How long does it take AI to cite a new page?

Across four of my own sites, the first citation on Bing's AI surfaces arrived anywhere from the same day to 64 days after tracking began. The fast sites shared one habit: they told the index the moment something shipped. The slow one waited to be found.

That is the short version. The honest version has caveats, a negative result, and one number I did not expect. All of it comes from my own properties, measured with the same instrument, so I will also tell you exactly what this data cannot say.

I track every AI citation Bing reports for my sites. The three sites in my citation study logged 26,767 of them through July 13, 2026, across reporting windows that open between December 2025 and April 2026, and a brand-new fourth domain joined tracking in July. That corpus already produced a playbook on what gets cited at all. This note answers the narrower question I get most often: once a page exists, how long until an AI answer actually cites it?

The scoreboard: same day to 64 days

Four sites, one instrument, four very different onsets. Per my house rule, sites are described by role, not name, and every number carries its measurement window.

Measurement notes, because they matter. The counter is Bing Webmaster Tools' AI Performance reports, the only public instrument I know that reports AI citations per site per day. It blends Bing's AI surfaces, Copilot included, and does not break out engines. "Lag" in the first three rows is measured from the first day of Bing reporting, not from the page's publish date, so it conflates page age with indexing speed. The new-domain row runs on a different clock: from its first crawlable content on June 30 to its export opening on July 3, already showing citations. And the two fastest rows are censored in the flattering direction: each export opens with citations already on the board, which means the true onset could be even faster, or the export could simply have started after citing began. Treat every row as a bound, not a stopwatch.

What the fast sites had that the slow one lacked

The gaming site, the fastest of the four, has one habit the slow site did not: a publish-event ping. Its weekly data refresh ends by submitting changed URLs to IndexNow, the protocol that tells Bing's index "this page changed, come look." The ping fires the moment content ships, wired into the same pipeline that regenerates the pages. Not a daily cron. Not a manual submission. Part of publishing itself.

The family storytelling site had no ping mechanism of any kind for the first months of its life, and it waited 64 days.

I want to be careful here, because this is where most articles in this genre start lying. That comparison is two different sites, on different topics, with different ages and different demand. The attribution to IndexNow is an inference, and my own research notes flag it as exactly that. What I can rule out is the boring explanation: crawl access. The slow site's robots.txt had allowed every major AI crawler since November 26, 2025, and indexing was enabled January 4, 2026, seven weeks before its first citation. The door was open. Nobody walked in for another seven weeks.

The cleanest test I have: a brand-new domain, day zero

Cross-site comparisons are weak, so in late June I got a better test for free. I launched a brand-new site, a reference library on a fresh domain with no history, no backlinks, and no authority, and built it the way the study said to: crawlable pages, sitemap, open robots, and an IndexNow key live from June 30, the first day real content existed.

Bing's export for that domain opens on July 3, already showing 12 citations that day. From first crawlable content to first AI citation: three to four days, on a domain that did not exist the week before.

Same caveat as above, stated plainly: the export opening on July 3 may reflect when the site was enrolled in reporting rather than when the first citation happened, so the true number is "no more than 3 to 4 days." Still. The previous new-domain data point in my set took 64 days. This one, built with the pipeline from day zero, took less than a week.

The result that keeps me honest: the retrofit that changed nothing

Here is the part that would not survive an affiliate blog's edit. If the ping explains fast onset, then adding it to the slow site should help the slow site. On June 21, 2026, I added IndexNow to the family storytelling site, alongside a visible FAQ on June 20 and an entity authority page on June 23. A real intervention, on the same site, with a before and after.

Nothing bent. The site ran 28.1 citations per day across the 14 days before the changes and 30.8 per day across the 21 days after, with the after window running through July 12, 2026. Flat, within noise.

The tempting stat, and I watched it form in my own dashboard, is that 38 percent of the site's lifetime citations landed after the changes shipped. That number is real and it means nothing. A roughly flat daily rate mechanically piles up share-since-launch as days pass. If you ever catch a vendor selling you "X percent of results came after we started," ask for the daily rate instead.

