ChatGPT does not rank pages. It runs a multi-stage pipeline: decide whether to search, query Bing, generate sub-queries, fetch pages live, read them as a language model, and pick which ones to cite. Understanding each stage is the only way to influence the outcome.
This is the single reference guide for everything we know about how ChatGPT search works and what it takes to get cited. It pulls together findings from two peer-reviewed studies, 19,556 queries, 10,293 crawled pages, and five experiments across 8 industry verticals. Every number is sourced. Every recommendation traces back to published data.
If you are looking for a broader view of AI search optimization across all platforms, see our GEO pillar guide. This post focuses entirely on ChatGPT.
🔍 HOW CHATGPT DECIDES WHETHER TO SEARCH THE WEB
Every conversation starts with a decision the model makes before it ever touches the web: does this prompt need external information, or can ChatGPT answer from its training data alone?
This decision is binary. Either it searches, or it does not. If it does not search, your content is invisible for that query. No optimization can change that. The trigger rate varies by query type, and this variation is the single most important number in ChatGPT SEO.
From our dataset of 19,556 Google Autocomplete queries mapped to ChatGPT behavior across 8 verticals (Lee, 2026a):
| Query Type | Share of Autocomplete Queries | Web Search Trigger Rate (API) | Citation Opportunity |
|---|---|---|---|
| Discovery ("best X for Y") | 31.2% | ~73% | High |
| Review-seeking ("X reviews") | 2.0% | ~70% | High |
| Comparison ("X vs Y") | 2.3% | ~65% | High |
| Validation ("is X good") | 3.2% | ~40% | Moderate |
| Informational ("what is X") | 61.3% | ~10% | Very Low |
The API rates above are for automated/developer access. The web browser interface behaves differently: across 391 brand and product queries tested through the web UI, 42% triggered a web search overall (Lee, 2026a). The gap between API and web UI exists because the web interface has additional retrieval paths, including broader Bing organic integration.
The trigger mechanism is not random. ChatGPT evaluates each prompt against its parametric knowledge. Time-sensitive queries ("best CRM for 2026"), product comparisons, and entities the model lacks confidence on all push the trigger rate higher. Generic knowledge questions ("what is a CRM") almost never trigger a search because the model's training data is sufficient.
The Bottom Line: Discovery queries trigger web search at 7 times the rate of informational queries. If your content library targets only "what is" and "how to" keywords, you are competing in the 10% trigger zone. The 73% zone belongs to discovery and comparison content. Content strategy for ChatGPT starts with intent mapping, not keyword mapping.
🌐 THE BING PIPELINE: HOW CHATGPT DISCOVERS URLS
Once ChatGPT decides to search, it does not use its own web index. It queries Bing's API. This is the single most important architectural fact about ChatGPT search.
ChatGPT has no proprietary crawl index. Every URL it considers for citation was first discovered through Bing. If Bing has not indexed your page, ChatGPT cannot find it.
| Bing Index Status | ChatGPT Visibility |
|---|---|
| Page indexed in Bing | Eligible for discovery |
| Page not indexed in Bing | Completely invisible to ChatGPT |
| Page indexed but buried past page 3 | Unlikely to enter candidate set |
| Page indexed with canonical issues | May be deduplicated away |
But Bing is the gatekeeper, not the decision-maker. Our research found that ChatGPT's top-3 Bing URLs matched actual citations only 6.8% to 7.8% of the time (Lee, 2026a). Being in Bing's results gets you through the door. Your position within those results barely matters for which URLs ChatGPT ultimately cites.
Domain-level alignment tells a different story. Domain-level overlap between Bing's top results and ChatGPT's citations ranges from 28.7% to 49.6%. ChatGPT draws from the same pool of domains that Bing surfaces, but picks different specific pages.
