How Long Do ChatGPT Citations Last? Half of Last Year's Top Finance Pages Now Get None
How long do ChatGPT citations last? Going by the first study I've seen that tracked the same pages for most of a year, not very long. Of the 215 finance pages ChatGPT cited most in October 2025, half get zero ChatGPT citations today, the group as a whole kept 14% of its citation volume, and a lot of the slots they gave up went to the banks and card issuers themselves (Green Flag Digital).
Joe Robison at Green Flag Digital published the rerun on September 21. He'd pulled ChatGPT's most-cited finance pages last fall, and eleven months later he went back to the same pages with the same method to see what was left. It gets at something most AI visibility reports skip over, which is what happens to a citation after you've earned it, so I want to walk through what he found and what I'd do with it.
The short version
Of 215 top-cited finance pages from October 2025, 107 get zero ChatGPT citations now, and only 8 gained citations (Green Flag Digital).
Pages with "2024" or "2025" in the URL did far worse. Together they kept 0.3% of their citations, against 16% for everything else.
Brand-owned pages (banks, card issuers, airlines) went from 1.8% of the top citations to 22.4%, while Wikipedia and Reddit went from 17.9% to zero in this sample.
Backlinks and rankings line up with citations only weakly, so the fix is mostly about owning current, specific answers, with link building a distant second.
How the study worked
Robison didn't query ChatGPT himself. He used Ahrefs Brand Radar, which logs the pages ChatGPT cites across a large set of prompts, and asked it for every page cited in a ChatGPT answer that also cited at least one of seven finance anchors (NerdWallet, Bankrate, Bank of America, Chase, Credit Karma, The Points Guy and Investopedia). The first pull was October 23, 2025. He reran the same query on September 21, 2026, and tracked the 250 most-cited pages from the first round. He dropped NerdWallet URLs that had changed structure during the year, since a moved page looks like a dead one, which left 215.
So this is finance, in the United States, on one vendor's prompt set. I'll come back to what that does and doesn't let you conclude. For now, keep in mind that "citation volume" here means how many times Brand Radar saw a given URL cited across its prompts, which is about as close as any of us can get to a share-of-voice number for a single page.
Half the pages went to zero
Of the 215 pages, 107 get no ChatGPT citations in the new data, which is almost exactly half. Another 100 still get some citations, and only 8 went up. Added together, the whole group kept 14% of the citations it had last October.
The pages that held up were the evergreen ones that get updated. Bankrate's auto loan rates page went from 193 citations to 342, and its best high-yield savings page went from 102 to 198 (Green Flag Digital). Those are pages whose whole job is to be current.
The URL finding is the one I'd forward to whoever runs your content calendar. Twenty of the 215 pages had 2024 or 2025 in the URL. Sixteen of those twenty went to zero, and as a group they kept 0.3% of their citations, against 16% for the pages without a year in the address. My guess (and it's only a guess) is that a year in the URL gives the engine a cheap way to decide a page is stale before it reads a word of it.
ChatGPT citations, 11 months later
Half of the top pages now get zero
Green Flag Digital tracked the 215 finance pages ChatGPT cited most in October 2025 and checked them again in September 2026. Pick a view.
Source: Green Flag Digital, ChatGPT Citation Study (Joe Robison, Sept. 21, 2026), using Ahrefs Brand Radar data from Oct. 23, 2025 and Sept. 21, 2026. U.S. finance pages only.
I wrote back in the spring about the February drop in ChatGPT citations, when the volume came back within a few weeks and the names attached to it didn't. That one played out over five weeks, and this looks like the same kind of turnover stretched over eleven months, with a lot more of it.
Who took the citations
The shift in who gets cited matters more to a brand marketer than the decay does, I think. In October 2025, publishers, affiliates and other sites made up 80.3% of the top citations in this sample, Wikipedia and Reddit had 17.9%, and the banks, card issuers and airlines had 1.8%. In September 2026 the brands had 22.4%, government and regulator pages had 5.3%, and Wikipedia and Reddit had none (Green Flag Digital).
Chase alone went from 4 cited pages to 63 across its subdomains. The Wall Street Journal went to 23, mostly its Buy Side guides.
Who ChatGPT cites for finance questions
The brands took a fifth of the slots
Share of top citations by source type, October 2025 against September 2026. Tap a source type to follow it across both years.
October 2025
September 2026
Source: Green Flag Digital, ChatGPT Citation Study (Sept. 21, 2026), Ahrefs Brand Radar data. Brand-owned pages are banks, card issuers and airlines. U.S. finance only.
