Key takeaways
- Profound's 12-month study of 883,000 pages found the median page loses half its citation share 11 days after peaking, and 78% of cited pages halve within two weeks.
- Different studies measure different things. Scrunch and Stacker found a roughly 4.5-week half-life for groups of sources, and Promptwatch Data found ChatGPT keeps citing landing pages and docs for months.
- Content type matters most. News fades fastest, while landing pages, docs, comparisons and listicles hold up.
- Being cited by many engines predicts staying power, and a dip on every engine on the same day is often a platform change, not your content.
- Check decay with a repeatable routine: baseline, weekly tracking by page, a cause check, then refresh or replace.
The short answer
AI citations do not last long, but "how long" depends on what you measure. If you measure one page's share of citations in a topic, the answer from Profound is brutal: 11 days from peak to half. If you measure whether ChatGPT cites a page again at all, the answer from Promptwatch Data is much kinder: months for the right kinds of pages.
Both are true. They are answering different questions, and mixing them up is how people end up either panicking or getting complacent.
This guide covers what the main studies found, why the numbers differ, which of your pages are most at risk, and a practical routine for checking your own pages for decay.
What citation decay actually means
Citation decay is the drop in how often an AI answer engine cites a page after that page has peaked. In classic SEO, a ranking is something you hold. A page earns links over months and sits in position three for a year. AI answers do not work like that. The engine re-decides what to cite on every response, using a fresh retrieval step each time, so a citation is closer to a lease than a deed.
That matters because decay is quiet. The page still loads. Backlinks are intact. Google traffic might even be steady. But the answers in ChatGPT, Perplexity or AI Overviews have moved on to other sources, and nothing in a standard rank tracker tells you.
What the research says
Three datasets are worth knowing. They do not measure the same thing, so here they are side by side.
| Study | What it measures | Sample | Headline result |
|---|---|---|---|
| Profound (Sep 2025 to Sep 2026) | Days for one page to fall to half its peak citation share and stay there | 883,000 pages, 7 engines | Median 11 days after peak; 42 days from first citation |
| Scrunch and Stacker (Sep 2025 to Mar 2026) | Weeks until a cohort of sources loses half its citations | 3.5 million citation events | About 4.5 weeks overall; ChatGPT 3.4, Perplexity 5.7 to 5.8 |
| Promptwatch Data (Jul 6 to Oct 4, 2026) | Share of ChatGPT-cited pages still cited again later | Every classified page ChatGPT cited on July 6 | News halves in 36 days; product pages, reviews and how-tos in 74 to 77 days; landing pages and docs likely beyond 90 |
The figures are not interchangeable. Profound tracks a smoothed share of citations inside a category, so a page can fall to half its share while still being cited. Promptwatch's measure asks whether the page shows up again at all. An 11-day half-share and a 75-day half-life for product pages can both be accurate.
Profound: most citations are single events
Profound measured 1.19 million page and engine lifecycles across ChatGPT, Perplexity, Claude, Google AI Overviews, Google AI Mode, Gemini and Microsoft Copilot. Their main findings:
- 78% of cited pages fall to half their peak share within two weeks of peaking. Most arrive in one burst and are never cited at that rate again.
- The remaining 22% last longer, with a median half-life of 27 to 29 days.
- Among those longer-lasting pages, a Claude citation is 2.8 times as likely as a ChatGPT citation to still be relevant eight weeks later.
- Decay on one engine barely predicts decay on another. Cross-engine correlations run between 0.03 and 0.09.
- Pages cited by five or more engines were 5.5 times as likely as pages cited by one to still be above half their peak share after eight weeks.
That last one is worth sitting with. It reads like a content quality signal. Pages that several engines independently want to cite tend to be the substantive ones, and they also take longer to build up, with a median rise of 46 days versus 19 days for single-engine pages.
Two caveats from Profound's methodology are worth knowing. Shares are smoothed over a trailing 14-day window, so a single-burst page can hold its share for about two weeks by construction, and that is why so many land right at or below 14 days. And pages tied to a date or event fade first, which fits what you would expect.
Scrunch and Stacker: a month for a cohort
The Scrunch and Stacker work looked at groups of sources that first appeared in the same week and asked how long until they lose half their citations. The answer was about 4.5 weeks overall, with ChatGPT turning over fastest and Perplexity holding on longest. Scrunch is also a tool you can use to monitor this yourself.
