What the paper studied
Generative search engines such as ChatGPT, Gemini and Perplexity now decide which web pages users see as sources, and a growing generative engine optimization (GEO) industry advises publishers on how to earn those citations. This paper tests what actually drives citation.
The authors collected answers daily for 90 days (24 June to 21 September 2026) in two contrasting markets: 651 consumer questions about a US fiber internet provider and 55 questions about proposal software for architecture, engineering and construction firms. In total, 158,099 answers and 468,037 citations. Pages were compared only with other pages of the same website, so a site's overall authority could not masquerade as a page-level effect. The questions were then replayed through the engines' APIs to expose the searches each model ran, the pages it retrieved and the pages it cited, and controlled experiments tested how brand names and source order change answers.
Key findings
- Matching the question is the only content property that holds up. Within a website, pages whose best paragraph matches the question earned 15.5% more citations per standard deviation, and title match added 14.1%.
- Popular GEO tactics only look effective across websites. FAQ blocks (+18.1%), summary boxes (+20.9%), author information (+14.9%), schema.org markup (+7.5%) and faster page speed (+11.7%) appear helpful when comparing different sites, but none is significant within a website. The authors attribute this to those practices clustering on sites that are already heavily cited.
- Length, headings, lists, tables and outbound links show no effect.
- Page features explain little. About 87% of citation variation lies between pages of the same website, and page features explain only 2.8% of it for ChatGPT.
- Citation requires retrieval. Every one of the 22,315 URLs ChatGPT cited in the replay had been returned by its own search. Which pages that search retrieves explains 32% of within-website variation, against 2.8% for page features. ChatGPT consults about 26 pages per question and cites about 3, citing 43.8% of first-ranked results but only 3.5% of results ranked 11 to 20.
- Engine search is not Google or Bing search. ChatGPT rewrites about four keyword-style sub-queries per answer, and they overlap only 2.7% between runs on the same day. For those sub-queries, 97% of the pages ChatGPT retrieved were absent from Google's top 50 and 99% from Bing's top 30. Even Gemini, which uses Google Search grounding, cited pages that were 93% absent from Google's top 50.
- The model's own memory matters. ChatGPT writes a brand into its searches 67% of the time when it already knows the brand, against 11% otherwise. 46% of its brand mentions are not supported by any cited page that names the brand. In controlled tests, revealing brand names reordered rankings of products with identical facts.
- Answers drift quickly. Two weeks later, only 40% of ChatGPT's cited set was still shared; after ten weeks, 14%.
- The pattern replicates for small B2B vendors, where a cited third-party page naming a vendor raised the odds that ChatGPT named it by a factor of 39.7.
Why it matters for hospitality
The study does not use hotel data, but the mechanics apply directly to how travelers now shortlist hotels in AI assistants. Visibility depends far more on whether a property's pages are retrieved by the engine's own rewritten searches, and on whether the model already knows the brand, than on technical page features. A strong Google ranking is not a reliable proxy for being cited, and a hotel the model has never heard of is less likely to be searched for by name. Third-party coverage that names a property can matter a great deal, especially for smaller and independent brands.
Practical takeaways
- Write pages around the questions guests actually ask, with titles and paragraphs that answer them directly.
- Do not rely on FAQ markup, schema packages, summary boxes or page speed alone to win AI citations; within a website, the study found no association with more citations.
- Do not treat classic search rank as a measure of AI visibility. Track citations per question and over time, since cited sets shift within weeks.
- Invest in being named on credible third-party pages and in building brand recognition, since models favor brands they already know.
Limitations
The results are associations, not causal effects. Both datasets cover English, US-market commercial questions from organizations using the authors' monitoring platform, and neither is a travel market. The replay uses APIs, which may differ from consumer apps, and Perplexity could not be replayed. About a third of cited URLs could not be analyzed. The controlled experiments use smaller models and constructed inputs, and because engines change continuously, the findings describe 2026 behavior.