A while back I started watching a thing I hadn't been watching before. Not whether a client's page got cited in an AI answer. Whether it stayed cited.
So I'd check a prompt, see a client's page show up in the response, note it, and come back a few weeks later. Sometimes it was still there. Plenty of times it wasn't. Same page, same prompt, gone. Something newer had taken its seat.
That gap is the whole point of this piece. We've spent a year teaching people how to get cited by AI at all. The next question, the one almost nobody's asking yet, is how long that citation lasts before the engine drops it and reaches for something fresher.
Honestly, the first time you watch a page you were proud of disappear from an answer set, it reframes how you think about the work. The citation was never the finish line. It was a seat you have to keep earning.
What is AI citation shelf life?
AI citation shelf life is how long your page keeps showing up in AI answers before the engine drops it and replaces it with something else. Not whether you get cited. How long the citation lasts.
Here's the number that makes the point. In Writesonic data reported by Foundation in June 2026, 44% of cited pages appeared exactly once and then disappeared from the answer set entirely. Cited, then gone, never seen again. That's not a published Writesonic study, to be clear. It's vendor data shared with and reported by Foundation, so treat it as a relayed figure and a spring-2026 snapshot, not a permanent law. But the direction it points is the part that matters, and it lines up with stronger primary research I'll get to in a second.
This is the reframe. Frequency is the vanity metric. "We got cited" feels like a win, and it's cheap, and it's common, and roughly half the time it never repeats. Durability is the real asset. The question worth tracking isn't "did we show up." It's "are we still showing up four weeks from now." If you're treating that first citation as the goal, you're celebrating a rental and calling it ownership.
How long does an AI citation actually last?
Weeks, not months. That's the honest headline, and the best primary data we have on it comes from the citation half-life study Scrunch ran and co-published on Stacker.
Worth saying up front that any single snapshot of AI behavior ages fast, which is a problem that runs well past citations. It's the same reason the published AI traffic benchmarks don't agree with each other, and why a figure quoted without its date is close to useless.
That study, published March 2026 on Stacker, looked at more than 3 million citation events across 120,000-plus domains over a 26-week window from September 2025 to March 2026. The overall half-life of a citation came out to roughly 4.5 weeks. Meaning citation activity for a typical source drops by half in about a month.
It splits hard by platform:
- ChatGPT: about 3.4 weeks. The shortest. It cycles its sources fastest.
- Gemini: about 4.6 weeks.
- Perplexity: about 5.7 weeks. The longest. It holds onto a source for most of a quarter.
So where you get cited changes how long the citation lives. A page that fades from ChatGPT in a month can hang on in Perplexity for closer to six weeks. Platform can matter as much as anything you do to the page itself.
The relayed vendor data points the same direction. In the same Foundation report from June 2026, Writesonic's data put the typical citation lifespan at 11 to 15 days and found Perplexity holding sources roughly twice as long as ChatGPT. Different study, different method, different numbers. But the shape is identical, and that Perplexity-versus-ChatGPT gap is the strongest cross-check in the whole picture: Scrunch's 5.7 weeks versus 3.4 weeks is about a 1.7x spread, which squares cleanly with "about twice as long."
Be clear-eyed about the precision here. The exact numbers vary by study, by platform, and by the month you measured. They'll keep shifting as the models update. I wouldn't bet a strategy on any single decimal. What I'd bet on is the direction, because every credible source agrees on it: AI citations decay in weeks, not months. The seat you won in March is up for grabs again by April. What's steadier than any single seat is whether you own the topic even as individual citations turn over.
Google's AI Overviews doesn't play by these rules
The half-lives above all come from chat surfaces. Google's AI Overviews behaves differently, and the difference matters more than it first looks.
Ahrefs' July 2025 study of roughly 17 million cited URLs compared the age of content cited by AI assistants against the age of content ranking in organic search. Every assistant leaned fresher, with one exception. AI Overviews cited content averaging 1,432 days since publication, against 1,416 days for the organic results it sits above. That's 16 days older.
Sixteen days out of nearly four years is about one percent. Read that as a preference for older content and you'll draw the wrong conclusion. Read it properly and it says something more useful: on freshness, AI Overviews is indistinguishable from the organic results underneath it.
So shelf life isn't a property of AI search. It's a property of the surface. Chat surfaces cycle their sources fast, and a citation there is genuinely perishable. Google's AI surface shows no freshness tilt at all, which means a page holding its seat there is holding it for the reasons organic search has always rewarded.
That splits the refresh question in two, which is where most freshness advice goes wrong.
