How long does an AI citation last?

Parse followed 16,164 repeated queries for 82 days. Consecutive answers shared only about one third of their cited domains, and roughly half of the first-run citations disappeared by the next run. The citations that survived tended to recur for weeks. Citation tracking therefore needs both repeat-run volatility and persistence, not a single observation.

How long does an AI citation actually last?

There is no single half-life, because citations split into two populations. About half of the domains an AI cites for a query on a given run are one-time appearances: they show up once and are gone by the next execution. The other half form a durable core that the model keeps reaching for, run after run, for weeks. So the honest answer to "how long does a citation last" is "either about a day or about a month, rarely in between."

A single citation check provides limited information because about half of the observed sources will not appear on the next run. The useful measurement is persistence. Parse's index covers 4.76 million AI responses, more than 603,000 brands, and 57.3 million citation observations, allowing repeat citations to be separated from one-run appearances.

What we measured

We used the citation arrays Parse stores for every tracked AI answer. Between February 2 and April 25, 2026, Parse re-ran the same set of 16,164 queries on ChatGPT and Google AI Overviews, executing each one on a median of 57 separate days across the 82-day window. That produced 921,107 ChatGPT answers carrying 17.3 million source references (an average of 18.8 cited domains per answer) and 921,103 Google AI Overviews answers carrying 9.2 million references (10.0 per answer).

For each query, we expanded its cited domains on every run and tracked how the cited set changed over the query's own sequence of runs. Indexing on runs rather than calendar dates is deliberate: Parse's daily cadence is uneven, so a week with more runs would otherwise look like more persistence. Every number below is measured against a query's actual re-executions, not the calendar.

:::

Half of a query's citations do not survive the next run

Start with consecutive runs. For each pair of back-to-back executions of the same query, we measured the overlap between the two cited-domain sets. On ChatGPT, the median overlap was 0.30 (mean 0.32); on Google AI Overviews, 0.33 (mean 0.36). In plain terms, when AI answers the same question two runs in a row, roughly 70% of the domains it cited the first time are not the domains it cites the second time. That turnover is not uniform across categories: AI citation volatility by industry shows software, AI, and commerce churn their sources hardest, while travel and financial services hold steadier.

Tracked forward from a query's first run, the drop is immediate and then it stalls. Of the domains ChatGPT cited on a query's opening run, 52.6% were still being cited on the next run. Google AI Overviews held only 43.9%. So between roughly 47% (ChatGPT) and 56% (Google AI Overviews) of a query's citations evaporate in a single step. This is why a one-shot citation audit is misleading: half of what it captures is gone almost as fast as you can read it.

If you want to know when AI changes its answer about your brand, start with a free brand check — it takes a minute.

The surviving half is a durable core

Most losses happen at the first repeat run. Sources that remain after that are more likely to persist. The table below tracks the share of a query's first-run citations still present at later runs. Runs are a little over one day apart on average, so the last column represents roughly six weeks.

ChatGPT loses nearly half of its citations at the first repeat run, remains near 46% through run 14, and declines again by six weeks. Persistent citations still disappear over time, but they last much longer than one-run appearances. This makes repeat frequency more useful than a single citation count.

Which sources persist across runs?

Source type predicts persistence, and the pattern is clean. For each query-domain pair on ChatGPT, we measured the share of the query's runs on which the domain stayed cited, plus how often a pairing was a single-run appearance ("one and done") versus cited on at least half the runs ("sticky"). Read this as which kinds of sources, once cited, tend to stay.

Reddit citations appeared only once 6% of the time, the lowest rate of any class, while Wikipedia had the highest share present in at least half of a query's runs at 21%. By comparison, 44% of less frequently cited domains appeared on one run and never returned. Review and directory platforms were closer to the less frequent group in this sample, though it contained fewer than 1,800 pairings. Persistence is concentrated in a few sources also ranked in the domains AI models cite most.

ChatGPT versus Google AI Overviews: who churns faster?

Both surfaces churn hard, but they churn differently. ChatGPT keeps more of its citations through the first re-run (53% versus 44%) and holds a notably stable mid-range, with 46% of first-run citations still cited three weeks later. Google AI Overviews drops more at the first step but then decays more gently, so by about six weeks its survival (35%) actually edges past ChatGPT's (29%). One way to read it: ChatGPT has a larger sticky core that erodes late, while Google AI Overviews trims faster up front and then settles into a smaller but steadier set.

The practical implication is that a blended, cross-platform citation number hides two different decay clocks. If you watch only one surface, you are calibrating to that surface's churn rate, which is why we treat ChatGPT and Google AI Overviews as separate measurement channels rather than averaging them into a single score, the same reason one AI visibility score is misleading.

Does this still hold on ChatGPT Search and Google AI Mode?

The 82-day panel covers the legacy ChatGPT and Google AI Overviews products. The current web-grounded surfaces, ChatGPT Search and Google AI Mode, only began collection in late May 2026, which is too short a window for a six-week survival curve. But we can still measure run-to-run overlap, and it confirms the pattern. From May 24 to June 26, 2026, consecutive runs of the same query shared a median of just 0.18 of their cited domains on both ChatGPT Search and Google AI Mode, even lower than the legacy surfaces.

So the newest surfaces churn their citations at least as aggressively. Cadence on these products is sparser (a median of four to six runs per query so far), which inflates run-to-run turnover, so treat this as directional confirmation rather than a clean comparison. The direction is unambiguous: citation volatility is a structural feature of AI answers, not an artifact of one model generation.

What this means for how you monitor AI visibility

Three things follow. First, do not rely on one citation check. About half of the sources in one run may be absent from the next. This is the same reason a visibility drop may be noise rather than a durable change: one run is a sample of a variable result.

Second, measure persistence, not presence. The citations worth chasing are the ones that survive repeated runs, and those are concentrated in Wikipedia, Reddit, and a small set of high-authority publishers. Earning a one-run mention on a long-tail blog is close to worthless for durable visibility.

Third, repeat a fixed query set on a regular schedule, as described in the weekly AI visibility review. A one-time check cannot distinguish a persistent citation from a one-run appearance.

How we measured this, and the caveats

The panel is Parse's first-party citation data: the cited-domain arrays attached to 921,107 ChatGPT and 921,103 Google AI Overviews answers, executed February 2 to April 25, 2026, across 16,164 repeated queries. A citation is one source reference inside a result's citation array. The survival and overlap curves use a fixed 12% deterministic sample of the query panel for tractability; the panel-wide volume figures use the full set.

Four caveats. Run cadence is uneven, so every metric is indexed on each query's own run sequence, not the calendar. "Still cited" is point-in-time at a given run, so a domain that drops out and later returns counts as present when it reappears; this measures whether a source is being cited at that run, not unbroken presence. The source-class buckets use a curated list of domain slugs, and the review-and-directory and Wikipedia buckets carry small samples that we read with caution. Finally, this is the legacy ChatGPT and Google AI Overviews collection; the current surfaces are reported only on run-to-run overlap because their window is too short for survival analysis.