How Long Do AI Search Citations Last? A Study of Citation Source Retention Across 320,000 Answers

If our page is cited once in AI search, how long does it stay there? To find out, Anymorph asked the same queries every day from October 16 to 23, 2026, across ChatGPT, Gemini, Google AI Overviews, and Naver AI Tab, and compared 324,215 answers. We matched answers to the same query from one day to the next and a week later to measure how many cited pages and sites remained. We focused on Korean-language queries and compared them with US English queries collected using the same method.

Summary of Findings

Research Method

What We Measured

All figures combine multiple measurement projects; results for individual companies are not included. The data consists of AI search answers collected by Anymorph.

Measurement Method

• Page retention rate: The share of pages cited in an earlier answer that were also cited in a later answer when the same query was asked of the same engine one day later (or three days or one week later). We counted only queries with answers on both days.
• Domain retention rate: We counted retention in the same way at the site (domain) level. If a different page from the same site was cited, we counted the site as retained.
• Identifying the same page: We compared URLs after removing advertising and referral-tracking parameters and page location markers (#). This ensured that the same page was counted once when it appeared under different URLs from one request to another.
• Sources: For ChatGPT, we counted inline citations linked beside answer text. For Gemini, Google AI Overviews, and Naver AI Tab, we counted the sources displayed in the answer. Shopping product cards in Naver AI Tab were excluded.
• One-week analysis: We used only queries with an answer every day from October 16 to 22. The one-week retention rate compares answers from October 16 and 23.
• Site types: We manually classified around 650 public platforms and frequently cited domains, and classified government and institutional addresses (.go.kr, .or.kr, .gov), university addresses, and media outlet addresses using rules. The rest were left unclassified.

Naver AI Tab displays far more sources in each answer than the other engines. Longer source lists are more likely to overlap the next day, so the results are best read by looking at changes within each engine rather than comparing absolute values across engines.

For Korean-language queries, 14.4% to 45.6% of pages remained cited the next day

Six out of seven pages cited today in Korean ChatGPT answers were missing from answers to the same queries the next day. Even when looking at sites rather than individual pages, only 23.8% of domains remained the next day. Across the eight days, Korean ChatGPT's next-day page retention rate ranged from 11% to 18%. This was a recurring pattern, not a change limited to a particular day.

Korean ChatGPT results changed more often than English results

To check whether this difference was due to the mix of industries, we separately compared only the 9 projects measured in both Korean and English. Even for queries in the same industries and on the same topics, next-day page retention for Korean ChatGPT was about half the English rate. In the same projects, Gemini (28.5% Korean, 32.4% English) and Google AI Overviews (16.3%, 15.8%) showed little difference between the languages.

A week later, only Naver AI Tab had fallen below half its next-day rate

ChatGPT, Gemini, and Google AI Overviews changed substantially by the next day, but their retention rates did not fall much after that. One-week retention was 74.2–87.5% of the next-day rate. This suggests a split between the many pages that change daily and the smaller number of pages that are cited consistently. Naver AI Tab had the highest next-day retention, but it declined over time, with one-week retention at 48.0% of the next-day rate.

The number of pages cited for a query over a week was 3.5 to 5.4 times the number cited in one day

We counted only queries that received an answer every day for a week. For Korean ChatGPT, the number of pages cited for one query over a week was 5.4 times the number appearing in a single day's answer, and 81.5% of those pages were cited on only one of the 7 days. A one-day measurement reveals only part of the pool of pages eligible to be cited for a query. Conversely, even pages not cited today have a chance to appear during the week, with new opportunities arising every day.

Pages cited on 4 or more days accounted for a large share of citations despite being few in number. In Naver AI Tab, 15.3% of pages accounted for 41.7% of citations over the week; in Gemini (Korean), 9.2% of pages accounted for 27.6%.

Sites held their positions, but pages changed

When the same site was cited the next day, in Korean Google AI Overviews, one time out of three (36.5%) a different page from that site was cited. The figure was 31.1% for ChatGPT as well. When several pages on the same site matched a query, one of them was selected on any given day.

Blog and video platforms kept their place, but the posts changed

Platform retention is the rate at which the same platform was cited the next day in answers to queries that cited that platform today. In 64.2% of cases where Korean Google AI Overviews cited Naver Blog, it cited Naver Blog again the next day, but it selected the same post again only 17.2% of the time. For YouTube, the platform remained 59.0% of the time, while the same video remained 19.3% of the time. Naver AI Tab retained Naver Blog and Naver Cafe positions more than 80% of the time, but the same post was cited again only 39.0% and 48.2% of the time, respectively. Among the platforms in the table, place information from Naver Map lasted longest, with the same location cited again 66.0% of the time.

Sources cited earlier in an answer were more likely to remain the next day

These figures are for Korean-language queries. Across all four engines, pages cited earlier in an answer were more likely to be cited again the next day. The first source in Korean ChatGPT remained at a rate of 19.7% the next day, twice the rate for sources 6–10 (9.3%). The same difference appeared in English ChatGPT: 36.8% for the first source, compared with 15.6% for sources 6–10.

