What Are Toxic Backlinks? (And How to Spot Them)

Search the term and you land in the middle of an argument that never ends. One side says bad links quietly tank your rankings, so you’d better buy a tool that scores them. The other side says Google ignores spam completely, so you should just relax. And after talking past each other for years, nobody can close the case because both are missing the same thing.

Toxic backlinks are low-quality, manipulative, or spammy links pointing at your site that could be read as an attempt to game rankings. “Toxic” is not Google’s word. Google says link spam. And Google’s response to link spam is neither a penalty nor a shrug.

It’s a third state that explains almost every confusing thing you’ve read on this topic. This guide covers what that third state is, how to spot the pattern that triggers it, and why the tool score you’re about to trust is measuring the wrong object.

A toxic backlink is any inbound link that looks less like a genuine editorial vote and more like an attempt to manipulate your rankings, and that could plausibly hurt you if a search engine treats it as manipulation. The usual suspects are familiar: links from private blog networks (PBNs), link farms, comment and forum spam, paid links passing PageRank, sitewide footer links, and exact-match anchor text at unnatural scale.

Toxic is a vendor word, not a Google word. Google’s documentation talks about link spam, unnatural links, and link schemes. There is no toxicity flag in Google Search Essentials. That gap between the industry’s vocabulary and Google’s own matters more than it sounds, because it’s the reason the whole category is so confused. We named a metric, built tools around the metric, and then argued about whether the metric describes anything real.

It also helps to remember where the fear came from. In the Penguin era, roughly 2012 through 2016, bad links genuinely could sink a site, and manual actions for unnatural links were common enough that a whole cleanup industry grew up around them.

Penguin 4.0 changed the game in 2016 by folding link evaluation into Google’s core algorithm and shifting toward devaluing spam rather than demoting sites for it. The tools and the toxic vocabulary were built for the old world and never fully updated for the new one. A lot of what still gets sold as urgent link cleanup is fighting a threat that mostly stopped existing a decade ago.

Ahrefs drew the sharpest distinction in the SERP, and it’s worth building on rather than reinventing. A spammy link is low-quality and machine-made: comment spam, scraper sites, auto-generated directories. A manipulative link is one you (or someone) placed to move rankings: a bought guest post, a PBN link, an exact-match anchor you engineered. Toxic gets used as a catch-all for both, which is where the confusion starts.

Spammy links are mostly noise Google’s systems filter out. Manipulative links are the ones tied to real risk, because they signal a pattern of intent. When a tool paints both red and slaps a single number on them, it flattens a distinction that decides whether you should care at all.

Tools needed something to sell, and “we found risk on your profile” sells better than “your links are mostly fine.” A 0 to 100 score is a product. It gives an agency a slide, a client a reason to act, and a vendor a recurring reason to log in. None of that makes the score wrong. It does mean the word was chosen for its commercial shape, not its accuracy, and you should read every toxicity number with that in mind.

This is the section the rest of the argument depends on. The “does Google penalize or ignore” debate has stayed unresolved because both answers are partly right and neither is complete. There’s a documented third behavior sitting between them.

Start with what Google has said in public. In the December 2022 link spam update, Google described using SpamBrain to “neutralize the impact of unnatural links on search results.” Read that word carefully. Neutralize, not punish. The links stop counting. They don’t help you, and they don’t hurt you, they simply get switched off as ranking signals. No manual action fires, no notice lands in Search Console, nothing tells you it happened.

This is the whole basis of the “Google just ignores spam” position, and as far as it goes, it’s true. If a spammer points a thousand junk links at your homepage, SpamBrain is designed to discount them and move on. For the average site with no manipulation in its history, that’s the end of the story.

The story gets more precise in Google’s internal documentation. The May 2024 Google API Content Warehouse leak, a dump of internal ranking documentation, contains a document-level attribute named badbacklinksPenalized, described like this:

“If this doc is penalized by BadBackLinks, we should not use the improvanchor score in Mustang ascorer.”

This does not say the document gets demoted or pushed down, but that the document stops receiving its improvanchor score, the boost your anchor text normally earns you, inside Mustang (one of Google’s scoring systems). Alongside it sit related attributes: AnchorSpamDays (the rate at which spam links were created), AnchorSpamCount (spam phrases found in anchors), and AnchorSpamDemoted (the total count of demoted anchors), plus fields tracking when a spam spike started and ended.

