# How ChatGPT Does Query Fan-Out: Inside Ahrefs' 66.6M Prompt Search Study

OpenAI's search architecture inside ChatGPT has fundamentally shifted away from user-generated forum threads toward primary brand domains and licensed news publishers. A massive empirical analysis of 66.6 million free, logged-out ChatGPT responses running on GPT-5.6 Luna between August and September 2026 reveals that the assistant triggers a live web search on only 14.4% of prompts—roughly one in seven user interactions—while relying on parametric model weights for the remaining 85.6%.

When ChatGPT does invoke its browsing pipeline, it executes an average of 1.6 background fan-out queries per search-enabled response, generating 15.0 million total search queries across the 9.6 million responses that accessed the live web. Rather than firing five or six broad keyword variations as early retrieval-augmented generation prototypes did, GPT-5.6 Luna issues compact, highly targeted search strings designed to verify live pricing, product specifications, and official documentation.

The most consequential shift for search marketers is ChatGPT's programmatic reliance on the `site:` search operator to bypass third-party aggregators. Exactly 6.8% of all fan-out queries explicitly constrained results to a single domain across roughly 116,000 unique websites. Among the 160 most frequently targeted domains, 75.9% of `site:` queries pointed directly at official brand properties—led by `apple.com` (9,442 fan-outs), `ikea.com` (7,361), `nike.com` (5,210), `adidas.com` (4,619), and `samsung.com` (4,432)—compared to just 11.9% for multi-brand retailers and 1.9% for community platforms such as Reddit and Facebook.

That structural preference for first-party verification coincides with an 88% six-month collapse in Reddit's citation share inside ChatGPT answers. In parallel, explicit modifier injection grew sharply between August and September 2026: the word `"official"` appeared in 13.1% of English fan-out queries in September (up from 10.1% in August), `"price"` jumped from 6.0% to 8.9%, and review-intent fan-out modifiers expanded by up to 47% month over month.

According to the [Ahrefs 66.6M-prompt ChatGPT query fan-out study](https://ahrefs.com/blog/query-fan-out-chatgpt/) authored by Louise Linehan, OpenAI also routes news and current-events retrieval through an internal tier codenamed `labrador` that prioritizes 17 licensed publishing partners, including The Associated Press, Financial Times, Reuters, Condé Nast, Axel Springer, Hearst, Dotdash Meredith, News Corp, The Atlantic, Vox Media, and Le Monde. Across the broader dataset, those 17 partner organisations saw their combined ChatGPT citation share more than double from 7.5% in March 2026 to a peak of 15.8% in July, stabilizing at 12.5% in September 2026.

## Fast Facts: ChatGPT Query Fan-Out Metrics (Aug–Sept 2026)

- **Dataset Scale**: 66.6 million logged-out GPT-5.6 Luna responses analyzed alongside a controlled 20,000-prompt US benchmark.
- **Web Search Trigger Rate**: 14.4% of responses (9.6 million) initiated a web search; 85.6% answered purely from model training data.
- **Fan-Out Density**: 15.0 million total fan-out queries, averaging 1.6 queries per web-enabled response.
- **Domain Operator Concentration**: 6.8% of fan-out queries used `site:` across ~116,000 domains, with 75.9% of top-domain `site:` searches targeting official brand sites (`apple.com`, `ikea.com`, `nike.com`).
- **Modifier Acceleration (Aug vs. Sept 2026)**: `"official"` rose from 10.1% to 13.1% of English fan-outs; `"price"` climbed from 6.0% to 8.9%; `"review"` modifiers surged up to 47%.
- **Reddit vs. Licensed Publishers**: Reddit citations dropped 88% over six months, while OpenAI's 17 licensed publisher partners (`labrador` tier) doubled citation share from 7.5% to 15.8% before settling at 12.5%.

