Semrush Confirms Poisoned Google SERPs; 12 of Our 24 Repeat Fetches Failed to Agree

Ras Al Khaimah, United Arab Emirates, 28 October 2026.

Semrush has become the latest SEO data company to say publicly that Google is returning altered results to some automated searches. In an update on its blog dated 27 October 2026, the company said some results pages look normal but differ from what a typical user would see, called it an industry-wide problem rather than one specific to Semrush, and said it now filters and re-collects suspect data. It follows research from AccuRanker, first published on 19 October and updated through 26 October, and two reports from Nozzle that we covered on 13 October and on 25 October. Google has not commented.

To see what this looks like from the outside, we fetched the same eight Google searches three times within five minutes on 28 October, twice through one SERP data API and once through another. Half of the 24 fetches shared fewer than half of their top-10 domains with the other two fetches of the same keyword.

What changed

• Semrush is now on record. Its update describes two forms of the problem: a page where the whole first page is clearly wrong for the query, and a page that looks right at first but carries loosely matched results lower down. Its example was a Brazilian search for a baby stroller model that returned a dental supply page, because both share the Portuguese word "carrinho".
• No size estimate from Semrush. The update gives no share of keywords affected. Semrush says it drops clearly invalid results, has changed some data sources and re-collects doubtful results, and asks customers to compare unexpected swings with their Search Console data.
• AccuRanker has put numbers on Google's recent steps. It reports that detection tightened on 15 October (mostly automated Android sessions), on 20 October (hardest on desktop) and again on 23 October across both devices. On 24 October, 40% of its first desktop fetches came back poisoned before its checks, and 3.23% of stored desktop keyword-days still failed its after-the-fact test. By 25 October that fell to 0.6% on desktop and 1.0% on mobile.
• The first fetch is the weak point. AccuRanker says Google still poisons 4% to 6.5% of its first fetches, against 1.6% to 3.3% before 23 October, and that a second fetch with another session came back normal about three times in four. More than nine in ten poisoned runs for a keyword lasted one day.

Timeline: what the trackers saw and what Google logged

We set the vendors' published dates against Google's Search Status Dashboard. The dashboard lists no ranking update since the September 2026 spam update finished on 8 October.

Date (2026)	Event	Source
24 September	Google's September 2026 spam update starts (16:15 UTC). AccuRanker's data first shows unusual results late the same day.	Google dashboard; AccuRanker
25 September	30.2% of desktop and 25.7% of mobile page-1 results fail AccuRanker's check, its worst day.	AccuRanker
8 October	Spam update rollout complete.	Google dashboard
15 and 20 October	Google tightens detection, first on automated Android sessions, then hardest on desktop.	AccuRanker
22 October	Nozzle publishes research on a second fault: pages that answer only part of the query.	Nozzle; our 25 October report
23 to 24 October	Larger step on both devices. 40% of AccuRanker's first desktop fetches poisoned on 24 October, the same day trackers reported a volatility spike.	AccuRanker; our volatility report
27 October	Semrush publishes its update. Trackers spike again with no Google update logged.	Semrush; our 27 October check

Speculation, not confirmed: some of the swings tools reported on 24 October, the poisoning peak, may reflect collection problems rather than ranking changes. Our 14-keyword check on 27 October found no extra churn in live results.

Our test: three fetches, eight keywords, five minutes

We requested Google's United States desktop results for eight common searches three times: fetches 1 and 2 about a minute apart from one SERP data API, and fetch 3 from a second API. For each fetch we asked how many of its top-10 organic domains appeared in either of the other two. AccuRanker treats a page as poisoned when its top 10 keeps fewer than half of the sites from neighbouring days, so we used the same 50% line, applied across fetches minutes apart rather than days.

Pairwise, the picture by keyword was mixed:

Keyword	Shared, fetch 1 and 2	Shared, fetch 1 and 3	Shared, fetch 2 and 3	In all three	What stood out
how to change a tire	6	7	6	6	Stable across all three
best running shoes	7	3	4	3	Same-API fetches near identical
crm for small business	7	0	0	0	Same-API fetches agreed; the second API shared none of their domains
mortgage rates	3	2	3	2	Fetches 1 and 2 led with state rate pages for a national query
noise cancelling headphones	1	1	8	1	Fetch 1 the odd one out
project management software	3	3	1	1	Fetch 2 leaned to "free" tool pages
baby stroller	1	0	2	0	All three disagreed
sourdough bread recipe	0	1	1	0	Three different sets of recipe sites

• Only one keyword was stable everywhere. For "how to change a tire", every fetch kept at least 86% of its domains in the others.
• Agreement did not prove accuracy. For "crm for small business", the two fetches from one API shared seven domains, so a check that compares a fetch with its own history would have passed them. The second API returned a different set entirely. AccuRanker says its history check cannot help where a keyword lacks enough history, and our case suggests a repeated fetch from the same source can repeat the same answer.
• No off-topic swaps. Every result we saw was about the query's subject. The disagreement was over which sites ranked, not which topic, which is the harder form to catch by eye.