A caveat you should read as confidence, not nerves. The leak is unverified by Google. It may be stale, and some of it may describe systems that were never in production. An attribute name in a documentation dump is not proof of live ranking behavior. What we do know: ex-Googlers who reviewed the documents considered them authentic, and Google has never disputed their authenticity. So we treat this as strong evidence, we do not treat it as gospel, and we mark exactly where our confidence ends. That’s the honest way to use a leak, and it’s more than most of the SERP is willing to show you.

Why does the distinction between demotion and suppression matter so much in practice? Because they behave differently when you try to fix them. A demotion is a scored penalty against your page, the kind of thing that historically cleared once you cleaned up and filed a reconsideration request.

Losing an anchor advantage is quieter and more stubborn than earning a demotion. There’s no message telling you it happened, no reconsideration flow for it, and no obvious “before and after” to measure against, which is exactly why sites that suffer it spend months chasing a ranking dip they can’t name. The attributes tracking spam spike start and end dates hint that the state can lift on its own once the offending pattern ages out, but you won’t get a confirmation email when it does.

Put the public statement and the leaked attribute together and a third behavior appears, one that isn’t a penalty and isn’t indifference. Google doesn’t push your document down for bad links. It stops letting your anchor text push it up.

That’s the reframe we want you to take from this article. A penalty is something you receive. This is something you silently stop earning. Your anchor-text advantage, the relevance signal you’d normally get from links describing your page, discreetly stops applying to that page. The page can still rank on everything else it has going for it. It just loses one lever it used to pull. If anchor signals can be suppressed this quietly, the unambiguous brand signals that a suppression can’t touch are what you want carrying the load.

Once you see anchor suppression as the mechanism, the contradictions in this topic collapse into one story:

One mechanism, four “unexplained” observations. That’s why it’s worth getting right.

Theory is cheap. The good news is that two people actually ran the test, disavowed links, and measured what happened, and the newer of the two is only weeks old.

In August 2024, Joshua Hardwick at Ahrefs took 129 links that tools had flagged as toxic or potentially toxic, pointing at three of his blog posts, and disavowed them. He watched traffic for 20 days, from July 26 to August 15, 2024. The combined result was a 7.1% traffic decline.

That 7.1% number has since been quoted around the web as proof that disavowing hurts you. It’s a misleading way to read it. The result was mixed per page: one post dropped about 12%, another about 8.25%, and a third actually rose 12.82% over the same window. A 20-day, three-page, 129-link test with one page improving is not evidence that disavow causes damage.

Hardwick’s own conclusion was blunter and more useful: blindly disavowing “toxic backlinks” reported by SEO tools is unlikely to have much positive impact, if any. When you see the 7.1% figure cited without the spread, treat the source as careless.

Then Cyrus Shepard at Zyppy ran the extreme version. In June 2026, he disavowed more than 10,000 backlinks from 1,473 domains, meaning every single link pointing at his site, including clean editorial links from Moz, Ahrefs, Semrush, and Backlinko. He left it for roughly two months and watched.

His summary: “Apparently, nothing. Nada. Not a damn thing.” No collapse, no recovery, no measurable movement from nuking his entire link profile through the disavow file. This is the single strongest piece of primary evidence on the whole topic, and as of mid-2026 it’s the freshest, which is exactly why so much of the ranking content on “toxic backlinks” hasn’t caught up to it yet.

Think about what that test actually rules out. If disavowing every link, good and bad alike, produces no measurable change, then the disavow file simply is not the lever most people believe it is. It doesn’t restore a suppressed document score, and on a healthy site it doesn’t remove some hidden drag either, because there wasn’t any drag to remove. A tool that changes nothing when you pull all the levers is a tool that was never touching the mechanism people bought it for.

Two experiments, one lesson. Disavowing links that tools flag as toxic did close to nothing, in a small careful test and in a scorched-earth one. That fits the mechanism perfectly. If the damage (when there is any) comes from a document-level anchor-suppression flag triggered by a pattern, then pulling individual links out of your profile with a disavow file was never going to be the lever that fixes it. You can spend a week grading links and still be measuring the wrong object.

Here’s where the keyword’s promise gets delivered, on terms that actually match how Google behaves. Stop asking “is this one link toxic?” The suppression flag is a state triggered by a pattern across your profile.

Before any tool, answer one question honestly. Did you, or anyone working for you, deliberately build manipulative links? Bought guest posts, PBN placements, paid link packages, exact-match anchors engineered at scale?

Your own history is the highest-signal input you have, and it’s free. If the answer is a clean no, the odds that you’re sitting on a real problem drop hard, and most of what the tools flag is noise you can move past. If the answer is yes, you already know where to look, and no toxicity score is going to tell you more than your own records will. Nobody’s guide leads with this, and it’s the most useful 30 seconds you’ll spend.

