How Long Does an AI Citation Last? 78% of New Citations Didn't Repeat for the Same Prompt Within Two Weeks

You open your AI visibility report and there it is: an assistant cited your integration guide in an answer about your category. Two weeks later you run the same prompt and the link is gone, and nobody can say whether something broke.

Probably nothing broke. We tracked 2,047 prompt-and-assistant pairs, each run about once a day from September 2025 to October 2026, and followed every URL after its first citation. A new citation usually turns out to be a one-off, not a position you keep.

The short version: In an internal Genezio research covering 922,000 new citations, 78% of newly cited URLs were not cited again by the same prompt on the same assistant within the next two weeks. For third-party pages, the presence that remained fell by half in about 45 days. Pages owned by the brand or its competitors faded much more slowly. Most of the difference in decay between periods tracked how many links answers contained (our interpretation, supported by a capped re-analysis). So a single "we got cited" snapshot tells you very little.

How likely is a new citation to show up again?

Finding: Most new citations don't repeat in the short term: that's the 78% above, measured on all 922,000 new citations.

Evidence: We counted a citation as "new" only on the first day a prompt/assistant pair cited a URL, and only if that pair had already been running for at least 21 days. That rule filters out URLs that only look new because tracking had just started. Then we checked whether the pair cited the same URL again in the next two weeks.

What this means: If your page is cited once, the likeliest outcome is that the same prompt won't cite it again soon. That isn't losing it forever (a URL can come back later), but one sighting isn't a win you can report.

How fast do third-party citations fade?

Finding: For earned (third-party) pages, presence halves in about 45 days. Across all source types pooled, the figure is 62 days.

Evidence: We measured "presence": the share of new URLs that the same pair cites at least once in a 14-day window, starting some number of days after the first citation.

The "half-life" is the day when that presence falls to half of what it was in days 1–14. Note the baseline: it's half of the first two weeks, not half of 100%. We followed a fixed group of 308,000 new citations (first cited between 20 October 2025 and 26 July 2026) for at least 61 days, so the same URLs are compared at every point. Even URLs that a pair had cited from its first days of tracking decayed, with a half-life of 75 days (95% interval: 71–79 days).

The so-what: A 45-day half-life on earned citations means that the review, listicle or forum thread that got you cited this quarter is likely to carry much less weight next quarter. A third-party citation is closer to a campaign than an asset.

Does it matter whether the page is on your own site?

Finding: Yes. Ownership was the one large, stable factor in the data. Owned pages decayed much more slowly in every time period we checked.

Evidence: "Owned" means pages on the site of the brand or a competitor; "earned" means third-party pages. We compared how much presence remained around day 76 (days 76–89, as a share of days 1–14):

Table: Owned pages kept about 1.5 to 2 times as much of their early citation presence as third-party pages by day 76.

This also explains apparent industry differences. Banking, for example, showed a 75-day half-life on all sources but 51 days on earned sources alone, because 42% of its classified new citations went to owned pages. In a separate internal analysis of 15,255 cited pages (August–September 2026), company websites were the largest cited source type in every industry and assistant combination we measured, at 64% to 86% of citations per answer (an average share per answer, weighting companies equally; "company websites" there also includes other sellers' sites, so it isn't the same category as "owned" here).

The implication: Owned pages decline too. But if you want citations that persist, your own docs, pricing and product pages are where persistence shows up in this data.

Why did my citation numbers change when I changed nothing?

Finding: Much of the month-to-month change came from answer size: how many links an assistant cited per prompt per day.

Evidence: Picture a raffle: draw 58 winners instead of 31 and every ticket's odds go up, though nobody bought more tickets. The assistant that supplied most of the third-party citations in our data (78% of them; a different 78% from the headline), which we'll call the high-volume assistant, went from 43 URLs per prompt per day in May 2026 to 31 in June, 35 in July, 48 in August and 58 in September.

Third-party pages first cited in May–June (followed up during the short June answers) had a half-life of 31 days (95% interval: 29–36). Those first cited in July (followed up during the longer August–September answers) had 55 days (51–59). That's a 24-day gap. When we re-ran the analysis keeping at most 20 random citations per pair per day, the half-lives became 37 (31–43) and 43 (41–47) days. The gap shrank to about 6 days, and the intervals overlap. A second random draw gave 38 and 44.

What this means: Our reading is that most of the period effect comes from answer size: when an answer cites more URLs, each URL has a higher chance of being cited again. It isn't the only factor, since a small gap remains. But if your dashboard shows citations "improving" in September, check links per answer before you credit your content team.

Does one assistant, country or industry hold citations longer?

Finding: Not in a stable way. We tracked seven assistants, and none decayed faster or slower in every period.

