How pages become sources, and why they stop
Between changing a page and an AI answer quoting it, six things have to happen. None of them are instant, and the thing they produce does not last.
The chain
| What happens | How long |
|---|---|
| You publish or change something | day 0 |
| A crawler fetches it | 1 to 14 days |
| It enters the index | 1 to 7 days after that |
| It becomes a candidate for a question | 2 to 6 weeks |
| An AI assistant uses it as a source | weeks |
| Enough answers use it that a rate moves | 4 to 12 weeks |
These timings are our working estimates, not measured and not documented by the vendors.
AI assistants run on different clocks
- Perplexity
- Retrieves on every query, so new content can appear quickly.
- ChatGPT
- OpenAI says ChatGPT may send searches to Bing, and that a site must let its own crawler, OAI-SearchBot, read it to be included. So being slow to appear in Bing is one route that waits. Submitting the URL rather than waiting on a crawl shortens it.
- Google AI Overviews
- Standard Google indexing, so a page has to be indexed and snippet-eligible before it can be quoted.
Days
1 to 3 weeks
4 to 8 weeks
On these estimates, a flat ChatGPT result two weeks after a change may be lag rather than failure.
These timings are our estimates too, not vendor documentation. What we measure is which assistant named you, on which question, on which day.
Sources get dropped
A study of 3.5 million citation events across more than 120,000 domains and six AI platforms, run between September 2025 and March 2026, found a median half-life of 4.5 weeks. Half the pages used as sources in a given week are no longer used a month later. ChatGPT churns fastest at 3.4 weeks. Perplexity holds longest at 5.8. Google’s surfaces sit between them.
Scrunch AI with Stacker, 3.5 million citation events, 120,000+ domains, six platforms, September 2025 to March 2026. Method and figures.
A separate study found that only 30% of brands stayed visible from one answer to the next, and only 20% across five consecutive runs.
AirOps, The 2026 State of AI Search. The report. Its sample size is not disclosed, so treat the figures as directional. We are giving them because an unstated sample with that noted is more useful than silence, and less misleading than implying a rigour we have not checked.
Neither of these is our measurement. Both come from companies that sell AI-visibility monitoring, which is a reason to read them carefully rather than a reason to dismiss them, and we have linked the method for each so you can. Our own work touches AI answers too, so the same caution applies to everything on this site: that is what publishing our own numbers, including the flat ones, is for.
When we hold enough clients to compute a half-life from our own daily collection, we will publish ours beside these, including if ours disagrees.
What that means if you check yourself
If only one brand in five stays visible across five consecutive runs, a single check tells you where you stood in one sample. It cannot tell you whether that was your typical position or an outlier, and there is no way to tell which from inside the one answer.
That is why we sample daily rather than monthly, and why our rates say how sure we are wherever the samples allow. A monthly check on a signal with a 3.4-week half-life is sampling noise and calling it a trend.
What speeds it up
- Submit the URL rather than wait for a crawl
- It tells the search engine the page changed, instead of waiting for a crawler to notice.
- Publish and update with real dates
- Ahrefs found AI assistants used pages 25.7% fresher than Google's organic results, ChatGPT most of all and AI Overviews not at all.
- Answer the question in the first lines
- An assistant lifts what it can lift. A page that opens with context and answers in paragraph six gives it nothing to take.
- Refresh pages already used as sources
- The decay happens whether or not anyone acts on it, so the pages working today are the ones worth keeping current.
Ahrefs, 16.975 million URLs from AI answers and Google’s organic results, published 28 July 2025.
None of that says which sources an assistant reaches for, and the assistants disagree with each other more than they agree. Only 2.7% of the domains they use are shared across five of them. What gets a business used as a source sets out what each one reads, with the published evidence.
Corrections
The page said Perplexity retrieves with no knowledge cutoff. We could not check that at its source, so the phrase is cut rather than kept.
The chain and clock timings had no source, and the clocks were called documented retrieval behaviour. No vendor documents them, so both are now labelled as our estimates.
The page said the ranges summed to ninety days and that a flat ChatGPT result at two weeks is lag. Both rested on those estimates, so the first is removed and the second is hedged.
The page said submitting a URL cuts the wait to hours from one to three weeks. No source we found states it, and the figure is removed.
The page said Perplexity uses fresh sources disproportionately. Ahrefs found ChatGPT favours fresh pages most, and Perplexity least of the assistants it studied apart from AI Overviews.
The page said every rate shows how sure we are and results are never pooled across assistants. Reports show how sure we are only where the samples allow, and also give a pooled rate.