By Tom Whitfield

Google served misleading search results pages on up to 25% of tracked queries starting September 20, according to a new analysis by enterprise rank-tracking firm Nozzle, which accuses the company of deliberately "poisoning" data collected by rank trackers and AI scrapers.

The anomalies have persisted. Since October 1, the non-standard SERPs have comprised between 5% and 11% of results per day, Nozzle found.

What does 'SERP poisoning' mean?

Nozzle draws a hard line between a refused request and a falsified one. The firm stated:

"That sequence is why we call it poisoning. Google is not refusing these requests; it answers them with a complete page whose results are wrong, and nothing on the page says so. A refused request tells a scraper to try again, while a poisoned one gets stored as real."

Nozzle's read: the practice targets AI scraping and grounding — bulk collection of Google results to feed LLM-generated answers — and rank trackers, which collect SERPs the same way, get caught in it.

The original analysis, published via Search Engine Journal with commentary by editor Roger Montti (writing as martinibuster), adds a further hypothesis. Montti suggests the poisoning may also confound spammers attempting to reverse engineer how pages rank.

How widespread is the disruption?

Very, but unevenly. The analysis found the misleading results vary across countries, industries and search types. Nozzle identified four common traits:

• Artificial SERP volatility
• More user generated content (UGC)
• Fewer AI Overviews
• Stable advertising and shopping results

The volatility numbers are stark. Normal search results retain 88.5% of the same websites across repeated searches for identical keyword phrases. Poisoned results retain only 20.7%.

To a rank tracker, that dilution reads as extreme ranking turbulence: a rotating cast of sites briefly appearing for the same keywords. Most poisonings are fleeting, per Nozzle's data:

• 80% lasted for one search result
• 15% lasted two successive results
• 5% lasted three or more

In plain words, four in five poisoned responses hit once and vanish, which makes them hard to detect and easy to log as real.

Which sites gain and lose?

Facebook, Instagram, LinkedIn, Quora and Reddit appear more frequently in poisoned SERPs. MapQuest, Yelp and YouTube appear less often. The analysis offers no explanation.

Local search offers no escape. Nozzle stated:

"Local targeting makes no difference. City and postal-code SERPs look worse, at 8.4% and 10.7% against 7.8% for whole countries, only because they lean toward mobile and 'near me'. The same phrase tracked nationally and locally, in the same country on the same device, comes out the same: 7.5% and 7.4% across 21,508 such pairs."

That sample size — 21,508 matched pairs — is the strongest evidence in the analysis that poisoning is systemic rather than query-specific.

Where did the AI Overviews go?

Poisoned SERPs cut expensive search features. AI Overviews appear in 34.9% of poisoned results versus 54.2% in the baseline. Nozzle attributes this to inference cost: generating an AI Overview costs more than serving ten blue links.

The feature-level shifts, per Nozzle's data:

• "People also search for": 52.7% in poisoned SERPs versus 87.2% baseline
• Image packs: 44.2% versus 23.0% — nearly twice as often
• Forum blocks: 27.0% versus 20.5%

Why image packs multiply and Related Searches shrink remains open. The analysis speculates only: Google may be minimizing actionable data, or cheaper features may simply survive the cut.

Two further findings round out the picture. Google displays fewer pages of results on some desktop searches, and ad plus shopping results remain largely untouched — suggesting the poisoning spares revenue-generating placements.

What it means for marketers

If Nozzle's interpretation holds, rank-tracking reports have been partially unreliable since late September, and any SEO benchmarking built on scraped SERPs carries noise that looks like genuine volatility. Marketers comparing quarter-over-quarter ranking data should treat September 20 onward with caution, because 5-11% of daily results may be fabricated pages that never reached real users.

The analysis also raises a harder question for the industry: if Google can detect and selectively alter scraped results at scale, every third-party SERP dataset — including those grounding AI answers — is now contestable at the source.