Seven AI platforms. Seven different ways of finding and citing your content. ChatGPT sends 3-5 sub-queries to Bing. Perplexity runs a three-layer reranking pipeline across 59 citation factors. Claude searches Brave. Gemini pulls from Google's Knowledge Graph. Google AI Mode fires 8-12 fanout queries. And the content each platform favors is dramatically different: ChatGPT prefers authoritative lists (41% of citations), while Claude overwhelmingly prefers databases (68%). If you are treating AI visibility as a single optimization problem, you are solving the wrong problem.
This post breaks down exactly how each platform retrieves, ranks, and cites content. Every claim is sourced. The goal: give you enough architectural understanding to build platform-specific strategies instead of hoping one approach works everywhere.
Why Retrieval Architecture Matters
Here is the foundational insight. In January 2026, Rand Fishkin published a study (SparkToro, 600 respondents) that tested whether AI platforms agree on brand recommendations. The headline finding: when you ask "what are the best X" across platforms, the exact brand lists are almost never the same. Less than 1 in 100 responses produce the same ordered list.
But Fishkin found something else. The visibility percentages are consistent. If a brand appears in 70% of ChatGPT responses for a category, that number holds across repeated tests. The list order is noise. The appearance rate is signal.
AI brand lists are random (less than 1/100 produce the same order). But visibility percentages are consistent. City of Hope appeared in 97% of responses. Bose held at 55-77%. This means you should measure visibility rate, not ranking position.
This has a direct implication for optimization: each platform's retrieval architecture determines which sources get surfaced consistently. Understanding the architecture tells you what to optimize. Here are all seven.
How Does ChatGPT Find Content?
ChatGPT uses Bing (and recently Google) for real-time web retrieval, breaking each user prompt into 3-5 sub-queries. It dominates AI referral traffic at 87.4% share (Conductor, Feb 2026).
ChatGPT's share of all AI-referred website visits (Conductor)
Sub-queries generated per user prompt for web search
Citations from first 30% of content (Kevin Indig)
The Retrieval Pipeline
When a user asks ChatGPT a question that needs current information, the system generates 3-5 distinct sub-queries (what practitioners call "fanout") and sends them to its search backend. OpenAI originally relied exclusively on Bing. As of early 2026, ChatGPT also queries Google, giving it access to both indexes simultaneously.
Each sub-query retrieves a set of web results. ChatGPT then synthesizes these into a single response, citing the sources it drew from. The sub-queries are visible if you use tools like Quolity.ai, a Chrome extension that intercepts ChatGPT's network requests to show you the exact queries being fired.
Where ChatGPT Looks in Your Content
Kevin Indig's analysis of 1.2 million AI citations identified what he calls the "ski ramp" pattern: 44.2% of citations come from the first 30% of a page's content. The middle 40% contributes 31.1%, and the last 30% just 24.7%.
This means your lead paragraph and first few sections carry disproportionate weight. Answer-first structure is not just a writing preference. It is a citation architecture decision.
What ChatGPT Prefers to Cite
Jesse Bailyn's research on AI citation sources breaks down ChatGPT's citation preferences by content type:
| Content Type | ChatGPT Citation Share |
|---|---|
| Authoritative lists | 41% |
| Awards/rankings | 18% |
| Reviews | 16% |
| Expert recommendations | 13% |
| Databases | 12% |
ChatGPT gravitates toward curated lists and rankings. If your brand appears on "Best X for Y" lists from authoritative publishers, ChatGPT is more likely to surface you.
Recent Developments
Two things changed ChatGPT's retrieval behavior in February 2026. First, Agent mode (also called "Operator" or "deep research") lets ChatGPT autonomously browse websites, fill forms, and execute multi-step tasks. This goes beyond search retrieval into direct website interaction. Second, Instant Checkout (launched February 16, 2026) enables in-chat purchases, meaning ChatGPT now accesses product catalogs and shopping data as part of its retrieval pipeline.
