If a client site dropped out of the top 10 in late September and came straight back, the likeliest explanation is not the spam update everyone blamed. It is that Google handed your rank tracker a page that was quietly wrong.
On 10 October, Nozzle published Poisoned Google SERPs, an analysis of 27.5 million results pages collected between 20 August and 9 October across 49 countries. The finding: since 20 September, Google has been answering a share of automated collection requests with complete, normal-looking pages whose results have been swapped out. No error, no captcha, nothing in the HTML to mark it, just the wrong results, stored by every tool that received them as though they were real.
What the data shows
Since 13 September, 8.0% of collected results pages came back poisoned, against 0.17% in the three weeks before. The sequence matters: requests started failing in volume on 13 September, and from 20 September the ones that got through came back wrong. The peak ran 27 to 29 September, at roughly a quarter of collected pages.
Inside a poisoned page, the organic results retain 20.7% of the sites on that keyword's previous page, against 88.5% normally. Social platforms and forums fill the space: Reddit's top-10 presence rises from 20.8% to 28.8%, Facebook from 28.3% to 33.8%. On desktop keywords that normally run to page 8 or beyond, half of poisoned pages stop paginating at page 4 or earlier.
The detail that gives the game away is which parts survive. AI Overview sources keep 42.3% of their previous occupants, but shopping keeps 92.2% and ads keep 83.7%, more than the 70.6% they hold normally. The surfaces advertisers pay for stay intact while the organic ones are replaced.
The consequence is blunt: where a tracked site was normally a page-1 fixture, it was missing from the top 10 on 58.6% of poisoned pages, against 2.7% in the baseline. The share of pages poisoned ranged from 18.8% in Austria to 3.0% in the UK, with legal the worst industry at 15.9%.
Not one vendor's bad week
A rank-tracking company reporting that rank tracking is hard deserves a second source. There is one.
Morningscore published a post-mortem across more than 4 million searches. From 22:10 UTC on 24 September, searches began returning truncated pages: fewer than 40 results where about 50 was normal, with 7.9 of the previous top 10 missing on average. At the 30 September peak, around 60% of their searches came back that way.
Two details there carry more weight than the percentages. They ran the same searches in an ordinary browser on consumer connections at the same time and got full results, so real searchers almost certainly never saw these pages. And of their five data providers, two never saw the problem at scale; when it returned on 2 October, one provider got truncated results for up to 72% of searches while another got 1 to 3% for the same keywords.
Nozzle confirms that independently: one collection route returned 1 to 7% poisoned pages while another returned 9 to 30% on the same days. The variable is not the keyword or the market, it is the infrastructure the request travelled through. That is why two tools could report contradictory positions for the same site and both be faithfully reporting what they received.
Nozzle's Derek Perkins described a three-part approach to Barry Schwartz: wide IP blocking aimed at residential proxy networks, a sharp rise in captchas, and bad data injected into results served to suspect requests. The third is the one that matters, because the first two announce themselves. A blocked request tells a tool to retry; a poisoned one gets written to a database and charted.
Google has not commented. The reasonable reading, which both vendors state as inference rather than fact, is that none of this targets SEO tools: it targets bulk collection of results to ground AI answers, and automated querying generally. Rank trackers use the same pipes, so they catch the same treatment.
Why the spam update hid it
Google's September 2026 spam update started on 24 September and was confirmed complete on 8 October after nearly 14 days, the longest spam rollout of the year. The volatility reported during it clustered around 25 to 27 September and 4 to 6 October.
Compare that with the poisoning curve: 11.1% on 25 September, rising through 24.9% on the 27th to 27.0% on the 29th, then 8.6%, 10.5% and 9.5% on 4, 5 and 6 October. The reported volatility dates and the contamination dates are the same dates.
The spam update was real and moved real rankings; Glenn Gabe documented drops he verified against a site owner's traffic. But volatility indices are themselves built by scraping results at scale, so a measurable share of the movement attributed to the update was movement in the collection layer, not the index. An agency that spent early October explaining a client's spam-update hit may have been explaining a 48-hour data artifact.
What to do if you publish at scale
Programmatic and AI-assisted page libraries are exposed in a specific way, because poisoned pages demote exactly this kind of result: commercial, content-dense pages vanish and forums replace them. New batches are also judged in their first few weeks, precisely the window where a one-day drop means nothing. Four in five poisoning spells last a single collection, and 77.9% recover within three days.
Read Search Console first, tracker second. Impressions and clicks come from real searches and were never affected. If a tracked position collapsed while impressions held steady, nothing happened to the page. Advanced Web Ranking's methodology write-up adds the right caveat: use Search Console to confirm direction, never to adjudicate a position on a given day, because average position blends every query, device, location and AI Overview citation into one number.
Compare against the keyword's own history, not yesterday, which may have been poisoned too.
Treat simultaneous drops as a data event. Many keywords falling together and recovering at the next check is a collection problem; a real loss holds and shows up in clicks.
Do not rewrite or deindex anything on tracker data from 20 September to 9 October. That window is contaminated unevenly, depending on your tool's provider. Re-measure first.
Ask your vendor which routes they collect through and whether they discard suspect pages. Morningscore now rejects truncated results and refetches; Advanced Web Ranking retries pages it judges structurally thin before storing a position. That is the difference between a tool that missed a week and one that recorded a week of fiction.
The broader point is one the industry has been slow to accept: there is no longer a single Google results page to rank on. There is the page a person is shown and the page an automated request receives, and since September those two have diverged far enough that one is actively unreliable.
FAQ
Did my rankings actually drop in late September 2026?
Possibly not. If the drop appeared across many keywords on the same day and recovered within one to three days, the likeliest cause is a poisoned or truncated results page served to your tracking tool. Check Search Console impressions and clicks for those dates: if they held steady, your visibility to real searchers did not move. A genuine loss persists across several checks and shows up in click data.
Is Google deliberately targeting SEO tools?
Google has not commented, so nobody outside the company can answer definitively. Both vendors who published data read it as collateral damage from enforcement aimed at bulk collection to ground AI answers, not an attack on rank tracking. The effect is the same either way, because rank trackers collect through the same automated methods that enforcement is built to catch.
How do I tell a poisoned results page from a real ranking change?
Five signals, strongest first. Search Console still shows impressions and clicks. The drop recovers within one to three days. Most of the keyword's usual page-1 sites are missing at once, not just yours. Social platforms and forums are unusually prominent. Pagination stops early on a keyword that normally runs deep. None is conclusive alone, and the page carries no error or captcha to give it away.
Should I stop using rank trackers?
No, but stop reading them as single authoritative positions. They remain the only source of competitor context, share of voice and results-feature occupancy, none of which Search Console reports. Use them for trends and ranges, and cross-reference anything that would trigger a decision.
Does this affect AI Overview and AI Mode visibility tracking?
It affects the collected view of it. AI Overviews appeared on 34.9% of poisoned pages against 54.2% normally, and the sources cited in them retained only 42.3% of their previous occupants against 68.8%. Any tool measuring AI Overview presence by collecting results pages at scale was reading a thinner page than real searchers saw, so treat September and early October AI visibility trends with the same caution as organic positions.