If you’re going to look, look at the patterns that map to documented behavior, not cosmetics. The unit of analysis isn’t the link, it’s the referring domain and the pattern across your profile. Four patterns are worth your attention:

Those four are the signals that track real, documented attributes. The rest is decoration.

Notice what they have in common: none of them can be read off a single link. Anchor distribution is a shape across your whole profile. A velocity spike only exists relative to your baseline. Network affiliation is defined by how domains relate to each other, not by any one of them in isolation. Relevance drift is a judgment about fit.

Every signal that maps to something Google appears to actually measure is a pattern-level signal, which is the tell that pattern is the right unit and the individual link is the wrong one. A tool that grades links one at a time is grading the wrong thing by design, no matter how many markers it stacks into the score.

Every guide lists these as “warning signs.” Almost none tell you they mostly don’t matter on their own:

Any one of these in isolation is not a reason to touch your disavow file. They only start to matter when they cluster into a pattern, and by then it’s the pattern you’re reading, not the individual flag.

Toxicity scores exist, and they’re not useless, they’re just widely misunderstood. Moz’s Spam Score is built from a couple dozen signals. Semrush’s Toxicity Score weighs dozens of markers. SE Ranking runs its own 0 to 100. They routinely disagree about the same link, and the reason is simple: they measure different bundles of signals, and none of them can see Google’s document-level flag, because that flag isn’t public.

So use the number as a flashlight, never as a verdict. A high score is a prompt to go look at the pattern around that link, whether it’s part of a cluster, a spike, or an overoptimized anchor set. It is not a sentence handed down. Treat a “toxicity score” as one signal that helps you surface patterns worth a human look, and you’ll get real value from it. Treat it as a diagnosis and you’ll disavow good links off a false positive, which is the exact mistake that makes the whole category disreputable.

You don’t need a paid tool to do the meaningful version of this. Open Search Console, go to Links, and export your top linking sites and top linking text. Twenty minutes:

If those three come back clean, and you didn’t build manipulative links, you’re done. Close the tab.

A paid tool speeds this up and adds coverage Search Console can’t give you, most usefully a fuller picture of referring domains and their overlap. That’s a fair reason to use one. What it can’t give you is a verdict, because the thing that would let it deliver a real verdict, Google’s document-level flag, isn’t visible to anyone outside Google. So the workflow is the same whether you pay or not: surface the pattern, then read it with your own judgment and your own memory of what you built.

Here’s the contrarian core, stated plainly: for most sites, the right amount of disavowing is none. The tool is a chainsaw, and most people reaching for it have a weed problem, if they have a problem at all.

The decision tree is short:

You don’t have to take our word that disavow is overused. Google’s own people have said as much. Gary Illyes, at PubCon, put it directly: “If it were up to me, I would do away with the disavow link tool because, often enough, it hurts more than it helps many sites.” He said it in 2019 and things haven’t changed, his reasoning was exactly the failure mode above: people disavow perfectly good links because a third-party tool threw a false positive.

In 2024, John Mueller said he was sure Google will remove the disavow tool at some point, and that most sites simply don’t need it without a link-scheme history or a manual action. He’s also called disavow work done without a manual action a billable waste of time.

However, the tool is still live in 2026, and Mueller confirmed in early 2026 that an undocumented domain: directive in the disavow file still works, which is not the behavior of a feature about to be switched off. So, Google’s staff would remove it, and Google hasn’t, meaning that its removal isn’t on the horizon.

If you’re in one of the two “yes” cases, work at the domain level, not the URL level (for anything that’s clearly a network), and disavow only what you’re confident is manipulative. Never bulk-upload a tool’s “toxic” export wholesale. That’s how a clean editorial link ends up in your disavow file, and unlike a bad link, a disavowed good link is a real loss because you’ve told Google to stop counting a vote you earned.

It’s rare, and it’s mostly not the threat it’s sold as. The same anchor-spam attributes that track when a spam spike started and ended suggest Google models this exact scenario, a sudden unnatural burst of links you didn’t build, and treats it as noise to discount rather than a reason to punish you. If a competitor points junk at your site, the realistic outcome is that SpamBrain neutralizes it and nothing happens. Watch your profile, keep your own link-building clean, and don’t lose sleep over it.

The permanent fix isn’t a tool or a monthly audit. It’s building a link profile that can’t trigger the flag in the first place, because it’s made of the kind of links Google is designed to reward.

