Poisoned SERPs: Why Your Rank Tracker's AI Overview Drop May Be Fake

If your rank tracker showed a sudden drop in late September or early October 2026, the drop may not be real. Since September 2026, some third-party rank trackers have been getting Google result pages that look normal but list the wrong websites. Nozzle, an enterprise rank-tracking company, calls these "poisoned" SERPs (SERP means search engine results page).¹ On those poisoned pages, AI Overviews show up less often than usual. So if you report AI Overview presence to a client or a boss, part of the "loss" in your tracker may be fake.

What happened to rank tracking data since September 2026

A rank tracker is a tool that searches Google for your keywords automatically, saves the results page, and records where your site appears.

Nozzle published its research on October 10, 2026, using data through October 9.¹ Its main finding is simple to state. When Google suspects a request is automated, it does not always block it. Sometimes it answers with a full, normal-looking page that has the wrong results on it. In Nozzle's words, "Google is not refusing these requests; it answers them with a complete page whose results are wrong, and nothing on the page says so."¹

A poisoned page looks like any other page. Nozzle describes it as "No error, no captcha, the same template, the same build of Google's frontend as a normal page."¹ (A captcha is the "prove you're human" test some sites show.)

Search Engine Journal covered the research on October 14, 2026.² Search Engine Roundtable reported on October 12 that several rank-tracking vendors had posted notices about Google data problems.³ Neither article includes an official Google statement, and Google has not publicly confirmed that it serves wrong results to scrapers.

A short timeline

One note on the peak. Search Engine Journal described poisoned results reaching "as high as 25%."² Nozzle's own daily figures for September 27 to 29 go slightly higher, up to 27.0%.¹ Both describe the same thing: for a few days, about a quarter of the pages Nozzle collected were wrong.

What Nozzle actually measured

Every number below depends on how Nozzle decided a page was "poisoned."

• Sample. 63.1 million SERPs from August 20 to October 9, 2026, across many countries, devices, industries and locations. Pages with no organic results were left out.¹
• "Usual sites." For each keyword, Nozzle listed the sites that appeared in its top 10 on at least 60% of searches between August 20 and September 9. This was the baseline period.¹
• The rule. "A SERP is poisoned when fewer than half of its keyword's usual sites appear anywhere on it," and those usual sites later come back.¹

That last part matters. A page only counts as poisoned once the normal sites return. So a real, lasting ranking change is not counted as poisoning. It also means the most recent days in any dataset are still "filling in," which Nozzle says itself.¹

The headline numbers

All figures from Nozzle's research, published October 10, 2026.¹

• Overlap. If you search the same keyword twice on a normal day, about 88.5% of the sites stay the same. On a poisoned page, only 20.7% of the previous page's sites are still there. So a poisoned page looks like the results were shuffled and swapped out.
• AI Overviews. On normal pages, an AI Overview appeared 54.2% of the time. On poisoned pages, it appeared 34.9% of the time. That is a gap of 19.3 percentage points.
• Your own site. For keywords where a tracked site was one of the usual sites, that site was missing from the top 10 on 58.6% of poisoned pages. On normal pages, that happened only 2.7% of the time.

How long poisoning lasts

This is the most useful finding for anyone writing a report. Poisoning is usually short.

• 80% of poisoning spells last a single SERP. 15% last two, and 5% last three or more.¹
• The next search has the usual sites back 55.9% of the time.¹
• 77.9% of poisoned SERPs have their usual sites back within three days. The median is two days.¹

So a single bad reading is very likely to go away. A drop that stays for several days and several checks is much more likely to be real.

Which sites fill the gap

When the usual sites drop out, something takes their place. Nozzle found that social and forum sites (often called UGC, or user-generated content, sites) gained share on poisoned pages.¹

Device barely mattered (7.7% on desktop, 8.3% on iPhone, 7.1% on Android), but the route a request took did: from September 30 to October 9, one route returned 1% to 7% poisoned pages and another 9% to 30%.¹

What other rank-tracking vendors reported

Nozzle is not the only company that noticed something wrong. Several vendors described related problems in their own words.

