How long does an AI citation last

Ask the same question twice and the AI rebuilds its source list from scratch. About half of the domains cited on a query's first run are gone by the very next run. The half that survives, though, tends to keep showing up for weeks.

That split is the whole story. Bimodal means the sources fall into two groups instead of clustering around an average: sources that appear once and disappear, and sources that stick. There is no smooth decay curve and no tidy half-life number. A one-time citation check cannot tell you which group you are looking at.

Measure how much two consecutive runs of the same query share, and the answer is: not much. On ChatGPT only about 30% of cited sources carry over from one run to the next, so roughly 70% of the set churns each run. Google AI Overviews is marginally steadier at about 33%.

This is the run-to-run instability marketers feel but rarely quantify. Most of what an AI cites today is not what it will cite on the next pass of the same question.

Track a query's first-run citations forward and the pattern is distinctive. The big drop happens immediately, from 100% to about 53% at the next run. Then survival plateaus near 46% for two weeks before sliding to 40% at three weeks and 29% by four weeks.

So the danger is not a slow fade. It is the first re-run, where half your sources fall away, and then a second cliff around the four-week mark for the survivors.

Google AI Overviews loses more first-run citations immediately: only 43.9% survive to the next run, against ChatGPT's 52.6%. But it then declines far more gently, holding 34.6% all the way out to four weeks.

The two engines cross over. By roughly six weeks the source you won on Google is more likely to still be cited than the one you won on ChatGPT (34.6% versus 28.9%).

Plot both survival curves together and they cross. ChatGPT starts higher, holding about 53% at the next run, but slides to 28.9% by four weeks. Google AI Overviews starts lower at 43.9% yet declines far more gently, holding 34.6%. The engine that looked more stable in the first week is the less stable one by the second month.

Persistence is not random. It tracks the kind of source. On ChatGPT, Wikipedia and Reddit-style Q&A both average a 0.27 presence rate (cited on about 27% of a query's runs), while everything in the long tail averages just 0.10.

Wikipedia is the stickiest by far. Reddit is the most reliable in a different way: only 5.8% of Reddit citations are cited just once, the lowest of any kind of source. The long tail is the opposite, with 44% of its citations appearing exactly once.

Once a Wikipedia page is cited for a query, 21% of the time it goes on to be cited on at least half of that query's runs, the highest of any kind of source measured. If you can earn the Wikipedia reference for a query, it stays cited far longer than a normal citation does.

Reddit and forum Q&A is the most repeat-reliable kind of source. Just 5.8% of its citations are cited only once, against 17% for Wikipedia and 44% for the long tail. When AI reaches for Reddit on a question, it tends to keep reaching for it.

Most of the churn lives in the long tail. Ordinary sites that are not Wikipedia, a major forum, or a top news brand average a 0.10 presence rate, and 44% of their citations appear exactly once. They get cited once, then the next run replaces them with something equivalent. A single appearance here is closer to a coin flip than a lasting position.

The 82-day panel runs on the older ChatGPT and Google AI Overviews engines. On the newer ChatGPT Search and Google AI Mode engines, run-to-run overlap is lower still: a median of 0.176 and 0.182 respectively.

The instability is not a side effect of how the older data was collected. If anything, today's answer engines swap citations more aggressively than the panel did.

These figures come from 16,164 queries re-run a median of 57 times over 82 days. That repeated set of queries is what we call the panel. It captured 17.3 million source references on ChatGPT and 9.2 million on Google AI Overviews, an average of 18.8 and 10.0 references per answer. The repeated re-runs are what make survival measurable at all.

Check more than once. A single snapshot cannot tell a transient citation from a durable one, and half of all first-run citations are transient.

Weight by the kind of source. A Wikipedia or established-forum citation is worth far more than a long-tail one, because it is far likelier to persist.

Mind the cliffs. Expect the steepest loss at the first re-run, and a second drop for ChatGPT survivors around the four-week mark.

We took 16,164 queries that were re-run repeatedly over an 82 day window and treated each source reference in an answer as a citation. Persistence is counted by each query's own run number, not the calendar, because the number of runs varies week to week. Consecutive-run overlap is the share of cited domains that two back-to-back runs have in common, counted as shared domains divided by all domains named across the two runs. First-run survival is the share of a query's first-run citations still cited at a given later run. For persistence by kind of source, each cited domain is grouped using a curated list (Wikipedia, Reddit and forum Q&A, major news, long tail, and so on).

A citation counts as present at a given run if it is cited on that run; sources that drop out and return are counted as present when they reappear, so this measures whether a source is cited at a point in time, not unbroken presence. Figures counted by run number are computed on a fixed, repeatable sample of the panel to keep the computation manageable. Figures describe an observed sample of answers from the engines in scope and are not projections.