The Half-Life of a Correction

Here is a familiar scenario. A company discovers that a widely repeated figure about it is wrong. They correct it at the source — updated site, updated record, a note to the outlet that originated the error. They consider the matter closed.

Eighteen months later the wrong figure is still the answer. Not everywhere. Just in the places that count.

A correction does not take effect. It propagates, unevenly, through a network of sources that copy from each other with varying lag and varying willingness to update.

The curve has a characteristic shape. There is a fast phase, where sources that pull directly and automatically from the corrected origin update within days. Then a long slow phase, where sources that copied the old value into static text update only if a human decides to. Then a tail that never resolves at all, because the source is abandoned, or its editors have moved on, or the wrong figure has become load-bearing in someone else's argument.

Half-life is the useful measure: how long until half the retrievable instances reflect the correction. The two cases behave very differently. A well-structured fact anchored in a machine-readable record resolves relatively quickly, because the sources consuming it re-derive on their own. A narrative claim that spread through commentary does not, because every instance requires a person to decide to act — and the asymptote never reaches zero. Measure your own half-lives rather than assuming either case.

The distinction that governs everything is whether a source is an anchor or an echo.

An anchor is consulted directly by retrieval systems and other publishers. Structured records, primary registries, the entity's own canonical pages, the small number of institutional references that everyone defers to. Anchors are few, and they update cleanly, because they exist to be updated.

An echo copied from an anchor at some point in the past and then froze. Echoes are numerous, individually low-authority, and collectively capable of keeping a dead fact alive forever through sheer redundancy.

Correcting an anchor changes the future. Correcting an echo changes one document. Almost everyone spends their effort on echoes, because echoes are where they encountered the problem, and because chasing them feels like progress.

Three failure modes account for most of it.

The correction is not machine-readable. A statement in a paragraph on a page is not a correction as far as a retrieval system is concerned; it is prose. If the structured record still says the old thing, the old thing is what gets retrieved, regardless of how emphatically the prose disagrees.

The correction is not more citable than the error. If the wrong figure lives in a well-linked institutional source and the right figure lives only on your own site, retrieval systems will resolve the conflict in favor of the source that looks more independent. Being right is not sufficient. Being right in a citable location is.

The correction arrived after the claim was absorbed rather than merely stored. A figure that has been repeated across enough independent-looking sources stops behaving like a citation and starts behaving like consensus. Consensus does not update on evidence at anything like the same rate.

Work anchors first, in order of retrieval frequency, and do not touch an echo until every anchor is clean. Make the correction structured before you make it eloquent — the parseable version is the one that travels. Give the corrected value somewhere independent and citable to live, because self-published corrections are discounted by design.

Then measure. Re-issue the queries that surfaced the error and watch the proportion of corrected instances over time. If the curve is flat after sixty days, you corrected an echo and mistook it for an anchor.

The goal is not a perfect internet. The tail never fully clears, and pretending otherwise is how retainers become permanent. The goal is that the sources doing the actual work of describing you are right, and that you can prove it with the same queries anyone else would run.