Key takeaways
- An AI citation is not a position you hold. Profound's 2026 study of 1.19 million page and engine lifecycles found the median page loses half its citation share within 11 days of peaking.
- About 78% of cited pages are single bursts. The other 22% are sustained pages with a median half-life of 27 to 29 days.
- Engines do not decay in sync. Cross-engine correlations run between 0.03 and 0.09, so you need per-engine, per-page curves.
- Content type matters: in Promptwatch Data, landing pages and docs stay cited longest, while news loses half its pages within 36 days.
- Read the curve as a trigger for refresh work, not a calendar. Decay with no competitor replacing you and no on-page change is the signal to act.
What a citation trend actually is
A citation trend is the history of how often one URL gets cited by one AI engine over time. Plot it and you usually get the same shape: a first citation, a ramp-up, a peak, then decay. The vocabulary is borrowed from epidemiology and radioactive decay, and it fits better than you would expect.
Four numbers describe nearly every curve:
- First cited date: the first day an engine used the page as a source.
- Rise: how many days it takes to get from first citation to peak.
- Peak: the highest point of the (usually smoothed) citation share.
- Half-life: days from peak until the page falls to half that level and stays there.
The reason this matters is that most teams still think of citations like rankings. In SEO, you win a position and hold it for months. The 2026 data says AI search does not work that way. Authority there expires more than it accumulates.
What the 2026 research says about shape and speed
Most citations are single events
Profound's September 2026 study measured 883,000 pages across ChatGPT, Perplexity, Claude, Google AI Overviews, Google AI Mode, Gemini, and Copilot over twelve months. Each page on each engine counted separately, which gave 1.19 million lifecycles. The headline: the median page loses half its citation share 11 days after peaking, and the median time from first citation to half-peak is 42 days.
The split underneath is what I find most useful. Of 1.11 million pairings observed for at least two weeks past peak, 78% had a half-life of 14 days or less. Those are burst pages: cited in a short window, never at that rate again. The rest are sustained pages, with a median half-life of 27 to 29 days and a median rise of 36 to 46 days.
The explanation is the size of the source pool. A human saw ten blue links. An answer engine draws from a pool orders of magnitude wider, so the odds that any specific page gets picked a second time are low.
Engines do not move together
Among sustained pages, a Claude citation is 2.8 times as likely as a ChatGPT citation to still carry half its peak share eight weeks later. Perplexity and Claude return to a page repeatedly, so their citations spread across weeks. ChatGPT, Google AI Mode, and Gemini tend to cite a page intensively for a few days and then stop.
The practical consequence: a page decaying on ChatGPT tells you almost nothing about Claude. Profound found cross-engine correlations of only 0.03 to 0.09. One blended "AI visibility" line hides most of the story.
Breadth predicts longevity
| Engines citing the page | Median rise | Still above half-peak at day 56 |
|---|---|---|
| 1 | 19 days | 1.2% |
| 2 | 31 days | 2.0% |
| 3-4 | 33 days | 3.5% |
| 5 or more | 46 days | 6.8% |
Pages cited by five or more engines are 5.5 times as likely as single-engine pages to still be above half-peak after eight weeks, and they take more than twice as long to reach peak. Slow rise and broad citation go together. Fast spikes on one engine are usually the ones that vanish.
Date-tied pages fade first
Profound classified 133,000 sustained pages by purpose. Reporting pages (anything tied to a date or event) came last on all seven engines. The other families hold at 1.4 to 1.6 times their rate, and reporting pages also peak faster, a median rise of 34 days against 41 to 43.
Content type changes the curve
Profound's measure is decay of smoothed citation share. A different, complementary measure comes from Promptwatch Data. In its half-life of content in ChatGPT report, every page ChatGPT cited on July 6, 2026 was tracked for 90 days, and the chart shows how many of those pages were still cited on or after each day.
| Content type | Still cited at 30 days | Still cited at 60 days |
|---|---|---|
| Landing pages | 78% | 69% |
| Docs | 76% | 67% |
| Comparisons | 72% | 63% |
| Listicles | n/a | 60% |
| Product pages, reviews, how-tos | n/a | about 55% |
| News | n/a | 41% |
A few things stand out. Docs have the smallest drop after day 0: ChatGPT cited 93% of them again later, against 77% of news pages. Almost a quarter of news pages were never cited again after day 0, and half were gone after 36 days. Product pages, reviews, and how-to guides reach their half-life after 74 to 77 days.