One more data point from Lily Ray's February 2026 study of 11 websites: when a site loses Google organic rankings, its ChatGPT citation rate drops by up to 49%. The correlation is not perfect, but it is strong enough to confirm that Google organic performance feeds into ChatGPT's citation pipeline, likely through Bing's indexing of Google's ranking signals.
How Does Perplexity Find Content?
Perplexity is built as a RAG-first (Retrieval-Augmented Generation) search engine, generating approximately 6 fanout queries per prompt and running results through a three-layer reranking pipeline. It holds 15.1% of AI referral traffic (SE Ranking).
The Retrieval Pipeline
Metehan Yesilyurt's infrastructure analysis revealed the technical architecture. Perplexity retrieves web results and then applies three reranking layers:
- L1: Initial retrieval from web search (broad recall)
- L2: Semantic reranking (relevance filtering)
- L3: XGBoost model across 59 citation factors (final ranking)
That L3 layer is what makes Perplexity distinct. It uses a machine learning model trained on 59 factors to decide which sources deserve citation. These include domain authority, content freshness, topical relevance, and source diversity.
Freshness Is Critical
Content recency carries outsized weight on Perplexity. Kevin Indig's data shows a 3.2x citation boost for content under 30 days old. Mike Lafferty's testing found a 2-3 day decay cycle where content loses citation priority rapidly after publication.
This makes Perplexity the most freshness-sensitive platform. If you are publishing evergreen content and never updating it, Perplexity will deprioritize you.
What Perplexity Prefers to Cite
Bailyn's data shows a very different preference profile than ChatGPT:
| Content Type | Perplexity Citation Share |
|---|---|
| Authoritative lists | 64% |
| Databases | 21% |
| Reviews | 8% |
| Expert recommendations | 5% |
| Awards/rankings | 2% |
Perplexity is list-obsessed. Nearly two-thirds of its citations come from authoritative list content. Awards and rankings, which ChatGPT values at 18%, barely register on Perplexity at 2%.
Reddit is also a major citation source for Perplexity. Averi's research found Reddit accounts for 46.7% of top 10 citations on Perplexity. Q&A threads and comparison posts drive the most citations.
Recent Developments
Samsung's Galaxy S26 shipped with "Hey Plex" integration in early 2026, giving Perplexity a hardware distribution channel. Perplexity also dropped its advertising model and went subscription-only, which may reduce the commercial bias in its citation decisions.
How Does Claude Find Content?
Claude uses Brave Search for real-time web retrieval and has a 200K token context window that allows it to process significantly more content per query than competitors. Despite holding only 0.17% of AI referral traffic, Claude users spend an average of 19 minutes per session, the longest engagement of any AI platform.
The Retrieval Pipeline
Anthropic confirmed in February 2026 that Claude's web search feature runs through Brave Search. This is architecturally significant because Brave maintains its own independent web index, separate from Google and Bing. Content that ranks well on Google or Bing is not guaranteed to rank on Brave, and vice versa.
Claude generates 4-6 fanout queries per prompt and tends toward conservative, depth-focused retrieval. It shows a clear preference for academic and reference-quality sources. The 200K token context window means Claude can ingest and reason over much longer documents than other platforms, potentially favoring detailed, comprehensive content.
Three Crawlers
Claude operates three distinct web crawlers:
| Crawler | Purpose | Robots.txt |
|---|---|---|
| ClaudeBot | Training data collection | Respects robots.txt |
| Claude-User | Live user queries | Ignores robots.txt |
| Claude-SearchBot | Brave Search integration | Respects robots.txt |
The distinction matters: blocking ClaudeBot prevents your content from entering training data, but it will not stop Claude from citing you in live search results (Claude-User ignores robots.txt).
