GPT-5.5 Alters ChatGPT Citations

A new analysis from SISTRIX suggests that citation behaviour in ChatGPT responses shifted noticeably following the rollout of GPT-5.5, with changes in which sources are referenced and how frequently they appear.

The study, which focused on German-language ChatGPT outputs, found that publishers and service-based websites based in Germany saw an increase in citations, while a number of international aggregator platforms experienced declines. Interestingly, Reddit stood out as a major exception, recording the largest overall rise in citations despite the broader trend towards localisation.

SISTRIX examined 3.8 million ChatGPT responses in total, using 38 daily samples of 100,000 responses. The research compared citation patterns in the four days before and after the GPT-5.5-related change, which the company describes as a kind of “core update” for ChatGPT, similar in concept to Google’s core search updates.

However, the firm also stresses that the findings show correlation rather than direct causation, meaning the model update cannot be definitively proven as the cause of the changes in citation behaviour.

According to the data, citation patterns were relatively stable prior to the update, typically fluctuating by around 1–2% per day. Following the GPT-5.5 transition around 22–23 May, this volatility increased sharply, reaching around 47%.

At the same time, the average number of sources cited per response fell slightly, dropping from 30.9 before the change to 28.4 afterwards. This suggests not only a shift in which domains were referenced, but also a reduction in overall citation volume per response.

Looking more closely at the winners and losers, SISTRIX identified clear patterns across different types of websites. Reddit recorded the strongest growth overall, gaining over 7,000 additional citations per 10,000 responses, a rise of approximately 59%. It was already one of the most frequently cited platforms before the update, and its position appears to have strengthened further.

German-language media outlets also saw strong gains across the board. Major publishers such as Welt.de, FAZ.net and Bild.de all recorded significant increases, alongside improvements for digital platforms in sports and streaming. Sites such as DAZN, Sky and Kicker saw particularly strong percentage growth, suggesting increased visibility for locally relevant entertainment and sports content.

Specialist tools and reference platforms also benefited from the shift. Services such as Mapbox, OpenStreetMap and JustWatch all recorded notable increases in citation share, indicating stronger representation for structured, utility-based resources.

In contrast, several international aggregator and job-related platforms experienced sharp declines. Websites such as Indeed, Tripadvisor, Glassdoor, Expedia and Rome2Rio all saw reductions in citation frequency, in some cases by more than 50%. Global technology platforms were also affected, with YouTube, Wikipedia, Google, Facebook and LinkedIn all recording declines in visibility within ChatGPT responses.

Within Germany, career and comparison websites such as Kununu also saw reduced citation levels, highlighting a broader downturn for certain types of informational and listing-based platforms.

SISTRIX identified three clear categories of winners in the data: German publishers and media outlets, Reddit, and specialised tools or data-driven services. On the other hand, the main losers included international aggregators, global tech platforms, and local career or comparison sites.

The firm suggests that this pattern reflects a degree of localisation in ChatGPT’s citation behaviour, with German-language responses increasingly drawing on native publishers and locally relevant services following the GPT-5.5 change.

However, Reddit remained a clear outlier. Despite being largely English-language, it continued to gain traction, reinforcing the idea that user-generated content remains highly influential across a wide range of topics.

SISTRIX also notes that it is not possible to confirm whether this shift was an intentional outcome of the model update, only that the timing of the change aligns with a noticeable redistribution in citation patterns.

The findings add to a growing body of research suggesting that ChatGPT’s citation behaviour is not static, but changes depending on model version and configuration. Previous studies have already shown differences in how many domains are cited across different GPT versions.

While this analysis focuses specifically on German-language data, SISTRIX indicates that similar trends may also be present in other markets, although further research is still ongoing.