Getting into the cited set and growing inside it are different problems. The ping seems to matter at the front door. It did nothing once the site was already inside.

My current read, held loosely: publish-event pings correlate with fast onset for new pages and new sites, and do not lift the ongoing rate of an already-cited site. Onset is a discovery problem. Volume is something else.

Volume is a demand problem, not a speed problem

Two findings from the same corpus frame what happens after day one.

First, once a page is in, it compounds without any new supply. Between pulls on July 1 and July 13, 2026, the storytelling site added 535 citations across the exact same 18 URLs, zero pages added, zero dropped. The interesting fight is getting a URL into the cited set at all.

Second, the biggest volume event in my data came from shipping fresh pages the day demand arrived. The gaming site's game launched a new season, and on July 8 and 9 the pipeline regenerated every season-dependent page and pinged the index. Daily citations went from 327 per day across June 30 to July 8, to 925 per day across July 9 to 12, with a peak of 1,870 citations across 19 pages on July 9. Earlier in June the same site's baseline was 134 per day. The spike then decayed through 851, 513, 468, and 525 per day, averaging 589 across July 10 to 13, well above the old baseline.

The confound is the point. A season launch is a demand spike, people were suddenly asking those questions, and the data cannot separate fresh supply from new demand. It does not need to. The operational lesson is that the pipeline shipped fresh answers the same day the questions surged, and the index was pinged the moment they shipped. Fresh supply, timed to demand, with the ping wired in. That is the whole trick, and no piece of it works alone.

What this data cannot tell you

Four honest limits, so you can weigh the rest.

• Every number is Bing's AI surfaces. Bing Webmaster Tools is the only counter reporting day-level AI citations, so "cited by AI" here means Bing's blended AI experiences, Copilot included. No engine breakdown exists in the export.
• IndexNow does not reach most of the AI world. Submissions fan out to Bing, Yandex, Seznam, and Naver. Google ignores the protocol, and the crawlers behind ChatGPT, Claude, and Perplexity discover pages on their own schedules. Any onset effect measured here is Bing-only by construction. The protocol details are at indexnow.org.
• Fast onset on one surface does not generalize. An independent corpus instrument tracking Google AI Overviews and ChatGPT mentions showed the gaming site at zero across the same twelve months in which Bing counted thousands of citations for it. That instrument also ingests ChatGPT answers in roughly monthly batches, with a lag of two weeks or more, so it cannot measure day-level onset at all. Different engines, different diets, different clocks.
• This is a case study, not a trial. Four sites, no control group, no p-values. The one same-site intervention returned a null. I publish the numbers because nobody else publishes any, not because n=4 settles anything.

What I would do on publish day

1. Wire the ping into publishing, not beside it. The submission fires when content changes, from the same pipeline that changes it. A daily cron resubmitting your sitemap is better than nothing and worse than the real thing.
2. Ship when the questions surge. The 1,870-citation day happened because the pages regenerated the same day a season launched. If your field has scheduled demand spikes, publishing cadence should track them, not your content calendar.
3. Plan a burn-in before judging anything. My own measurement designs now exclude the first two weeks minimum. A page that looks dead at day 10 may simply not have been found yet.
4. Measure onset and volume separately. "How fast was it cited" and "how much is it cited" respond to different levers. The retrofit null is what happens when you optimize one and measure the other.
5. Write your prediction down first. The only reason I trust my own numbers is that the claims were registered before the pulls. If you change something, write what you expect to happen and by when. Flat is a finding too.

Questions I get about AI citation speed

Where to go deeper

The full playbook from this dataset, How to get cited by AI, covers the upstream questions: which topics can earn citations at all, why page shape beats page depth, and how to build answers an engine can lift. The free GEO course walks the same ground in five short video lessons, including the access audit and how to measure your own AI visibility.

If you run a business where being described accurately by AI matters, that measurement problem is what I build at CredibilityOS. And if you want a second set of eyes on where your own site stands, book a clarity call. One conversation, no pitch.