The method, and why it matters here
The Historic export records, for every link, the date it was first indexed and the date it was lost. That is enough to build a survival curve rather than an average, and the distinction is not academic.
Averaging the age of links that died throws away every link still working, which understates lifespan badly. The correct treatment is Kaplan–Meier, the same method used for clinical trials, because the question is structurally identical: how long does the thing last, given that most of the subjects have not died yet? Links still live are censored at their last-seen date — they leave the risk set rather than being counted as either survivors or losses.
One observation per referring domain, not per link. That choice turns out to matter more than the choice of statistical method, and the last section on this page explains why.
The headline
Median survival: 1,080 days. Just under three years.
| Elapsed | Referring domains still linking |
|---|---|
| 1 year | 80.9% |
| 2 years | 62.8% |
| 3 years | 49.3% |
| 5 years | 29.7% |
| 10 years | 9.5% |
Across the whole profile, 49.6% of every link ever recorded is now flagged lost.
The practical version of that number is a replacement rate. If half your referring domains disappear within three years, then holding your current position requires acquiring roughly a sixth of your profile every year before you have grown by anything at all.
Most link building proposals model acquisition and nothing else. They present a treadmill as a staircase. An honest link budget has a maintenance line in it, and until now there has not been a defensible number to put there.
It also reframes what happens when a program stops. The common expectation is that rankings hold and then slowly drift. What this suggests is more active: the asset base is eroding on a measurable schedule whether or not anyone is adding to it.
Links from strong domains last about four times longer
This is the finding I have not seen published anywhere, and the one I would most like other people to test against their own data.
Segmenting the survival curve by the Trust Flow of the referring domain produces a clean, monotonic relationship — every band, in order, with no exceptions.
| Referring domain | Domains | Median life | Alive at 3 yrs | at 5 yrs |
|---|---|---|---|---|
| Trust Flow 0 | 15,421 | 859 days | 40.7% | 22.2% |
| Trust Flow 1–10 | 4,031 | 1,531 days | 62.8% | 40.1% |
| Trust Flow 11–20 | 1,318 | 1,911 days | 72.6% | 52.6% |
| Trust Flow 21–30 | 647 | 2,223 days | 78.5% | 59.5% |
| Trust Flow 31–40 | 397 | 2,375 days | 79.9% | 59.9% |
| Trust Flow 41–60 | 323 | 2,555 days | 83.5% | 60.8% |
| Trust Flow 61+ | 123 | 3,353 days | 85.2% | 69.9% |
A link from a Trust Flow 61+ domain lasts a median of 9.2 years. A link from a Trust Flow 0 domain lasts 2.4. That is close to a fourfold difference in how long the asset exists at all.
The commercial implication is direct. Links are priced on the authority they pass today; nothing in the standard pricing model accounts for how long the link will survive. If a strong link both passes more equity and persists four times longer, then the honest measure is cost per year of survival — and on that basis the premium at the top of the quality range is far better value than the headline price suggests, while the bargain end is considerably worse than it looks.
Put plainly: cheap links are expensive. They are weaker while they last, and the replacement cycle comes round more than twice as often.
Three medians from one dataset, and which one is right
The same file produces three different answers to “how long does a link last,” depending on decisions that look technical and are not.
| Method | Median | What it is actually measuring |
|---|---|---|
| Naive, lost links only | 1,135 days | The age of links that died. Ignores every link still working. |
| Kaplan–Meier, per link | 1,255 days | Handles censoring, but a few sitewide placements dominate the count. |
| Kaplan–Meier, per domain | 1,080 days | Handles censoring, and one domain counts once. The right answer. |
The naive figure was my own first pass, and it is wrong in a way worth naming: it conditions on the event having happened. It answers “among links that died, how old were they,” which is not the question anyone is asking.
The per-link Kaplan–Meier figure fixes the censoring and introduces a worse problem, covered next.
The trap: same data, opposite answers
Run the Trust Flow segmentation per link instead of per domain and the relationship inverts. Trust Flow 61+ comes out with a median of 851 days — the worst band on the board, worse even than Trust Flow 0.
That result is an artifact, and a very believable one. A handful of large sites contributed tens of thousands of sitewide links each. When one of them re-templates, tens of thousands of “observations” die on the same day and drag the whole curve down with them. The per-link view is measuring a few CMS migrations, not link durability.
Weighting by domain fixes it, because a domain that sent 219,159 links and a domain that sent one each count once.
I am including this because it is the single most useful thing in the dataset for anyone who reads SEO research. The same file, analyzed honestly twice by someone trying to get it right, produced opposite conclusions. The difference was not the statistics — both runs used Kaplan–Meier correctly. The difference was the unit of analysis, which is the kind of decision that gets one line in a methodology note, if it gets mentioned at all.
When you read a study that reports a per-link figure on link data, ask what proportion of the links came from the top few domains. If the author does not say, the number may be measuring something other than what the headline claims.
How the losses are distributed
Among the 645,202 links that did die, how long they lasted first:
| Survived | Share of lost links |
|---|---|
| 0–30 days | 0.2% |
| 31–90 days | 2.3% |
| 91–180 days | 3.2% |
| 181–365 days | 6.6% |
| 1–2 years | 22.9% |
| 2–3 years | 13.7% |
| 3–5 years | 25.5% |
| 5–10 years | 23.3% |
| 10 years+ | 2.3% |
Only 2.5% of links that eventually died did so within the first 90 days. 87.7% lasted more than a year.
This cuts both ways and the honest version says both. A link is a far more durable asset than the “links decay constantly” framing suggests — the early attrition everyone worries about barely exists. And a link is emphatically not permanent: half the profile is gone by year three.
One practical consequence: judging a link building campaign at ninety days tells you almost nothing about link durability, because almost nothing has had time to die yet.
Questions
Why use Kaplan-Meier rather than a simple average?
Because most links have not died yet. A simple average of the links that did die answers a different question and biases the result downward. Kaplan-Meier handles the still-alive cases by removing them from the risk set at their last-seen date rather than counting them as either successes or failures.
Does this mean I should only buy expensive links?
It means the price-per-link comparison everyone runs is missing a variable. A link that lasts four times longer is a different asset from one that does not, even at the same authority. Whether the premium is worth paying depends on the prices in front of you, but you should at least be dividing by expected lifespan before deciding.
Would the relationship hold on a different site?
Unknown, and that is the honest answer. One profile establishes a plausible baseline and gives you something specific to test. If you have Majestic access and an aged domain, the method pages describe exactly how to reproduce it.
Why is Trust Flow measured now rather than when the link was acquired?
Because the export only carries the current value. A domain that was strong in 2011 and is Trust Flow 0 today is banded as Trust Flow 0, which will bias the low bands' survival downward by an unknown amount. It is the most important limitation of this finding and it is stated in full on the caveats page.