What Claude Prefers to Cite
Claude's citation preferences are dramatically different from every other platform:
68% databases. Claude overwhelmingly cites structured data repositories, knowledge bases, and reference databases. This is nearly the exact inverse of ChatGPT (41% lists, 12% databases). If you are optimizing for Claude, structured reference content and database listings matter far more than getting on "best of" lists.
Recent Developments
Claude Cowork launched in February 2026, letting Claude work alongside users on extended projects. This extended interaction model means Claude may access and cite content multiple times within a single session, further rewarding depth and comprehensiveness over shallow list-style content.
How Does Gemini Find Content?
Gemini retrieves content through Google's Knowledge Graph and Google Search, using dense vector matching for semantic similarity. It holds approximately 6.4% of AI referral traffic.
The Retrieval Pipeline
Gemini's architecture gives it a unique advantage: direct access to Google's Knowledge Graph. This massive structured database contains billions of entity relationships (people, companies, products, concepts) with verified attributes. When Gemini processes a query, it can pull authoritative entity data directly, not just web pages.
On top of the Knowledge Graph, Gemini runs standard Google Search retrieval with dense vector matching. Dense vectors represent content as high-dimensional numerical embeddings, letting Gemini match queries to content based on meaning rather than keyword overlap.
Multi-Modal Content Preference
Gemini has a strong preference for YouTube content, unsurprisingly given Google's ownership of the platform. Video transcripts, YouTube descriptions, and video content are weighted more heavily in Gemini's retrieval than on other platforms. Multi-modal content (text plus images, diagrams, or video) also appears to receive a retrieval boost.
What Gemini Prefers to Cite
| Content Type | Gemini Citation Share |
|---|---|
| Authoritative lists | 49% |
| Databases | 31% |
| Reviews | 10% |
| Expert recommendations | 6% |
| Awards/rankings | 4% |
Gemini sits between ChatGPT and Claude. It values lists (49%) but also gives meaningful weight to databases (31%). This dual preference means both list-oriented and database-oriented content strategies can work on Gemini.
Technical Note: Opaque Grounding URLs
One implementation detail worth knowing: Gemini's API returns grounding URLs as opaque redirects through vertexaisearch.cloud.google.com. These are not the actual source URLs. You must resolve them (follow the redirect) to get the real source domains. If you are building any monitoring or analytics on Gemini citations, this redirect layer can cause source attribution errors if not handled.
How Does Google AI Mode Find Content?
Google AI Mode fires 8-12 fanout queries per user prompt, more than any other platform. It uses a specialized Gemini version and pulls from Google's Shopping Graph for commercial queries. Google AI Mode has reached approximately 75 million users across 53 languages.
Most sub-queries of any AI platform per prompt
Google AI Mode reach across 53 languages
Users who click traditional results below AI Overviews (Pew)
The Retrieval Pipeline
Google AI Mode (formerly AI Overviews) represents Google's full-stack AI search experience. When triggered, it runs a special version of Gemini that generates 8-12 sub-queries against Google's index. This is 2-3x more fanout queries than ChatGPT or Claude. More fanout means a wider net of sources are considered, but it also means the synthesis step is more aggressive in filtering.
For commercial queries ("best running shoes for flat feet," "CRM software for small business"), AI Mode accesses Google's Shopping Graph, a structured database of product information, prices, availability, and reviews. This means e-commerce visibility in AI Mode requires product data feeds and structured markup, not just great content.
The Binary Visibility Problem
AI Mode creates a binary outcome: you are either cited in the AI response, or you are invisible. Only 8% of users click the blue links below AI Overviews, according to a Pew Research study. This means traditional "page 1" rankings are increasingly irrelevant if you are not included in the AI-generated answer itself.
With only 8% of users clicking traditional results below AI Overviews, Google AI Mode is the most binary platform. You are either cited in the AI answer or effectively invisible. There is no "page 1, position 7" fallback anymore.
