A new study tracked 23 million cited sources across ChatGPT, Gemini, Perplexity and Google AI Overviews between April and June 2026. The finding making the rounds: 44% of cited pages appeared exactly once and then disappeared, and the typical citation lasted 11 to 15 days before churning out of the answer. The number is striking, and it is worth taking seriously. But almost everyone quoting it is stopping at the symptom. The data tells you that citations evaporate. It does not tell you why some evaporate in a day and others hold for months. That second question is the one that matters, and it has a structural answer.
The number worth quoting carefully
The study comes from Writesonic, shared through a B2B marketing newsletter, and the headline figures are clear enough: 44% of cited pages were one-and-done, the median citation lifespan sat at 11 to 15 days, and Perplexity held its sources roughly twice as long as ChatGPT. Two related findings sharpen the picture. After a model update in May, the share of ChatGPT citations going to third-party sites jumped sharply, and in roughly 90% of cases, when a new brand entered an answer, an existing one was dropped. The seats are fixed. The rotation is constant.
A word of discipline before building on it. This is one study, over a two-month window, on one provider's index. It is a strong, useful data point, not a law of physics. Treat the exact figure as indicative rather than eternal. What makes it valuable is not the precise number of days. It is what the pattern reveals about how AI visibility actually behaves.
Share is a snapshot. Durability is the asset.
The pattern exposes a measurement gap most of the market has been ignoring. Citation share – the percentage of queries where you appear this week – is a point-in-time number. It has the same flaw as judging a portfolio by a single trading day. It tells you where you stood. It says nothing about whether you are holding ground.
This is a distinction we have always drawn. The AI Visibility Index measures presence: how often you surface, right now, across engines. It is the snapshot, and it is genuinely useful. But a healthy-looking snapshot can hide constant churn – a rare spike in a sea of one-and-done appearances. If 44% of citations never come back, then appearing once is not the same as being in the answer, and being in the answer this week is not the same as being in it next month. Presence over time is a different signal from presence today, and the study makes that gap impossible to ignore.
Getting cited once is luck or noise. Getting cited again, and again, is structure. The first is a snapshot. The second is an asset.
Why citations evaporate
Here is the question the data raises and does not answer: why do some citations last weeks while others vanish after a single appearance? The agency reflex is to treat it as a publishing-cadence problem – refresh more often, post more, chase more surfaces. That is treating churn with volume. It misses the mechanism.
A citation that appears once and disappears was never anchored. The model surfaced the page because it happened to match a query in that moment, not because it had a durable structural reason to resolve your brand to that answer. When the rotation moved, nothing held it in place. A citation that persists is one the model keeps returning to because the underlying source is unambiguous: it resolves cleanly to a known entity, it carries signals the model can corroborate, it is specific enough to be the right answer rather than a plausible one. That is not a function of how often you publish. It is a function of how you are built.
This is exactly what the Citability Score measures. It is the structural precondition for being cited – and, it turns out, for being cited again. Durability is not a separate property you optimize with a content calendar. It is citability observed over time. The pages that hold are the pages that were citable in the structural sense to begin with. The ones that flicker in and out were riding the rotation, never anchored to it.
Fixed seats make structure decisive
The finding that 90% of new entries displace an existing brand is the part that turns this from interesting to urgent. If the seats were unlimited, durability would be cheap – you would simply accumulate. But the seats are fixed, and every new competitor signal is a candidate to take yours. To stay cited you have to do two things at once: remain trustworthy enough that the model retrieves you again, and be hard enough to dislodge that a newer signal does not quietly replace you. Both are structural properties. Neither is solved by posting more frequently. A brand tracking only whether it appears is measuring the first half of the game and is blind to the second – the slow displacement happening underneath a citation share that still looks fine.
Measure the why, not just the what
The study is a gift, because it makes a previously abstract argument concrete. Citations have a shelf life. Most expire fast. The seats are fixed and contested. All of that is the what, and it is now measured. The why – whether a given citation will hold or evaporate – is structural, and it is knowable before the churn happens, not after.
That is the work. Measure durability, not just frequency: the same prompts, run over time, to see which citations from your domain hold and which flicker out. Then ask why the durable ones stick, because the answer is always structural – clean entity resolution, corroborated signals, specificity a model can trust. This is what Brand Radar tracks over time and what the Citability Score explains at the root. The number everyone is quoting this week says citations have an 11-day shelf life. The more useful sentence is the one underneath it: the shelf life is set by how well you are built to be understood, and that is the only part of this you can actually change.