# Graphite Study of 957K Queries Finds ChatGPT Rarely Names Brands Missing From Training Data

A large-scale empirical study by SEO platform Graphite reveals that getting mentioned on the third-party web pages ChatGPT cites is rarely enough to earn a brand recommendation in the final AI answer unless the model already recognizes the brand from its pre-training weights. Analyzing 3,115 entity-question pairs alongside 957,000 background fan-out search queries generated from 22,953 non-branded user prompts, researchers found that ChatGPT injects specific brand names into 59.6% of its internal retrieval searches before reading a single live web page.

For the past eighteen months, Answer Engine Optimization (AEO) and Generative Engine Optimization (GEO) strategies have largely revolved around a single tactical assumption: if a software vendor or consumer brand secures placements on the top comparison listicles, Reddit threads, and editorial reviews cited by retrieval-augmented generation (RAG) pipelines, the LLM will synthesize those pages and recommend the brand. Graphite's dataset challenges that linear model of AI search discovery.

When a user submits a purely non-branded commercial prompt—such as asking for the best enterprise payroll software or cloud observability platform—ChatGPT does not simply pass that generic phrase to a search index. Instead, the model consults its internal parametric memory first, selects a shortlist of brands it already associates with the category, and fires off background fan-out queries that explicitly name those incumbents.

As a result, live web retrieval frequently operates as a verification and citation-gathering step for brands the model has already chosen rather than an open discovery mechanism for emerging challengers. Even when a newer brand appeared prominently on at least 15% of the exact web pages ChatGPT cited in its footnotes, the brand's median visibility in the generated response remained stuck at 11% if the model lacked prior parametric recall in no-search tests.

## Fast Facts: Graphite 957K Fan-Out Query & ChatGPT Citation Study

- **Research Scope & Dataset**: Graphite evaluated 3,115 entity-question pairs across 57 commercial comparison prompts and inspected 957,000 background fan-out queries generated from 22,953 non-branded prompts, as detailed in [Search Engine Journal's breakdown of the Graphite study](https://www.searchenginejournal.com/chatgpt-brand-mentions-cited-pages-graphite/594379/).
- **Pre-Retrieval Brand Injection**: Although 0% of the 22,953 user prompts contained a brand name, 59.6% of ChatGPT's internal fan-out search queries included specific brand names (classified using Claude Sonnet 4.5).
- **The Parametric Recall Bottleneck**: For brands with low baseline model recall (named in fewer than 25% of no-search responses), appearing substantially on 15% or more of cited web pages yielded a median ChatGPT search visibility of just 11%.
- **Citation Rate Surge in Logged-Out ChatGPT**: The share of logged-out ChatGPT responses displaying web citations jumped from 56.6% during the week of August 3 to 88.5% during the week of September 28.
- **Gemini Citation Comparison**: Over the same August-to-September tracking window, logged-out Google Gemini responses containing citations shifted slightly from 79.0% to 75.5%.
- **Strategic Implication**: Listicle placement and RAG citation capture only convert reliably into ChatGPT brand recommendations when paired with sufficient corpus-wide entity authority to register in base model weights.

## Empirical Breakdown: Parametric Memory vs. Live RAG Citations

The table below synthesizes the core quantitative findings from Graphite's multi-part study across query fan-out behavior, cited-page correlation, and platform citation rates.

| Research Dimension | Dataset / Methodology | Measured Metric | Operational Takeaway for AI SEO Teams |
|---|---|---|---|
| **Fan-Out Brand Injection** | 957,000 fan-out queries from 22,953 non-branded user prompts (labeled via Claude Sonnet 4.5) | **59.6%** of background search queries contained a brand name | ChatGPT decides which brands to investigate *before* querying the live search index |
| **Low-Recall Cited Page Impact** | 3,115 entity-question pairs (57 comparison prompts); brands on >=15% of cited pages but <25% no-search recall | **11%** median visibility in ChatGPT with web search enabled | Securing placements on cited third-party listicles rarely overcomes weak pre-training recall |
| **ChatGPT Logged-Out Citations** | Weekly longitudinal tracking (Week of Aug 3 vs. Week of Sept 28) | Rose from **56.6% to 88.5%** (+31.9 percentage points) | Nearly nine in ten logged-out ChatGPT answers now trigger web grounding and footnote attribution |
| **Gemini Logged-Out Citations** | Weekly longitudinal tracking (Week of Aug 3 vs. Week of Sept 28) | Shifted from **79.0% to 75.5%** (-3.5 percentage points) | Gemini maintains steady three-quarters citation density while ChatGPT aggressively expanded grounding |

## How Query Fan-Out Locks Out Challenger Brands During RAG

To understand why appearing on a cited web page fails to guarantee a mention in ChatGPT's output, growth engineers must look at the two-stage architecture of modern conversational search: query decomposition (fan-out) followed by grounded synthesis.