Do not compare these numbers with Profound's 11 days. They measure different things. Promptwatch's is cohort survival (is the page cited again at any level?), so a page cited once at a trickle still counts as alive. Profound's is how long a page holds a meaningful share of its peak. Both are true. A page can be technically still cited and practically gone. The Promptwatch window also ends at 90 days, so the final-week drop is partly an artifact, and true half-lives for landing pages and docs are probably close to or beyond 90 days.
Reading each stage of the curve
Ramp-up
A fast ramp-up (days) on a single engine is the burst signature. It often follows a news cycle, a launch, or a pile of fresh coverage. Expect it to fall just as quickly. A slow ramp across several engines looks different and is the better sign: the page is being discovered through multiple retrieval paths, and Profound's data says those pages last longest.
What to do during ramp-up:
- Check whether the spike is one prompt or many. One prompt is fragile.
- Confirm the AI crawlers can actually read the page and are returning to it.
- Link to it internally from related pages while it is getting attention.
Peak
Peak is a moment to learn from, not to celebrate. Record which prompts, engines, and competing sources were present at the top. That snapshot is your baseline for diagnosing decay later. Note also what type of page it is. If it is a news or event page, you already know roughly how long you have.
Decay
Decay is where interpretation gets tricky, because a falling line has more than one cause. Three are common:
- Normal rotation. Profound's earlier volatility research found 40 to 60% of domains cited for the same query changed within a month. Some of your decline is just how engines behave.
- Replacement. A competitor or newer page took your slot in the same answers.
- Staleness. Facts, prices, or examples on your page are out of date and nothing has been touched.
BrightEdge's weekly tracking adds a wrinkle: 96.8% of cited domains showed zero change week to week, but among the roughly 3% that moved, 87% were declines, and most were binary, from cited to not cited. So decay in aggregate looks gradual, but at the prompt level it can be a cliff.
A simple three-part test for real decay: weekly citations fall for several consecutive weeks, no competitor obviously replaced you, and the page has not changed since its last uptick. Any one alone can be noise. Together they say act.
Is freshness the lever?
Partly. A July 2026 Seer Interactive study of 7,683 pages and 47,097 citations found 75% of AI-cited pages had been updated within the past year, with Gemini at 78%, ChatGPT at 73%, and Perplexity at 65%. But the interesting detail is that most of that freshness is maintenance. Among pages with readable publish and update dates, 72% looked fresh by last update, against 42% by original publish date. Pages cited in all four months of the study were only 68% fresh, with a median update age of 5.6 months. Pages cited in just one month were 86% fresh.
Read that as: spikes come from very recent content, and durability can tolerate older pages if they are kept up.
Discovered Labs' analysis of 2 million citations adds that engines disagree on age. Median cited-content age was 5.1 months on Claude, 6.0 on Google AI, 7.8 on Gemini, and 8.0 on ChatGPT. If you calibrate a refresh cadence to the most demanding engine, you stay competitive on all of them.
What about independent practitioners? Ken Imoto logged citations weekly for nine weeks across five engines and found his peak around week three, falling to 46% of peak by week nine, while Google clicks for the same pages stayed essentially flat. It is a small sample, and it is his own blog, but the shape matches the large studies: AI citations and Google rankings run on different clocks.
Turning curves into a refresh routine
A quarterly "republish everything" calendar wastes effort on pages that decay slowly and misses pages that fell off a cliff. Trigger on behavior instead.
Step 1: track by page and by engine
Brand-level visibility can stay flat while one important page collapses. You need first-cited date, rise, peak, half-life, and last-cited date per URL, per engine. Without tooling, a fixed panel of 15 to 20 prompts per priority page, run every one to two weeks across the main engines, is a workable start. A single check of a single prompt is useless because AI answers vary run to run.
Step 2: separate decay from rotation
One dip is not a signal. Several consecutive weeks of decline, no replacement, and no page change is.
