Introducing web search monitors

Google's Poisoned Search Results in November 2026: What Developers Should Do

ยท

TL;DR

  • What Nozzle reported: In the November 3 update to its study, the rank tracker Nozzle found page one or page two "poisoned" on 27.6% of the desktop Google results pages it judged on both pages since September 13, against 0.28% before. The daily share peaked at 68.6% on November 2.
  • What that measures: The results Google returned to Nozzle's automated rank-tracking requests. It does not measure what people see in Google Search, and Google has not commented publicly.
  • If a tracked ranking drops: Compare the keyword with its usual sites over the previous weeks, check page two, see whether the drop lasts, and look at Search Console before you report a loss.
  • If your product reads search results: A response can be complete and still wrong. Validate results against the query and the source page, not the status code.
  • Where we fit: Our Web Search API returns ranked results with source URLs and optional page content for applications and AI agents. It is not a Google rank tracker, and we have not measured it as a fix for what Nozzle observed.

What Nozzle found in its November update

Nozzle first published Poisoned Google SERPs on October 10, 2026 and updated it on November 3 with data through that day. The study draws on a sample of the search engine results pages (SERPs) Nozzle tracks for its customers from August 20 to November 3 across 58 countries on desktop, iPhone, and Android. Nozzle judged 68.7 million results pages, 42.5 million of them since September 13.

MeasureBaseline (Aug 20 to Sep 9)Since Sep 13Worst day
Page one or page two poisoned, desktop results judged on both pages0.28%27.6%68.6% on Nov 2
Page one poisoned, every results page in the study0.17%15.0%34.8% on Oct 27

The sequence matters as much as the totals. Nozzle says requests to Google started failing in large numbers on September 13, and the requests that got through were normal until September 20, when poisoned pages first appeared. The share eased in mid-October and then climbed past the September peak of 43.5% to reach 68.6% on November 2.

The denominator matters too. The 27.6% and 68.6% figures cover desktop results where Nozzle could judge both page one and page two, and a result counts if either page is poisoned. Page one alone, across every device in the study, sits at 15.0%. All of it describes results returned to an automated rank tracker. None of it tells you how often a person typing the same query sees the same thing.

What "poisoned" means here

Nozzle's definition is specific. For each keyword it finds the "usual sites": the ones that appeared on a page in at least 60% of that keyword's results from August 20 to September 9. A page is poisoned when fewer than half of those usual sites appear anywhere in the results. A results page counts as poisoned when page one or page two is.

What makes this hard to catch is that the page looks fine. Nozzle describes complete, normal-looking pages with no error, no CAPTCHA, and the same template, where the organic results keep 30.0% of the sites from the keyword's previous results, against 88.2% in the baseline. Social platforms take their place, Facebook most of all.

This is not the security meaning of the term. SEO poisoning usually describes attackers pushing malicious pages into search results. Nozzle's use of the word says nothing about malware or hacked sites. It describes a results page whose usual sites have been swapped out.

Nozzle reads the pattern as Google's response to AI scraping, with rank trackers caught in it. That is Nozzle's interpretation. Search Engine Land reported on November 20 that Google has not commented publicly on poisoned results and has not confirmed whether they are intentional, while vendors including Semrush and AccuRanker have published their own responses. Treat the numbers as one vendor's well-documented observations, not as a confirmed Google policy.

How to tell a poisoned reading from a real ranking change

A tracked site that falls off page one is the symptom most teams will notice. Where a project's own site was one of the keyword's usual sites, Nozzle found it outside the top 10 on 63.7% of poisoned results, against 2.6% in the baseline. Before you treat a drop as real, run the checks Nozzle recommends:

  1. Do not trust a complete page. A poisoned result has no error and no CAPTCHA. Nothing in the HTML marks it.
  2. Compare against the keyword's history. Use the usual sites from the previous weeks, not yesterday's result, which may have been poisoned too.
  3. See whether it lasts. Two in three spells are a single results page. Nozzle found the usual sites back within three days on 45.1% of poisoned results and within two weeks on 89.0%. A real ranking change stays.
  4. Check page two. It is poisoned one and a half times as often as page one, mostly under a page one that looks fine.
  5. Check Search Console. If a site "disappears" from your tracker while Search Console still shows impressions and clicks for the same queries, nothing happened to the site.

Two cautions keep this honest. A change in which domains rank is not proof of poisoning on its own, because rankings also move for ordinary reasons. And real losses happened in the same window: Search Engine Land notes that genuine spam update losses overlapped with poisoned readings from September 20 to October 8. If both your tracker and Search Console show a sustained decline, investigate it as a real change.