• Sistrix wrote on October 7, 2026 that since September 2026 it has seen Google return divergent results with different websites for some queries. It switched off on-demand updates and now re-collects a keyword until it gets a result that passes its quality checks. It says its top 10 rankings and Visibility Index remain reliable.⁴
• Morningscore wrote on October 8, 2026 (updated October 14) that from September 24 its searches started coming back cut short. At the peak, about 60% of its searches returned fewer than 40 results instead of about 50. Its tracker recorded the missing sites as ranking drops, and it says "most of the drops this caused weren't real."⁵ It is now testing a check that confirms a sudden disappearance with a second search.⁵
• SE Ranking, DataForSEO and Advanced Web Ranking also posted notices or articles, according to Search Engine Roundtable.³

Advanced Web Ranking had already argued in September, before the poisoning wave, that tracked rankings and what real users see were drifting apart. It recommended reporting "positions as ranges and trends, not points" and comparing tracked positions with Search Console trends.⁶

What is measured and what is interpretation

It is easy to read this story as "Google is attacking SEO tools." The evidence supports something narrower. Here is how the claims split.

A few caveats worth keeping in mind:

• This is observational data from one company's systems. Nozzle's numbers describe what its own requests got back. Other trackers use other routes and may see more or less poisoning.
• The lower AI Overview rate on poisoned pages is a correlation. It tells you that poisoned pages tend to show fewer AI Overviews. It does not tell you why.
• Google's rules do not allow scraping for rank checking. Google's spam policy says sending automated queries, including "scraping results for rank-checking purposes," without express permission violates its spam policies and Terms of Service.⁷ That does not prove what Google is doing now. It does explain why tool makers expect Google to keep pushing back.

Why AI Overview reports are most at risk

A lot of teams now report two things from rank trackers: where a site ranks, and whether it appears in Google's AI Overview for a keyword. Both depend on one saved page per keyword per day. Poisoning hits both numbers, but the AI Overview number is hit twice.

1. The AI Overview may not appear at all. On Nozzle's poisoned pages, the AI Overview showed up 34.9% of the time instead of 54.2%.¹ So a tracker may record "no AI Overview" on a keyword that normally has one.
2. If it does appear, the sources may change. Poisoned pages kept 42.3% of the previous AI Overview sources, compared with 68.8% normally.¹ So even when the AI Overview is there, a site that is usually cited may be missing.

In other words, a single scraped snapshot can make a brand look like it lost an AI Overview citation when nothing changed for real users. If that snapshot goes straight into a client report, the report is wrong.

It also cuts the other way. If your tracker shows AI Overviews on fewer of your keywords this month, it may just have hit more poisoned pages.

Before you report a drop: a five-step check

This is a simple procedure for anyone who reports rankings or AI Overview presence. It uses Nozzle's findings on how long poisoning lasts and what it looks like. You don't need special tools for most of it.

1. Check more than once. Run the keyword again, ideally two or three times on the same day. 80% of poisoning lasts one search, and the next search has the usual sites back 55.9% of the time.¹ If the citation or ranking comes back on a re-run, the first reading was probably bad data.
2. Wait for it to persist. Only treat a drop as real if it shows up across at least two days in a row. 77.9% of poisoned SERPs recover within three days.¹ For a high-stakes report, three days is safer.
3. Look for the poisoned signature. Compare the page with the keyword's usual top 10 over the past few weeks, not just yesterday. Nozzle warns that "Comparing against yesterday alone is fragile, because yesterday may have been poisoned too."¹ Warning signs:
   • Fewer than half of the usual top 10 sites are on the page.
   • Facebook, Reddit, Instagram, LinkedIn or Quora suddenly fill several spots.
   • The AI Overview vanished on a keyword that almost always has one.
   • The page has fewer results than usual (the cut-short pattern Morningscore saw).⁵
4. Compare with Search Console. Google Search Console reports what real users saw. An impression means a user "has seen (or potentially seen) a link to your site" in Search.⁸ If your tracker says a page vanished but Search Console still shows steady impressions and clicks for that query, the drop is very likely in the tracker. Nozzle says it plainly: "If a site "disappears" and Search Console still shows impressions and clicks, nothing happened to the site."¹
5. Label what you can't confirm. If a drop is recent and you can't confirm it yet, say so in the report. "Pending: one day of data" is more honest than a red arrow.

