How Google detects paid links

Google has not published its method. What is known is that a machine learning system called SpamBrain looks for sites that buy links and sites that sell them, that patterns across many links matter more than any single one, and that human reviewers and spam reports add a manual layer.

In short

  • Google is reported to have said in December 2022 that SpamBrain detects both sites that buy links and sites used to pass outgoing links, and neutralises the links it finds.
  • Detection works mainly on the selling site, because one publisher that sells links exposes every buyer on it.
  • The 2024 leaked Google documentation describes fields that record spikes of similar anchor text and the quality tier of the linking page, but it does not show how they are weighted.
  • Google said of its August 2026 spam update that link spam is handled by separate systems, which suggests link detection runs continuously.
  • In Editorial.link's March 2026 survey of 518 SEOs, 56% said they believe Google cannot reliably identify paid links.

Google has never published how it detects paid links, and nobody outside the company can see a payment. What it does is infer payment from patterns. The known parts are a machine learning system called SpamBrain, signals in the link graph and in anchor text, and a manual layer of human reviewers and spam reports. This lesson separates what Google has stated, what leaked documents describe, and what practitioners infer. Treat the three differently, because a statement from Google, a field name in a leaked document and a working rule from people who vet publishers carry very different weight, and most of the confident claims you will hear on this subject blur them together.

What Google has said

With its December 2022 link spam update, Google is reported to have said that SpamBrain can detect both sites that buy links and sites used for the purpose of passing outgoing links. Links found this way are “neutralized”, meaning any credit they passed is lost. Two points follow from that statement. Detection covers both ends of the transaction, so a buyer’s exposure depends partly on what the seller’s site looks like. And the usual result is not a penalty. Since Penguin 4.0 in September 2016, Google has said it devalues spam links instead of demoting the whole site, which means the ordinary consequence of a detected paid link is that it stops doing anything.

There has been no named link spam update since December 2022. When Google ran its August 2026 spam update, it said the update “does not target link spam” and that link spam is “handled by separate systems”. The reasonable reading is that link detection now runs all the time. That reading is an inference, not a statement from Google. Its practical meaning, if it is right, is that there is no update date to watch for and no quiet period between rollouts in which a batch of bought links is safe. Google’s link spam updates timeline has the full sequence.

The seller is the weak point

A buyer’s link profile can look tidy. A seller’s site rarely does. A publisher that sells links builds a visible record: article after article linking to unrelated commercial pages with keyword anchors, across niches that have nothing in common. The buyer makes one purchase and sees one article, but the publisher repeats the transaction with every customer, and all of those transactions sit together on the same domain where any crawler can read them.

Once a site is classed as one that sells links, every link it gives can be discounted at once. That is why one bad publisher list exposes many buyers, and why the same inventory sold on several marketplaces is worth less each time it is sold. The patterns below are the ones on a selling site that anyone, including a classifier, can observe, and they work as a checklist when you look at a publisher before paying.

  • A public “write for us” or “advertise” page with a price.
  • Many posts with commercial outbound links and exact-match anchors. A share above 60% is a working red flag in publisher vetting.
  • Outbound links to casino, loan or pharmacy pages from unrelated content.
  • Five to fifteen posts a day across unrelated categories.
  • Rankings for topics that do not match the site’s stated subject.
  • An organic traffic collapse after a spam or core update.

None of these needs inside knowledge to spot, which is the point. If a person with a browser and a backlink tool can see them in a few minutes, there is no reason to assume a system built for the purpose cannot. How to vet a publisher turns that list into a scoring method.

Signals at the buyer’s end

On the receiving site, the signals are about shape and timing, not individual links. Anchor text is the first. A natural profile is mostly brand and URL anchors, because people who link of their own accord tend to name the company or paste the address. A profile heavy with exact-match commercial phrases is the classic sign of built links, and the lesson on anchor text covers it in detail. Timing is the second signal. A sudden burst of links with similar anchors, with no news event to explain it, stands out, and link velocity covers what is normal.

Source overlap is the third, and it is the hardest for a buyer to see from the inside. If most of your new links come from sites that also link to the same set of unrelated commercial sites, the group is connected in the link graph even when the sites look independent. Analysts describe this graph relationship as the largest link footprint. Imagine a pet food subscription service whose newest links all come from blogs that also link to the same loan broker and the same casino: each blog may have its own design and owner, yet the shared customers tie them together. Relevance is the fourth signal. Links from pages and sites with no topical connection to the target are easier to discount.

What the leaked documents describe

In 2024, internal Google API documentation leaked. It lists fields and modules, not weights, and Google has not confirmed how any of them are used. Analyses by iPullRank and Hobo describe several that relate to this subject. In each case there are two layers to keep apart: what the documents contain, and what analysts infer from it. The first layer is documented. The second is interpretation.

