GPT-5.5 rewires ChatGPT citations, per SISTRIX

SISTRIX, tracking German-language ChatGPT responses, says the citation mix shifted sharply when GPT-5.5 quietly rolled out. Search Engine Journal reports the SEO data firm is comparing the disruption to a Google core update: same domains, different winners, no announcement from OpenAI.

That framing matters. Core updates in classical search are the moments when visibility gets redistributed wholesale, and entire categories of publishers learn overnight whether the algorithm still likes them. SISTRIX is arguing that model upgrades inside ChatGPT now function the same way, with one important difference: there is no Search Console, no rollout notice, and no named update for marketers to point at when their citations evaporate.

A core update nobody announced

The mechanics, as SISTRIX describes them, are straightforward. After GPT-5.5 began serving answers, the set of domains ChatGPT pulls from to ground its responses moved. Some sources that had been reliable citation winners lost share; others gained. The composition of the answer changed even when the question did not. That is the textbook definition of a ranking shift, dressed in different clothes.

What it is not, on the available evidence, is a deliberate editorial choice by OpenAI to favour any particular type of publisher. Model upgrades adjust retrieval behaviour, embedding quality, and the way the system weights candidate sources. The citation reshuffle is a downstream effect of those mechanics, not a published policy. Which makes it harder to game and harder to appeal.

One caveat worth keeping in view: the SISTRIX dataset is German-language. Whether the same redistribution holds in English, French, or Arabic responses is unverified. Citation behaviour in ChatGPT has varied by language before, particularly where the underlying corpus is thinner. Treat the German numbers as a leading indicator, not a global verdict.

What changes for brands tracking AI visibility

For communications teams at banks, multilaterals and industrial groups that have started measuring share of voice inside ChatGPT, the practical lesson is uncomfortable: your baseline can move without warning, and the vendor will not tell you when. A drop in citations next month is as likely to reflect a silent model swap as anything you did or failed to do.

That argues for three adjustments. First, measure continuously rather than quarterly. Monthly snapshots will conflate model changes with content changes and leave you debugging the wrong variable. Second, track citation share against a stable basket of competitor domains, not in absolute terms; relative position survives model churn better than raw counts. Third, keep a dated log of OpenAI model versions alongside your visibility data, so that when a step-change appears you can attribute it correctly.

For multilateral institutions and policy bodies, the stakes are particular. These organisations rely on ChatGPT and its peers to surface their research, statistics and definitions when a user asks a substantive question. If a model update demotes the World Bank's data portal in favour of a secondary aggregator, the institution's authority in the answer layer thins, even as its underlying authority on the web is unchanged. The citation is the visibility. Losing it is not a ranking inconvenience; it is a quiet erosion of who gets quoted as the source of record.

Financial services firms face a related but distinct problem. Compliance teams have spent two years pressing communications to monitor what LLMs say about their products. A core-update-style reshuffle means that monitoring has to be version-aware. "ChatGPT cited us in October" is not a durable claim if GPT-5.5 has changed the rules in November.

The wider point

SISTRIX's analogy to a core update is the most useful framing yet for what model upgrades do to brand visibility. Google trained the industry to expect periodic, named, semi-explained disruptions to organic rankings. OpenAI, Anthropic and Google's own Gemini team are now delivering the same disruptions with none of the scaffolding: no update names, no guidance, no rollback windows. The asymmetry favours the platforms.

The brands that will hold citation share through the next several model versions are the ones treating AI visibility as a measured, instrumented discipline rather than a campaign output. Everyone else will discover their position the way SISTRIX's German publishers did: by checking the data after the fact and finding the floor had moved.