Key takeaways
- The median AI citation half-life is about 4.5 weeks across platforms. Google AI Overviews cluster around 4.3 to 4.8 weeks, ChatGPT churns fastest at ~3.4 weeks, and Perplexity holds longest at ~5.7 weeks (Scrunch/Stacker analysis of 3.5M citation events).
- In a 23-site field audit from April 2026, the median (query, URL) citation lasted just 19 days in Google AI Overviews, and only 11% of pairs stayed continuously cited across a 30-day window.
- Decay isn't only a page-level problem. Citation slots themselves shrink overnight (ChatGPT's average citations per response dropped ~27% after the GPT-5.3 rollout), and entire domains can lose most of their citation share in days.
- Refresh cadence is now an operational requirement: roughly monthly for Google AI surfaces, biweekly for ChatGPT, six-week cycles for Perplexity.
- Distributed and syndicated content roughly doubles citation durability (4.5 weeks to ~10 weeks half-life in the Stacker network study).
Citation decay is a lease, not a deed
For most of SEO history, ranking behaved like property. You earned a position, and unless something changed, you kept it. Rankings decayed, sure, but over months and years, gradually.
AI Overviews don't work that way. A citation is re-evaluated on every query, against a fresh retrieval pass, and there is no presumption of incumbency. The best mental model I've seen comes from a May 2026 field study by Ranko, which tracked citation persistence across 23 client sites: citation is not a stable state but a continuously re-evaluated one, and the operational story looks more like lease renewal than one-time acquisition.
That study deserves more attention than it got. Capturing AI Overview cited URL sets twice daily across April 2026, they measured how long a page stays in a query's cited set once it gets in. The results:
- Median persistence: 19 days
- 25th percentile: dropped out within 7 days
- Only 11% of (query, URL) pairs remained continuously cited across the full 30-day window
Three drivers explained most early drop-out: a fresher competitor publishing on the same intent, structural drift on the cited page itself, and what the authors call "composer topic drift" inside the engine's understanding of the query.
The uncomfortable implication is that your monthly citation-rate dashboard can look flat while half the underlying URLs churn out underneath it. A site cited on 60% of basket queries this month and 58% next month might have lost half of last month's specific cited URLs along the way. The aggregate hides the churn.
The numbers: how long citations actually last
Several independent datasets converged on similar answers in 2026, which is worth something. When Scrunch, Digital Authority Partners, SISTRIX, and Ranko all land in the same range from different methodologies, the finding is probably real.
Citation half-life by platform
The most-cited primary source is the Scrunch and Stacker collaboration, which analyzed 3.5 million citation events across 120,000+ domains, 8 industries, and 6 AI platforms between September 2025 and March 2026, using survival-curve analysis:
| Platform | Approximate citation half-life | Notes |
|---|---|---|
| ChatGPT | ~3.4 weeks | Fastest churn; rewards very recent content and real-time browsing |
| Google AI Mode | ~4.3 weeks | Tracks with the other Google surfaces |
| Google AI Overviews | ~4.7–4.8 weeks | Cites content ~16 days older than standard organic results on average, but still recency-driven |
| Gemini | ~4.8 weeks | Mild bias toward current-year content |
| Perplexity | ~5.7–5.8 weeks | Most durable; its index retains sources longer once cited |
| Microsoft Copilot | ~34% retained after 4 weeks | Runs on Bing's index; no clean half-life published |
| Claude | No published figure | Leans on stable, well-established sources; freshness is a trust signal more than a ranking lever |
A half-life is a decay rate, not a hard cutoff. A page cited in 10% of relevant AI answers at peak, with a three-week half-life, falls to roughly 5% three weeks later, then 2.5% three weeks after that, until it's effectively invisible.
Retention and churn studies
Other datasets sharpen the picture:
- Digital Authority Partners tracked 1,127 URLs across 5 platforms over six weeks and found an average 28-day citation retention rate of 33%. By platform: Perplexity 44%, Copilot 34%, ChatGPT 31%, AI Overviews 27%, Gemini 11%. A stricter cut of the same data found only 10.6% of URLs persisted across the full 28-day window. The gap between those two figures is methodological (average retention vs. strict same-URL-every-wave persistence), but either way, most citations cycle out within a month.
- SISTRIX's AI Research Index analyzed 82,619 prompts and 1.5 million snapshots over 17 weeks. Google AI Mode swapped 56% of cited domains every week; ChatGPT Search swapped 74% weekly, rising to 85% at the URL level. Crucially, churn held steady at 54–59% across all 17 weeks with no sign of stabilization. This isn't launch-phase wobble. It's a structural property of retrieval-augmented systems.
- A Profound analysis of 240 million citations found 40–60% of cited domains rotating month-to-month for the same query, and 70–90% rotating over six months.
- Search Engine Journal's figure, cited in multiple 2026 playbooks: 70% of pages cited in AI Overviews change citation status within 2–3 months.