The documentation contains fields counting spam-phrase anchors, the days over which they appeared and whether they were demoted. From that, analysts infer that spikes of similar anchors are measured, and that anchors inside the spike can be discounted. It also contains Penguin-era anchor statistics, including counts of dropped anchors, which analysts read as suspect anchors being ignored instead of the page being punished. That inference matches what Google has said publicly about devaluing links since 2016, which is part of why it is taken seriously.

Three further items complete the picture. The leak documents a source quality bucket tied to the index tier of the linking page, and the inference is that links from low-tier pages may count for little or nothing. It documents a demotion for anchor mismatch, from which analysts infer that off-topic anchors or sources can be discounted. And it documents a measure of how focused a site is on its topic, read as a sign that the relevance of the linking site is assessed. All of these line up with the buyer-side signals above, but lining up is the most that can be claimed. The leak and antitrust documents on links explains the status of this evidence.

The manual layer and the two outcomes

Algorithms are not the only route. Human reviewers are the second: Google’s webspam team can apply a manual action for “unnatural links to your site” or “unnatural links from your site”. Unlike algorithmic devaluation, this appears in Search Console. Spam reports are another input. Google accepts reports from the public, including competitors, and says it uses them to improve detection. Then there is public evidence. Seller price lists circulate as spreadsheets, and marketplaces list their inventory to any registered user. Nothing stops a reviewer from reading the same lists a buyer does.

That leaves two possible outcomes, and they differ in every respect that matters to a site owner. In the algorithmic outcome the links are ignored. There is no notification. The symptom is that links stop working and rankings plateau or slide, and the fix is to stop the tactic and earn real links. In a manual action a reviewer applies a penalty, and a Search Console message tells you so. The symptom is a sharp drop for pages or the whole site, and the fix is to remove or disavow, then file a reconsideration request.

The two can meet at the extreme. Google’s Gary Illyes and John Mueller are reported to have said that when a profile is overwhelmingly manipulative, the algorithms can distrust the site’s links as a whole. For the manual route, see link penalties and recovery. The difference in notification is what makes the algorithmic outcome hard to manage: a manual action announces itself, while ignored links give no sign beyond results that fail to arrive.

What Google probably cannot see, and what this means in practice

Detection is not complete, and it is honest to say so. Practitioners largely agree: in Editorial.link’s March 2026 survey of 518 SEOs, 56% said they believe Google cannot reliably identify paid links. An Ahrefs study of the top 110,000 sites, published in May 2020, found the sponsored attribute on about 0.01% of referring domains, so most paid links carry no label. The survey figure is a measure of what practitioners believe, which is different from a measure of what Google can do, and it should be read that way.

A single paid placement on a real publication with editorial standards, real readers and a relevant article looks the same as an earned link. What gets caught is scale and repetition: the same sellers, the same anchors, the same pace. This is also why the usual cost of buying links is wasted money, not a penalty. The risk sits in the pattern, so the decisions that matter are the ones that decide whether your links form a pattern with other buyers’ links.

In practice that means judging the publisher’s outbound profile before your own link, because you inherit its record. It means avoiding inventory that every buyer can see, such as a link marketplace list sold unchanged to all. Keep anchors mostly brand and URL. A sponsored attribute removes the policy risk, at the cost of ranking credit, which makes it the right choice when the placement is wanted for its readers. And do not treat absence of a penalty as proof that links are counting, since the usual outcome of detection is silence.

How to buy backlinks safely applies this to purchasing decisions. Terms such as SpamBrain, manual action and neutralised link are defined in the glossary.

Common questions

Can Google tell if a link was paid for?

Not by seeing the payment. Google infers it from patterns, such as a site whose articles link out to unrelated commercial pages with keyword anchors. A single well-placed paid link on a real site is hard to tell apart from an earned one.

What is SpamBrain?

SpamBrain is Google's machine learning system for spam detection. Google is reported to have said in December 2022 that it is used to detect sites that buy links and sites used to pass them on.

What happens when Google detects a paid link?

Usually the link is ignored and passes no credit, with no notification. If paid links dominate a profile, a human reviewer can apply a manual action for unnatural links.

Does a sponsored attribute protect me?

Yes. A paid link with rel="sponsored" or rel="nofollow" is allowed under Google's spam policies. It is treated as advertising, not as a ranking signal.

Do competitors report paid links?

They can. Google accepts spam reports from the public and says it uses them to improve its detection systems. A report does not guarantee action on a single site.

Vendors to look at

  • BazoomEditor's pick

    Sponsored content and link marketplace, managed service

  • MotherlinkEditor's pick

    Backlink services, guest posts, niche edits, full SEO