Government and public institution pages and brand and company site pages lasted longer

These are next-day page retention rates by site type for Korean-language queries. Types accounting for less than 1% of an engine's citations are marked “–”. Government and public institution pages and brand and company site pages lasted longer than the overall engine average in every engine with reported figures. In Gemini, review and ranking platforms (39.8%) lasted the longest. Retail and commerce pages changed faster than average in ChatGPT, Gemini, and Google AI Overviews; for Korean ChatGPT, only 5.1% remained the next day. In English ChatGPT, academic and medical pages (37.1%) and media and news pages (35.4%) also lasted nearly twice as long as retail and commerce pages (18.7%).

What to Do

Judge by weekly citation rates, not by a single day's results

Of the pages cited by Korean ChatGPT over a week, 81.5% were cited on only one day. If you judge a page's performance solely by whether it was cited or not on any one day, you will mostly be seeing chance variation.

• Measure important queries every day for each engine, and use the number of days cited out of 7 as a weekly metric.
• After updating a page, compare a period of the same length before and after the change.

Cover a query with multiple pages

The same site was cited on the next day through a different page 36.5% of the time in Korean Google AI Overviews and 31.1% of the time in ChatGPT. Even within the same site, engines selected a different page suited to that day's query.

• For each key query, create pages on your own site from different angles, such as definitions, comparisons, how-tos, and frequently asked questions.
• Link the pages to one another to make it easier for engines to find other pages on the same topic.

Put a paragraph that answers the question directly at the top

Sources cited earlier were more likely to remain the next day. The first source in Korean ChatGPT remained twice as often as sources 6–10.

• Write the answer to the question in one or two sentences in the first paragraph of the page.
• Include facts in the body that can be carried straight into an answer, such as figures, conditions, and reference dates.

Put evidence and official explanations on your own site

Brand and company site pages lasted longer than average across all four engines, while retail and commerce pages changed faster than average in three engines. Government and public institution materials also lasted longer.

• Describe product and service standards, ingredients, specifications, and pricing policies on explanatory pages on your own site. Do not keep the information only on retailer product pages.
• When citing statistics from public institutions or official standards, include a link to the source.

Publish blogs and videos regularly

In Google AI Overviews and Naver AI Tab, Naver Blog and YouTube positions remained the next day 59.0–86.6% of the time, but the same post or video was selected again only 17.2–39.0% of the time. In Naver AI Tab, same-page retention fell below half over the course of a week.

• On Naver Blog and YouTube, do more than publish one carefully crafted post or video; split the same topic into multiple pieces and publish them regularly.
• If you are targeting Naver AI Tab, publish a new post every week. Businesses with physical locations should keep their Naver Map place information up to date.

Use different measurement schedules for different engines

Korean ChatGPT pages changed most often among the four engines, and the share of pages remaining the next day was about half that of English ChatGPT. In the Korean market, do not infer Korean ChatGPT results from English-language measurements or a single day's data. Measure Korean-language queries separately every day.

Limitations

• This was an 8-day observation from October 16–23, 2026. Retention over periods longer than a week was not covered, and results may change if the engines' models or search methods change.
• The queries were general questions asked by Anymorph for its measurement projects. They do not exactly represent the distribution of queries across the entire market. Naver AI Tab had a smaller sample than the other engines, with 10 measurement projects and 1,087 queries.
• Sources were counted according to how each engine displays them: inline citations for ChatGPT and displayed sources for the other engines. Since the length of source lists varies by engine, absolute comparisons across engines are limited.
• Same-page matching was based on URLs with tracking parameters removed. Pages with identical content but completely different URLs were counted as different pages.
• We manually classified around 650 public platforms and frequently cited domains by site type, and classified government, institutional, university, and media outlet addresses using rules. Across engines and languages, 14.7–38.4% of citations were unclassified.
• A citation means that a source was used in an answer. This study did not determine whether it led to a click or purchase.
• Differences by type and position reflect observed distributions; the study did not experimentally establish their causes.

Conclusion

1. AI search citations change every day. For Korean-language queries, the share of pages still cited the next day was 14.4–45.6%, with Korean ChatGPT changing most often.
2. A small number of pages remain consistently cited. In ChatGPT, Gemini, and Google AI Overviews, one-week retention remained at 74.2–87.5% of the next-day rate. Sources cited earlier, along with pages from government and public institutions and brand and company sites, lasted longer.
3. Sites hold their positions. Platforms and sites often remained the next day, but the pages within them changed frequently.

To maintain citations in AI search, it is more important to build a set of pages that covers a query and publish consistently than to focus on the ranking of a single page. Measure performance by tracking the same queries on each engine every day and checking the share of days cited over a week.

Appendix: Research Data