What makes a citation last?
Here's where it gets useful. If most citations fade in a month, what separates the ones that stick around?
Foundation's read on the durable citations, the ones that kept getting re-cited, is that they almost always carry one or more of four things. This isn't a measured percentage. It's practitioner judgment, a pattern someone watched closely enough to name. But it's a good pattern, and it matches what I see, so it's worth leaning on.
Foundation's synthesis (June 2026): durable AI citations almost always carry at least one of four things. Original data. Named-expert commentary. Deep specificity. Strong third-party validation. Pages with none of these tend to be the ones that show up once and vanish. The ones that keep their seat almost always have at least one, and the best have several.
There's a thread running through all four: a durable citation is one the engine can't easily replace. And replaceability is exactly where freshness comes in. Ahrefs research from July 2025 found AI-cited content averages 1,064 days since publication against 1,416 days for organic results, close to a year fresher. ChatGPT runs furthest ahead of the pack: its end-of-answer citations are about 458 days newer than organic, its in-text references about 393 days newer. That 1,064-day average covers ChatGPT, Perplexity, Copilot and Gemini. AI Overviews isn't in it, for the reason covered above. Stale pages get swapped for newer ones. Freshness and shelf life are two views of the same behavior: a page that never gets refreshed slides out of the answer set, and a page nobody else vouches for is the easiest one to drop.
That last attribute, third-party validation, is the one most teams underweight, and it's where the off-site and earned side of AI citations does the heavy lifting. A single page on your own domain is a single thing that can be replaced. Mentions distributed across the open web are harder to dislodge, because they teach the model an association, not just a URL.
What to do about it, so your citations don't expire
None of this is exotic. It's the same content discipline that earns the first citation, pointed at keeping it. A handful of moves:
- Publish original data and research. First-party numbers and patterns from real work give an engine a reason to keep coming back to you as the source. This is the durability move that pays off most, and it's the one most sites skip.
- Put real names and expertise on the work. Bylines matter. Anonymous content is easier to drop than a named expert's judgment. Make it obvious who's speaking and why their take is worth quoting.
- Get specific. Name the use cases, the industries, the company sizes. The concrete details generic content avoids are the details that keep a page getting pulled for the questions it actually answers.
- Build off-site validation. Earned mentions and a distributed presence across the open web outlast a single page on your own domain. This is the durability lever hiding in plain sight, and it's the off-site, earned-citation work most content plans ignore.
- Refresh on a cadence, and set it per surface. Decide up front whether a piece is evergreen, refreshed quarterly, or year-stamped, and then actually do it. But make that call against the surface you're trying to hold. If the page's job is staying cited in chat answers, you're up against a weeks-long half-life and a real freshness tilt, and a genuine refresh cycle earns its keep. If the page's job is holding a spot in AI Overviews, freshness buys you close to nothing, and that effort belongs on whatever got the page ranking in the first place. Ahrefs' own read on their data is worth repeating here: the average AI-cited page is still 2.9 years old, Google is the platform least moved by freshness, and most brands get more from publishing new work than from pouring resources into constant updating. Publish-and-forget still loses you the chat surfaces. Refresh-everything-constantly just costs you the time.
Two of those have peer-reviewed work behind them. A 2024 study out of KDD tested nine different content changes against a benchmark of 10,000 queries, measuring how much of a generated answer each source ended up owning. Adding relevant statistics moved that number 34%. Adding quotations from credible sources moved it 41%. Keyword stuffing moved it backwards.
Two cautions on reading that. It ran against a simulated answer engine rather than live ones, and its headline metric tracks how much of an answer a source occupies, not whether that source still gets cited next month. So it isn't durability evidence. It's evidence that the things making a page worth citing once are the things you can actually control.
The bigger frame here is that this is one discipline, not five tactics. Earning durable AI citations is the same job as showing up in AI search in the first place: be genuinely worth citing, and be current. Those two levers earn the first citation and they keep it. And before any of this matters, you have to know whether AI can see and cite you at all. "Are we still cited?" is the question that comes right after "are we cited?" Both of those are also what protect the commercial payoff, because a citation only sends you a ready buyer off an AI answer for as long as it lasts.
A citation you earn once and never reinforce is a rental, not an asset. It expires. The data is consistent on that, even as the exact numbers move: AI citations decay in weeks, and roughly half of them never come back after a single appearance. Durability is the real win, and durability comes from being genuinely worth re-citing, not from getting picked up once and going quiet.