We cannot say which fetch matched what a person in the United States saw: APIs can query Google from different places, and Google varies results for ordinary users too. The test shows something narrower: on 28 October one automated snapshot of page 1 was not a reliable record of where a site ranked.

Why Google might be doing it

Google has not explained it. Its spam policies class scraping results for rank-checking without express permission as machine-generated traffic that violates the policies and the Terms of Service. Semrush cites community reporting that the aim is to limit large-scale data collection for AI systems; that remains unconfirmed.

What this means for SEO teams

• Treat one-day drops as unverified. AccuRanker's data says most poisoned runs last a day. Wait for a second day, or a second fetch, before reporting a ranking loss.
• Use Search Console as the reference. Semrush now gives the same advice. If impressions and average position hold while a tracker shows a crash, the tracker is the likelier fault.
• Ask vendors what they check. Does the tool refetch suspect pages with a fresh session? Does it compare against history, and what happens on keywords with no history? Has it reprocessed data from 24 September onward?
• Repeat single-pull research. SERP overlap studies and AI Overview counts built from one fetch per keyword since late September may carry noise.
• Annotate the period. Mark 24 September onward in dashboards, and 24 October as a bad day for desktop data.

Methodology

On 28 October 2026, between 05:20 and 05:25 UTC, we requested Google organic results for eight keywords (United States, English, desktop, top 10 organic positions). Fetches 1 and 2 came from the same commercial SERP data API, roughly one minute apart; fetch 3 came from a second commercial SERP API using a country-level United States location. We kept organic results only, so some fetches had six to nine organic positions after SERP features. Each URL was reduced to its registrable domain (www dropped; subdomains such as cooking.nytimes.com counted as nytimes.com and us.amazon.com as amazon.com), and duplicate domains within a fetch were counted once. For each fetch we computed the share of its unique domains that appeared in either other fetch of the same keyword and flagged values under 50%. Topic relevance was judged by hand from titles and snippets. The sample is small and not random, covers one short window and two sources, and has no real-user baseline. Vendor figures come from the companies' published pages.

Conclusion

With Semrush joining AccuRanker and Nozzle, three rank tracking companies now say Google is feeding some automated searches results that real users would not see, and none of them claims to have solved it. AccuRanker's figures show the problem rising and falling with each step Google takes, with 24 October the worst recent day for desktop collection.

Our small test found that the problem is easy to see and hard to resolve: half the fetches disagreed with the others within five minutes, and two that agreed with each other did not agree with a third source. Until Google says what it is doing, rank data from any single fetch is best read as a sample, with Search Console as the record.

Questions and answers

What are poisoned SERPs?

They are Google results pages served to some automated searches that look normal but list a different set of organic results from the ones users see. Semrush describes two forms: a page that is clearly wrong for the query, and one that looks right but carries loosely matched results lower down.

Has Google confirmed it is serving altered results?

No. Semrush, AccuRanker and Nozzle all say Google has not commented. Google's spam policies say scraping results for rank-checking without express permission violates its policies and Terms of Service.

How common are poisoned results now?

AccuRanker reports that 0.6% of desktop and 1.0% of mobile page-1 results still failed its check on 25 October, down from 3.23% on desktop on 24 October and far below 30.2% on 25 September. It says 4% to 6.5% of its first fetches are poisoned before its own checks and refetches.

How can I tell if a ranking drop is real?

Compare it with Search Console impressions and average position, which Semrush also recommends, and wait for a second day of data. AccuRanker says more than nine in ten poisoned runs for a keyword last a single day.

What did the SaaS Link Building test find?

We fetched eight Google searches three times within five minutes on 28 October 2026, using two SERP data APIs. In 12 of the 24 fetches, fewer than half of the top-10 domains appeared in either other fetch of the same keyword. Only "how to change a tire" was stable across all three.

###