ChatGPT Trusts Directories. Gemini Trusts Your Site. Perplexity Trusts Experts.

If you have been treating your AI visibility strategy as one problem to solve, the data says you are likely solving it correctly on one platform and doing nothing useful on the other two.

Yext analyzed 17.2 million citations across ChatGPT, Perplexity, and Gemini and published the per-platform breakdown of which source types each platform actually trusts. The numbers are not subtle.

**ChatGPT:** 49% of citations come from third-party directories -- Yelp, BBB, G2, Capterra, industry registries.

**Gemini:** 52% of citations come from brand-owned websites.

**Perplexity:** Trusts niche expert directories specific to the industry category, not the general-purpose directories that work for ChatGPT.

Three platforms. Three completely different source pools. If a business is running one optimization strategy across all three, it is probably getting the right answer for one platform.

Why the platforms diverge structurally

This is not a quirk or a temporary difference. It is a direct consequence of how each platform retrieves information before generating an answer.

In our 2026-04-22 investigation of platform citation divergence (knowledge/platform-divergence-data.md, updated 2026-06-09), we tracked the retrieval architecture behind each platform's citation behavior.

ChatGPT generates what researchers are now calling "query fan-out." A single user prompt is decomposed into sub-queries, with 91% of those sub-queries being unique to that expansion. ChatGPT is not looking for your website. It is casting a wide net across dozens of related angles and pulling from third-party sources that have documented entities across all of them. For businesses, those are the directory listings that have been consistently crawled and indexed.

Perplexity runs the inverse. It retains 88% of the original prompt words and runs approximately one query per prompt. Its retrieval layer behaves more like a live web search than a parametric training database. What surfaces depends on what ranks in Perplexity's real-time crawl -- which is why niche expert directories (legal registries for lawyers, medical directories for healthcare, tech databases for SaaS) outperform general consumer directories on this platform. The source pool is specific to the category, not universal.

Gemini's brand-site trust reflects its Google lineage. It evaluates authority at the domain level, weighs structured schema markup and E-E-A-T signals, and draws heavily from Google properties -- Google Business Profile, Maps, Reviews. For Gemini, what you have on your site matters more than what third parties say about you.

Three architectures. Three citation patterns. Three different fix sequences that do not substitute for each other.

The 11% overlap confirms it at scale

The practical implication of this divergence is documented across three independent studies. Averi's 680 million citation dataset, Whitehat SEO's 118,000-response analysis, and Passionfruit's cross-platform study all arrive at the same number: only 11% of domains are cited by both ChatGPT and Perplexity.

When three methodologically different studies land on essentially the same figure -- confirmed across hundreds of millions of data points -- you have a stable finding, not a sampling artifact.

The platforms are nearly independent in their source selection. That independence is the architecture point above, observed at scale.

The practical consequence: a business that audits only on ChatGPT and builds directory coverage based on that audit will not automatically improve on Perplexity, because 89% of Perplexity's source pool is different. A business that optimizes its website for Gemini -- the dominant recommendation in SEO-influenced AEO guides -- will not see the same movement on ChatGPT, because ChatGPT's citations come primarily from third-party sources rather than the brand site.

The citation rate gap between the two is also notable. Averi's B2B data shows Perplexity cites brands at 13.05% and ChatGPT at 0.59% -- a 46x difference. A business invisible on Perplexity and present on ChatGPT is not necessarily failing AEO. It may be in a category where Perplexity's niche directory coverage is thin, or where the brand's expert-directory presence is weak.

The diagnosis requires per-platform data, not an aggregate score.

Entity fragmentation cuts across all three platforms

The source-architecture picture covers the question of which sources each platform trusts. There is a second problem that affects all three: whether those sources describe the same entity consistently.

In our 2026-04-23 investigation of entity disambiguation (knowledge/wikidata-entity-disambiguation.md, session 9 update), we documented a failure pattern we call entity fragmentation. When a brand is described differently across directory listings -- G2 says "AI audit tool," Capterra says "SEO visibility platform," Crunchbase says "marketing analytics software" -- LLMs may treat these as distinct entities rather than one coherent business.

The result is that the entity confidence score splits. Each description competes with the others, and no single representation reaches the threshold needed for reliable citation. The brand has presence, but fragmented presence -- and the platforms read it as ambiguity rather than authority.

The inverse is what actually works: when G2, Capterra, Crunchbase, and LinkedIn all use the same semantic terms to describe the business, the confidence score concentrates rather than fragments. Stanford AI Index 2024 found that 18% or more of LLM outputs involving brand entities contain hallucinations or misattributions. Entity fragmentation is one of the structural causes -- the model has competing descriptions and hedges toward the safer option, which is often no mention.

This is not a schema fix or a website fix. It is a cross-source consistency problem. And it cuts across all three platforms because the fragmented descriptions flow into each platform's respective source pool.

The platform-specific fix sequence

If a Sourcepull audit shows weak scores on a specific platform, the per-platform breakdown points to different starting points.

**ChatGPT weak:** Third-party directory coverage is the first lever. BBB, Yelp, G2, or the industry-equivalent general directories. The goal is breadth of entity mentions in sources ChatGPT's training pipeline indexed. Schema and brand site improvements help Gemini more than they help ChatGPT.

**Gemini weak:** Brand site structure is the starting point. LocalBusiness schema, complete structured markup, strong E-E-A-T signals, current Google Business Profile. If the site is well-structured and GBP is populated, Gemini typically converts faster than the other two platforms.

**Perplexity weak:** Identify the niche expert directories your specific category uses. For a law firm, legal registries and bar association profiles. For healthcare, medical directories. For SaaS, niche tech databases. These outperform general directories on Perplexity and are under-prioritized in most AEO guides that treat directory presence as a single undifferentiated category.

**Across all three:** Check entity consistency. If the business description varies meaningfully across the top 10 directory listings, unifying that language is a cross-platform fix that benefits citation confidence on every platform simultaneously.

A note on the Yext data

The 17.2 million citation study was explicit in its methodology: ChatGPT, Perplexity, and Gemini only -- Claude was not included. We noted internally that Yext is a directory and listings company, so their research methodology has an incentive to emphasize directory presence. The source-type breakdown is still the most granular public data available on platform citation architecture, and it aligns directionally with the 11% overlap finding from independent studies that have no commercial interest in the finding. The convergence between the Yext source-type data and the three-study overlap figure is what makes the platform divergence case strong.

Treating the Yext percentages as directional rather than precise is the right read. The qualitative conclusion -- ChatGPT draws heavily from directories, Gemini draws heavily from brand sites, Perplexity from niche expert sources -- has corroboration from both architecture analysis and behavioral data.

---

Signal Check at sourcepull.ca runs the full four-platform query set and returns per-platform citation data. For businesses running a single-platform strategy, the per-platform breakdown often makes visible a gap they were not aware they had.