Since mid-September 2026, many marketing teams have watched their rankings collapse in their tracking tool, only to see them reappear a few days later. According to Nozzle's study published on 10 October 2026, some of those drops never happened in real search. Several tool vendors report that some automated tools now receive results pages from Google that look complete but in which the usual sites have been swapped for others.
The industry has given the phenomenon a name, "poisoned SERPs", and several vendors have since published their own findings. It coincides with a Google spam update, which makes September and October reports harder to read. It also affects, by extension, the tools that monitor AI Overviews, the AI-generated summaries Google shows above its results.
This article brings together what vendors have measured, explains why their figures differ so widely, and sets out a method for telling a real drop from bad data before you report it to a client or to your leadership.
Key takeaways
- Several rank-tracking vendors confirm that Google has been returning altered results to their collection bots since September 2026.
- The share of affected pages varies widely by vendor, because each one measures the problem with its own method and its own defences.
- A poisoned page almost always looks like a real one: the keyword's history, or a second fetch, remains the surest way to expose it.
- Search Console remains the reference for checking a drop, because its data comes from real searches rather than from a bot.
- Tools that track AI Overviews from Google's results pages face the same risk as conventional rank trackers.
What tool vendors have observed since September 2026
According to Nozzle, the first signs date from 13 September 2026, when requests sent to Google started failing in large numbers. A week later, bots began receiving pages that loaded normally but listed different sites from usual.
Nozzle, which sells a rank tracker, has published one of the most detailed quantified studies of the problem, written by Derek Perkins and updated on 21 October 2026. It draws on a sample of 56.1 million results pages tracked for its customers, on desktop and mobile, across more than fifty countries. A page counts as poisoned when fewer than half of the sites that usually appear for a keyword are still there. According to the update, the share of poisoned first pages rose from 0.17% before 13 September to 8.7% since.
In a report, a site missing from a poisoned page appears to have lost its position, even though nothing changed in real search. The same study found that a Nozzle customer's own site was missing from the top 10 on 60.8% of poisoned pages, compared with 2.7% before. Most episodes are short, with the usual sites returning within three days in 67.9% of cases. According to Nozzle, only the placements advertisers pay for, namely ads and shopping results, hold steady on these pages.
Other vendors have confirmed the pattern, though at different magnitudes. In an update published on 22 October 2026, Ahrefs describes an industry-wide problem and reports that between 5% and 20% of results are affected, depending on the data provider. Semrush told Search Engine Roundtable it had been seeing inaccurate results since late summer, with what it considers a limited effect on tracked rankings. As of 28 October 2026, Google had made no public statement on the matter.
Why do the measurements differ so much between tools?
Each vendor measures the problem on its own data, collected with its own safeguards. A tool that filters and re-fetches suspicious pages keeps far fewer of them than one that stores them as they arrive.
The sharpest contrast is between Nozzle and AccuRanker. In his analysis of 19 October 2026, AccuRanker's Peter Emil Tybirk compares each first page with the readings from the day before and the day after, on a fixed sample of one tracked keyword in 200. On that basis, the share of poisoned pages in its stored data peaked on 25 September 2026 at 30.2% on desktop. It then fell to an average of 1.47% between 1 and 9 October, which AccuRanker puts at about a fifth of Nozzle's figure over the same days. AccuRanker notes that its data goes through checks before it is stored, a difference that may account for part of the gap.
The table below brings together the published figures. They cannot be compared directly, since methods, samples and periods differ, and none has been independently audited.
| Source and date | What is measured | Published magnitude |
|---|---|---|
| Nozzle, 10 October 2026, updated 21 October | Poisoned first pages since 13 September | 8.7%, against 0.17% before |
| AccuRanker, 19 October 2026 | Poisoned first pages, average from 1 to 9 October | 1.47% |
| Ahrefs, 22 October 2026 | Results affected, depending on the data provider | Between 5% and 20% |
| SISTRIX, 7 October 2026 | Reliability as assessed by the vendor | Top 10 and Visibility Index considered reliable |
Geography adds another layer of variation. In Nozzle's data, France is among the least affected markets, at 5.3% of pages poisoned. Belgium, on the other hand, reaches 15.9%, close to the level Nozzle records in Switzerland. Agencies that track Belgian or Swiss clients should read those reports with particular care. AccuRanker, for its part, sees similar rates across markets, a reminder of how much these figures depend on each tool's own defences.