What Google AI Mode Prefers
Google AI Mode inherits many of Gemini's citation preferences but adds strong E-E-A-T (Experience, Expertise, Authoritativeness, Trustworthiness) signals from Google's existing search quality systems. Content that already ranks well in traditional Google Search has a baseline advantage, but AI Mode also pulls from sources that Google Search might not rank highly if they have strong entity signals in the Knowledge Graph.
How Does Copilot Find Content?
Microsoft Copilot uses a dual-lane retrieval system: BM25 (keyword matching) plus dense vector search, powered by Bing's index. Copilot holds approximately 2% of AI referral traffic.
The Retrieval Pipeline
Copilot's retrieval is technically interesting because it runs two parallel search strategies:
- BM25 lane: Classic keyword matching (term frequency, inverse document frequency). This is the same algorithm Bing has used for years. It rewards content with exact keyword matches.
- Dense vector lane: Semantic embedding similarity. This lane matches content based on meaning, even if the exact keywords do not appear.
Results from both lanes are merged and reranked before being fed to the language model for synthesis. This dual-lane approach means Copilot can find content that matches either by keywords or by semantic meaning, giving it broader recall than systems that rely on only one approach.
Microsoft 365 Integration
Copilot's deep integration with Microsoft 365 (Word, Excel, PowerPoint, Outlook, Teams) means it also has access to enterprise data. For B2B companies, this is significant: Copilot may surface your content not just in web search but within the tools your buyers use daily. Microsoft's documentation, GitHub repositories, and LinkedIn content also receive preferential treatment in Copilot's retrieval.
What This Means for Optimization
Because of the BM25 lane, traditional keyword optimization still works on Copilot. Do not abandon keyword targeting for Copilot the way you might for a purely semantic platform. Use both: clear keyword-rich headings (for BM25) and semantically comprehensive content (for the dense vector lane).
How Does Grok Find Content?
Grok prioritizes X/Twitter social signals and real-time conversation data. It has access to 300+ million X users as a data source. Grok holds less than 1% of AI referral traffic.
The Retrieval Pipeline
Grok's retrieval architecture is unique because it has privileged access to X (formerly Twitter) data. While other platforms can crawl X's public pages, Grok has direct API access to the real-time firehose of posts, replies, threads, and engagement signals.
This means Grok weights social proof differently than any other platform. A brand that is actively discussed on X, by verified accounts, in threads with high engagement, will have a retrieval advantage on Grok that does not translate to other platforms.
What Grok Prefers
Grok shows the strongest preference for:
- Recency: Real-time X data means Grok's retrieval is the most time-sensitive of any platform
- Social authority: Verified accounts, high-follower users, and viral threads carry weight
- Conversational content: Natural discussion and debate format, not polished marketing copy
- Trending topics: Grok can identify and surface content aligned with current X trends
When Grok Matters
For most B2B companies, Grok's sub-1% traffic share makes it a low priority. But for consumer brands, media companies, and anyone in categories with active X communities (tech, crypto, gaming, sports, politics), Grok can be an early warning system. If your brand is discussed positively on X, Grok will amplify that. If it is discussed negatively, Grok will amplify that too.
What Is the Same Across All Platforms
Despite the architectural differences, some optimization strategies work everywhere. Here is the universal checklist based on data from all seven platforms.
1. Structured Data (Schema Markup)
Every platform benefits from structured data. Schema.org markup (Organization, Product, Article, FAQ, HowTo) gives retrieval systems clean entity signals. According to Pacestack's benchmark of 243 sites, only 46% have any Schema.org markup. This is low-hanging fruit.
2. Content Freshness
Every platform rewards recently updated content, though the degree varies. Perplexity is the most freshness-sensitive (3.2x boost for content under 30 days). Even ChatGPT and Gemini, which rely heavily on training data, factor publication dates into their retrieval when using web search.
Action: Add datePublished and dateModified to your Article schema. Update key pages regularly with new data or insights.
3. Answer-First Structure
The ski ramp pattern (44.2% of citations from the first 30% of content) holds directionally across all platforms. Put your key claims, data points, and conclusions at the top. Expand below.
