SISTRIX: ChatGPT replaces 74% of its cited sources every week

The largest public measurement of citation drift re-sampled 82,619 prompts weekly for 17 weeks, from December 17, 2025 to April 8, 2026, producing 1,548,213 snapshots across six countries and three surfaces — Google AI Mode, Google AI Overviews and ChatGPT Search. It was published by Johannes Beus at SISTRIX on May 1, 2026.

The headline: Google replaces 56% of the sources in its AI answers every week. ChatGPT replaces as much as 74%.

The structure underneath is the useful part. For 86% of prompts there is a stable core of a few domains, and everything outside that core rotates at 89% per week. So the answer to “is AI citation stable” is that it is two different things at once — a small persistent set, and a churning remainder that is most of the visible surface.

That resolves what looks like a contradiction in the same study. On AI Overviews specifically, 53% of prompts saw not a single source change across all 17 weeks. Aggregate weekly churn and per-prompt permanence are both correct; they describe the tail and the core.

Two findings that should change how visibility is reported.

AI Overviews and AI Mode cite different domains 83% of the time. Both are Google, on the same query, in the same week. A citation-share number measured on one is not a reading of the other, and a tool that reports “Google” as a single surface is averaging two populations that agree less than a fifth of the time.

Only 1.4% of news articles cited as sources remained permanently in the citation set. For a publisher, an AI citation is not an asset that accrues. It is closer to a placement that expires — and the 17-week window means the study cannot tell you whether the other 98.6% left in week two or week sixteen.

Brands do better than articles. For 43% of brand queries the brand’s own domain was present in all 17 weeks — while the co-citations alongside it rotated at 70% per week. The brand holds; the company it keeps in the answer does not.

What this record independently observed, before it had the word for it. Three separate items on this wire measured drift without naming it. Kevin Indig’s Search Signals analysis found AI Mode’s publisher mentions falling nearly by half in thirty days while retailer mentions tripled, across 15 of 20 verticals. Promptwatch traced Reddit’s 86.4% four-day collapse in ChatGPT citations to a change in how the engine assembles its background queries. And Profound shipped a feature called Citation Decay on August 13 that tracks week-over-week counts per URL — a vendor building the instrument this study argues is necessary.

Those are three different mechanisms — the engine’s preference changing, its retrieval method changing, and a tool built to watch either. SISTRIX supplies the base rate they were all sampling from.

The caveats are the usual ones and they matter here. SISTRIX sells visibility tooling and published this itself; nobody has replicated it. The window closed on April 8, before the model changes this wire has covered since, and the glossary’s own standing caveat is that citation share moves when an engine swaps the model underneath. A drift rate measured on one set of models is not a constant.

What survives all of that is the shape: AI visibility is a distribution to move, not a position to hold. This publication’s own glossary has said citation share is model-sensitive and probabilistic since it was written. This is the first time the record can say by how much.