Part of Google's wider push against automated scraping
These fake results are the latest in a series of measures Google has taken since 2025 to curb automated collection of its pages. SEO tools bear the consequences, whether or not they are the main target.
The earlier steps are documented by vendors and in court filings. According to Google's complaint against SerpApi, covered below, the search engine introduced SearchGuard in January 2025, a JavaScript check designed to filter out automated software. The SISTRIX note cited below says Google has returned only ten results per request since September 2025. Tracking a keyword beyond the first page has required many more requests ever since. The same note says that since June 2026, result links have gone through an internal redirect address, which a tool has to follow to find the destination page.
Google has also taken the fight to court. According to IPWatchdog, it sued SerpApi, a company that resells automatically collected search results, on 19 December 2025, accusing it of circumventing SearchGuard. The complaint describes a volume of requests running to hundreds of millions a day. Speaking to Search Engine Roundtable, Derek Perkins described the methods Google appears to be combining: large-scale IP blocking, more frequent CAPTCHAs and the injection of incorrect data.
SISTRIX offers a reading of Google's motives. In a note published on 7 October 2026, Johannes Beus, the company's founder and CEO, argues that Google mainly wants to make it harder for new competitors such as ChatGPT to access its index. He also sees an effort to stop its results being used to train open-weight AI models. "Analysis tools like ours aren't the target, but they are caught up in it," he writes. This is an interpretation that Google has neither confirmed nor denied.
How can you tell a real drop from bad data?
The most reliable check is to compare the drop shown by your tool with Search Console data for the same query and date. A position that holds in Search Console most likely points to a poisoned reading.
Search Console has a clear advantage here, because its data comes from searches that real users actually ran. Its performance report shows the average position in search results and the number of clicks for each query and page. These figures do not depend on any automated collection, unlike a results page downloaded by a bot. The newest data is still preliminary and may change over the next few hours, so it pays to wait a little before drawing conclusions.
The timing is what makes this tricky. According to the Google Search Status Dashboard, the September 2026 spam update began on 24 September and ran for almost fourteen days. It overlaps with the first wave of poisoned pages, since the peaks published by AccuRanker, on 25 September, and by Nozzle, on 30 September, both fall within its rollout. A drop seen in late September could therefore stem from either cause, and only a keyword-by-keyword check can settle it.
Several signs can guide the diagnosis before you even open Search Console. The table below draws on the recommendations published by Nozzle and AccuRanker.
| What you see | Likely explanation |
|---|---|
| A one-day drop that reverses the next day or the day after | Very likely bad data |
| Several usual competitors vanish on the same day for the same keyword | Poisoned page, with the usual sites swapped out in bulk |
| Local results showing businesses from another town | Poisoned page, according to AccuRanker |
| A drop confirmed by Search Console, in both average position and clicks | A real decline worth analysing |
| A decline that persists across several fetches and several days | A real decline, possibly linked to the spam update |
A second fetch is another useful test, if your tool lets you refresh a keyword. According to AccuRanker, a repeat request returns normal results about three times in four. A position that reappears after a refresh was most likely never lost. Finally, we recommend looking at trends over several days rather than at a single reading, which is enough to rule out most episodes, since they are usually short.
What this means for AI visibility tracking
AI Overviews appear on Google's results pages, so the tools that monitor them collect the same pages as rank trackers. In principle, they face the same risk of poisoned data.
AI visibility tracking works in much the same way as rank tracking. A tool regularly asks an engine the same questions, records the answer and logs which sites are cited, as we described in our article on how to prove your GEO is working. In Google's case, that collection runs through the very pages the search engine has been altering for bots since September.
