I Measured AI Citation Half-Life for 90 Days: 6 of 10 Sources Vanish by Week 4

Last time I tracked citation decay, I tracked my own pages, which is exactly the kind of self-absorbed measurement you would expect from someone who runs a personal blog and reads the analytics like a horoscope. This time I picked pages that were not mine and watched what happened to them. Ten URLs. Three engines. Ninety days.

The protocol in one sentence: I locked one fixed query, logged the top-cited URLs from ChatGPT (browsing on), Claude (web search on), and Perplexity (Sonar Pro) on day zero, then re-ran the same query every Monday for thirteen weeks to see which of the original ten URLs still showed up in any engine’s citations that week.

Six of the ten were gone by week four. Not “cited less.” Gone. As in, the URL that led every answer on day zero did not appear in any of the three engines by the time I checked on day twenty-eight. The decay shape was not smooth. It was a cliff.

This is the writeup of what I saw, plus the postmortem for each vanishing source. The pattern that killed them is the interesting part.

The setup

I chose one query, kept it fixed, and let the engines pick their own sources. The query was a real intent question in a domain I don’t publish in (privacy tech, chosen because I had zero horse in that race). Every Monday morning I asked ChatGPT with browsing on, Claude with web search on, and Perplexity on Sonar Pro. Three retries per engine, separate sessions, no chat history.

Day zero was the first Monday. I logged the top-cited URL from each engine for that week. Across the three engines, there were ten unique URLs cited on day zero. That is not that many (the engines overlap more than you’d think), but it was enough to have a cohort worth tracking.

Then I locked the cohort. For thirteen weeks I kept asking the same query and logging whether each of those ten URLs still appeared. That is the only thing I measured. Not “how prominent,” not “which position.” Just: does this URL still show up in any of the three engines’ cited sources this week, yes or no.

The cliff appears in week four.

What the curve actually looks like

Normalized to peak week, here’s the count of the original ten URLs still being cited by any of the three engines each week.

Week:      0    1    2    3    4    5    6    7    8    9   10   11   12
Alive:    10   10    9    8    4    4    3    3    3    3    2    2    2

Nine survive to week two, eight to week three. Between week three and week four, four more die at once — the largest single-week drop in the entire window. That takes the cohort down to four survivors, and from that point on the survivors mostly hold, with slow erosion.

That week-three-to-four cliff is not what I expected. The prior lit reports a smooth exponential decay with a median citation half-life around 4.5 weeks, and 3.5 million citation events analyzed by Scrunch confirm the same 4.5-week median. What I saw was closer to a step function. Slow attrition, slow attrition, then almost half of the cohort dies in a single week.

The half-life-of-4.5-weeks framing is still approximately right. Six of ten by week four is a hair over half. But the shape is not exponential. Individual URLs stopped being cited in a single week rather than fading gradually.

The six that vanished, and why

I did the postmortem for each of the six that vanished by week four. This part is the whole reason I ran the experiment, because “why did this one die” turned out to be a small number of repeating reasons.

URL 1 — SEO blog post, publication date late 2024. Died because a competing 2026 blog post from a domain with higher authority shipped in the same week. Both engines started citing the new one. This is the “fresher answer to the same question” case, and it was the modal failure — three of the six died this way.

URL 2 — Company documentation page. The page still exists and is still correct. It got beaten by a first-party primary source (an official spec page) that hadn’t been indexed on day zero. Once the primary source was crawled, the derivative page lost its citation. Same category as URL 1, but the winner was a more authoritative source rather than a fresher article.

URL 3 — Wikipedia section. Cited by all three on day zero, gone by week four from all three simultaneously. I could not reproduce this one. My best guess: an internal ranking change on the engine side, because the Wikipedia page itself did not change. This is the one I have the least confidence in explaining, and I want to be honest about that instead of confabulating a reason.

URL 4 — Reddit thread from 2024. Died between weeks three and four in what I think was a training/index freshness rotation. ChatGPT and Perplexity both stopped citing it. Claude never cited it at all, so no signal from that engine. The thread itself wasn’t deleted; it just aged out.

URL 5 — News article with a stale date. Publication date was more than 18 months old on day zero. Survived for three weeks, then two of the three engines stopped citing it in the same week. The topic itself hadn’t changed. The date on the page had.

