Since September 2026, rank tracker data has been unreliable, and it is not your site’s fault. An independent study by Nozzle covering 29.9 million Google search results pages (SERPs) found that after tracker requests to Google started failing en masse on September 13, 2026, a measurable share of the SERPs served to automated tracking systems began showing swapped-out results — social sites and forums standing in for the usual organic listings. As of the study’s October 21, 2026 update, 8.7% of page-1 SERPs tracked since September 13 were poisoned, and the rate peaked at 43.5% on September 30. Google has not commented on any of it. If your monthly report is built on tracker positions, the numbers may be describing a view of Google that your customers never see.

What does “SERP poisoning” actually mean?

SERP poisoning is when a search engine returns different results to automated rank trackers than to regular users. In Nozzle’s study (published October 10, 2026, updated October 21), a poisoned result page looks normal at a glance — but only about 22.3% of the sites that usually appear for a given keyword are still there, and social sites and forums take their place. On 60.8% of poisoned SERPs, the site being tracked missed the top 10 entirely.

The core measurement, with its as-of dates: initially (October 10), 8.0% of page-1 SERPs tracked since September 13 were poisoned, versus a 0.17% baseline before that date. The October 21 update puts page 1 at 8.7% and 22.6% counting page 1 or 2, with page 2 hit roughly 1.5× harder than page 1. One number deserves a warning: 67.9% of poisoned SERPs returned to normal within three days. That is real, but it is not “resolved” — the poisoning rate resurged to 35.7% on October 21. Anyone telling you this problem has passed is working from a mid-October dip.

What happened in September 2026, and what do the numbers show?

The sequence is: blocking failures first, then poisoned results. Requests from tracking systems to Google began failing at scale on September 13, 2026; the first poisoned SERPs appeared September 20. (The poisoning itself has no documented date before September 20.)

What changedWhenMeasured by
Tracker requests to Google started failing en masseSept 13, 2026Nozzle study, 29.9M SERPs
First poisoned SERPs observedSept 20, 2026Nozzle study
Peak: 43.5% of tracked SERPs affectedSept 30, 2026Nozzle study
Initial figure: 8.0% of page-1 SERPs poisoned (baseline 0.17%)Oct 10, 2026Nozzle study
Dip to 4–6%Oct 14–15, 2026Nozzle study
8.7% page-1; 22.6% counting page 1 or 2; back to 35.7%Oct 21, 2026Nozzle study update
Poisoned SERPs retain ~20.7% of usual sites vs 88.5% normal; fewer AI Overviews (34.9% vs 54.2%); ~80% of poisoning lasts a single pullOct 14, 2026Search Engine Journal
Vendors confirm poisoned results; “Google hasn’t commented publicly”Oct 27, 2026Semrush

Two details from the study are worth keeping. First, the poisoning is real but often brief per pull — Search Engine Journal’s October 14 analysis found roughly 80% of poisoning lasted a single SERP pull, which is why a single day of tracker data looks mostly normal. Second, targeting made no difference in Nozzle’s data: identical keyword pairs showed 7.5% poisoning for national targeting versus 7.4% for local. Do not let anyone sell you on local SERPs being hit harder — the data does not support it.

Why is Google doing this?

Nobody knows for certain, and the honest statement of that is this: Google has not commented publicly on SERP poisoning as of October 27, 2026 — a fact Semrush confirmed on that date, with no Search Central or developers.google.com statement in sight. The poisoning is documented by a large-scale independent measurement and confirmed by multiple rank-tracking vendors, but Google has not explained it.

The most detailed account of the mechanism comes from Derek Perkins of Nozzle, reported by Seroundtable on October 12, 2026: a three-pronged approach — (1) wide IP blocking, (2) more captchas, and (3) injecting bad data into suspect SERPs. The sequence matches: blocking failures on September 13, then injected results from September 20. Nozzle’s own framing — “It is Google’s response to AI scraping, and rank trackers are caught in it” — is its interpretation of the pattern, not a Google confirmation, and nothing in this post should be read as one either.

