ChatGPT replaces 74% of its cited sources every week

Last quarter you ran an AI visibility audit. Your brand showed up in 12 ChatGPT responses for the prompts you care about. You called it a baseline. You set a quarterly cadence to check again.

That number was a snapshot of a mirage. Most of the sources ChatGPT cited that week are no longer cited today.

A SISTRIX study published in April 2026 tracked 82,619 prompts across 17 weeks in six countries, generating more than 1.5 million citation snapshots. The headline finding: ChatGPT Search replaces 74% of its cited domains week over week. Google AI Mode swaps out 56%. Even Google AI Overviews, the most stable surface, churns 5% of its sources every seven days.

Pair that with Profound’s analysis of 240 million ChatGPT citations — which found that 70 to 90% of cited domains for any given query are completely different six months later — and you arrive at an uncomfortable conclusion. AI search isn’t a ranking system you measure quarterly. It’s a moving target you have to watch in real time.

The fixed core and the carousel

The most useful finding from the SISTRIX data isn’t the drift rate. It’s the structure underneath it.

For 86% of all prompts SISTRIX studied, the citation set has two parts:

A fixed core — a handful of domains that show up week after week, almost regardless of model updates or query phrasing
A carousel — the rest of the cited sources, rotating at roughly 89% per week

The split is platform-specific and brutal. ChatGPT Search cites only 3 to 4 domains per response, and most of those slots churn. Google AI Mode cites 14 to 16 domains with similar churn on the non-core portion. Only AI Overviews — which leans heavily on its existing search index — gives you 11 cited domains with 8 of them stable.

If your brand isn’t in the fixed core, you’re competing for a seat on the carousel. That seat lasts a week.

Drift rates by platform

The SISTRIX numbers, side by side:

Platform	Cited domains per response	Weekly source replacement	Practical implication
Google AI Overviews	~11 (8 stable)	5%	A stable, small set. Hard to break in, but visibility persists
Google AI Mode	14-16	56%	More slots, but more than half rotate every week
ChatGPT Search	3-4	74%	Tiny citation set, near-total weekly turnover

The country results matter too. AI Mode drift hit 54% in the US, 56% in Germany, 59% in Italy, and 57% in France. Across 17 weeks, the rate never trended down. SISTRIX concludes this isn’t a temporary effect from new model launches — it’s a structural feature of how these systems retrieve and cite.

Why drift happens

Three architectural reasons explain why AI citation sets churn this hard.

Retrieval pulls a different sample every time. When ChatGPT runs a query against the open web, it issues fan-out subqueries, samples a few results from each, and synthesizes an answer. The candidate pool is huge. The cited set is tiny. Small changes in query routing or candidate scoring shuffle the output dramatically. We covered this retrieval mechanic in detail — the system reads less of the web than people assume, but it reads different parts of it each time.

Freshness is weighted heavily. AI platforms favor recently updated content because stale sources are reputational risk. A page that gets a substantive refresh can leapfrog older, better-linked pages overnight. Seer Interactive’s data showed 65% of AI bot traffic targets content under a year old, and academic work on retrieval ranking has shown that timestamp signals alone can move passages dramatically in the ranking order, even when the underlying content is equally relevant.

The training and retrieval indices update constantly. ChatGPT’s cited web sources fell about 20% after a model transition in early 2026 — meaning the model itself decided to pull from a smaller, different set of sources. You can’t stabilize against a moving model.

The result: AI citations are not durable rankings. They’re more like impressions in a programmatic ad auction — re-bid every time the query runs.

What most teams get wrong

Three patterns show up in audit work, and all three lead teams astray.

Mistake 1: Treating a single audit as a baseline. A weekly drift rate of 74% on ChatGPT means a one-time citation count has roughly a one-in-four chance of being repeatable next week. If you set goals against that number, you’re measuring weather, not climate.

Mistake 2: Optimizing for the carousel and calling it victory. Many of the wins teams claim — “We got cited in ChatGPT for our category prompt!” — are carousel placements. They feel like progress. They evaporate.

Mistake 3: Ignoring the fixed core. The brands in the fixed core barely move regardless of what their competitors do. Those slots get earned through structural authority: Wikipedia presence, mainstream press coverage, dense interlinking from other cited domains. Most teams ignore those investments because they don’t fit the GEO playbook they bought.

The teams that are winning AI search treat citation drift as the central problem, not a measurement footnote.

The two strategies that work

If the fixed core is small and stable and the carousel is large and churning, you have two real options.

Play for the fixed core. This is a multi-quarter play. The fixed core for any commercial prompt tends to overlap with the 15 or so domains that already control most AI citation share — Wikipedia, Reddit, YouTube, established trade publications, a handful of category-defining sites. Getting into the fixed core means getting cited by those domains, getting a Wikipedia entity, accumulating mentions across editorial outlets, and building structured data that AI systems can verify against multiple sources. It’s slow, expensive, and durable.

Game the carousel. This is a weekly play. If most ChatGPT slots churn every seven days, then frequent, substantive content updates and fresh PR mentions can put you in the rotation. Carousel wins don’t compound the way SEO wins did. But they generate traffic now, and they create the citation footprint that eventually helps you petition for fixed-core status. This is the citation velocity model we wrote about — recent momentum beats accumulated authority, at least in the rotating tier.

Most brands need both. Run a fixed-core campaign aimed at structural authority over 12 to 24 months. Run a carousel campaign aimed at weekly citation cycling. Measure them differently because they pay off differently.

What this changes about measurement

A few practical adjustments if you’re operating with this in mind:

Move from quarterly snapshots to continuous monitoring. With a 74% weekly replacement rate on ChatGPT, any audit older than two weeks is junk data. Quarterly visibility decks are storytelling, not measurement.
Separate fixed-core and carousel metrics. Track which of your citations have repeated across 4+ consecutive weekly snapshots — that’s your real fixed-core footprint. Track the rest as carousel activity. The two require different strategies and should not be averaged together.
Watch the stable core for your category, not just your brand. Knowing which 3 to 8 domains anchor ChatGPT’s responses for your prompts tells you exactly where to invest in earned media. If a publication is consistently cited week over week for “best CRM for startups,” that’s a fixed-core slot you can target with pitches.
Stop celebrating one-off citations. A single Tuesday where you appeared in three ChatGPT answers is noise. Three consecutive weeks of repeated citations on the same prompt is signal.
Tie content updates to drift cycles. If your competitive set turns over every 7 to 14 days, your refresh cadence should run at least monthly for high-value pages. Quarterly refreshes hit a citation set that no longer recognizes them.

The bigger picture

Search Engine Land and several research firms have spent the past year treating AI search like a slightly weirder version of Google. It isn’t. Google’s index churns at a glacial pace by comparison. SEO professionals built careers on the assumption that hard-won rankings stick around. That assumption doesn’t translate.

The SISTRIX data, the Profound data, and the model-transition citation drops all point to the same thing: the citation surface of AI search is structurally unstable, and the only durable position is the fixed core. Everything else is rented week-to-week.

Teams that internalize this stop measuring AI visibility the way they measured SEO. They stop running quarterly audits as anything more than narrative. They invest in structural authority for the long arc and citation velocity for the weekly cycle. And they monitor continuously because they have to — the data they collected last month is no longer describing the system they’re trying to win.

Stop guessing about your AI search presence. Start your free RivalHound trial and get real data.