The 68% training data factor
A widely repeated industry claim says 87% of ChatGPT citations come from Bing. Our 400-query replication with fan-out sub-query tracking found the real number is 27% (Lee, 2026a; see our full Bing replication study). The breakdown:
| Citation Source | Share of All Citations |
|---|---|
| Matched Bing's top 20 results | 27% |
| Matched Google's top 20 results | 13.4% |
| Neither search engine (training data) | 68% |
Two-thirds of what ChatGPT cites does not come from any search engine at all. These citations were determined months or years ago when the model was trained. No amount of Bing optimization will influence them.
The Bottom Line: Submit your XML sitemap to Bing Webmaster Tools. Verify your pages are indexed. But understand that Bing indexation is necessary, not sufficient. Training data accounts for 68% of all citations, and only 27% trace back to Bing's live results.
🔄 FAN-OUT QUERIES: CHATGPT'S HIDDEN SUB-QUERY SYSTEM
When a user asks ChatGPT a complex question, it does not send a single query to Bing. It generates multiple reformulated sub-queries and sends them in parallel. This "fan-out" mechanism is one of the most important differences between ChatGPT search and traditional search engines.
We observed ChatGPT generating 2 sub-queries 92% of the time, with discovery queries averaging 3.63 sub-queries per prompt (Lee, 2026a).
| Fan-Out Metric | Value |
|---|---|
| Sub-queries per complex prompt | 2 (typical) |
| Discovery queries (highest fan-out) | 3.63 average |
| Comparison queries | 4 to 6 sub-queries typical |
| Simple informational queries | 1 or 0 (often no search at all) |
| Time window for parallel queries | Under 2 seconds |
For example, a prompt like "What project management tool should a marketing agency use?" might decompose into:
- "best project management software marketing agencies 2026"
- "project management tools for creative teams comparison"
- "Asana vs Monday vs ClickUp agency workflows"
- "project management pricing small agency"
Each sub-query returns its own Bing results. The candidate URL pool is the union of all result sets. This means your page can be discovered through queries you never explicitly targeted.
The Bottom Line: Single-keyword optimization is not enough. The fan-out mechanism rewards topical breadth. Cover multiple dimensions of your topic (features, pricing, comparisons, use cases, alternatives) so your page enters the candidate pool through multiple sub-queries. For the full breakdown of how fan-out works, see ChatGPT Fan-Out Queries Explained.
🤖 THE THREE OPENAI BOTS (AND HOW THEY DIFFER)
OpenAI operates three separate bots that interact with your website in completely different ways. Each serves a different purpose, follows different rules, and has different implications for your content.
| Bot | Purpose | Respects robots.txt | When It Runs |
|---|---|---|---|
| GPTBot | Training data collection | Yes | Background crawling |
| OAI-SearchBot | Search index for ChatGPT Search citations | Yes | Background indexing |
| ChatGPT-User | Live page fetching during conversations | No (since Dec 2025) | Real-time, per conversation |
GPTBot crawls pages to collect training data for OpenAI's models. Blocking it prevents your content from being used in future model training but has zero effect on whether ChatGPT cites you during live conversations.
OAI-SearchBot builds the index specifically for ChatGPT Search results (the inline citations users see). Blocking it can reduce your visibility in search-style results, but will not prevent ChatGPT-User from fetching your pages during browsing sessions.
ChatGPT-User is the bot that actually matters for citations. When ChatGPT decides to search during a conversation, this bot fetches candidate pages in real time. As of December 2025, it ignores robots.txt entirely. OpenAI's rationale: ChatGPT-User is "a technical extension of the user," not an autonomous crawler.
| ChatGPT-User Behavior | Detail |
|---|---|
| JavaScript execution | None. Content must be server-side rendered. |
| robots.txt compliance | Ignores all directives since December 2025 |
| Fetch timing | Real-time during conversation |
| Rendering | Reads raw HTML only. No CSS or JS processing. |
| Cookie/auth walls | Cannot bypass. Paywalled content is invisible. |
| Concurrent page fetches per prompt | 5 to 15 (from merged fan-out results) |
The Bottom Line: You can block training (GPTBot) and search indexing (OAI-SearchBot), but you cannot block conversational fetching (ChatGPT-User). Server-side rendering is non-negotiable: if your content loads via JavaScript, ChatGPT-User sees an empty page.
