Generative search engines such as ChatGPT, Gemini and Perplexity now decide which Web pages users see as sources, and a growing industry of generative engine optimization advises publishers on how to earn these citations. We study what shapes citation in two 90-day collections of answers to commercial questions in contrasting markets: 158,099 answers with 468,037 citations. Comparing pages of the same website, content matters in one way: pages whose title and paragraphs match the questions people ask are cited more, whereas FAQ blocks, structured data, summary boxes and page speed appear effective only across websites. Page properties nevertheless explain little of the variation between pages of a website. Replaying the questions through the engines' APIs shows why. Every page ChatGPT cites is one its own search retrieved, and which pages that search retrieves, for sub-queries the model rewrites on every answer and that rarely coincide with Google's or Bing's results, explains far more of the variation than page properties (32% against 2.8%). Content acts mainly through this step, by matching the engine's queries in meaning. The model's own knowledge shapes it too: ChatGPT writes brands it knows into its searches, prefers websites it knows, and names brands without any supporting page, and controlled experiments show that brand names reorder its rankings of products with identical facts. Pages are thus cited when searched; brands are also named when remembered. The findings replicate in a market of small B2B vendors. We release data and code.
Figures & tables
ChatGPT
Gemini
Perplexity
Answers
58,590
26,069
58,590
Citations a
116,530
263,300
16,358
Pages per answer
2.0
10.1
0.3
Answers without a citation
30.8%
15.6%
95.8%
URLs / websites
15,833 / 2,190
23,640 / 5,452
3,498 / 1,067
Citations to brand websites
64.3%
16.8%
6.8%
Table 1. The main dataset per engine (24 June – 21 September 2026, 651 queries). a A URL counts once per answer.
Feature
All pages
Brand
Third party
Best paragraph matches
+15.5 (5)
+5.2 (2)
+25.7 (5)
Title matches a query
+14.1 (5)
+7.4 (4)
+23.6 (5)
Word overlap with queries
+9.1 (4)
+11.9 (3)
+12.8 (5)
First paragraph matches
+5.9 (4)
+7.8 (4)
+3.1 (3)
Page age
−5.1 (4)
−10.6 (3)
−4.2 (3)
Word count
+0.9 (1)
+2.9 (1)
−3.0 (2)
Table 2. Effect of each page feature on citations, comparing pages of the same website (% more citations per SD; yes/no features: page has the marker vs. not). In parentheses, the number of the five tests passed; bold marks features that pass all five.
Figure 1. Effect of recommended practices on citations: across all pages with the practice alone (grey), across all pages with all other features (orange), and comparing pages of the same website (blue), with 95% intervals. Left: third-party pages; right: all pages. Dot plot of the estimated effect of FAQ blocks, structured data, summary boxes, author boxes and page speed under three comparisons; the effects shrink to near zero within websites.
Pages consulted / cited per question
26 / 3
Cited at rank 1 / 2 / 6–10 / 11–20 (%)
43.8 / 35.5 / 12.2 / 3.5
Choice among consulted pages explained: rank / + content
26.8% / 32.5%
Same question, same day: query / cited-page overlap
2.7% / 15%
Retrieved pages absent from Google top 50 / Bing top 30
97% / 99%
Within-website variation explained: retrieval
32% [24, 42]
page features / shuffled retrieval
2.8% / 0.05%
Table 3. What the replay shows (ChatGPT; 95% CI in brackets).
ChatGPT
Gemini
Same source first − last, cited (points)
17.6 [15.1, 20.3]
− 0.4 [ − 2.3, 1.6]
Brand names shown: change vs repeat
2.1 × [1.8, 2.6]
3.3 × [2.5, 4.5]
Publisher names shown: change vs repeat
1.3 × [1.1, 1.6]
1.6 × [1.3, 2.0]
Random names: spread in top-3 rate (points)
41
18
Held-out pairs, collected answers: score per brand and question / per brand
74 / 64%
77 / 67%
Full rankings predicted: one score per brand / other order
82 / 81%
77 / 79%
Table 4. Controlled experiments on the answer step (models without search unless stated; 95% CI in brackets) and brand order in the collected answers.
Consumer
Niche B2B
Title matches a query, all / brand pages
+14.1 (5) / +7.4 (4)
+26.4 (5) / +33.6 (5)
Best paragraph matches, all pages
+15.5 (5)
+12.9 (5)
FAQ block: across / within websites
+18.1 / −3.4
+22.5 / −1.9
Variation between websites
13%
9%
Website authority / page links (per SD)
+12.1 / +22.7%
−1.3 / +16.3%
Rank 1 / ranks 11–20 cited
43.8 / 3.5%
57.1 / 8.3%
Table 5. Core results in the two markets. Page effects are within websites (% per SD; yes/no markers: has vs. not; in parentheses, tests passed). Replay results for ChatGPT.
