According to Profound, a page cited by AI apps typically loses half of its citations 11 days after its peak (median; 78% of pages lose half within two weeks), pages cited by many apps hold on longer, and Claude Sonnet 5.5 cites newer pages (the top 10% most-cited pages have a median age of 107 days versus 164; titles containing "2026" 38% → 50%).

The number that stopped me

On 14 October 2026 I read Profound's post "The half-life of an AI citation is 11 days", published on 30 September.

It is easy to think of an AI mention as a position you win and then hold for a while. Eleven days says otherwise.

Profound sampled about 883,000 pages over a year, September 2025 to September 2026, across seven apps: ChatGPT, Perplexity, Claude, AI Overviews, AI Mode, Gemini and Copilot. What they found:

• Median half-life: 11 days from a page's citation peak. 78% of pages lose half their citations within two weeks.
• Pages cited by many apps last longer. On day 56, 6.8% of pages cited by five or more apps were still above half their peak, against 1.2% of pages cited by a single app.
• Comparison and list articles hold on longest; news and press releases fade fastest.

The second post, a day earlier

"What's changed with Sonnet 5.5", from 13 October 2026, compares Claude Sonnet 5.5 with Sonnet 5. Among the top 10% most-cited pages, median age dropped from 164 days to 107. The share of titles containing "2026" rose from 38% to 50%. And per query that triggers a search, Sonnet 5.5 cites slightly fewer pages: 6.19 against 6.93.

One new model, and the pages it picks got almost two months younger.

How I read it

If Profound's numbers hold, being cited by AI is a flow, not a status. A press release gets a burst of citations and then goes quiet. A data page that is refreshed regularly, carries a date in its title and gives apps a reason to come back is more likely to keep its place.

I map it onto the two rungs I measure:

1. Crawlers come back. A page that really changes gives crawlers a reason to fetch it again. Google and Bing both say they use the <lastmod> date in your sitemap when it reflects the real change date, not the date the sitemap was generated.
2. Apps keep naming you. This is where Profound's data applies, but only as a correlation.

Something you can do now: if you have comparison pages, price tables or lists, update them on a genuine schedule, show the update date on the page, and let <lastmod> carry that same date. Do not bump the date without changing the content; Google and Bing both say they may ignore dates they cannot trust.

What I do not take for granted

This is an external-source claim, and the source sells tools for tracking citations, so a picture of fast-moving citations suits its business. More importantly, Profound measured how fast citations decay, not whether updating a page keeps it cited longer. The advice above is my inference, not a measured result. Nor is there anything specific to Japan. The limits section below sums up those three points.

I want to test this on my own pages: a data page updated on a fixed rhythm, watching how crawlers return and whether apps keep mentioning it. When there are numbers, they will be on the pw research pages.