Link Velocity and Other Metrics That Mislead

Attribution is hard, so the industry generates proxies. Some are useful. Several are folk metrics that acquired authority through repetition and have no documented basis at all.

Here’s a pass through the common ones, sorted roughly by how much they deserve.

Link velocity

The claim: links acquired per unit time. Acquire them too fast and you look unnatural; a “natural velocity curve” keeps you safe.

Where it came from: the reasonable observation that a sudden spike of thousands of links from nowhere is a pattern spam produces, plus a large amount of extrapolation.

What’s actually knowable: search engines evidently do detect link-spam patterns — that’s what link-spam policies are for. What has never been documented anywhere authoritative is a rate threshold, a safe curve, or a formula. The specific numbers you’ll see quoted (“no more than 10% month over month”) have no published source. They are made up.

There’s also a measurement problem underneath. What you’re plotting isn’t when links were created, it’s when your tool’s crawler found them. Twenty links discovered in one week may have accumulated over three months. A velocity chart is substantially a picture of crawl scheduling, and treating a discovery spike as an acquisition spike is a category error.

Verdict: the underlying intuition — a sudden mass of low-quality links from a coordinated set of sources looks bad — is sound, and it’s about pattern, not rate. The metric as commonly used, a line chart with a target slope, measures your crawler. Don’t set targets on it. It’s worth glancing at for anomalies, in the same way you’d glance at any time series, and worth nothing as a KPI.

Anchor-text ratios

The claim: keep exact-match anchors under some percentage or risk a penalty.

Where it came from: exact-match anchor stuffing genuinely was a dominant manipulation tactic, and profiles built that way genuinely do look distinctive.

What’s actually knowable: anchor text is a real signal, and heavily over-optimised anchor distributions are a documented spam pattern. The specific percentages circulating — 1%, 5%, 20% — are again unsourced.

The measurement problem is worse than for velocity. Ratios computed over link instances are dominated by whichever sitewide footer happens to exist, so a single template placement can push a profile from “1% exact match” to “60% exact match” without a single new editorial decision. If you compute the same ratio one-per-domain you get a completely different picture. Any anchor statistic that doesn’t say which basis it used is uninterpretable. (See referring domains versus total backlinks.)

Verdict: worth looking at as a distribution, computed one link per referring domain. A profile where nearly every domain uses the same commercial phrase is a real finding. A percentage compared against a threshold someone posted in 2013 is not.

“Toxicity” and spam scores

The claim: a vendor score, typically 0–100, for how harmful a link is.

Where it came from: the disavow era, when a wave of manual actions created enormous demand for a tool that would tell people which links to remove.

What’s actually knowable: these are proprietary classifiers built on observable features — the source’s own link profile, anchor patterns, page characteristics, network co-occurrence. Vendors document the general inputs and never the weights. Crucially, they are the vendor’s opinion of risk, not a measurement of any signal Google exposes. No search engine publishes a per-link toxicity value, so there’s nothing for the score to be calibrated against.

They are also, structurally, biased toward false positives. A classifier that flags a harmless link costs the vendor nothing; one that misses a harmful link generates complaints. The incentive runs one direction.

Verdict: useful as a sorting mechanism when you have 30,000 links and need somewhere to start reading. Actively dangerous as a decision rule — disavowing on a score alone means removing links a machine guessed about. (Whether to disavow at all is backlink.so’s territory; the measurement point here is simply that the score is a prior, not an observation.)

Trust-flow-style two-axis metrics

The claim: a pair of scores — one for volume, one for quality — whose ratio reveals manipulation.

Where it came from: a genuinely interesting idea, that propagating “trust” from a seed set of known-good sites behaves differently from propagating raw link count, so the divergence between the two is informative.

What’s actually knowable: the concept has real research behind it. The implementations are proprietary, seed-set-dependent, and computed over one vendor’s crawl like everything else. The ratio-based rules of thumb are folklore layered on top.

Verdict: conceptually the most interesting metric in this list. Same caveats as every vendor composite: comparative within one tool, not an absolute property of a site.

Links-to-page ratios and “link gap”

The claim: you need N links to outrank a competitor, derived from what the current top-ten pages have.

What’s wrong with it: it reads a correlation backwards. The pages ranking top-ten have more links partly because they rank top-ten and get seen. Copying the observed link count of a winner doesn’t reproduce whatever made it win, and the calculation ignores every non-link factor.

It’s also computed on tool-specific counts, so the same “gap” changes size depending which tool you opened.

Verdict: fine as a rough sense of the competitive neighbourhood — “the pages here have hundreds of referring domains, not five” is genuinely worth knowing before committing a quarter. Useless as a target. “We need 47 more links” is a sentence with no defensible derivation.

The pattern

Every metric in this list has the same three-layer structure:

1. A real observation at the bottom. Spam does look different. Anchor stuffing did happen. Trust propagation is a real idea.
2. A vendor implementation in the middle — proprietary, crawl-dependent, uncalibrated against anything the search engine exposes.
3. A folk threshold on top, unsourced, repeated until it sounded like documentation.

Layers one and two are usable if you keep them straight. Layer three is where the misleading happens, and you can spot it by a simple test: ask where the number came from. If the answer is “everyone knows” or a blog post citing another blog post, you’ve found layer three.

The alternative isn’t to abandon metrics. It’s to treat them as what they are — ordinal, tool-specific, useful for comparison — and to build your actual decisions on a base rate you collected yourself, which is the argument in how to tell whether a link did anything.