A 1.3-million-link survival analysis shows half of all backlinks vanish within three years, and almost never because the source page died. Here's the data.
SEO consultant Bill Hartzer downloaded the complete Majestic export for his own domain, every link it had earned naturally since 2006, 1,301,839 of them from 22,260 referring domains, and ran a proper survival analysis on it rather than an average. The headline number: median survival of 1,080 days, just under three years. After one year, 80.9% of referring domains are still linking. After three years, 49.3%. After ten, 9.5%.
That's not a site that did something wrong. Hartzer never bought a link in twenty years. This is what attrition looks like on a large, organic, entirely natural backlink profile, and it lines up with the largest independent study on the same question: Ahrefs' own analysis of 2,062,173 websites found 66.5% of links since January 2013 have rotted, with a further 74.5% lost once you count everything that stops counting for ranking purposes. Two very different datasets, a single-site deep audit and a multi-million-domain sweep, pointing at the same conclusion: losing roughly half your links over a few years isn't a symptom. It's the baseline.
The assumption worth retiring: links die because pages die
Ask most people why backlinks disappear and the answer is some version of "the page got deleted" or "the site went down." Hartzer checked this directly against his own 645,202 lost links by looking at what the source page was doing at the moment the link was lost. The result: 78.0% of the time, the page downloaded fine, it simply no longer had the link on it. Another 19.4% had changed what they canonicalize to. Only four links, out of 645,202, were lost to an outright 404.
Links mostly die because someone edited the page, a redesign dropped a link module, or a migration changed what a page points to, not because the page itself vanished. The practical consequence is specific: if you're monitoring for dead backlinks by watching for 404s or down domains, you're checking for the thing that almost never happens while missing the thing that almost always does. A link can disappear from a page that's still live, still ranking, still maintained by a real person, which is also why Hartzer's post flags this as good news for link reclamation specifically: the page you'd be asking to re-add a link to is a going concern, not a corpse.
Stronger domains hold their links longer
Segmenting the same survival data by the referring domain's Trust Flow produces a clean, step-wise relationship. A link from a Trust Flow 0 domain has a median life of 859 days and a 22.2% chance of still being there at five years. A link from a Trust Flow 61+ domain has a median life of 3,353 days, about 9.2 years, and a 69.9% chance of surviving five years. That's close to a four-fold gap in how long the asset lasts, layered on top of whatever gap already exists in how much authority it passes while it's there.
The compounding effect matters more than the headline number. A cheap, low-quality link isn't just weaker while it's active, it's also gone sooner, which means the true cost difference between a strong link and a weak one is wider than a snapshot comparison of their authority scores would suggest.
Junk domains are the default, not the exception
Of the 22,260 domains that ever linked to Hartzer's site, 15,421 of them, 69.3%, sit at Trust Flow 0: scrapers, expired-domain churn, auto-generated directories, comment spam, sites that copied a post wholesale. He didn't build any of them and didn't buy any of them. They accumulate on their own, because that's what happens to any domain that's been publishing for two decades.
If you've ever run a backlink audit on your own site and felt a jolt of alarm at the pile of garbage domains in the list, this is the calibration point: that pile is the normal residue of existing online, not evidence that a competitor attacked you or a past agency did something shady. The same dataset puts a number on what a genuinely good year of natural link acquisition looks like, too. Hartzer's domain earned 446 referring domains at Trust Flow 41 or higher across twenty years, which works out to roughly 22 a year. Anyone offering fifty TF 50+ links in a single quarter is offering more than double that site's best annual output, which is a useful gut check against vendor pitches that promise results no organic profile produces.
What this means for a replacement rate, not a one-time campaign
The number that should change behavior is the practical flip side of "half gone in three years": if a profile loses roughly a sixth of its referring domains every year just from ordinary attrition, standing still requires replacing that sixth annually before any real growth happens at all. A link building effort that runs for a quarter and then stops isn't a program that plateaus. On this data, it's a program that starts reversing the moment it stops, because the churn underneath it doesn't pause along with the campaign.
This is also where source durability becomes a real variable, not just a quality signal. This blog has already covered why losing even your single best backlink tends not to move rankings much, so the point here isn't that any one lost link is a crisis. It's that the aggregate churn rate is something you can plan a budget and a cadence around, the same way you'd plan for customer churn in a subscription business, rather than treating link building as a project with a finish line.
A directory listing is a useful example of the kind of link that resists this decay better than most editorial links do, for a structural reason rather than a lucky one: it doesn't depend on a single blog post surviving its author's next redesign. It lives on a page whose entire purpose is to keep listings current, which is a different durability profile than a mention buried in a 2019 roundup post that nobody's opened since.
What to check
Pull your own backlink export and look at what share of your referring domains you've earned in just the last twelve months versus what's aged past three years. If the recent slice is thin, the data here says you're not looking at a stable profile, you're looking at one that's already decaying faster than it's being replaced, and it will show up as a slow decline rather than a single alarming event. That's also the more useful lens for interpreting how long it took to earn those links in the first place: acquisition is already slow, which makes an undetected attrition rate the more expensive problem of the two.
