AI citations turn over fast. In 2,545 answers from ChatGPT, Gemini, Perplexity and Claude to Noma’s own tracked prompts, two in three cited URLs appeared in a single run only, and just 36.1% of the URLs an answer cited were cited again by the same engine for the same prompt within a week. The citation half-life was under seven days.
This page sets out what we measured, how, and what it does not show. The data comes from the tracking Noma runs for Nurdd, the company that builds Noma, so we can publish it; no customer data is included.
What is citation half-life?
Citation half-life is the time after which half of the sources an AI answer cites are no longer cited when the same question is put to the same engine again. It is a way to describe how quickly the pages behind an answer turn over. The term is also in our AEO glossary.
Key findings
- Two in three citations were one-offs. Of 8,375 combinations of a URL, a prompt and an engine, 5,590 (66.7%) were cited in one run only.
- The half-life was under a week. Of the URLs an answer cited, 36.1% were cited again in a run 1 to 7 days later. A week is the shortest gap our runs can measure, so the true half-life may be shorter.
- Decay slows after the first week. 31.7% were cited again after 8 to 14 days, 28.5% after 15 to 30 days and 26.1% after 31 to 60 days. A core of sources keeps coming back.
- Domains last longer than pages. 41.8% of cited domains were cited again within 7 days and 34.2% after 31 to 60 days.
- Engines were close. The one-run share was 63.1% for Claude, 65.1% for Gemini, 66.4% for ChatGPT and 67.5% for Perplexity.
How long do cited URLs keep being cited?
| Days between runs | URLs cited again | Domains cited again |
|---|---|---|
| 1 to 7 | 36.1% (of 15,223) | 41.8% (of 13,186) |
| 8 to 14 | 31.7% (of 10,035) | 38.8% (of 8,759) |
| 15 to 30 | 28.5% (of 18,556) | 35.6% (of 16,456) |
| 31 to 60 | 26.1% (of 5,349) | 34.2% (of 4,926) |
Each row compares two runs of the same prompt on the same engine, that many days apart, and asks what share of the URLs (or domains) cited in the earlier run were cited in the later one. The counts in brackets are the citations in the earlier runs.
Method
- Data: 2,545 completed answers for the project Noma tracks for Nurdd, from ChatGPT, Gemini, Perplexity and Claude; 1,823 of them listed sources. Read from a snapshot of Noma’s database taken on 1 November 2026, the same data as our study of 2,545 AI answers.
- Prompts: 96 tracked prompts, each asked of each engine on up to 11 dates.
- Dates: 15 run dates between 6 August and 22 October 2026, irregularly spaced; all but four answers were collected by 26 September.
- Matching: URLs were compared after removing “www”, trailing slashes, fragments and tracking parameters (utm_*, ref and srsltid), in lower case.
- Span: for each URL, prompt and engine, the first and last run in which it was cited.
Limits
- One company’s prompts, in one category (marketing technology). Other categories may differ.
- Runs were not daily, so changes within a week are not visible, and the half-life can only be placed below seven days.
- The data cannot tell a source dropped for good from ordinary answer-to-answer variation: an engine asked the same question twice in a row may cite different pages.
- Engines change their models and search systems over time; this covers August to October 2026.
What this means for brands
- Never judge a page from one answer. A citation seen once is more likely gone next time than not. Track the same prompts on a schedule and count recurrences.
- Aim for the sources that recur. About a quarter of cited URLs were still cited a month or more later. Those are the pages and sites worth earning a place on.
- Build presence across a domain. Engines returned to domains more often than to pages, so several good pages on a trusted site hold up better than one.
- Report trends, not snapshots. See how often AI answers and their citations change for how many runs to sample.
Related research
Ahrefs published its own study of citation stability on 29 October 2026, How Stable Are AI Citations? 10 Things We Learned Tracking 37,500 Citations, using a different sample and method; read the two side by side rather than comparing numbers directly. For why engines cite what they do, see how AI engines choose citations.
Track your own citations
Noma by Nurdd asks the engines your buyers use the questions they ask, keeps every answer with its sources, and shows which sources recur for your category. Run a free brand audit to see what AI answers cite about you today.
See what AI engines say about your brand
Free · no account · or start a 3-day trial
Questions people ask
What is citation half-life?
Citation half-life is the time after which half of the sources an AI answer cites are no longer cited when the same question is put to the same engine again. A short half-life means the sources behind an answer turn over quickly, so a single check says little about which pages an engine relies on.
How long do AI citations last?
Not long, in our sample. Across 2,545 answers from ChatGPT, Gemini, Perplexity and Claude, 66.7% of cited URLs appeared in one run only for a given prompt and engine. Of the URLs an answer cited, 36.1% were cited again in a run 1 to 7 days later, so the half-life was under a week.
Are domains more stable than individual URLs in AI answers?
Yes, a little. In the same sample, 41.8% of cited domains were cited again 1 to 7 days later, against 36.1% of URLs, and 34.2% after 31 to 60 days against 26.1%. An engine often returns to the same site while picking a different page from it.
Which AI engine has the most stable citations?
In our sample the differences were small. The share of URLs cited in only one run was 63.1% for Claude, 65.1% for Gemini, 66.4% for ChatGPT and 67.5% for Perplexity. Perplexity cited the most URLs, so it also had the most one-off citations.
What should brands do if AI citations change so often?
Measure over several runs, not one. Track the same prompts on a schedule, count how often each source is cited across runs, and treat sources that recur as the ones to win. Domain-level presence is steadier than any single page, so being cited from several pages on a trusted site helps.