URL 6 — Personal blog post. Author had deleted the post between weeks three and four. Genuine 404. I could see this one coming because when I clicked the URL myself in week three, it was already gone; the engines just took one more week to notice. That one-week lag is interesting on its own. An engine can be citing a dead URL for a full week after it dies.

If I zoom out and squint at those six postmortems, they collapse into four causes:

1. A fresher competitor for the same question shipped (URLs 1, 5). This is the freshness clock hurting the incumbent.
2. A more authoritative primary source got indexed (URL 2). This is the authority signal hurting the derivative.
3. The engine’s own index rotation (URLs 3, 4). Not something the URL owner can influence.
4. The URL is actually dead (URL 6). Trivially, a 404 doesn’t get cited for long.

The one I keep chewing on is cause 3. Two of the six died for reasons I can’t attribute to the page or its competitors, only to something inside the engine. Which means even if you write a perfect page and it does get cited, roughly a third of the mortality risk is out of your hands.

The four that survived

Small n, so I want to be honest that this is anecdote-sized, not study-sized. But the four survivors did have something in common.

• All four were first-party primary sources — official docs, a canonical spec page, and two vendor pages by companies that actually built the thing.
• Three of the four had recent updates visible on the page (dateline, changelog note, or “last updated” field).
• All four appeared in at least two of the three engines on day zero, not just one.

None of that is a discovery. Freshness plus authority plus multi-engine coverage is the boring answer everyone has been giving for a year. What the experiment added for me is that when I actually watched the failure mode play out, the boring answer looked less like a curve and more like a filter: the pages without those three properties dropped off a cliff in a single week, and the pages with them mostly held.

The GSC control curve

I also pulled Google Search Console traffic estimates for the ten URLs from a third-party tool, because I wanted to know if the AI citation death was tracking with search-traffic death or was independent.

It was independent. All ten pages held their search traffic within their normal weekly noise band across the entire 90 days, including the six that had already been dropped from AI citations. Not one of them ranked lower on Google in any meaningful way during the window.

This is the same divergence I saw the last two times I ran this kind of experiment. AI citations and Google rankings move on different cycles. If you only track the Google side, the AI citation dropoff will not show up in your weekly metrics. Ahrefs’ AI Overview CTR analysis, the one that got quoted for weeks about the 34.5% CTR drop, is worth reading because it frames the CTR loss as decoupled from the ranking curve, which is the same shape I keep hitting.

What I actually took away

Three things.

Week four is the trip wire. If you are tracking AI citations for an owned or partner URL, the interesting datapoint is not day zero and not day ninety. It is week four. That is when the cohort thins. Waiting for a smoother monthly cadence to check will hide the event that matters.

“Half-life” undersells the shape. The 4.5-week median from the industry benchmarks is a real number, but it summarizes a step function, not a smooth exponential. Half of the cohort dies in a couple of specific weeks, and the survivors mostly hold. Schedule your review at the cliff week, not on a monthly cadence that averages over it.

Three of the four survivor properties are things you can influence. First-party status, visible freshness, and multi-engine day-zero coverage are all decisions, not luck. The one you can’t influence (the engine’s own index rotation) is a fixed cost, not a strategy. Do the three you can and accept the one you can’t.

I’m going to re-run this experiment on a different query in Q4 to see if the cliff-at-week-four pattern reproduces or if I was unlucky with one query’s freshness cycle. If it holds, week four is the number I will start telling people to schedule the review against. If it doesn’t, I will come back and say so, because “a citation curve is a cliff, not a slide” is the kind of claim that needs more than one 90-day window to earn.

One more thing worth calling out: this matters for engineers, not just marketers. If your team ships a public API doc and cares whether coding agents can cite it accurately (I looked at the agent side in ChatGPT Codex vs Claude Code), the same decay curve applies to your doc pages. Framed under the LLMO Framework: Citability tells you whether an engine will cite you at all, but Authority and Coherence are what decide whether you sit in the surviving four or the vanishing six a month later. I have been mostly measuring the first pillar. This experiment is what convinced me to start measuring the other two on an actual cadence instead of hoping.

If you want the measurement loop I keep leaning on (the five-prompt, three-retry, monthly-cadence setup with the GA4 segment regex and the Python visibility script), chapter 10 of LLMO: AI Search Optimization walks through it, including the Python code for running the check across ChatGPT, Claude, and Perplexity on a weekly cron. This post is what happened when I pointed that loop at other people’s URLs instead of my own.