A note on Search Console's limits. Search Console's average position is not the same thing a rank tracker measures. It is the topmost position your link held, averaged across searches, and a link must get an impression for its position to be recorded.⁸ An AI Overview counts as a single position, and all links inside it share that position.⁸ So Search Console is good for checking direction (are impressions and clicks steady or falling?). It is not good for settling an exact position on a single day, which Advanced Web Ranking also points out.⁶

Worked example: one week of tracker readings

This is an illustrative composite, not a real client or study. The brand and numbers are invented to show how the check works. The pattern is built to match Nozzle's published findings.

Harbor Lane Kitchenware (invented) tracks 50 keywords. In a normal week, its own pages are cited in Google's AI Overview on 12 of those keywords. The tracker saves one snapshot per keyword per day.

During the week, two things happen at once:

• Some snapshots are poisoned. On those, the AI Overview is missing or Harbor Lane is not cited.
• From Thursday, Harbor Lane really does lose its citation on one keyword. This is a real loss.

The team uses a multi-run rule:

• Each cited keyword is checked three times a day.
• A citation counts as present if it shows up in at least two of the three runs.
• A loss is only reported once it holds for two days in a row. Until then, it is marked "on watch."

Here is how each row is built. The single-snapshot number is the 12 normal citations, minus any poisoned snapshots, minus any real losses. For Thursday, that is:

12 − 1 poisoned − 1 real = 10

Now compare what each method would tell the client.

What a single-snapshot report says

• Wednesday looks like a crisis. Citations fell from 12 to 9. That is a drop of 3 ÷ 12 = 25%.
• Saturday shows the same 25% drop again.
• The weekly average is (12 + 10 + 9 + 10 + 11 + 9 + 11) ÷ 7 = 72 ÷ 7 ≈ 10.3 cited keywords.

What the multi-run rule says

• Monday to Wednesday hold steady at 12. The poisoned snapshots fail the re-run test, because the citation comes back on the other runs.
• Thursday flags one keyword "on watch." Friday confirms it. Citations fall from 12 to 11. That is a drop of 1 ÷ 12 ≈ 8.3%.
• The weekly average is (12 + 12 + 12 + 12 + 11 + 11 + 11) ÷ 7 = 81 ÷ 7 ≈ 11.6 cited keywords.

So the single-snapshot report shows a 25% drop on two different days. The multi-run report shows one real loss, an 8.3% drop, confirmed on Friday. The real loss is still caught. It just isn't buried under fake ones.

The cost is more checks: 36 a day instead of 12. You can cut that by only re-running keywords that look like they dropped. On Wednesday, that means just the 3 missing ones.

The same team can then check Search Console. If impressions and clicks for the Wednesday "drops" stayed steady, that backs up the multi-run result. For the Thursday keyword, a real fall in clicks would support the real loss. (Search Console won't tell you directly whether the AI Overview cited you, since links inside it share one position.⁸ But it does show whether real users still saw and clicked your page.)

What to tell clients and stakeholders now

If you have already sent reports covering late September or early October 2026, it's worth a second look. Morningscore suggests checking rankings from September 24 to October 2 "before you act."⁵ Nozzle's data shows the worst days fell on September 27 to 29.¹

Some practical points for the next report:

• Flag the data window. Tell readers that Google rank data from late September 2026 was unreliable across several tools. Name the dates.
• Report trends and ranges. A single day's AI Overview count is fragile right now. A rolling multi-day number is sturdier.
• Ask your tool vendor what they changed. Some now re-collect results that fail checks, and some are testing second-search confirmation.⁴⁵ Knowing whether your tracker does this tells you how much to trust a single day.
• Pair tracker data with Search Console. Use it to check direction, not exact positions.⁶⁸

Check your AI visibility with more than one snapshot

A single scraped page can make a brand look like it lost AI Overview citations overnight. Ripplix's free AI Visibility Report shows where your brand is named and cited across AI answers, so you can see the wider pattern before you act on one bad reading.

Get your free AI Visibility Report →