+ Research

Inside ChatGPT's Search-to-Citation Funnel: Where Pages Lose Visibility

How ChatGPT moves from search to citation. 548,534 pages retrieved across 15,000 prompts analyzed to separate the retrieval decision from the citation decision, with rank-controlled six-model decomposition and DA × rank tier cross-stratification.

Inside ChatGPT's Search-to-Citation Funnel: Where Pages Lose Visibility: hero figure

Getting a page cited in ChatGPT involves three separate decisions: which pages Google ranks for the query, which pages ChatGPT retrieves through its internal fan-out, and which retrieved pages ChatGPT chooses to cite in the answer. Each of those decisions is governed by different signals, and treating them as one obscures where a page is actually losing visibility.

This study separates all three across 15,000 prompts and 548,534 retrieved pages, capturing the behavior end users see (ChatGPT’s real consumer interface, not the developer-facing endpoints that surface a different retrieval pipeline). Every signal is tested under enough statistical controls to rule out the most common spurious findings.

Research question

How does ChatGPT move from search to citation? Specifically, what share of retrieved pages get cited, how much does fan-out contribute, and which on-page and authority signals predict citation when retrieval is held constant?

Dataset

15,000 prompts × 3 runs against GPT-5.1 (ChatGPT, consumer interface).

DimensionValue
Unique prompts15,000
Query types8
Runs per prompt3
ChatGPT response captures (15,000 × 3)45,000
Retrieved pages analyzed548,534
Authority providersthird-party DA and backlink APIs

Two analytical populations were maintained separately:

PopulationDefinitionUsed for
Population 1Every URL in Google’s top-20 for any original or fan-out querySERP-rank vs. citation analyses
Population 2Only pages ChatGPT actually retrieved (cited or not)On-page signal analyses, logistic regression, random forest

Mixing the two populations without flagging it is treated as a methodological error.

Methodology

15,000 prompts × 3 runs against ChatGPT’s real consumer interface (not the API), yielding 45,000 response captures. For every response: cited URLs, seen-but-not-cited URLs, and fan-out sub-queries were captured. Google top-20 SERPs were collected for the original query and every fan-out sub-query.

Each retrieved page (cited or not) was analyzed for 24 content quality signals across structure, readability, formatting, schema, and evidence-density families. Domain authority and backlink data were joined from third-party authority providers.

Statistical analysis: logistic regression, random forest, VIF screening for collinearity, Cohen’s d, Pearson / Spearman / point-biserial correlations, Mann-Whitney U, DA and backlink bucket stratification, and a rank-controlled six-model decomposition (rank-only, authority-only, content-only, rank+authority, content+authority, all combined).

A finding was reported only after surviving multiple-method agreement. Effects that appeared in logistic regression but disappeared under DA stratification or VIF correction were excluded.

The retrieval-to-citation funnel

Across the full 15,000-prompt dataset:

MetricValue
Pages retrieved per query36.6
Pages cited per query5.5
Overall retrieval-to-citation ratio6.7:1
Share of retrieved pages never cited85%

Eighty-five percent of pages that enter ChatGPT’s retrieval pool are dropped at the citation gate. The 6.7:1 ratio is the headline funnel statistic, and it is consistent across query types within a narrow band.

DA 80 is the empirical inflection

Below domain authority 80, on-page signals are near-random predictors of citation. Above DA 80, content signals begin to matter.

DA bucketContent-only AUC
0 to 200.506
20 to 400.521
40 to 600.524
60 to 800.526
80 to 1000.607

The legacy “DA 40 threshold” that appears in older SEO literature is not supported by the bucket-level data. The signal floor is essentially flat from DA 0 through DA 80, then jumps sharply at DA 80.

Beyond rank 20, content signals matter more

Inside the top-20 Google ranks, rank dominates. Beyond rank 20, the picture inverts.

Rank tierContent-only AUCNotes
Top-20 ranks0.5295small content lift over rank-only
Beyond rank 200.5542larger content lift (1.8× the top-20 content lift over chance)
DA 80 to 100 beyond rank 200.621highest content-only AUC in the dataset

The interaction between DA and rank tier surfaces the most actionable finding in the study: high-DA sites ranking beyond position 20 are where content quality has the largest marginal effect on citation.

Rank-controlled six-model decomposition

Treating “rank only” as the baseline understates content’s contribution because most rank-only models also implicitly include authority. The decomposition:

ModelAUC
Rank only (true)0.6078
Authority only0.5303
Content only0.5394
Rank + authority0.6319
Content + authority0.5525
All combined0.6346

The historical “rank only” figure (0.6319) was actually rank + authority. The corrected rank-only number is 0.6078. The all-combined model adds only 0.0027 over rank + authority, which is the universal content lift inside the top-20 ranks.

Signals that survived multi-method triangulation

Universal positive signal:

  • Heading count. Positive in all 8 query types. 12× stronger for DA 80-100 vs. DA 0-20. Coefficient doubles beyond rank 20.

Universal negative signal:

  • Domain backlinks. Universally negative across all query types. Cited pages come from domains with fewer backlinks. Niche relevance beats bulk authority.

Strong stage-specific signals:

  • Title-query word overlap. Pages with 50%+ overlap earn a 2.2× citation lift over pages with under 10% overlap.
  • Heading-query word overlap. Positive in all 8 query types, scaling with overlap percentage.

Fan-out behavior

ChatGPT generates fan-out sub-queries internally during browsing. The fan-out layer drives a third of citations on average.

Fan-out patternCitation rate
Near-verbatim (fan-out closely matches original query)12.62% (best)
Added-year (fan-out adds a year qualifier)10.34% (worst)

84.8% of queries generate exactly two fan-out sub-queries. The “sweet spot” is two-fan-out structure with near-verbatim relationship to the original query.

Commercial fan-out is structurally different

Commercial queries use 1.76× more split-queries (fan-outs that fragment the original query into independent sub-queries) than informational queries. But commercial queries cite 2 to 3 percentage points lower per fan-out pattern.

The pattern signals a citation selection problem rather than a discoverability problem. Commercial pages are surfacing in the retrieval pool at expected rates but losing the citation gate more often.

Discussion

Three patterns are worth surfacing.

Rank is necessary, content is amplifying. Inside the top-20 Google ranks, rank dominates. Beyond rank 20, content quality is where the marginal lift sits. The “high DA + beyond rank 20” cell (AUC 0.621) is where the legacy SEO playbook breaks down most cleanly.

The DA 40 threshold was an artifact. Below DA 80, content signals are near-random. The empirical inflection sits at DA 80, not DA 40. Authority floors that older SEO programs treated as threshold checkpoints do not survive bucket-level analysis.

Citation selection differs from discoverability. Commercial queries surface in the retrieval pool at expected rates but cite at lower rates per fan-out pattern. The bottleneck for commercial brand visibility is not retrieval; it is the citation gate downstream.

Limitations

  • ChatGPT-specific. Generalization to Claude, Perplexity, or other AI-search engines is not claimed.
  • 24 on-page signals only. Off-site signal layer (link graph, third-party brand mentions, community presence) is outside the scope of this study and covered separately by Borrowed Visibility: How AI Engines Discover Brands Through Third-Party Content.
  • Authority data quality. DA and backlink data come from third-party providers. Provider-specific noise propagates through DA-stratified analyses.
  • Effect sizes are associations. Causal claims would require an intervention study.

This research was first published as an AirOps report.