Nozzle's study gives a first indication of the effect on AI Overviews. On a poisoned page, the AI Overview shows only 37.5% of the sources it cited on the keyword's previous reading, compared with 68.1% between 20 August and 9 September. To our knowledge, no study has yet measured what share of AI Overview readings is affected, and it would be unwise to infer a rate.
Our own tracking is not entirely immune to this risk, and the difference between engines matters. Our citation tracking queries ChatGPT, Perplexity, Gemini and Claude through their APIs, which are not directly targeted by the measures described here. For AI Overviews, however, our readings rely on a specialist provider that checks Google's results pages, so they can be affected in the same way as rank trackers. That is why we treat the sudden disappearance of an AI Overview citation as a signal to confirm over several readings, not as a finding.
The situation also highlights a known limit of measurement in answer engines. Each engine picks its own sources, as we showed in our article on why ChatGPT and Perplexity cite different sources. Reliable measurement combines several readings, several engines and several weeks, and it rests on what the site controls directly. An AI visibility audit of your pages, which analyses their content rather than Google's results, is not exposed to this kind of bad data.
Our recommendations for your September and October reports
If you run an agency, review the reports sent to clients since mid-September 2026 before drawing any conclusions from them. Any sharp drop since 20 September calls for a check in Search Console, especially for Belgian or Swiss clients. When you present those reports, flag the disruption in one sentence, without alarm, and base your conclusions on Search Console clicks and average position rather than on tracker readings alone.
If you manage SEO in-house, avoid reacting to an isolated drop with a page redesign or a content rewrite. Poisoned episodes often last a single reading, and a rushed change could harm a page that had in fact lost nothing. Wait until a decline is confirmed over several days and in Search Console before starting any work.
Finally, ask your tool vendor three specific questions: how it detects poisoned pages, how much of its data has been affected since September and whether it checks a page before storing it. AccuRanker, which has detailed its method, says its checks keep most of these pages out of its data. SISTRIX, for its part, has chosen to focus its analyses on the top 10, which in its view drives visibility and clicks.
At the end of October 2026, nothing suggested that Google intended to back down. AccuRanker had its worst day on 24 October 2026, when 40% of its first desktop fetches came back poisoned before its checks. A cautious way of reading the data, grounded in Search Console and multi-day trends, will probably remain useful for several months.
Frequently asked questions
What is a "poisoned" SERP?
It is a Google results page that looks complete and normal, but where most of the sites that usually rank have been replaced by others. Nozzle, which uses the term in its study of 10 October 2026, flags a page when fewer than half of a keyword's usual sites still appear on it. The page itself rarely gives it away, so comparing it with nearby readings is usually the only way to spot it.
Do ordinary searchers see these fake results too?
There is no evidence of that so far. Ahrefs wrote on 22 October 2026 that the different results are shown to suspected SERP data providers, and every available study is based on automated queries. None of them measures what ordinary users see.
My site dropped in late September 2026. Is it bad data or a penalty?
It could be either, because Google's spam update began on 24 September 2026, while poisoned pages were spreading. Check the same query and date in Search Console: if average position and clicks held steady, the drop in your tracker most likely comes from a poisoned page. A drop that Search Console confirms and that lasts several days points to a genuine change.
Are AI Overview tracking tools affected?
They can be, since AI Overviews appear on the same results pages that Google is altering for automated collectors. According to Nozzle, an AI Overview on a poisoned page shows only 37.5% of the sources cited on the previous reading, compared with 68.1% between 20 August and 9 September 2026. Readings taken through the ChatGPT, Perplexity or Gemini APIs run through other channels and are not directly targeted by these measures.
Should I switch rank trackers?
Not necessarily. It is more useful to ask your vendor how it detects poisoned pages, how much of its data has been affected and whether it checks a page before storing it. AccuRanker, SISTRIX and Ahrefs all say they check results before publishing them. AccuRanker states that its checks keep most poisoned pages out of its data.