🏆 HOW CHATGPT SELECTS WHICH PAGES TO CITE
After discovering URLs through Bing, decomposing the query via fan-out, and fetching page content live, the model reads the fetched content and decides which sources to cite. This is where the process diverges most sharply from traditional search ranking.
ChatGPT does not use PageRank, backlink counts, or domain authority scores. It reads your page as a language model and evaluates whether the content adds value to its response.
The 6 page-level features from Experiment M
Our most recent study tested 10,293 pages across 250 queries on 3 AI platforms (ChatGPT, Perplexity, Google AI Mode), controlling for Google rank position so that page features are not confounded by website reputation (Lee, 2026c). Six features predicted citation in all four Google position bands:
| Rank | Feature | Effect Size (Cohen's d) | Direction | What It Means |
|---|---|---|---|---|
| 1 | Comparison structure | 0.43 (medium) | Positive | Pages with comparison tables, "A vs B" sections, or feature grids get cited more, even for non-comparison questions |
| 2 | Query term coverage | 0.42 (medium) | Positive | Cited pages contain 100% of the search query's key words; uncited pages cover about 75% |
| 3 | Blog/opinion tone | -0.30 to -0.37 (medium) | NEGATIVE | First-person writing ("I think," "in my experience") suppresses citation; reference-style writing wins |
| 4 | Primary source score | 0.27 (small-medium) | Positive | Pages that create original data get cited more than pages that aggregate others' work |
| 5 | Word count | 0.20 (small) | Positive | Cited pages median ~2,000 words vs ~1,400 for not-cited (within same Google position band) |
| 6 | Subheading depth | 0.19 (small) | Positive | Cited pages have roughly 2x as many H3 subheadings |
Three things stand out:
Comparison structure is the strongest content signal in the entire study. It works for all question types, not just comparison questions. Comparison tables are packed with specific, extractable facts that AI can easily pull out and cite.
Blog tone is the strongest negative signal. Pages written like personal blogs get cited less than pages written like reference guides. Same information, different tone, different outcome.
Word count reverses once you control for website identity. The earlier cross-domain finding (cited pages = 1,799 words, shorter) was confounded by prestigious domains having shorter, denser pages. Within the same position bands, cited pages are actually ~42-52% longer. Target ~2,000 words.
What about commonly recommended "AI SEO" tactics?
Several commonly recommended optimizations showed no effect once we controlled for Google position:
| Feature | Common Advice | What the Data Showed |
|---|---|---|
| Page speed | "Make your site faster!" | Not significant in any position band (p > 0.39) |
| Author bylines | "Add author bios for trust!" | No consistent effect |
| Readability scores | "Write at an 8th grade level!" | No effect |
| Product schema | "Add Product markup!" | No effect within position bands |
| Review schema | "Add Review markup!" | No effect within position bands |
| Popup/modal elements | "Remove popups!" | p = 0.606, not significant |
Page speed is the most dramatic reversal. It appeared to be the most important factor in our first experiment (65.8% of the model). After we controlled for which website a page was on, it completely disappeared.
The website-level factor: topical authority
The single biggest predictor at the website level is how many different searches your site ranks for. Among 2,542 websites that appeared in Google's top 20 (Lee, 2026c):
| Google Top-20 Appearances | AI Citation Rate |
|---|---|
| 1 question | 33.8% |
| 2 to 3 questions | 72.9% |
| 4 to 7 questions | 87.3% |
| 8 to 15 questions | 95.2% |
| 16+ questions | 100% |
Websites that rank for 4+ related questions have an 87%+ citation rate. This is not just more lottery tickets: per-appearance citation rates also increase (0.67 citations per slot for 1-appearance sites vs 2.04 for 8+ appearances). AI platforms genuinely trust websites with broader topical coverage.
