When a Metric Moves but Nothing Changed

Link metrics move for two entirely different reasons: something happened to your site, or something happened to the measurement. The second kind is common, produces movements of the same size as the first, and looks identical in a dashboard.

There’s a cheap test that separates them most of the time, and running it before you investigate saves a great deal of wasted work.

The five vendor-side events that move your numbers

Index refresh and expansion. A vendor crawls more, or crawls differently, and previously unseen links enter the index. Everyone’s counts rise. Scores derived from those counts shift, sometimes down, if the population grew faster than your slice of it.

Recalibration of a score. Authority-style scores are normalised against a population. If the population or the normalisation changes, every score moves. Vendors usually announce significant changes; the announcement rarely reaches the person reading the dashboard.

Model retraining. For scores documented as machine-learned predictions, retraining can change outputs with no change in the underlying graph at all. This is a real distinction between metric designs — see Domain Rating versus Domain Authority.

Policy changes on dedup, link lifetime, or attribution. How long dead links persist, whether subdomains are folded in, what counts as a distinct link. Each is a vendor decision subject to revision, and each moves totals across the entire customer base at once.

Your own plan or filter changed. Export caps, row limits, and default filters are part of the instrument. A plan upgrade can raise a “toxic link percentage” without a single new link — the mechanism in sampling, export caps, and missing rows.

The cohort test

The test is one question: did everyone move?

Pull the same metric, from the same tool, on the same date, for four or five comparable domains you don’t control — competitors, or just sites in your sector. Then look at the direction and rough size of their movement over the same window.

Everyone moved the same direction by a similar amount. Vendor event. Stop investigating your site.
Only you moved. Site event. Now investigate.
Mixed, no pattern. Probably ordinary index churn plus small real changes. Usually not worth pursuing unless the magnitude is unusual.

This works because vendor-side events are systematic across the index and site events aren’t. It requires you to be storing a comparison set over time, which is the small amount of setup this test needs and the reason most people can’t run it when they need to.

A hypothetical: your domain score drops from 44 to 41 over a month and someone asks what went wrong. You check five competitors: 52→49, 47→45, 61→57, 38→36, 55→52. All down two to four points. Nothing went wrong. The report sentence is “network-wide movement consistent with a vendor recalibration; our relative position is unchanged.” (Illustrative figures, not measurements.)

Signatures of specific causes

Beyond the cohort test, a few movements have recognisable shapes.

Step change on a single date, no gradual approach. Vendor event. Real link accumulation is gradual; index policy changes are instantaneous.

Referring domains up, score flat. Discovery of low-value links, or the log scale absorbing the change. Both boring.

Score up, referring domains flat. Recalibration, or one existing linking domain’s own score rose. Nothing you did.

Total backlinks up dramatically, referring domains flat. A sitewide or template link on one domain. The distinction is the point of referring domains versus total backlinks.

Everything drops on the day you changed subscription. Your instrument changed. Check filters first, always.

One metric moves in one tool only. Vendor-specific. Cross-check in the second tool before writing anything.

What to store so the test is available

The test only works retroactively if you kept the data. A minimal panel, pulled monthly, is enough:

Your domain: referring domains (all attributes and followed-only), total links, domain score, page score for your three or four most important URLs.
Four to six comparison domains: referring domains and domain score.
The vendor, date, and every filter setting used.

That’s a small table. Adding it costs a few minutes a month and converts “I think something happened” into an answerable question. It’s also the backbone of the reporting habits in building a link report you can defend.

The reporting language for a vendor event

Don’t hide it, and don’t treat it as an embarrassment.

“Domain score moved from 44 to 41 this month. The comparison set moved by a similar amount in the same direction over the same window, which is the pattern of a vendor-side recalibration rather than a change in our link profile. Referring domains are up over the period. No action recommended.”

Four sentences, and they pre-empt the question that would otherwise generate a week of work.

The failure mode worth avoiding is the opposite habit: explaining every movement with a site narrative. A metric that only ever moves for reasons you can tell a story about is a metric you’re not reading. Some months the honest answer is “the measurement changed, we didn’t.”

What nobody outside the vendors knows

When recalibrations happen unless announced, how large they typically are, how often models retrain, and how much of any given month’s movement in any given account is index growth. Vendors publish changelogs of varying detail; reading the one for your primary tool is one of the higher-value habits in this whole area, and almost nobody does it.

The durable rule is simpler than the mechanisms behind it: before you explain a movement, check whether it happened to everyone.