Resoneo claims that ChatGPT responses began listing about 20% fewer websites after what it identifies as the early March transition to GPT-5.3 Instant.
THE analysis comes from French consultancy SEO and draws on data from Meteoria, an AI visibility tracking platform that monitored 400 prompts daily for 14 weeks, producing 27,000 comparable responses.
The average number of unique domains per response increased from 19 before the transition to 15 after. The average number of unique URLs per response decreased from 24 to 19.
The URL per domain ratio remained at 1 throughout the monitoring period. The data suggests that ChatGPT doesn’t visit as many sites per response, but digs deeper into each one.
Fewer domains now share the same citation area in each answer, meaning cited sites take up a larger share of each answer.
Server logs back up the model
Independent analysis of Jérôme Salomon’s journals at Oncrawl confirms these results. By tracking ChatGPT-User bot activity across multiple websites, its data shows that crawl volume has stabilized at a lower level. Some pages are no longer crawled at all and the crawl frequency of pages still visited has decreased.
Resoneo ties the change to the default ChatGPT experience, now driven more by GPT-5.3 Instant, which the company says triggers fewer web searches and citations than previous behavior. Oncrawl’s server log data shows the lowest crawl pattern over the same period.
Previous data has shown how AI platforms cite sources differently than traditional research. An analysis of the SE ranking of 129,000 domains found referring domains to be the strongest predictor of the citation probability of ChatGPT, with a threshold effect at 32,000 referring domains.
A Search Atlas report also showed low overlap between Google rankings and ChatGPT citationswith a median domain overlap of approximately 10-15%.
Why it matters
A 20% drop in cited domains per response means fewer websites competing for visibility in each ChatGPT response. The total area of citations decreased, but the sites that continued to appear maintained the same depth of exploration.
For anyone tracking ChatGPT referral traffic, the model transition in early March is a date range worth checking in your analytics.
Looking to the future
Resoneo’s analysis notes that GPT-5.4 Thinking reintroduces search deployments and uses site operators: to target trusted domains, but these behaviors were not captured in the quantitative data set, which covers GPT-5.3 Instant and below.
It is not yet clear whether the citation surface continues to shrink or widens again with newer models.
