SISTRIX: ChatGPT’s Move to GPT-5.5 Redistributed 47% of Citations in 48 Hours

News, June 16, 2026. ChatGPT's citation pattern moved fast. SISTRIX's Johannes Beus, in analysis published June 1 and relayed by Search Engine Journal on June 2, reported that 47% of ChatGPT citations went to different domains within 48 hours around the May 23 switch from the "GPT-5 mini" identifier to "GPT-5.5" — against a normal daily variance of just 1–2%. SISTRIX named the event "the ChatGPT Core Update," a deliberate nod to Google core updates.

The reshuffle was measured on 800,000 responses (comparing May 18–21 against May 26–29), part of a larger corpus of 3.8 million German-language responses and 100M+ mentions. Sources per response slipped from 30.9 to 28.4. This is the raw event; the conceptual frame — why a model update behaves like a search core update, and what to do about it — is in our pillar on AI citation core updates.

What SISTRIX Measured

SISTRIX runs a standing panel of prompts against ChatGPT and logs which domains the engine cites. On a typical day, the set of cited domains barely moves — Beus puts the baseline churn at 1–2%. Around May 23, that churn jumped to 47%: nearly half of the citations landed on different domains than they had days earlier. The same window coincided with the model identifier flipping from "GPT-5 mini" to "GPT-5.5," and with average sources per response dropping from 30.9 to 28.4.

The measurement rests on 800,000 responses compared across two four-day windows, inside a corpus of 3.8 million German-language responses. That is a large sample, but a German-only one — a limit SISTRIX states plainly, and one worth keeping in view before mapping the exact percentages onto an English-language site.

Winners and Losers

The redistribution was not random. It favored community and editorial sources and punished aggregators.

The shape is consistent: original or community-sourced content gained, while pages that mostly repackage other sites' inventory lost. That tilt is the part of the finding most likely to travel beyond the German dataset.

Correlation, Not Causation

SISTRIX is careful here, and so are we. The 47% reshuffle coincided with the GPT-5.5 identifier appearing, but coincidence is not proof. From outside OpenAI, there is no way to tell whether the model swap alone drove the change, or whether a parallel update to the retrieval layer or system prompt ran at the same time. Beus says this explicitly: the data shows a strong correlation, not a confirmed cause.

That distinction matters for how you react. Treat the model update as the most likely trigger, not as a settled mechanism — and resist the urge to rewrite your strategy around a single 48-hour event in one language market.

What It Means for You

The headline isn't a particular winner or loser — it's the volatility itself. A single model update redistributed nearly half of ChatGPT's citations in two days. Citation position in an AI engine is not a standing you earn once; it is recomputed every time the underlying model or retrieval stack shifts. Treating it as permanent is the mistake.

The practical move is to monitor citations continuously rather than spot- check them, and to build the source qualities that survived this reshuffle: original data, editorial depth, and a clean, extractable page — the opposite of thin aggregation. For the full framework on how AI citation core updates work and how to position for the next one, read AI citation core updates.

FAQ