Building content that stays cited is a discipline, not a one-time fix. It's original data, named expertise, real specificity, off-site validation, and a refresh cadence you actually keep, aimed at the surfaces where refreshing changes anything. That's the kind of thing an engagement is built around, across search and AI as one job, because being found in Google and being surfaced by AI assistants are the same work.
If you're getting cited in AI answers but you're not sure the citations are sticking, or you don't know whether you're showing up at all, that's exactly the kind of thing worth a conversation. Keeping your seat in the answer is a discipline an engagement is built to run, across search and AI together. Take a look at what I help with, and if it sounds like your situation, let's talk.
Getting cited once is a rental. Staying cited is the asset, and the rent comes due in weeks.
Frequently asked questions
Weeks, not months. The best primary research, a Scrunch study co-published on Stacker in March 2026 across more than 3 million citation events, put the average citation half-life around 4.5 weeks. It splits by platform: ChatGPT about 3.4 weeks, Gemini about 4.6, Perplexity about 5.7. These describe spring-2026 behavior and will shift as models update, but the direction is steady: citations decay fast.
Not the way they do in chat answers. Google's AI Overviews shows no meaningful preference for fresh content. Ahrefs found it cites pages averaging 1,432 days since publication, 16 days older than the organic results it appears above, and a gap that small is no preference at all. The weeks-long half-lives described above are chat-surface behavior. Refreshing a page to hold an AI Overviews citation is effort aimed at a signal that surface doesn't appear to weigh.
AI citation shelf life is how long your page keeps appearing in AI answers before the engine drops it and replaces it with something else. It's a measure of durability, not frequency. Getting cited once is common and cheap. Staying cited is the part that's hard and the part that's worth it, since citations decay in weeks and many never repeat after a single appearance.
Most likely your citation simply expired. AI citations have a short shelf life, weeks on average, and chat engines tilt toward fresher content, so a page that isn't refreshed or reinforced gets swapped for something newer. Google's AI Overviews is the exception and shows no freshness tilt, so a disappearance there usually tracks a ranking change instead. In Writesonic data reported by Foundation in June 2026, 44% of cited pages appeared once and then vanished. Disappearing isn't a penalty. It's the default unless you earn the re-citation.
Per Foundation's read on durable citations in June 2026, the content that keeps getting re-cited almost always carries at least one of four things: original data only you have, named-expert commentary, deep specificity (real use cases, industries, company sizes), or strong third-party validation off your own domain. The common thread is that it's hard to replace. Generic, anonymous, single-source pages get swapped out first.
It depends on the surface. Ahrefs research from July 2025 found AI-cited content averages 1,064 days since publication against 1,416 days for organic results, with ChatGPT's end-of-answer citations running about 458 days newer. On chat surfaces, where citations decay in weeks, a real refresh cadence is how you keep your seat. Google's AI Overviews is the exception: it cites pages averaging 1,432 days old, 16 days older than the organic results it appears above, so freshness buys you almost nothing there. Refresh for the chat surfaces. And don't only change the publish date. Update the substance.
Sources
- Scrunch, "The Half-Life of AI Citations," co-published on Stacker (March 2026). Primary source for the citation half-life data: about 4.5 weeks overall; ChatGPT ~3.4, Gemini ~4.6, Perplexity ~5.7 weeks; 3 million-plus citation events, September 2025 to March 2026.
- Foundation (Ethan Crump), June 2026. Relay for the Writesonic vendor data cited here (44% of pages cited once then gone; 11 to 15 day typical lifespan; Perplexity holding sources about twice as long as ChatGPT) and for the four durability attributes, which are Foundation's editorial synthesis, not a measured statistic.
- Ahrefs, "Do AI Assistants Prefer to Cite Fresher Content," July 2025. Source for the freshness figures: 16.975 million cited URLs across seven surfaces. AI assistants average 1,064 days since publication against 1,416 for organic; ChatGPT's end-of-answer citations about 458 days newer and its in-text references about 393 days newer; Google's AI Overviews 1,432 days, or 16 days older than organic. The study reports the assistant-versus-organic gap as 25.7%, calculated against a 1,432-day baseline that its own platform table assigns to AI Overviews rather than to organic, so the day counts are quoted here in preference to the percentage.
- Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, "GEO: Generative Engine Optimization," KDD 2024. Source for the content-change figures: GEO-bench, 10,000 queries; Quotation Addition +41% and Statistics Addition +34%, both measured on Position-Adjusted Word Count, the authors' constructed proxy for how much of a generated response is attributable to a source, weighted by position. That metric is not a citation count and not a retention measure, and the tests ran against a simulated answer engine rather than live ones.