The practical brief
The finding
In the main ChatGPT analysis, retrieval measures explained 32% of within-website citation variation, compared with 2.8% for page features. [c2] [c3] [c4] [c5] [c8]
Why it matters
A page-formatting project may leave the larger discovery bottleneck untouched, while brand mentions can occur without a supporting citation.
| Market and analytical sample | Retrieval measures | Page features |
|---|---|---|
| US home internet: 6,178 pages on 362 websites | 32% of within-website variation in ChatGPT’s 90-day citation counts | 2.8% of within-website variation in ChatGPT’s 90-day citation counts [c4] |
| Niche B2B software: 629 replay pages | 30% of within-website variation in ChatGPT’s 90-day citation counts | 10.0% of within-website variation in ChatGPT’s 90-day citation counts [c7] |
What to try
BLURSOR’s practical interpretation
Our interpretation: evaluate question relevance, retrieval and brand mentions separately before expanding an AI-visibility budget.
Study boundary
These are mostly observational API findings in two commercial markets, not proof that particular edits increase citations or sales.
Before buying a formatting overhaul, identify the bottleneck
If you are paying to make your website appear in AI answers, do not treat FAQ blocks or structured data as demonstrated citation-growth tactics. This study found stronger evidence for relevance to questions and for retrieval—the engine finding a page before composing its answer—than for common formatting features.
The researchers collected 158,099 answers containing 468,037 citations across two 90-day datasets: US home internet and niche business software. Their author-operated platform queried search-enabled APIs daily, using default sampling and no system prompt. The tested models were GPT-5.4 nano, Gemini 3.1 Flash Lite and Perplexity Sonar, not necessarily the products customers encounter in consumer apps.
They compared pages within the same website to reduce the risk of crediting a page feature for advantages belonging to the whole site. They then replayed questions after collection to observe retrieval and ran controlled experiments on how models selected products.
Question relevance held up; formatting advantages did not
Across 12,038 analyzed pages on 1,120 websites, a one-standard-deviation improvement—a standardized step in the study’s relevance score—in best-paragraph match was associated with 15.5% more citations. Title match was associated with 14.1% more. These are comparisons between pages, not measured gains from rewriting them; title match was not positively associated with ChatGPT citations separately.
FAQ blocks, summary boxes, author information, schema.org markup and faster loading looked beneficial when comparing different websites. None had a statistically significant association when comparing pages within the same website. That does not mean these features are useless: it means their apparent citation benefits could reflect other website advantages.
For ChatGPT, replay retrieval measures explained 32% of within-website variation in citation counts, versus 2.8% for page features. Every cited ChatGPT URL in the replay had first been retrieved. The replay occurred later, so this is explanatory association rather than an intervention proving causality.
A realistic interpretation: measure discovery and mentions separately
A citation links a page; a brand mention merely names a business. In ChatGPT, 46% of mentions of brands absent from the question lacked a cited page naming that brand. This does not establish that those mentions were false, but it makes citation counts an incomplete visibility measure.
In constructed experiments, revealing brand names changed product rankings despite unchanged facts. That establishes an identity effect in those inputs—not superior product quality, recommendation accuracy or a proven way to influence model memory.
Our interpretation is to use the findings as a diagnostic framework, not an optimization recipe:
- Review whether important pages directly address relevant customer questions; test edits rather than budgeting around the reported associations.
- Track linked pages and unlinked brand mentions separately, recording which question and engine produced each.
- When API retrieval logs are available, distinguish pages not found from pages found but not cited. Do not assume API results reproduce consumer-app exposure.
- Keep useful FAQs, authorship and structured data for their other purposes; require separate evidence for promised citation gains.
What this study cannot tell your business
The smaller software dataset pointed in the same direction, but its retrieval-versus-page-feature difference was not statistically significant. It cannot establish that retrieval dominates by the same margin in every market.
The study does not measure sales, lead quality, product quality or answer accuracy. Unmeasured differences between pages remain, roughly a third of cited URLs could not be analyzed, and engines change continuously. Treat these results as evidence about particular systems and questions in 2026—not a forecast of returns from your next content project.
The source and evidence
Study limitations and disclosures
- Both authors, Mahe Chen and Nick Koudas, list University of Toronto affiliations. They operated the monitoring platform used for collection, and the paper thanks ktau.ai for data assistance. Questions were written for organizations using the platform. These relationships are material context; the supplied paper does not specify funding, ownership interests or a financial-conflict declaration.