4. E-E-A-T Signals
Experience, Expertise, Authoritativeness, and Trustworthiness are not just Google concepts. Every AI platform evaluates source quality. Author bios, citation of primary sources, transparent methodology, and third-party validation all help.
5. Entity Density
Kevin Indig's data shows a 4.8x citation boost for content with 15+ named entities per article. Named entities (companies, people, products, locations, technical terms) help retrieval systems understand what your content is about and connect it to the Knowledge Graph.
6. Original Data and Research
Content that contains original data, proprietary research, or unique analysis is harder to replicate and more likely to be cited. All seven platforms show preference for content that adds new information to a topic rather than summarizing existing sources.
7. Third-Party Mentions
Your own content is necessary but not sufficient. AI platforms cross-reference your claims against third-party sources. Being mentioned on Wikipedia, industry directories, Crunchbase, G2, Reddit, and in news coverage provides corroboration signals that boost citation probability across all platforms.
Frequently Asked Questions
Which AI platform sends the most traffic to websites?
ChatGPT at 87.4% of all AI referral traffic (Conductor, Feb 2026). Perplexity is second at 15.1% (SE Ranking). But most AI-referred visits arrive without referrer headers ("dark AI traffic"), making GA4 attribution unreliable. Loamly's detection finds that 80%+ of AI visits are classified as "Direct" in standard analytics.
Does optimizing for ChatGPT help on other platforms?
Partially. ChatGPT and Claude share a moderate correlation of 0.503 in our 2,014-company dataset. But ChatGPT vs Gemini drops to 0.175. The architectural differences explain why: ChatGPT uses Bing + authoritative lists, Claude uses Brave Search + databases, and Gemini uses Google Knowledge Graph + dense vectors. Universal tactics (schema, freshness, E-E-A-T) help everywhere, but citation preferences diverge significantly.
How many prompts do you need to measure AI visibility accurately?
At least 50 per platform. Fishkin's study showed that individual brand lists are random, but visibility percentages stabilize with sufficient sample size. Loamly's paid audit runs 50-100 prompts per platform across ChatGPT, Claude, Gemini, and Perplexity to get statistically meaningful visibility rates. Single-prompt tests are noise.
Can I block AI platforms from using my content?
Partially. You can block training crawlers (ClaudeBot, GPTBot) via robots.txt. But live search retrieval crawlers (Claude-User, ChatGPT-User) may ignore robots.txt for real-time queries. Blocking training crawlers removes you from the model's learned knowledge but will not prevent citation through web search. For most companies, being cited is the goal, not prevention.
Measure Your Visibility Across Platforms
Each platform finds and cites content differently. The only way to know how you appear on each one is to measure it.
Run a free multi-platform visibility report at loamly.ai/check. It queries ChatGPT, Claude, Gemini, and Perplexity with prompts relevant to your business and shows you exactly where you appear, what AI says about you, and where the gaps are.
If you need the full picture (50-100 prompts per platform, competitive analysis, citation chain mapping, and per-page GEO scores), the Loamly AI Visibility Audit covers all four major platforms in a single report.
Last updated: February 25, 2026
Stay Updated on AI Visibility
Get weekly insights on GEO, AI traffic trends, and how to optimize for AI search engines.
No spam. Unsubscribe anytime.
Related Articles
The 15 Search Queries Your Competitors Are Not Appearing In (And Why You Should Target Them)
Most companies optimize for 20-40 queries where they already win, missing hundreds of high-value queries where competitors haven't established dominance.
The AI Search Visibility Stack: 7 Essential Components Every Company Needs in 2026
AI search visibility needs 7 components: monitoring, content optimization, entity mapping, audience research, competitive intel, integrations, measurement.
Check Your AI Visibility
See what ChatGPT, Claude, and Perplexity say about your brand. Free, no signup.
Get Free Report