When a user asks a zero-brand question, the LLM orchestrator first decomposes the prompt into multiple sub-queries executed in parallel against the search index. Graphite's inspection of 957,000 fan-out queries from 22,953 non-branded prompts showed that in 59.6% of cases, the model wrote specific brand names directly into those sub-queries. If a user asks for "top B2B data enrichment APIs," the model's query planner—guided by its pre-training weights—might emit background searches for `"Clearbit pricing 2026"`, `"ZoomInfo API rate limits"`, and `"Apollo vs Lusha B2B data"`.

Because the retrieval query itself is anchored to pre-selected brands, two downstream filtering effects occur during synthesis:

1. **Targeted Extraction Over Broad Summarization**: When the search engine returns a "Top 10 B2B Data Tools" comparison page in response to a branded fan-out query, the attention mechanism of the reader model focuses on extracting pricing, features, and verification facts for the specific brand named in the sub-query. Other challenger brands listed on that exact same web page are treated as peripheral noise and omitted from the final synthesized answer, even though the URL appears in the footnote citations.
2. **Parametric Prior Dominance**: In Graphite's analysis of 3,115 entity-question pairs across 57 comparison prompts, researchers isolated brands that achieved substantial coverage (appearing on at least 15% of the cited URLs for a given prompt) but had low baseline recall when ChatGPT was queried without web search (appearing in under 25% of offline answers). Despite winning the RAG retrieval layer, those low-recall brands achieved a median visibility of only 11% when web search was turned on.

At the same time, the study confirms that web grounding is now the default behavior for anonymous users. Between the week of August 3 and the week of September 28, the percentage of logged-out ChatGPT answers containing external citations climbed from 56.6% to 88.5%, surpassing logged-out Google Gemini (which moved from 79.0% to 75.5%). The challenge for marketers is no longer whether ChatGPT searches the web, but how to influence the parametric priors that govern how it searches.

## Rebuilding AI Search Strategy Around Parametric and Retrieval Alignment

Because ChatGPT's retrieval pipeline blends pre-trained entity associations with live web grounding, SEO and content engineering teams must stop treating third-party listicle placements as a standalone silver bullet. Sustainable AI search share requires dual-layer optimization targeting both offline training corpora and live RAG extraction.

### The 4-Step Enterprise Implementation Playbook

1. **Benchmark No-Search Parametric Recall Before Buying Listicle Placements**: Before investing budget in affiliate listicle sponsorships or review roundups, query target commercial prompts against API checkpoints with web search disabled (`temperature=0.7`, 20 runs per prompt). If your brand appears in fewer than 25% of ungrounded responses, third-party listicle inclusions alone will likely yield ~11% visibility even when cited.
2. **Engineer Co-Occurrence With Category Anchors Across Pre-Training Corpora**: To enter the 59.6% of fan-out queries that inject brand names, your brand entity must co-occur repeatedly alongside established category leaders in high-weight training datasets (technical documentation, GitHub repositories, Wikipedia references, arxiv papers, original industry benchmark reports, and syndicated trade press). Publish head-to-head architectural comparisons and open datasets that force third-party writers to link your brand directly beside category incumbents.
3. **Publish Dedicated Comparison Nodes Matching Fan-Out Syntax**: Since nearly 60% of ChatGPT fan-out queries contain explicit brand names, structure your own domain's comparison and integration pages around the exact multi-brand query strings the model generates (e.g., `[Incumbent Brand] vs [Your Brand] API throughput benchmark`). When the model fires a branded fan-out query for an incumbent, a tightly scoped comparison page allows your brand to ride the incumbent's retrieval query into the context window.
4. **Optimize Machine-Readable Fact Density for the 88.5% Grounded Layer**: Once parametric recall crosses the 25% threshold, live web retrieval becomes the primary differentiator. Format product specifications, pricing tiers, and SLA metrics inside clean Markdown tables, JSON-LD schema, and `/llms-full.txt` bundles so the synthesis model can extract verifiable claims without ambiguity.

## Next Steps

- **Split AI Visibility Reporting Into Offline vs. Online Shares**: Update internal AEO dashboards to measure "Ungrounded Recall Rate" (API calls without web search) separately from "Grounded Citation Visibility" (ChatGPT Search / Perplexity responses) across your top 50 non-branded commercial prompts.
- **Audit Cited-But-Ignored URLs**: Identify third-party comparison articles where your URL or brand appears on a page cited by ChatGPT, yet your brand is omitted from the generated prose; prioritize co-occurrence PR and technical documentation links in those specific topic clusters to lift base-model association.
- **Capture Logged-Out Referral Traffic**: With logged-out ChatGPT citation rates now at 88.5%, ensure all comparison pages, benchmark studies, and documentation hubs track `utm_source=chatgpt.com` and server-side referrer headers to attribute assisted pipeline accurately.