- The within-website comparisons remain observational and cannot eliminate unmeasured differences between pages. Pages and links were measured after citation collection, and roughly a third of cited URLs were unavailable for content analysis. The page-speed comparison covered only 466 pages.
- The results concern selected API models, English US-market questions written for platform users, and two commercial markets—not a representative sample of consumer searches. The later replay excluded Perplexity, and API behavior may differ from consumer applications.
- The smaller B2B dataset consisted mostly of learning questions, limiting conclusions about comparison and purchase intent. Its retrieval-versus-page-feature difference was not statistically significant.
- Controlled experiments used small models and constructed inputs. Familiar providers could sometimes be identified from masked facts, so masking did not fully remove identity information. Gemini declined to choose in 9% of pairwise answers in the separate stability experiment.
- The crawler respected robots.txt and rate limits but used a browser user agent without identifying itself as a research crawler. Sites disallowing automated access, including Reddit, were not fetched, restricting available page content.
Evidence behind this briefing
[c1] The authors used their own monitoring platform to query search-enabled APIs daily with default sampling and no system prompt. The tested systems were GPT-5.4 nano, Gemini 3.1 Flash Lite and Perplexity Sonar—not necessarily their consumer applications.
Section 3.1, Answer Collection · Read the study
[c2] Within the same website, stronger semantic matches between pages and the question set were associated with more citations. Across the main analytical sample of 12,038 pages on 1,120 websites, a one-standard-deviation improvement in best-paragraph match corresponded to 15.5% more citations and title match to 14.1%. These are associations, not demonstrated effects of rewriting pages; title match was not positively associated with ChatGPT citations separately.
Section 4.1, Effects Within Websites; sample denominator in Table 2; engine qualification in Section 4.1 and Appendix B, Table 8 · Read the study
[c3] FAQ blocks, summary boxes, author information, schema.org markup and faster page speed appeared beneficial across websites, but none had a statistically significant association within websites. This does not establish that these features are useless; it undermines interpreting cross-website correlations as evidence that adding them earns citations. Page-speed analysis covered only 466 pages.
Section 4.2, Recommended Practices: Across vs. Within Websites; page-speed sample in Section 3.2 · Read the study
[c4] For ChatGPT, replay retrieval measures explained substantially more within-website variation in 90-day citation counts than page features: 32% versus 2.8%, across 6,178 pages on 362 websites. The difference was 29 percentage points, with a 95% CI of 20–38. This is explanatory association from a replay conducted after collection, not a causal intervention. Every cited ChatGPT URL in the replay had been retrieved; the result does not establish the same mechanism for every engine.
Section 6.4, How Much Retrieval Explains; retrieval prerequisite and replay scope in Section 6.1 · Read the study
[c5] Brand mentions and page citations are different visibility outcomes. In ChatGPT, 46% of mentions of brands absent from the question lacked a cited page naming the brand; brands elicited in no-search probes had 5.8 times the mention odds at equal citation support. This is an association with the model's elicited prior, not proof that unsupported mentions are false or that changing training exposure would cause visibility gains.
Section 6.5, The Model’s Prior; causal qualification in Section 8, A hypothesis: the prior as an entry point to retrieval · Read the study
[c6] In a controlled preference experiment using 60 five-product packets and 1,920 answers, revealing brand names changed top-three selection rates more than repeating identical prompts: 2.1 times as much for ChatGPT and 3.3 times for Gemini. Facts were held fixed, so this supports an identity effect on rankings in these constructed inputs—not a difference in product quality or recommendation accuracy.
Section 6.6, From Sources to Answer; Table 4; Appendix D, Brand and publisher identity · Read the study
[c7] The niche B2B replication covered 55 questions and 14,850 answers. Retrieval explained 30% of ChatGPT's within-website citation variation versus 10.0% for page features, but the difference was not statistically significant in the 629-page replay sample. This supports cautious cross-market comparison, not a definitive claim that retrieval dominates by the same margin everywhere.
Section 7, Generalisation: A B2B Niche Market; Table 5 · Read the study
[c8] The observational design cannot remove unmeasured page-level confounding. Pages and links were measured after citation collection, roughly a third of cited URLs were unavailable for content analysis, replay omitted Perplexity, and API behavior may differ from consumer applications. Generalization is restricted by English US-market client-written questions, small models, constructed experiments, a small second dataset and continuously changing engines.
Section 8, Limitations · Read the study