The Bottom Line: Intent selects the pool. Page features select the winner within that pool. Build topical authority by ranking for many related searches, then optimize individual pages with comparison structure, query term coverage, reference-style tone, original data, and clear subheading structure. For the full accessible breakdown of Experiment M, see What Gets You Cited by AI, Explained.
📊 CHATGPT CITATION STATISTICS: THE REFERENCE TABLE
Every key number in one place.
| Statistic | Value | Source |
|---|---|---|
| Total queries analyzed | 19,556 | Lee (2026a) |
| Industry verticals covered | 8 | Lee (2026a) |
| Pages crawled (expanded dataset) | 4,658 | Lee (2026a) |
| Pages crawled (position-controlled) | 10,293 | Lee (2026c) |
| Discovery query trigger rate (API) | ~73% | Lee (2026a) |
| Informational query trigger rate (API) | ~10% | Lee (2026a) |
| Overall web UI trigger rate | 42% | Lee (2026a) |
| ChatGPT top-3 URL overlap with Google | 7.8% (API), 6.8% (web UI) | Lee (2026a) |
| Domain-level overlap with Google | 28.7% to 49.6% | Lee (2026a) |
| Cross-platform URL overlap (all 4 platforms) | 1.4% | Lee (2026a) |
| Citations from training data (not from any search engine) | 68% | Lee (2026a) |
| Citations traceable to Bing top 20 | 27% | Lee (2026a) |
| Sub-queries per prompt (typical) | 2 (92% of the time) | Lee (2026a) |
| Sub-queries per discovery prompt (average) | 3.63 | Lee (2026a) |
| Strongest content predictor (comparison structure) | d = 0.43 | Lee (2026c) |
| Median word count (cited, position-controlled) | ~2,000 | Lee (2026c) |
| Median word count (not-cited, position-controlled) | ~1,400 | Lee (2026c) |
| Full model AUC (page features + position) | 74.0% | Lee (2026c) |
| Reddit citations via API | 0% | Lee (2026a) |
| Reddit citations via web UI | 17% | Lee (2026a) |
The Reddit citation paradox
Reddit dominates Google's organic results (51.8% of top positions) but receives zero citations through the ChatGPT API. Through the web UI, Reddit jumps to 17%. The API and web UI use different retrieval pipelines. The API strips Reddit entirely, likely due to content licensing constraints. For businesses competing with Reddit for visibility, the API channel is wide open.
🆚 HOW CHATGPT COMPARES TO OTHER AI PLATFORMS
ChatGPT's search pipeline is not universal. Each AI platform has its own architecture, and the differences are substantial enough that optimizing for one does not guarantee visibility on others.
| Dimension | ChatGPT | Perplexity | Google AI Mode | Claude |
|---|---|---|---|---|
| URL discovery | Bing API | Own pre-built index | Google index | Live fetch on demand |
| Crawler | ChatGPT-User (live) | PerplexityBot (background) | Googlebot | Claude-User (session) |
| Freshness | Real-time (live fetch) | High (3.3x fresher than Google) | Google's crawl schedule | On-demand only |
| Citation format | Inline with URL | Numbered footnotes | Integrated with search | Inline when web-enabled |
| robots.txt | Ignores (ChatGPT-User) | Often ignores | Respects (Googlebot) | Respects |
| Fan-out queries | Yes (2 to 4 sub-queries) | Yes (similar decomposition) | Yes (tied to Google) | Limited |
| Schema sensitivity | Moderate (cross-domain only) | Moderate | Moderate | Low |
| Gatekeeper | Bing index | PerplexityBot crawl | Google index | No persistent index |
The most striking finding: cross-platform citation agreement is essentially random. The overlap rate across all four major platforms was just 1.4% (Lee, 2026a). A page cited by ChatGPT has no higher probability of being cited by Perplexity or Claude for the same query.