## How Intent Splits Citation Share: Factual vs. Comparison Prompts

To isolate how query fan-out changes depending on buyer intent, Ahrefs evaluated a controlled sample of 20,000 US prompts split evenly between 10,000 factual product questions (such as warranty terms, dimensions, or API rate limits) and 10,000 comparative buying questions (such as best-in-class rankings or head-to-head software evaluations).

| Metric / Retrieval Behavior | Factual Product Prompts (10,000 US) | Comparison & Buying Prompts (10,000 US) | Global Dataset Baseline (66.6M Prompts) |
|---|---|---|---|
| **Responses Citing >=1 Official Brand Site** | 84.4% of all responses | 68.3% of all responses | Driven by 75.9% brand `site:` share |
| **Official Brand Share of Total Citations** | 63.4% of cited links | 46.4% of cited links | Top domains: Apple, IKEA, Nike, Samsung |
| **Responses Citing >=1 Review / Listicle Site** | 15.6% of all responses | 31.7% of all responses | `"review"` fan-out modifier up 47% MoM |
| **Primary Retrieval Operator Strategy** | `site:brand.com` + `"official"` / `"price"` | Category keywords + `"review"` / year | 1.6 fan-outs per search (14.4% trigger) |

The data exposes a sharp bifurcation in Answer Engine Optimization (AEO). When a user asks a factual or specification question, ChatGPT's fan-out generator appends `"official"` or `site:brand.com`, resulting in 84.4% of responses citing at least one official source and first-party domains capturing 63.4% of all citation slots. Third-party blogs attempting to rank for factual queries about another company's pricing or specs are structurally filtered out before synthesis even begins.

Conversely, when a user asks a comparative question, ChatGPT suppresses single-brand `site:` constraints and injects `"review"` or current-year modifiers. In those comparative prompts, official brand citation share drops by 17 percentage points to 46.4%, while independent review publications, testing labs, and industry analysts capture the remaining 53.6% of citations.

## Why First-Party Technical Architecture Now Dictates AI Visibility

Because ChatGPT actively constructs `site:yourdomain.com` queries and appends `"official"` or `"price"` when researching commercial entities, a brand's own domain architecture is now its primary AEO asset. If an enterprise hides its pricing behind interactive JavaScript calculators, gates specification sheets behind lead-capture forms, or blocks `OAI-SearchBot` in `robots.txt`, GPT-5.6 Luna's `site:` fan-out query returns empty or thin snippets—forcing the model to fall back to secondary aggregators or competitors.

### The 4-Step Enterprise Implementation Playbook

1. **Audit `site:` Indexability for Commercial Modifiers**: Run live search tests for `site:yourdomain.com official`, `site:yourdomain.com price`, and `site:yourdomain.com specs` across Bing and Google to verify that canonical product, pricing, and documentation pages return clean, machine-readable HTML snippets without JavaScript rendering dependencies.
2. **Publish Un-Gated Specification and Pricing Matrices**: Because `"price"` fan-outs jumped 48% in a single month (from 6.0% to 8.9%), replace opaque "Contact Sales" walls with transparent tier tables, token or seat unit pricing, and structured `Product` and `Offer` JSON-LD schema so factual prompts resolve directly to your domain.
3. **Reallocate Off-Site Budget From Reddit Seeding to Independent Reviews**: With Reddit citations down 88% over six months and `site:` queries hitting social forums just 1.9% of the time, shift earned-media budgets toward authoritative industry testing sites, benchmark repositories, and vertical publications that capture the 53.6% non-brand citation share on comparison prompts.
4. **Maintain a Synchronized `/llms.txt` and `/llms-full.txt` Index**: Expose a clean Markdown directory at the root of your domain summarizing every product specification, pricing tier, and canonical documentation URL so retrieval crawlers can ingest verified first-party facts in a single pass.

  Next Steps

1. Verify that `OAI-SearchBot` and `GPTBot` have full crawl access in your production `robots.txt` and Cloudflare WAF rules.
2. Create dedicated, crawlable HTML comparison and pricing pages that explicitly include your brand name alongside `"official"`, `"price"`, and `"review"` semantic headings.
3. Benchmark your domain's citation frequency across 50 factual prompts and 50 comparison prompts to measure whether your first-party pages hit the 84.4% factual citation baseline.