Before the study published, the warning had already circulated: a practitioner post on X from @rustybrick, a New York City development shop, on October 9, 2026 (Seroundtable’s X coverage that day was replies to it). Nick Swan put it bluntly: rank trackers and SEOs are the “casualties” of anti-AI-scraping measures. The vendor side corroborates the phenomenon — Semrush, Sistrix, AccuRanker, SE Ranking, AWR, Morningscore and Data for SEO have all confirmed seeing poisoned results — and Semrush’s working guidance is detect, exclude, recollect, and compare against Search Console. We cover the measurement side in the next section.

How should I measure rankings again?

Move the center of gravity of your measurement off the tracker. The primary instrument is Google Search Console — positions, impressions and clicks for your own property, which Google serves to your verified site rather than to a third party’s scraping infrastructure. GSC has its own documented counting quirks: the Performance report’s average position uses only the topmost result for each page, and different report dimensions count results differently (query-based versus impression-based), so compare periods within the same view, not across them. The counting documentation spells out the differences. If you are already reading GSC’s generative-AI report, how to read it without miscounting covers where those dashboards disagree.

Around GSC, keep three secondary instruments:

  • Real-browser spot checks. Manually search your five to ten money keywords in a normal logged-out browser, a few times a week, and log what actually shows. This is what a customer sees — it is the ground truth the tracker is failing to approximate.
  • Call tracking and revenue. For any business where the point of ranking is a call or a sale, the phone number that gets dialed and the revenue it produces is the ranking. Position is a proxy for this; during a poisoning event the proxy is broken.
  • Multi-day trends, never single-day numbers. Two effects compound here: the data says roughly 80% of poisoning lasts a single pull, and practitioners report that trackers now sample SERPs instead of counting them in full, so some daily swing is plain measurement noise. Daniel Skov Jacobsen, quoted via Seroundtable on October 9, 2026, put it as “GSC trend first.” Treat any single-day tracker movement as suspect by default.

If a client insists on rank data in the report, keep it — with the caveat printed on it: state that SERPs served to trackers have been altered since late September 2026, cross-check headline movements against real-browser spot checks, and present multi-day trends instead of daily positions. That is exactly the discipline our own reporting language follows: tracker positions are approximate, and GSC plus logged AI citations are the instruments of record.

What does this mean for reporting — and for AI visibility?

It makes explicit what was already true: rank position is a lagging, noisy proxy for the thing that actually matters. Before September 20, the proxy was noisy but honest on average. Now the noise is structured — it arrives as swapped results on a measurable share of pulls, peaking at 43.5% and still at 35.7% as recently as October 21. A client report anchored on tracker positions is anchored on an instrument that Google appears to be degrading, with no documented end date and no official explanation.

The reporting hierarchy that survives this is: GSC positions, impressions and clicks (with their counting caveats) first; real-browser spot checks second; call tracking and revenue third; tracker positions, if included at all, last and footnoted. For local businesses specifically, the honest dashboard is money-keyword positions from your own sources, calls, and revenue from calls — a rank tracker is not on that list.

And the bigger point is the one our whole practice is built on: the customer’s question is increasingly answered before any list of results. A site can sit at #1 in a clean SERP and still be absent from the AI answer — that is the gap the ranking-#1-but-skipped post works through. AI visibility is a separate metric from Google rank, and the poisoning makes the separation unavoidable rather than philosophical. If you have not yet set up the measurement side — the official GA4 and GSC sources plus logged AI-answer citations — measuring AI search traffic in 2026 is the setup guide, and the five numbers worth tracking are in AI visibility metrics: the five numbers that matter.

Where to go from here

Stop anchoring reports on tracker positions, pull your money keywords in a real browser this week, open GSC and look at multi-day impressions before you react to anything, and add the caveat to any client-facing rank table. If you would rather have the whole measurement stack rebuilt the way we run it — GSC as the instrument of record, browser spot checks on the money keywords, call tracking, and AI-answer citations logged alongside — that is exactly what we run end to end, and the full process is in our methodology.

Sources: Nozzle, “Poisoned Google SERPs” (2026-10-10, updated 2026-10-21); Seroundtable (2026-10-12); Search Engine Journal (2026-10-14); Semrush (2026-10-27); Google Search Console documentation. All accessed 2026-10-29. Google has not commented publicly on SERP poisoning as of 2026-10-27.