This means each platform requires somewhat different optimization:
- ChatGPT: Bing indexation is the gatekeeper. Fan-out queries reward comprehensive content. Training data accounts for 68% of citations.
- Perplexity: Needs PerplexityBot discoverability. Re-crawls 3.3x more frequently than Google, so freshness is weighted heavily.
- Google AI Mode: Google indexation is the gatekeeper. Google ranking signals carry partial weight, making it the most traditional-SEO-friendly platform.
- Claude: Citation only happens with web search enabled. No persistent index means every search starts from scratch. Content directness matters most.
The shared foundation (clean HTML, server-side rendering, comprehensive content, reference-style tone) applies everywhere. The discovery layer is platform-specific.
The Bottom Line: "AI SEO" is not one strategy. For the full head-to-head comparison, see ChatGPT vs Perplexity vs Gemini: What AI Platforms Actually Cite.
✅ THE OPTIMIZATION CHECKLIST
Based on all the research above, here is the priority-ordered checklist for getting your content cited by ChatGPT. Items are ordered by measured impact.
Priority 1: Target the Right Query Types (Highest Impact)
- Map your target queries to intent categories (discovery, comparison, review-seeking, validation, informational)
- Prioritize discovery and comparison content (73% and 65% trigger rates) over pure informational content (10%)
- For discovery queries: build comparison tables with specific recommendations
- For informational queries: create comprehensive, Wikipedia-style explainers
- Match content format to what ChatGPT cites for that intent type
Priority 2: Build Topical Authority (Website Level)
- Create a cluster of pages around your core topic, each targeting a different question
- Interlink all related pages to build deep internal navigation
- Aim to rank in Google's top 20 for 4+ related searches (87%+ citation rate)
- Track SERP co-occurrence across your topic cluster
Priority 3: Optimize the 6 Page-Level Features
- Add comparison sections: tables, "A vs B" breakdowns, feature grids (d = 0.43, strongest signal)
- Include the exact words people search for, early in the page (100% query term coverage)
- Write in third-person, reference-style tone (avoid "I think," "in my experience")
- Include original data: statistics, test results, percentages, sample sizes
- Target ~2,000 words of comprehensive, focused coverage
- Break content into lots of subsections with clear H3 headings (2x more than you think you need)
Priority 4: Technical Requirements
- Submit XML sitemap to Bing Webmaster Tools
- Verify ChatGPT-User is allowed (note: it ignores robots.txt since Dec 2025 anyway)
- Ensure server-side rendering (ChatGPT-User does not execute JavaScript)
- Confirm pages return meaningful HTML without JavaScript
- Add self-referencing canonical tags on every page
- Implement FAQ schema with high attribute completeness
Priority 5: Cover Fan-Out Sub-Queries
- Cover multiple facets of your topic (features, pricing, comparisons, alternatives, use cases)
- Structure content with clear headings that match potential sub-query phrasing
- Add FAQ sections addressing related questions
- Include specific numbers and comparison tables that match multiple sub-query angles
For a free assessment of your pages against these predictors, try our AI Visibility Quick Check. For a comprehensive audit, see our AI SEO services.
❓ FREQUENTLY ASKED QUESTIONS
Does my Google ranking affect whether ChatGPT cites me?
Google rank is the starting line, not the finish line. Pages in Google's top 3 get cited 57.2% of the time. Pages ranked 13 to 20 get cited 13.5% of the time (Lee, 2026c). So ranking higher helps, but even at position 1, 31% of pages are NOT cited. Within the same position band, the 6 page-level features determine who wins. There is essentially zero direct correlation between Google rank and AI citation (Spearman rho = -0.02 to 0.11). ChatGPT uses Bing for URL discovery, not Google.