Evidence: One assistant kept sources longer than the high-volume assistant in May–June (59 vs 29 days) but not in July (39 vs 57). Pooled averages mislead too: 55% of one assistant's new citations entered in July versus 27% for another. Countries behaved the same way. Great Britain decayed faster than the US in May–June (24 vs 49 days) but not in other periods. No industry differed in the same direction in every period.

Worth calling out: A "which assistant keeps citations longest" ranking built from one period probably measures the calendar, not the assistant.

Which third-party pages fade fastest?

Finding: Forum threads faded faster than average in two of three periods. A large discussion forum, a large social network and a large video platform were among the fastest-fading domains.

Evidence: In May–June and July, forum threads had half-lives of 16 and 25 days, against 31 and 55 for all earned pages (17 and 24 vs 37 and 43 with long answers capped). In February–April they were at the average. Among the largest domains, a large discussion forum, a large social network and a large video platform were among the fastest-fading of the largest domains (24 to 30 days). A professional social network and brand sites held up.

Page age is a weaker signal. Only 8% of new citations had a usable publish date. In May–June, third-party pages older than a year halved in 27 days (24–30) versus 55 days (43–67) for younger pages. Pooled, pages under a month old came back within the first two weeks 37% of the time, compared with 26% for pages over two years old (an 11-percentage-point gap). In other periods the difference was unclear.

The so-what: A forum thread can get you into an answer quickly, but don't plan around it staying there. Treat the page-age result as a hint, not as proof that updating pages keeps citations.

What does a sudden source shift look like?

Finding: The mix of cited sources can change almost overnight, with no change on your side.

Evidence: A separate internal research note looked at 3,179 prompts answered once a day on each of the 31 days of August 2026, on the same high-volume assistant. The share of answers citing a large discussion forum fell from 23.2% on 12 August to 4.1% on 14 August, a drop of about 19 percentage points in two days. Over the same window, the share citing a large document-sharing site rose from 1.7% to 7.0%, roughly four times higher. The forum share fell for all 103 brands with more than 200 answers. Links per answer didn't fall (around 46–49 links from about 18 domains, before and after); only the mix changed.

Of the document-sharing links, 98% pointed to user-uploaded PDFs. Many of the most-cited ones were old copies of company documents: fee lists from 2023 and 2025, app manuals, early-2025 phone offers, interest-rate lists.

Our interpretation: The assistant often answered questions about fees, offers and features from stale uploaded copies instead of the company's own site. We can't say what caused the shift, only when it happened. But snapshots from 12 and 14 August would tell two very different stories.

How to measure and act on AI citations

1. Stop reporting single citations as wins. Re-run the same prompts repeatedly (daily or weekly) and report presence over a window, not a one-day sighting.
2. Split owned and earned citations. They decay at very different rates, and mixing them makes industry or period comparisons misleading.
3. Log links per answer next to every citation count. If the number of links an assistant cites jumps, expect your citation counts to move with it.
4. Compare segments only within the same period. Don't rank assistants, countries or competitors from one month of data.
5. Invest in owned pages that answer buyer questions. Put current pricing, fees, specs and docs on your own site, where they're easy to find.
6. Audit stale copies of your documents. Search for old PDFs, price lists and manuals hosted elsewhere, and request removal where you can.
7. Treat citations as a diagnostic. The goal is that assistants recommend you accurately in buyer conversations; citations help you see why they do or don't.

Methodology

The main dataset is an internal Genezio research on citation decay: a 10% random sample of tracked prompts, covering 4.6 million distinct prompt/assistant/day/URL citations from 2,047 prompt/assistant pairs run about once a day between September 2025 and October 2026. Half-lives use 95% intervals from 400 bootstrap resamples of pairs, and segments needed at least 20 pairs. A 14-day window counted only if the pair ran on at least 5 days in it. Results held when we used a 28-day new-citation rule (62 vs 61 days pooled, 45 for earned under both) and when we merged URLs differing only by tracking parameters.

Limits matter here. The prompts are simulated, on topics of brands that use our platform, so they're not a random sample of the web, of industries or of real users, and they say nothing directly about clicks or revenue. Half-lives depend on the 14-day window and on answer size (a stricter "present" rule gives roughly 10–15% shorter half-lives), so compare segments with each other, not with figures from other studies. We used overlapping 95% intervals as a comparison criterion, which isn't a formal test. Results after day 61 describe only early entrants, which decay more slowly. Several segments are thin. "Decay" means an assistant stopped citing a URL, not that the page went offline. Five review rounds fixed estimator bugs and removed claims we couldn't support (for example, that product pages decay more slowly or that one assistant got faster over time).

If you want to see which sources assistants use for your brand across repeated, persona-based multi-turn conversations, and whether what they say is accurate, Genezio runs those tests.