Be a bit careful with the smaller numbers floating around. A Reddit post claiming 62% of month-one sources were gone by month three, and a claim that pages updated within two months earn about 28% more citations, are both third-party figures I could not verify. Treat them as directional at best.
Promptwatch Data: content type changes everything
Promptwatch Data followed every classified page ChatGPT cited on July 6, 2026 for 90 days and checked what share was still cited later. The results by content type:
| Content type | Still cited after 30 days | Still cited after 60 days | Notes |
|---|---|---|---|
| Landing pages | 78% | 69% | Half-life likely close to or beyond 90 days |
| Docs | 76% | 67% | 93% were cited again after day 0 |
| Comparisons | 72% | 63% | Keep over 60% after two months |
| Listicles | n/a | 60% | Likely long half-life |
| Product pages, reviews, how-tos | n/a | about 55% | Half-life of 74 to 77 days |
| News | n/a | 41% | Half gone after 36 days; almost a quarter never cited again |
One pitfall from that report: lines drop faster in the final week purely because the window ends and there are fewer days left for a page to be cited again. Do not read that last-week cliff as real decay.
Why the numbers differ, and which one to use
It comes down to the question. If your question is "will this page keep winning the same share of answers?", use Profound's framing: expect a spike and a fade, and plan for sustained pages to last four weeks or so. If your question is "will this page still be in the pool ChatGPT draws from?", Promptwatch's framing is more useful: evergreen commercial and reference pages stay in the pool for months.
In practice, you want both. Share is what drives how many answers you appear in. Presence is whether you have any shot at all.
Which pages decay fastest
The pattern across studies is consistent enough to act on.
- News and press releases fade first. Promptwatch Data shows half of news pages gone within 36 days. Profound found date-tied and event-tied pages fade first on every engine. Use them for bursts of visibility and do not expect them to compound.
- Product pages are cited a lot but do not stay cited on their own. In Promptwatch's citation share by content type report, product pages are the most cited type in ChatGPT at 22.4% (Sep 28 to Oct 4, 2026), yet they reach half-life around 74 to 77 days. Stale specs, prices and availability are the likely cause of losing them.
- Landing, docs, comparison and listicle pages are the durable ones.
- Single-engine winners are fragile. If only one engine cites a page, treat it as a burst waiting to end.
There is also a ceiling effect worth knowing about. Promptwatch's domain rank data shows sites with DR 46 to 75 earned almost half of ChatGPT citations in August 2026, so you do not need a huge domain to be in the pool. You do need to keep re-earning your place.
Is it decay, or did the platform change?
Before you rewrite anything, check whether the drop is yours. Platform shifts can look exactly like decay.
Promptwatch's citation drop report is the clearest example. After the GPT-5.3 rollout on March 4, 2026, average sources per search-enabled response fell from about 6.4 to roughly 4.7 to 4.9, around 27% fewer, across GPT-5.3, GPT-5.4 and GPT-5-Mini within a day. There was no recovery a month later. Every site saw fewer citations, and none of it was a content problem.
Other shifts in 2026 worth cross-checking against your dates:
- ChatGPT started using the site: operator at scale on August 8, 2026, per Promptwatch Data, which changed what gets retrieved.
- Reddit's share of ChatGPT citations fell from roughly 4% to 0.5% on August 14, 2026, according to Promptwatch's Reddit report.
A quick test: if the drop hits many unrelated pages on the same day, and your competitors lost citations too, suspect the platform. If it hits one page or one topic and a competitor page replaced you, suspect decay.
How to check your pages for citation decay
Here is a routine that works without special tooling, then how tools make it easier.
Step 1: build a baseline
List your 20 to 50 most important pages. For each one, record the prompts you expect it to be cited for, which engines currently cite it, and the date it was first cited if you know it. You cannot spot decay without a starting point.
Step 2: track by page, not just by brand
A brand-level visibility score hides decay. One strong page can mask five fading ones. You want per-page citation counts over time, per engine, since Profound found engines barely move together.
Step 3: measure peak and drop
For each page, note its peak week and compute how far it has fallen from that peak. A rough alert rule, borrowed from Profound's approach: flag any page whose share has fallen to half its peak for two consecutive weeks. Smooth over at least a 14-day window so daily noise does not trigger false alarms.