That lines up with what I found in my own data, from a different angle. In the AI Verdict Study I logged 32,089 citations from 432 buyer conversations across ChatGPT, Claude, Gemini and Perplexity. When the buyer asked an open question about the category, the brand's own site was 4% of the citations I could classify. When the buyer asked about the brand specifically, things like "what are the tradeoffs" and "will it still fit if my needs grow," the brand's own site jumped to 35% and 37%, and help center and support docs were 29% of everything cited from brand sites (Where AI Engines Go to Decide About Your Brand).
Put the two together and you get a pretty consistent picture. When somebody asks about your product, the engines increasingly go to you for the answer, and they'll take it from whatever page on your site answers it most directly. If that page doesn't exist, they'll take it from somebody else, and that somebody may well be your competitor's comparison page.
Where I'd be careful with these numbers
I like this study a lot, and I'd still push back on anyone who quotes "half of all ChatGPT citations disappear in a year" as a universal fact. Robison is upfront about the limits, and they matter.
Number one, it's one vendor's prompt set. If Ahrefs changed its prompts or the way it collects answers between the two dates, that would look exactly like a change in ChatGPT, and as Robison says, there's no way to separate the two from the outside. Two, it's finance only and U.S. only, and finance is a category where rates and terms change constantly, so freshness may count for more there than in, say, project management software. Then last, it starts from the most-cited pages, and things at the very top have the furthest to fall. A page in the middle of the pack might hold steadier.
I still believe the pattern, with the 50% figure belonging to finance pages on one tracker, and your number could be higher or lower. I went through the five things that move these numbers around in How to Read an AI Visibility Study if you want a checklist for this kind of thing.
The other result I'd put in front of anyone still budgeting for AI visibility like it's 2019 SEO is the correlation work. Domain Rating had no detectable link to citations once everything else was in the model. Ten times the backlinks or ten times the ranking keywords went with about 10% more citations, and links, rankings and authority together explained less than a fifth of the difference between pages (Green Flag Digital). So links help a little, and on this data they explain a small part of which page ChatGPT ends up reading.
How long do ChatGPT citations last for your brand?
You won't know until you check the same thing twice, a few months apart. A lot of AI visibility reporting is a single snapshot that gets treated as current long after it was taken. So this is what I'd do.
Number one, assume any citation you have today could be gone in a year. Pick the prompts and pages that matter most to your pipeline, write them down, and recheck them every quarter with the same prompts so you can see what moved. Robison recommends a quarterly refresh calendar for cited pages, and I'd add the prompt check to the same calendar so the two happen together.
Two, get the year out of your URLs. If you've got "best-savings-accounts-2025" style pages that still earn traffic, move them to a permanent address with a redirect and update the content in place. You can change a year in a title any time you update the page, where a year in the URL sticks around and, going by this data, costs you a lot more.
Three, make sure there's a clear, current page on your own site for every question a buyer would ask about your product. That means pricing, limits, who it's for and who it isn't, how it compares, and what happens when somebody outgrows it. The Chase number and my help-docs number point the same direction, which is that engines go to the brand for the answer when the brand has written the answer down somewhere they can find it.
Then last, keep earning mentions on other sites, because the first round of any AI shortlist still gets built mostly from pages other people wrote about your category. That part hasn't changed much since I wrote The Click That Didn't Happen, and it's slower work than updating your own pages, which is a good reason to start it now.
Two-minute self-check
Will your citations still be there next year?
Answer six questions about how you track and maintain the pages AI engines cite. You'll get a score and the fixes to start with.
If you'd rather see where you stand before building a quarterly process, the free Quick Check will give you a quick read. If you want the prompts, trackers and templates to run the quarterly version yourself, that's what the $199 Audit Kit is for.
What I'd like to hear from you
I'd love to see this kind of before-and-after data for categories outside finance. If you've got citation data on your own pages from late last year and you've rerun it since, tell me what held and what went to zero, especially if it doesn't match what Green Flag found. I'm also curious whether anyone has moved a dated URL to a permanent one and watched what happened to its citations afterward.
A lot of my book, Explainable, is about the idea that AI engines describe brands using whatever evidence is easiest for them to find and trust, and this study is a good reminder that the evidence has to stay current to keep counting.
Jarred Smith is the author of Explainable: Why AI Recommends Some Brands & Ignores Others, an Amazon bestseller on AEO, GEO, and SEO. He's a marketing leader with nearly 20 years of experience across healthcare, public media, retail, and environmental services. Find him at jarredsmith.com.