One more wrinkle: the median cited page in Google AI Overviews is 14 months old, per the Everything-PR Citation Source Index 2026. That sounds like it contradicts everything above, but it doesn't. Age and freshness-of-update are different things. Pages cited every single month tend to have a median update age of about 6 months, while one-time citation spikes come from pages roughly 2 months old. Engines don't care that your page is old. They care whether it's stale.
Why citations decay faster than rankings
Traditional SEO decay plays out over months to years. AI citation decay can start within days. There are four distinct mechanisms, and it helps to keep them separate because they have different fixes.
1. Recency as a cheap accuracy proxy
Retrieval systems re-rank sources on every query, and recency acts as a cheap proxy for accuracy. Fresher pages routinely displace older, otherwise-equal ones. Roughly half of all AI-cited content is under 13 weeks old, and content under 30 days old earns an estimated 3.2x more AI citations than older pages. Google's own AI surfaces are slightly more forgiving, citing content about 16 days older on average than what appears in standard organic results, but the recency bias is still there.
2. Fan-out changed the citation economy
AI Overviews use query fan-out, issuing multiple related sub-searches across subtopics, which means citations increasingly come from outside the direct top-10 results. Ahrefs measured this directly: top-10 organic results accounted for 76% of AI Overview citations in July 2025, but only 38% by March 2026. Moz's 2026 analysis of 40,000 queries found 88% of Google AI Mode citations are not in the organic top 10 at all. Ranking protects you less than it used to, which means the citation you hold is more exposed to displacement by any page the fan-out surfaces, not just your direct SERP competitors.
3. Platform-level volatility you can't control
This is the part most decay write-ups miss. Citation slots themselves are platform-controlled and can shrink overnight, independent of anything you did.
Around the GPT-5.3 rollout on March 4, 2026, average citations per ChatGPT search-enabled response dropped from ~6.4 sources to ~4.7–4.9, a roughly 27% reduction in citation slots, with no recovery a month later, per Promptwatch's citation drop data. If your ChatGPT traffic fell in March 2026, some of that was fewer slots, not worse content. Before attributing loss to decay, check whether the platform shrank the pie.
Slot counts differ wildly by engine, too. Promptwatch's sources-per-response data shows Google AI Overviews averaging ~10 sources per answer (roughly double ChatGPT's ~5) and staying notably steady over time, while Copilot swings from under 2 to nearly 17 sources within weeks. More slots per answer is one reason AI Overview citations feel slightly more stable than ChatGPT's.
Domain-level share can also collapse fast. Reddit's share of ChatGPT Search citations fell from ~3.8% to under 1% in a matter of days in August 2026, while its AI Overviews share declined only gradually over the same period, per Promptwatch's Reddit citation research. If you'd built your GEO strategy on Reddit citations, that was a bad week.
4. Format churn
Even content formats decay in and out of favor. In Google AI Overviews during July 2026, product pages overtook listicles as the most-cited format from July 28 onward, ending the month at 17.9% vs 16.2%, with listicles down from ~26% averages in Q1, per Promptwatch's AI Overviews citation-type data. A page can hold its quality constant and still lose citations because the engine's preference for its format shifted underneath it.
What actually extends citation life
The good news is that the same datasets pointing at the problem point at fixes. Here's what the evidence supports.
Refresh on a cadence matched to the platform
A page with a 3–4 week half-life needs a refresh roughly every month just to hold position. A page with a 10-week half-life can run on a quarterly cadence. Practical guidance from the Scrunch/Stacker-derived research:
| Platform | Half-life | Recommended refresh cadence |
|---|---|---|
| ChatGPT | ~3.4 weeks | Biweekly |
| Google AI Overviews / AI Mode / Gemini | ~4.3–4.8 weeks | Roughly monthly |
| Perplexity | ~5.7 weeks | Six-week cycle |
And a refresh doesn't mean a rewrite. A new statistic, a corrected claim, an updated example, or a visible timestamp change is enough to reset the freshness clock. For high-priority commercial pages (product, comparison, pricing), a 60–90 day cycle is the floor. Reference and definition pages can run annually.
Get distributed, not just published
The single biggest durability lever in the data: content on domains in the Stacker publisher network had a citation half-life of nearly 10 weeks vs 4.5 weeks for non-network domains, a 2.1x advantage that held across all 6 platforms and 8 industries. Earned, editorially syndicated distribution nearly doubles how long citations persist. The best measured combination was Perplexity plus editorial distribution, at a 12.3-week half-life.
The mechanism is corroboration. When the same claim, stat, or framework appears across multiple credible domains, engines keep retrieving it because multiple sources validate it. A page that exists only on your domain has one point of failure.
Fix the substance, not just the timestamp
Decay severity correlates with what Searchbloom calls Information Gain Score. Higher-substance content holds citations longer. News and trending content decays in days; technical reference content decays over months; brand pages decay slowest. Definitions and foundational frameworks survive past the 365-day mark; market analysis and tactical guides rarely do.
