ChatGPT responses referenced approximately 20% fewer websites following the early-March transition to GPT-5.3 Instant, according to a report from French SEO consultancy Resoneo.
The analysis utilized data from Meteoria, an AI visibility-tracking platform that examined 400 prompts daily over 14 weeks, yielding 27,000 comparable responses. Following the transition, average unique domains per response declined from 19 to 15, while average unique URLs per response fell from 24 to 19.
The URLs-per-domain ratio remained constant at 1 throughout the tracking period, indicating that ChatGPT is visiting fewer sites per response but exploring each one more deeply. Resoneo’s findings also revealed a decrease in the number of domains sharing the same citation surface, suggesting that sites that are cited take up a larger portion of each answer.
Jérôme Salomon of Oncrawl conducted an independent log analysis that corroborated Resoneo’s findings. His data indicated that ChatGPT-User bot activity has settled at a lower crawl volume, highlighting that some pages are no longer being crawled and that the crawl frequency for remaining pages has also decreased.
Resoneo attributed this change to GPT-5.3 Instant being more responsible for ChatGPT’s default experience, which has triggered fewer web searches and citations compared to earlier versions. Salomon’s server log data demonstrated the same lower crawl pattern observed by Resoneo.
Prior analyses have indicated differences in how AI platforms cite sources compared to traditional search engines. An analysis by SE Ranking of 129,000 domains found that referring domains were the strongest predictor of ChatGPT citations, with a significant threshold effect at 32,000 referring domains. Additionally, a report by Search Atlas revealed low overlap between Google rankings and ChatGPT citations, with median domain overlap between 10-15%.
The decline in cited domains per response suggests fewer websites competing for visibility within each ChatGPT answer. Despite the overall reduction in citation surface area, the sites that remain continue to maintain the same crawl depth.
For stakeholders monitoring referral traffic from ChatGPT, the early-March model transition is recommended as a critical date range for analysis.
Looking ahead, Resoneo noted that the upcoming GPT-5.4 Thinking model may reintroduce search fan-outs and utilize site operators to target trusted domains, although these behaviors were not included in the dataset covering GPT-5.3 Instant. The future of the citation surface—whether it will narrow or widen with new models—remains uncertain.
