.txt open, if there is no structured data, there is nothing for AI to read.

# Blocked (GEO impossible)
User-agent: GPTBot
Disallow: /

# Allowed (GEO possible)
User-agent: GPTBot
Allow: /
User-agent: Google-Extended
Allow: /
User-agent: anthropic-ai
Allow: /

Restructuring Alone Can Change Citation Rates – No New Content Needed

The most common misconception when starting GEO is that you need to create new content. While new content certainly helps, simply restructuring existing content can already change AI citation outcomes.

Pages with structured data are 36% more likely to appear in AI Overview (GrackerAI), and applying complete Schema pushes the ChatGPT visibility rate up to 80% (Search Engine Land). With only basic Schema, it is 20%.

Things you can do with existing blog posts:

• Add a 40-60 word TLDR summary at the top
• Include author bio and expertise credentials
• Add an FAQ section wrapped with FAQPage Schema
• Rewrite <meta name="description"> to be more specific, at least 100 characters

Content freshness also matters. More than three-quarters of pages highly cited by Perplexity were updated within the past month. Pages untouched for over three months get pushed down.

Reddit Appearing in Korean Queries Is Not a Coincidence

You may have noticed Reddit threads being cited in response to Korean-language queries. It seems odd, but there is a structural reason.

Reddit uses AI translation and hreflang tags to operate in 22 language versions. The infrastructure to match Korean queries is already in place. Reddit accounts for 6.6% of all Perplexity citations and 2.2% in Google AI Overview – both top-tier. Its share climbs even higher for subjective queries like “best XX” or “XX recommendations.”

This has significant implications from an Off-Site GEO perspective. There are domains that On-Site structuring alone cannot cover. For query types where AI prioritizes “real user opinions,” community and review platform influence is substantial.

One Strategy Cannot Cover All AI Platforms

Each platform trusts different things:

On-Site GEO is the baseline. Open robots.txt, add JSON-LD, structure FAQs – and you will see strong results on Gemini and Google AIO in particular.

To gain visibility on ChatGPT, you need to additionally ensure consistency across third-party listings. To target Perplexity as well, natural mentions in community discussions become necessary.

This series focuses on On-Site GEO. However, depending on which platform you prioritize, Off-Site work may be needed too.