Two structural findings worth acting on:
- Schema-marked pages are cited 2.3x more often in AI Overviews (Everything-PR Source Index 2026).
- Word count has essentially zero correlation with citation position (Spearman r = 0.04 in Ahrefs' study of 174,048 pages). The average cited page runs 1,282 words, and 53.4% of cited pages are under 1,000. Density beats length.
Also worth knowing: 44.2% of AI citations are extracted from the first 30% of a page. If your key claims are buried at the bottom, you're making the engine work harder than it needs to.
A practical retention workflow
Here's how I'd operationalize all of this if I ran a content team today:
-
Track citation churn weekly, not monthly. Monthly citation-rate dashboards hide the churn that matters. You need to see which specific (query, URL) pairs dropped out, not just your aggregate share. Remember the Ranko finding: 25% of citations vanish inside 7 days.
-
Segment your pages by decay risk. Commercial and tactical pages on a 60–90 day refresh cycle. High-value money pages monthly. Reference pages annually. Match the cadence to the half-life, not to your content calendar's convenience.
-
Watch platform-level changes before blaming your content. If citations dropped across the board on one engine in one week, check whether slot counts or format preferences shifted. The GPT-5.3 citation drop and the Reddit collapse are both examples of platform behavior masquerading as content failure.
-
Build corroboration deliberately. Every major claim on a money page should exist somewhere else credible too, through digital PR, expert commentary, or syndication. This is the 2.1x durability play.
-
Measure duration, not just acquisition. A page cited for 25 days produces more branded search uplift, referral clicks, and training exposure than two pages each cited for 7 days. Median citation duration is the metric to optimize.
Tools for tracking citation decay
You can't manage decay you can't see. A monitoring tool that tracks specific cited URLs over time, per platform, is now table stakes for any serious GEO program. Promptwatch is what we use for this: its citation trends classify every citation into content and source types with ramp-up, peak, and decay phases per cited page, which is exactly the decay curve you need, and its crawler logs show whether AI systems are even re-reading your refreshed pages.
A few alternatives worth comparing, depending on budget and needs:
| Tool | Starting price | Strength for decay tracking |
|---|---|---|
| Promptwatch | $95/mo | Citation trends with per-page decay phases, crawler logs, content gap analysis, automated refresh workflows |
| Otterly.AI | $29/mo | Cheapest entry point; daily tracking across ChatGPT, AI Overviews, Perplexity, Copilot |
| Peec AI | ~$80/mo | Cost-effective visibility tracking, 3,000+ customers |
| Scrunch AI | $250/mo | The team behind the half-life research; strong on citation persistence analytics |
| Profound | $99/mo | Enterprise-grade, huge citation dataset (3.25B citations analyzed) |
| Ahrefs Brand Radar | $199/mo + base plan | Good if you're already in the Ahrefs ecosystem and want AI + traditional in one place |
Otterly.AI
Profound
If you're comparing the full category, the GEO software directory at bestgeosoftware.com and the AI rank tracking tools directory at ai-rank-tools.com both maintain current listings.
Common questions
Do AI Overview citations ever last years?
Rarely, and only for specific content types. Definitions, foundational frameworks, and evergreen reference material can survive past 365 days. Market analysis, tactical guides, and anything price- or product-related almost never do. The 14-month median page age in AI Overviews reflects pages that have been maintained, not pages that were published once and forgotten.
Is citation decay the same as ranking decay?
No. Rankings decay over months to years; AI citations can start decaying within days and follow a per-platform half-life curve. Also, ranking protects citations less than it used to: only 38% of AI Overview citations come from top-10 organic results as of March 2026, down from 76% eight months earlier.
If my citations dropped suddenly, is it always decay?
No, and this is the trap. Check platform-level changes first. Citation slots shrank ~27% on ChatGPT overnight in March 2026. Reddit lost 86% of its relative citation share on ChatGPT in days in August 2026. Both would look like content decay on a naive dashboard. Segment by platform and check whether the whole pie moved before you rewrite anything.
Does updating the date alone help?
A visible timestamp change contributes to resetting the freshness signal, but the durable effect comes from substantive updates: new data, corrected claims, updated examples. Engines and readers both penalize fake refreshes eventually. The evidence points to update substance, with the timestamp as a supporting signal.
The bottom line
Citation decay in 2026 is fast, structural, and measurable. Median durations run 19 to 33 days depending on the dataset, half-lives cluster around 3.4 to 5.8 weeks depending on the platform, and weekly domain churn of 56–74% appears to be a permanent property of how these systems retrieve sources. None of that is a reason to despair. It's a reason to change what you measure (duration, not just acquisition), how often you refresh (monthly for Google surfaces, not annually), and where your claims live (distributed and corroborated, not isolated on one domain). The teams that treat citations as a renewable lease will keep them. The teams that treat them as a deed will watch their dashboards look stable while their portfolio quietly empties.