How quickly can ChatGPT see changes to my content?
Because ChatGPT-User fetches pages live during conversations, changes are visible almost immediately. If you update a page and someone asks ChatGPT a relevant question within minutes, the bot can fetch and cite your updated content. This is a significant advantage over platforms like Perplexity that rely on background crawling.
Should I block or allow OpenAI's crawlers?
Allow ChatGPT-User (the live fetcher) if you want citation visibility. You can separately block GPTBot (training data crawler) if you do not want your content used in model training. OAI-SearchBot (search indexer) is worth allowing for search-style results. Note that ChatGPT-User ignores robots.txt since December 2025, so blocking it only works through IP-level server blocks, which also prevents citation.
Can I optimize for ChatGPT and Google at the same time?
Yes, and the strategies overlap more than they conflict. Technical fundamentals (fast pages, clean HTML, proper canonicals, XML sitemaps) benefit both. The biggest difference is in content strategy: Google rewards keyword-optimized pages with strong backlink profiles. ChatGPT rewards comprehensive, well-structured pages that match query intent regardless of backlinks. The shared foundation is reference-style content with comparison structure, original data, and clear subheadings.
Why does Reddit get 0% of ChatGPT API citations but 17% in the web UI?
The API and web UI use different retrieval pipelines. The API pipeline strips Reddit entirely, likely due to content licensing constraints. The web UI pipeline incorporates broader source diversity, possibly influenced by Bing's organic results where Reddit occupies 51.8% of top positions. This split is consistent across all AI platforms: 0% API citations for Reddit across the board.
If cross-platform overlap is only 1.4%, do I need to optimize separately for each AI platform?
The foundation is shared: clean HTML, server-side rendering, comprehensive content, reference-style tone. Where the platforms diverge is in discovery (Bing vs Google vs proprietary indexes) and retrieval architecture. A practical approach is to optimize the shared foundation first, then add platform-specific layers: Bing indexation for ChatGPT, Google indexation for Google AI Mode, and crawlability for Perplexity's background bot.
What is the best content length for ChatGPT citation?
Within position-controlled analysis, cited pages have a median of ~2,000 words vs ~1,400 for not-cited pages (d = 0.20). The earlier cross-domain finding of 1,799 words for cited pages was confounded by domain identity. The practical target is around 2,000 words: long enough to be comprehensive, short enough to stay focused. Extremely long pages (5,000+) do not help more.
How do I track whether ChatGPT is citing my content?
There is no native dashboard. The most reliable method is server log analysis: look for requests with "ChatGPT-User" in the user-agent string. The volume of ChatGPT-User requests to a specific URL is a leading indicator of citation frequency. Our AI Visibility Quick Check can also assess your pages against the known citation predictors.
📚 REFERENCES
Lee, A. (2026a). "Query Intent, Not Google Rank: What Best Predicts AI Citation Behavior." Preprint v5, A.I. Plus Automation. DOI: 10.5281/zenodo.18653093. Dataset: 19,556 queries across 8 verticals, 4 AI platforms, 4,658 crawled pages.
Lee, A. (2026c). "I Rank on Page 1: What Gets Me Cited by AI? Position-Controlled Analysis of Page-Level and Domain-Level Predictors of AI Search Citation." A.I. Plus Automation. Paper: aixiv.science/abs/aixiv.260403.000002. Dataset: DOI: 10.5281/zenodo.19398158. 10,293 pages, 66 features, 250 queries, 3 AI platforms.
Aggarwal, P., Murahari, V., Rajpurohit, T., Kalyan, A., Narasimhan, K., & Deshpande, A. (2024). "GEO: Generative Engine Optimization." Proceedings of KDD 2024. DOI: 10.48550/arXiv.2311.09735. (Note: specific content-level features do not replicate on production platforms.)
Sellm (2025). "ChatGPT Citation Analysis." Industry report (400K pages analyzed).