Step 4: classify the cause
| Pattern | Likely cause | What to do |
|---|---|---|
| Many pages drop on one day, competitors too | Platform or model change | Wait, log it, do not rewrite |
| One page drops after a burst | Normal single-event decay | Accept it or build a follow-up asset |
| A page drops and a competitor page now gets cited | Competitor freshness or better answer | Refresh and compare |
| Page is cited but facts are outdated | Stale prices, specs, claims | Update and re-date |
| Crawlers stopped hitting the page | Technical or access issue | Check crawler logs and robots rules |
Step 5: refresh or replace
Refreshing makes sense for durable page types: landing pages, docs, comparisons, listicles and product pages. Replacing or republishing makes sense for news-like pages whose moment has passed. Update the facts that actually changed, not just the date stamp.
Tools that help
You can run the routine in a spreadsheet for a handful of prompts. Past that it gets tedious, and the tool should do per-page and per-engine tracking.
Promptwatch is the one I would reach for first, because it ties the pieces together: page-level citation tracking, IP-verified AI crawler logs so you can tell whether a bot stopped visiting, and Citation Trends that show first cited date, ramp-up, peak and decay half-life per cited page. That last one is nearly the exact measure this guide is about. It also has an Action Board and a Content Refresh tool that flags what to update, so the finding turns into a to-do rather than a chart.
Other options in the catalog, depending on what you need:
| Tool | Good for | Limitation to know |
|---|---|---|
| Profound | Enterprise tracking and the citation decay metric behind the 11-day study | Higher price point |
| Scrunch | Monitoring how AI assistants describe and cite you | Monitoring-focused |
| Otterly.AI | Cheap, simple prompt tracking | No crawler logs or content workflow |
| Peec AI | Straightforward AI search monitoring | No optimization layer |
| Rankshift | Citation and visibility tracking for AI answers | Tracking-oriented |
If you want to compare more options, the GEO software directory at bestgeosoftware.com lists many of them. For free checks, Google Search Console is still useful for seeing whether a page is losing classic search traffic at the same time.
What to do about it
Some practical moves that follow from the data:
- Treat landing, docs and pricing pages as long-term assets and keep prices and features accurate. This is Promptwatch's own recommendation from the half-life report.
- Do not assume product pages stay cited by themselves. Keep specs, availability and prices current.
- Invest in comparison pages and listicles, which keep more than 60% of their citations after two months.
- Use news and press for short bursts, and plan the next burst before the last one ends.
- Aim to be cited by multiple engines. Profound's 5.5x finding is the strongest argument for broad, substantive pages over pages tuned to one engine's quirks.
- Track per engine. A Claude citation lasts differently from a ChatGPT one, and one engine's decay tells you almost nothing about another's.
- Monitor continuously. Single-snapshot audits miss the decay entirely.
If you would rather hand this to specialists, an agency like 1001 SEO Media runs this kind of ongoing citation monitoring and refresh work as part of its AI search and GEO services.
Frequently asked questions
Do AI citations expire on a fixed timer?
No. They are re-decided on every response. What the studies show is that, on average, citations turn over within weeks, so winning once is not enough.
Is the half-life 11 days, 4.5 weeks or 75 days?
All three, depending on the measure. 11 days is a single page's share of citations falling to half (Profound). 4.5 weeks is a cohort of sources losing half its citations (Scrunch and Stacker). 74 to 77 days is when product pages, reviews and how-tos in ChatGPT drop to half still-cited (Promptwatch Data).
Which engine cites for the longest?
It depends on the study. Profound found Claude citations more durable than ChatGPT among sustained pages, and Scrunch and Stacker found Perplexity holding citations longest and ChatGPT turning over fastest. The datasets differ in method and period, so test on your own pages.
Should I republish old content or update it?
Update durable pages with real changes. For time-bound pages, publish a new one. Changing only the date does not fix stale facts.
Bottom line
Assume any single citation is a short burst, then build the pages that outlast it: docs, landing pages, comparisons and listicles, cited by more than one engine and kept accurate. Check them weekly, rule out platform changes before you rewrite, and refresh what has really gone stale. For tracking, start with a per-page view like the one in Promptwatch, and keep the research dates in mind, since all of these numbers move month to month.