SpamBrain: How Google Detects Spam in 2026

SpamBrain is Google's AI-based spam-prevention system, not a single model. It runs as a platform hosting specialized detectors, each trained for one abuse category. Google ships detector improvements without publishing the schedule (https://developers.google.com/search/docs/appearance/spam-updates).

Detectors run at two gates. At crawl time, SpamBrain classifies fetched pages and blocks detected spam from entering the index. At serving time, classifiers rerank live results to demote surviving spam. A page must pass both gates to rank.

Google's spam-policies guide lists nine categories SpamBrain enforces (https://developers.google.com/search/docs/essentials/spam-policies). Cloaking, doorway abuse, expired domain abuse, hacked content, hidden text and links, keyword stuffing, link spam, machine-generated traffic, and malicious practices.

Detection ran across the entire crawl queue daily as of Google's last published webspam update on 2025-12-10 (https://developers.google.com/search/docs/appearance/spam-updates). SpamBrain also screens URLs submitted through Search Console and sitemaps. Verified spam submissions get blocked before they hit the index.

Cloaking tops the detection list. SpamBrain compares HTML served to Googlebot against the version served to headless Chromium at render time. Mismatch scores flag the URL for immediate demotion or removal from the index (https://developers.google.com/search/docs/essentials/spam-policies). Detection runs inline, not in post-batch.

Hacked content triggers a separate detector. Code injection, page injection, and hidden redirects each leave distinctive markers in the rendered DOM. SpamBrain flags these markers automatically, then routes the page for further review by Google's trust and safety team (https://developers.google.com/search/docs/essentials/spam-policies).

Link spam detection focuses on patterns, not single links. SpamBrain scores outbound link velocity, anchor text distribution, and host-to-host co-occurrence across your backlink profile. Networks with thousands of cross-linked domains get clustered and demoted as a unit. The cluster penalty propagates across all linked properties.

Manual link cleanup does not restore lost ranking benefit after a link spam update fires (https://developers.google.com/search/docs/appearance/spam-updates). The penalty is permanent even if you disavow or remove the violating links. Future compliance matters. Past ranking transfer does not.

Google shipped three named spam updates in 2025. The August 26 to September 22, 2025 rollout targeted thin content and keyword stuffing, per Raptive's network-wide analysis published November 12, 2025 (https://raptive.com/blog/heres-what-we-learned-from-googles-august-2025-spam-update/). Posts under 500 words consistently underperformed.

Ad-density thresholds shifted in the same update. Sites running above 25% ad-to-content ratio saw ranking declines, per the same Raptive analysis. Stable performers sat near 20% desktop and 24% mobile. The ratio window tightened by roughly 4 percentage points year-over-year (https://raptive.com/blog/heres-what-we-learned-from-googles-august-2025-spam-update/).

Update impact concentrated in specific verticals. Sports publishers lost traffic to AI Overviews first, then to consolidation in major networks, per Raptive's same analysis (https://raptive.com/blog/heres-what-we-learned-from-googles-august-2025-spam-update/). Food and evergreen sites held steady. SpamBrain does not punish categories uniformly.

Domain authority shielded publishers. Sites with Domain Authority scores above 40 plus strong branded search volume saw ranking gains during the August 2025 spam update, per Raptive's network data (https://raptive.com/blog/heres-what-we-learned-from-googles-august-2025-spam-update/). Smaller publishers without branded search took the losses.

Manual actions remain the escalation path when SpamBrain's automated confidence is low. Google's spam policies document covers nine categories that trigger human review: cloaking, doorway abuse, expired domain abuse, hacked content, hidden text and links, keyword stuffing, link spam, machine-generated traffic, and malicious practices (https://developers.google.com/search/docs/essentials/spam-policies).

Reconsideration runs 4-12 weeks after submission. SpamBrain needs observation time to reclassify a domain as compliant. Old expectations of recovery within days are gone. Compliance is a continuous state monitored by the same detectors that flagged you originally.

Link spam updates carry a permanent ranking penalty (https://developers.google.com/search/docs/appearance/spam-updates). Any ranking benefit gained from violating links is lost the moment SpamBrain's classifier scores the network. Disavow files do not restore the lost signals. They only prevent future penalty stacking.

Your post-spam-update triage starts the morning after Google announces a new rollout. You open Search Console's Manual Actions report and scan for any new entries. You then pull your backlink profile from your third-party tool and sort by acquisition date for the last 30 days.

You score new domains against your existing anchor text distribution. Commercial anchor terms above 40% trigger a deeper look at host quality. SpamBrain's link detector weights anchor patterns heavily. You reject or disavow suspect domains before the next crawl lands. Skipping the audit today means filing reconsideration in 8-12 weeks.

You grep your CMS export for the policy signatures. White-on-white CSS, off-screen positioned text, font-size zero, links passing ranking credit without nofollow or sponsored. Any hit is removed and republished before the next index request fires. SpamBrain cannot flag what no longer exists at crawl time.

Note the gap. This post covers SpamBrain's architecture and recent update impact from three primary sources. Google's spam-policies documentation, Google's spam-updates documentation, and Raptive's August 2025 update analysis. SpamBrain's current detection thresholds, model architecture, and false positive rates remain undisclosed. Engineers operate against observable signals only.