Step 3: prioritize by content type
If a page is a news or event piece, accept that it is a burst asset and use it for short visibility, not long-term value. Spend refresh effort on landing pages, docs, comparisons, and product pages whose facts drift (specs, prices, availability). Promptwatch's data suggests docs, landing pages, and comparison pages are the durable assets, while product pages decay about as fast as how-tos.
Step 4: change content, not the date
A refresh that works changes facts, numbers, examples, or a whole section. Updating the timestamp alone does nothing useful. After the change, watch the same page and prompts for two to four weeks to see if it reopened a citation window.
Step 5: when refreshing is not enough
Similarweb's decay guide makes a fair point: page updates alone often stabilize a falling score without recovering it. Recovery usually needs new third-party coverage on the domains AI already cites for your topic, such as review sites, Reddit, and publishers. Look at who is being cited in your place and go earn presence there.
Tools for reading citation trends
You can start with a spreadsheet, but the page-level view gets heavy fast. Several platforms track this now.
Promptwatch classifies every citation by content type and source type and shows citation trends per cited page, including first cited date, ramp-up, peak, and decay half-life. It also logs IP-verified AI crawler hits, so you can tell whether a falling page is an access problem or a content problem, and its Content Refresh and Action Board turn declining pages into tasks. Pages you publish through its Content Agent can be followed through crawl, citation, and AI traffic. You can try it with the 7-day free trial or start with the free AI visibility report.
Profound launched its Citation Decay view in August 2026, showing rise, peak, and half-life for each URL an engine cites, and its research underlies much of this guide.
Similarweb's AI Brand Visibility tracks a Domain Influence Score and lets you sort prompts by period-over-period change to find where you lost ground.
BrightEdge covers citation volatility weekly across nine industries, which is useful for benchmarking whether your drop is category-wide.
| Tool | Page-level lifecycle view | Crawler data | Content action layer |
|---|---|---|---|
| Promptwatch | Yes (ramp-up, peak, half-life, content types) | IP-verified crawler logs | Action Board, Content Refresh, Content Agent |
| Profound | Yes (Citation Decay) | Not the focus | Mostly monitoring |
| Similarweb | Domain Influence Score and prompt deltas | No | Citation analysis for outreach |
| BrightEdge | Weekly volatility data | No | Enterprise SEO suite |
For more options, the GEO software directory at bestgeosoftware.com lists dozens of trackers.
Common mistakes
- Averaging across engines. With correlations near zero, a blended curve tells you nothing about any single engine.
- Treating a spike as an achievement. Burst pages are 78% of the population. Ask whether you have a second or third engine citing it.
- Refreshing by age instead of behavior. The oldest page is not necessarily the one decaying.
- Judging by one prompt. Use a fixed panel.
- Ignoring the citation rank. A page that is cited but sinking to the bottom of the source list is decaying in a way share counts can hide.
A quick worked example
Say a pricing comparison page peaked on ChatGPT 21 days after first citation and is now at 40% of peak after five weeks. Check the three conditions. Weekly citations dropped four weeks in a row. A competitor's comparison page now appears in the same answers, so it may be replacement. The page still lists last year's prices. Action: update prices and add the two new competitors' features, link to it from your newest posts, and pitch inclusion to the review sites ChatGPT cites on those prompts. Then watch it for a month against the same prompt panel.
Compare that with a launch announcement that peaked and fell to nothing in ten days. Nothing is wrong. It did its job as a burst asset, and the effort belongs on a durable page.
What to do this week
- List your 20 most important pages and the prompts each should answer.
- Pull first-cited, peak, and last-cited dates per engine. If you lack tooling, run the panel manually.
- Tag each page as burst or durable by content type.
- Pick the three durable pages that have decayed longest with no replacement, and do a real content refresh.
- Re-check in two to four weeks and log whether a new ramp-up started.
The pattern across all of this research is consistent: getting cited is a launch problem, staying cited is a retention problem, and the retention work happens at the page level, engine by engine.
For more on how AI engines pick and keep sources, see the Promptwatch Data hub, including its report on ChatGPT citation share by content type.