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Consumer (651)
Niche (55)
Customer Support
100
AEC Industry Focus
10
Equipment
100
AI Capabilities
10
Competitive Landscape
100
Knowledge Management
10
Speed
100
Proposal Automation
10
Pricing
100
Workflow Integration
10
Performance
98
Competitive Landscape
5
Appendix
Table 6. Questions per topic and intent.
Consumer market
Niche market
Feature
Effect
1
2
3
4
5
Effect
Tests
Title match
+14.1
∙
∙
∙
∙
∙
+26.4
5
Best-paragraph match
+15.5
∙
∙
∙
∙
∙
+12.9
5
Word overlap
+9.1
∙
∘
∙
∙
∙
+13.0
4
First-paragraph match
+5.9
∙
∘
∙
∙
∙
+0.1
0
Length (words)
+0.9
∘
∘
∘
∘
∙
+2.2
1
Appendix
Table 7. The five tests on all pages: 1 FDR significance, 2 stability selection, 3 double ML, 4 shape, 5 sign in ≥ 4 of 5 variants ( ∙ passed, ∘ failed). Effects in % more citations per SD ( † has vs. not). Bold: passes all five.
Feature
ChatGPT
Gemini
Perplexity
Title match
−0.2
+16.8 ∗
+4.6 ∗
Best-paragraph match
+9.5 ∗
+7.8 ∗
+4.8 ∗
Word overlap
+2.4
+5.3 ∗
+7.1 ∗
First-paragraph match
+0.7
+8.6 ∗
+0.2
FAQ block †
−1.7
−0.3
−2.7
Schema.org markup †
−0.0
+1.5
−0.5
Appendix
Table 8. Within-website effects per engine, main dataset (% per SD; † has vs. not; ∗ p<0.05 ).
AI answer engines generate answers from retrieved pages but cite only a few sources. This makes visibility depend not just on ranking, but on being cited. We study competitive Generative Engine Optimization (GEO): when two retrieved candidates compete, what makes one more likely to be cited first? We build a controlled two-document retrieval-augmented generation (RAG) testbed that injects exactly two candidate sources into the model context and measures which source is referenced by the first citation marker in the output. Across six LLMs we execute 252,000 trials, repeated paired comparisons under one factorial program over 18 content factors. In each trial the two sources differ in exactly one factor; we use brand anonymization and counterbalanced source order to separate content effects from position bias. Mixed-effects models show that topical relevance and list position are the biggest drivers of being cited first. Including explicit price information and a recent timestamp also helps consistently. Completeness and trust cues add smaller gains, while formatting-only edits have little impact. We release a reproducible evaluation protocol and a prioritized GEO checklist for practitioners, and we exercised it in an early internal pilot at Sprinklr, where teams reported positive qualitative feedback on workflow usability.
Production large language models retrieve and cite web pages alongside generated answers, yet the page-level features that predict citation frequency remain poorly characterised. We present an observational study of approximately 2 million LLM citations from four commercial engines (ChatGPT, Claude, Google AI, Gemini) over six months, joined to 10,000 crawled pages from nineteen B2B SaaS workspaces. Sixty-plus features are tested using a nine-method consensus framework combining mixed-effects regression with domain fixed effects, FDR correction, stability-selection Lasso, double machine learning, generalised additive models, and temporal hold-out replication. Four findings survive all checks. First, prompt-content alignment (Jaccard overlap between page tokens and the full workspace prompt corpus, including non-citing prompts) is the dominant page-level predictor (beta = +0.37, 95% CI [+0.33, +0.41], q ~ 10^-73). Second, the standard AEO checklist (FAQ blocks, structured data, Core Web Vitals) shows positive effects in pooled data that reverse or collapse to zero once domain fixed effects are applied: Simpson's paradox with practical consequences for the AEO literature. Third, domain-level AI authority exceeds the strongest non-alignment page-level feature by a factor of six in mean absolute SHAP value. We release the analytic pipeline as a methodological contribution.
Generative search engines answer local and commercial questions with a written answer and a few cited sources. A growing practice of generative engine optimization (GEO) monitors which sources and brands these answers show, typically from a single place. In this paper, we present a controlled audit of how the user's location changes the sources that generative engines cite. We ask 300 questions of ChatGPT with 19 declared locations, and of ChatGPT and Gemini from five network locations, each twice to measure run-to-run noise (31,200 answers). The engines take location from different channels. ChatGPT follows the location an application declares, down to the city: it writes the place into its own searches, and its search backend localizes results even when the queries name no place. It ignores the network. Gemini ignores the declared location and follows the country of the network, changing most of its sources even for questions that do not depend on place. The businesses the answers recommend change with the sources. Thus, visibility in generative search is specific to the user's location, and audits must control both channels.