How to Tell Whether a Link Did Anything

A good link landed six weeks ago. Rankings for the target page improved. Someone wants to know whether the link caused it.

The honest answer is that you cannot know, and anyone who tells you otherwise with a percentage attached is selling something. But “you cannot know” is not the same as “you have nothing,” and a report that stops at the first sentence is as unhelpful as one that overclaims.

Here’s how to build the strongest honest case.

Why isolation is impossible

To attribute causally you’d need a counterfactual: the same site, in the same index, over the same six weeks, without the link. That doesn’t exist.

What changed in those six weeks, besides the link:

The index. Google ships ranking changes continuously, most unannounced.
Your competitors. Their content, their links, their technical state.
Your own site. Other pages published, internal links added, a template tweak, a crawl-budget shift.
Demand. Query volumes and intent mixes move seasonally, and average position is sensitive to which queries fired.
The SERP itself. A new feature above the organic results changes clicks without changing rank, and changes rank measurement if the feature occupies a position.

Any of these can produce the movement you’re attributing to the link. Most weeks, several are happening at once.

This is the part worth internalising: the difficulty isn’t measurement precision, it’s that the question is causal and you have no control group. Better tooling doesn’t fix that.

What a defensible case looks like

You’re building circumstantial evidence, in the legal sense: several independent things that would each be unlikely if the link had done nothing.

1. Specificity of the target. Did the linked page move, or did the whole site? If rankings improved sitewide, the link is a weak explanation and something broader happened. If the linked URL moved and its unlinked siblings didn’t, that’s your strongest single piece of evidence.

2. Timing that fits the mechanism. Effects require the linking page to be crawled first. A ranking change the day after publication, before the source page has been recrawled, doesn’t fit — it’s a coincidence or something else. A change a few weeks later fits, but so do many other things, so timing alone proves little. Note the direction of the logic: timing that doesn’t fit is much more informative than timing that does.

3. A control set. Pick comparable pages on your own site that didn’t get links and track them over the same window. If the linked page rose 6 positions and the controls averaged +5, you’ve learned that something site-wide or index-wide happened. If controls were flat, your case is much stronger. This is the single most valuable thing you can add, and almost nobody does it.

4. Competitor movement. If the pages that dropped below you also dropped below three other sites, the market moved and you rode it.

5. Referral traffic, separately. A link can send actual humans. That’s directly measurable, doesn’t require any inference about rankings, and is often the more defensible half of the link’s value. Report it separately and don’t let it blur into the ranking claim.

The write-up

Language does most of the work here. Three registers, in increasing order of honesty:

Overclaiming: The DR 74 link from the industry publication drove a 6-position improvement and a 31% traffic increase for the target page.

Underclaiming: We acquired one link. Rankings improved. Attribution is not possible.

Defensible: The target page moved from position 14 to 8 over six weeks following the link, while five comparable unlinked pages on the same site stayed within one position of their baseline and the two competitors above us were unchanged. The link is the most plausible explanation available, though we can’t rule out index-side changes. The link also sent 340 referral sessions, which is measured directly.

The third is longer and it’s the only one you can defend in a room. Note what it does: states the observation, states the control, names the competing explanation it can’t eliminate, and separates the measured number from the inferred one.

Things that will trip you up

Average position is a treacherous metric. It’s averaged over whichever queries generated impressions, and that mix changes. A page can “improve” in average position purely because it stopped appearing for a batch of low-ranking long-tail queries. Track position for a fixed set of named queries instead, and check impression counts alongside.

Rank tracking has its own noise floor. Personalisation, location, device and index churn mean the same query measured twice can differ. Movements of a position or two are usually noise; treat them as such.

Regression to the mean. Pages get links because someone noticed them, and pages get noticed when they’re doing well. A page sampled at an unusually good moment will tend to look worse later regardless of what you did, and vice versa. If you only investigate links to pages that subsequently moved, you’ve built a filter, not a study.

Survivorship in your own reporting. If you write up the three links that preceded improvements and quietly don’t write up the eleven that didn’t, your internal evidence base is worthless. Log all of them, including the nothing-happened ones. The base rate is the most useful number you’ll ever collect about your own link building, and it’s the one nobody has because nobody logs the misses.

What to do at scale

For a single link, you’re stuck with circumstantial reasoning. For a programme, you have a better option: stop asking about individual links and look at the distribution.

Over a year, with fifty links logged and the outcome of each recorded honestly — including the nulls — you can say useful things. Which source categories preceded movement more often than others. Whether links to already-ranking pages behave differently from links to page-three pages. What fraction did nothing observable.

That’s not causal proof either. But a base rate built from your own site, on your own logging, beats any published correlation study, because it’s about you rather than about a scrape of the top ten results.

And it changes the question you’re answering. “Did this link work?” is unanswerable. “Of the links we built like this, what fraction preceded a measurable improvement?” is answerable, honest, and considerably more useful for deciding what to do next quarter.

Two related pieces: what Domain Rating actually measures on why the source’s score is a weaker predictor than it looks, and link velocity and other metrics that mislead on the metrics that tend to fill this vacuum when attribution gets hard.