When brands first start measuring AI visibility, the question they track is simple: are we showing up?
That question gives you a snapshot.
What it doesn’t tell you is whether you’ll still be showing up next week, or the week after that. So we looked into it. We ran consumer prompts across ChatGPT, Gemini, and Google AI Mode every week for 14+ consecutive weeks, recording every URL each engine cited. Then we measured how much the list changed.
So, how long does an AI citation actually last?
The short answer: not long. Citation sets are turning over between 70% and 81% every single week, depending on the engine.
Every week for 14+ weeks, we ran the same set of consumer prompts through all three engines and captured every URL cited in the responses. To measure stability, we calculated the overlap between each week’s citation list and the one before it.
A score of 1.0 means the citation list was identical week over week. A score of 0.0 means it was completely different. We broke that data down by engine, by funnel stage, and across time gaps so we could see not just week-to-week drift, but what happens two weeks out, and three or more weeks out.
Here’s the baseline finding: brands can expect roughly 70%-81% of their cited sources to turn over every single week.
ChatGPT (the most stable engine in the dataset) replaced roughly 70% of its cited sources from one week to the next. Gemini replaced 76%. Google AI Mode replaced 81%.
This isn’t a quirk of the data. Every engine is doing live retrieval every time a user asks a question. With each retrieval, the majority of cited sources change. Getting cited in a given week tells you that your content registered with that engine for that prompt. It tells you very little about whether you’ll be cited again seven days later.
The overlap between a citation set this week and the same citation set three weeks from now is significantly lower than the week-to-week number.
A URL cited by ChatGPT today has roughly a 15% chance of still appearing three weeks from now. For Google AI Mode, that number is 11%. The decay is gradual, citations tend to fade rather than drop off overnight, but the direction is consistent across all three engines and all time windows measured.
Not all queries behave the same way. Across every engine in the dataset, one pattern held: high-intent questions, the specific, factual ones buyers ask close to a purchase, produced meaningfully more stable citations than broad exploratory searches.
When a consumer asks a specific, factual question like:
“Does X have free returns?” or “How long does shipping take?”
the engines find the page that answers it and tend to return to that page consistently. These queries have a right answer, and the engine learns which source has it.
Broad awareness queries like “best shoes for running” pull from an extremely wide and shifting pool of sources. Google AI Mode answering an awareness query carries a stability score of 0.11, essentially a different set of sources every single week.
Your product pages and policy pages aren’t just good for SEO. According to the data, they’re your most defensible real estate in AI search.
AI visibility requires consistent, fresh content maintained over time.
A single citation reflects one week of one engine’s retrieval cycle, and the data shows that cycle turns over fast.
Decision-stage queries produce the most stable citation behavior in the dataset.
When someone asks a specific question about your brand, the engines consistently return to the same sources. Clear, accurate, easily-crawlable product and policy pages are your most defensible AI real estate.
Getting cited for broad category queries on Google AI Mode is nearly random from week to week.
ChatGPT, with a stability score of 0.30, is a measurably more reliable engine for sustained awareness presence.
The decay data also points to a practical publishing implication.
Overlap scores continue to fall at the two-week and three-week marks across every engine. A brand publishing on a consistent weekly or biweekly cadence maintains a continuously refreshed pool of citable content, while a brand that publishes in bursts sees its citation share erode in the gaps.
If citations change so frequently, does a single citation still have value? Yes, a citation tells you that your content was visible and trustworthy to a particular engine for a particular prompt in a given week. The limitation is that one citation is a sample of one week. What builds durable AI presence is the content infrastructure that earns citations consistently: technically sound pages and regularly updated material that engines can return to across multiple retrieval cycles.
Does citation instability affect large brands the same way it affects smaller ones? The data shows citation instability as a structural property of how these engines operate. Even well-known brands with strong organic search presence see 70%+ of their citation share turn over week to week.
How does Gemini compare to the other two engines for long-term stability? Gemini sits between ChatGPT and Google AI Mode at a weekly stability score of 0.24. Its decay follows the same pattern as the others, the further apart two weeks are, the less citation overlap there is. At three or more weeks apart, Gemini’s overlap score drops to 0.16, slightly above Google AI Mode’s 0.11 but still a long way from anything resembling a stable foothold.
What’s the most practical first step for brands taking this seriously? Start with decision-stage content. Audit your product pages and policy pages for clarity, accuracy, and structure. These are the pages the study identifies as returning the most consistent citations, across all three engines, because they answer specific buyer questions with a clear, findable answer.