What to do if your product depends on search results

Rank reports are one casualty. The broader lesson is for any system that reads search results and acts on them, including AI agents that search the web. Nozzle puts it this way: a refused request tells a scraper to try again, and a poisoned one gets stored as real. Retries and status codes do not catch a response that is well formed and wrong.

A few habits make that failure visible:

  • Log what you received. Store the query, timestamp, location, device, and the full list of result URLs. You cannot audit a bad week later without them.
  • Keep canary queries. Pick a small set of queries whose top sources are stable and well known. If most of their usual sources vanish at once, suspect the data before you suspect the web.
  • Check relevance before position. Have your application test whether each result answers the query, and do not assume the first result is the best source.
  • Read the source. Before an agent cites a page, fetch its content and confirm it supports the claim. A title and a snippet are not evidence.
  • Annotate, do not delete. Mark suspect periods in dashboards and client reports. Separate what the tool recorded from what happened to the site.

Rank tracking and retrieval are different jobs

A Google rank tracker and a web search API for an application answer different questions.

A rank tracker asks where a specific page appears for a specific Google query, location, and device. Its value depends on matching what a person would see, which is the assumption Nozzle's data calls into question for automated requests.

A retrieval pipeline asks which sources can support an answer. Position matters less than whether the sources are relevant, whether their content is fresh enough, and what each usable result costs once you count page fetches and retries. A healthy rank report cannot tell you that, and a useful retrieval response cannot tell you a Google ranking.

Keep the two apart when you investigate. Use rank-tracking history and Search Console for ranking questions. Use relevance and source checks for retrieval questions.

Where Context.dev's Web Search API fits, and where it doesn't

Our Web Search API is built for retrieval. It is not a rank tracker, so do not read a result's position as a Google ranking, and it does not detect or repair the substitutions Nozzle observed in rank-tracking data. We have not run it against Nozzle's study and make no claim about it.

What it does is return ranked results with titles, descriptions, and source URLs, with filters for domains, freshness, country, and language. In the same request you can ask for Highlights, the passages from each result page that are relevant to the query, or for the full page as Markdown. That gives your application the page content it needs for the "read the source" check without a second integration:

Shell
curl -s https://api.context.dev/v1/web/search \
  -H "Authorization: Bearer $CONTEXT_DEV_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "what did Nozzle report about poisoned Google SERPs",
    "numResults": 5,
    "freshness": "last_month",
    "highlightsOptions": { "enabled": true }
  }'

Each entry in results carries its url and title, and the passages arrive in highlights.highlights with a per-result highlights.code, so your code can tell a page that was read from one that was not.

Test it the way you would test any search source: with queries your customers actually send. Review relevance and content completeness, compare source dates with your freshness needs, and work out cost per usable result. A search costs 1 credit per 10 results. Highlights add 1 credit per 10 results, and Markdown adds 1 credit for each result page returned. Plans are on the pricing page and the full rules are in the credit documentation.

If you are also dealing with wrapped result links, our write-up of Google's /goto redirect change covers that separate issue.

FAQs

Is this a confirmed Google change?

No. Nozzle measured a pattern in the results returned to its rank tracker and attributes it to Google's response to AI scraping. As of November 20, 2026, Google has not commented publicly on poisoned results, according to Search Engine Land.

Does this affect what people see in Google Search?

Nozzle's data does not measure that. It covers automated rank-tracking requests. Nozzle's own advice points the other way: if Search Console still shows impressions and clicks for a site that vanished from a tracker, nothing happened to the site.

How long does a poisoned reading last?

Usually not long. In Nozzle's data, two in three spells are a single results page, 45.1% of poisoned results have their usual sites back within three days, and 89.0% within two weeks.

Should I pause SEO reporting?

Keep reporting and annotate the suspect data. Compare each drop with the keyword's history, check page two and Search Console, and avoid naming a cause Google has not confirmed.

Can a web search API replace Google rank tracking?

Not if you need Google positions. A web search API supplies sources for a product or an AI agent. Its ordering does not tell you where a site ranks on Google.

Does Context.dev detect or fix poisoned SERPs?

No. Our Web Search API returns ranked results with source URLs and optional page content for retrieval. It does not diagnose rank-tracking data, and we have not measured it against Nozzle's findings. Test it on your own queries before you rely on it in production.

The takeaway for engineering teams

Nozzle's study is a reason to check unusual rank-tracking data, not a verdict on any single ranking change. When a position drops, keep the query, timestamp, device, and both result pages, then compare them with the keyword's history and Search Console.

The lesson for products that consume search results is older than this incident: a well-formed response is not a correct one. Log what you receive, watch a few canary queries, and read the source before you cite it.

Start building with live web context.

Scrape, search, and research the web with one API. Bring the results straight into your product.

1,000free credits

Every month. No credit card required.

Start for freeRead the docs