The leak is instructive here too. Attributes like context2, fullLeftContext, and fullRightContext suggest Google reads the text surrounding a link, not just the anchor, and anchorMismatchDemotion suggests it discounts links where the source and target have nothing to do with each other. In plain terms, relevance is the thing Google is actually reading, links included. A relevant, editorially placed link inside a paragraph about your topic is worth more, and is structurally incapable of looking like spam.

So the standard is old and boring and correct — earn links you’d still want if search engines didn’t exist. Links from real publications, industry sites, and partners who mention you because you’re worth mentioning. Build those, in context, on topic, and the whole “toxic backlinks” question stops applying to you, because there’s no pattern to trigger anything.

Keep a simple record of every link you or an agency deliberately place, so that if a question ever comes up, you can answer “did you build them?” from a document instead of a guess. Vary your anchors naturally, because a clean profile is mostly branded and URL anchors with money-keyword phrases as the exception.

And resist the packaged-link temptation entirely, the cheap bulk placements that create the exact velocity spike and network footprint the anchor-spam attributes are watching for. A profile built this way doesn’t need auditing, because there’s nothing in it to audit. If you’d rather have that handled by people who do it for a living, our managed link building work is built exactly around the relevance and editorial context over volume.

The comfortable belief in this space is that toxic backlinks are a scoreable property of individual links, something a 0 to 100 number can diagnose and a disavow file can cure. It’s comfortable because it’s actionable, and it sells tools and audits. It’s also wrong, and admitting that is awkward for anyone whose business runs on the score.

What the documentation and the experiments point to is less tidy. Toxicity isn't a property of a link, but a state that lives on the page, triggered by a pattern. That pattern quietly switches off your anchor advantage instead of penalizing you. That’s why no penalty shows up, why real damage sometimes does, and why disavowing individual links moves almost nothing. You were grading the wrong unit the whole time.

For the overwhelming majority of sites, managing toxic backlinks is a solution to a problem you don’t have, and the hours would be better spent earning relevant links than scoring bad ones. Score the pattern if you must. Then go build something worth linking to.

Toxic backlinks are low-quality, spammy, or manipulative links pointing at your site that could be read as an attempt to game rankings, such as PBN links, comment spam, paid links, and exact-match anchors at scale. Toxic is an SEO-tool term. Google’s own word for the same thing is “link spam.”

Both. The term “toxic” is made up by SEO tools that needed a score to sell. The underlying mechanism is real but works differently than the label implies. Google doesn’t penalize a toxic link, it can suppress the anchor-signal boost on a document when it detects a spam pattern across many links.

Usually not directly. Google typically neutralizes spam links so they simply stop counting, with no penalty. The one real risk, per leaked documentation, is document-level anchor suppression: when a spam pattern trips a flag, your page can lose the ranking boost its anchor text normally provides. There’s no notification when that happens.

First ask whether you deliberately built manipulative links, your own history is the strongest signal. Then spend 20 minutes in Google Search Console reviewing your top linking sites and top linking text for identical money-keyword anchors, unrecognized domain clusters, and unexplained link spikes. Score the pattern across your profile, not individual links.

Probably not. Disavow only if you have a manual action for unnatural links, or a deliberate history of buying or building manipulative links. For everything else, including tool-flagged noise, leave it alone. Two disavow experiments found that removing tool-flagged toxic links did close to nothing.

Google’s own staff would like it gone. Gary Illyes said he’d remove it because it hurts more sites than it helps, and John Mueller expects Google to remove it eventually. But it’s still live in 2026, and Mueller confirmed an undocumented directive still works that year, so removal isn’t imminent. Don’t act out of fear it’s leaving.

Rarely. Real, effective negative SEO via bad links is uncommon. Google’s systems are built to neutralize sudden unnatural link bursts, and leaked attributes tracking when a spam spike starts and ends suggest Google models exactly this scenario. If a competitor points junk links at you, the realistic outcome is that those links get discounted and nothing happens.

If links were simply neutralized, there’s nothing to recover from, they never counted against you. If a suppression flag was triggered by a manipulative pattern, recovery depends on the pattern clearing over time, not on a disavow file. There’s no fixed timeline and no reset button, which is why prevention beats cleanup.

There isn’t one, because the number isn’t a verdict. Moz, Semrush, and SE Ranking all use different signals and routinely disagree about the same link, and none of them can see Google’s document-level flag. Treat a high score as a prompt to inspect the pattern around a link, never as a diagnosis to act on blindly.

There’s no clear evidence either way. AI Overviews draw heavily on established authority and relevance, so the same fundamentals apply: earn relevant, editorial links and build brand signals that hold up when other signals get discounted. Chasing or disavowing toxic links to influence AI visibility isn’t supported by any current evidence.