Your rank tracker may be reading a decoy page
Your best page fell out of the top ten on Monday and was back by Thursday. A new study says Google has been feeding scrapers believable fake results since September, so check before anyone rewrites a thing.
Probably not a real drop, if it reversed within a few days and Search Console stayed steady. Nozzle's October 10, 2026 study found Google serving rank trackers normal-looking pages with swapped results on about 8% of checks since mid-September. Cross-check with Search Console, wait three days, and make any tool that reads Google compare each pull with its history before acting.
The drop that wasn't
Picture a Monday rank report. Your best category page fell out of the top ten overnight, two product pages went with it, and someone on the team is already drafting new copy. By Thursday the pages are back where they were, and nobody can say what changed.
If that happened to you any time since late September, there is a decent chance nothing happened to your site at all. On October 10, 2026, the rank-tracking company Nozzle published a study of 27.5 million Google results pages collected for its customers across 49 countries, from August 20 to October 9. Its finding: since September, Google has been answering a share of automated searches with complete, normal-looking results pages whose results have been swapped out. Search Engine Roundtable flagged rank tracking breaking on October 12, and Search Engine Journal wrote up the study this morning.
Google has not confirmed any of this. What follows is Nozzle's measurement, not a Google statement. It is still enough to change how you should read a ranking chart this month.
What the study actually found
The sequence matters. On September 13, Nozzle's requests to Google started failing in large numbers. From September 20, many of the ones that got through came back with the wrong results. Nozzle calls these pages poisoned and reads them as Google's response to bulk scraping for AI answers, with rank trackers caught in the same net because they collect results the same way.
The numbers, as Nozzle reports them:
- 8.0% of pages since September 13 came back poisoned, against 0.17% before. The worst stretch was September 27 to 29, at about a quarter of all pages, and it has run at roughly 5 to 11% a day since October 1.
- A poisoned page keeps only 20.7% of the sites from the keyword's previous page, against 88.5% normally. Reddit, Facebook, LinkedIn and Quora fill the gaps.
- When a tracked site is normally on page one, it fell out of the top ten on 58.6% of poisoned pages, against 2.7% in the baseline.
- Most spells last one check. 77.9% of poisoned pages had the usual sites back within three days.
- Ads and shopping results barely moved.
The part worth sitting with is Nozzle's own line: a refused request tells a scraper to try again, while a poisoned one gets stored as real. Nothing on the page says it's wrong. No error, no captcha, the same layout.
Decision guide
Decoy page or real drop?
| Consideration | Looks like a decoy page | Looks like a real change |
|---|---|---|
| Search Console | Impressions and clicks steady on the day the tracker says you vanished. | Impressions and clicks fall on the same days as the tracker. |
| Duration | Back within one to three checks. | Still down after several days and a week of checks. |
| Who replaced you | Most of the keyword's usual sites gone at once, social posts and forums in their place. | One or two competitors moved up; the rest of the page looks familiar. |
| Your side | Nothing shipped: no migration, redirects, theme or feed change. | Something changed on the site shortly before the drop. |
A failed request at least shows up as a failure. A decoy shows up as a result.
Triage a ranking drop before you touch anything
So when the report says you fell, run four quick checks before anyone rewrites a page or calls an agency.
Does Search Console agree? Search Console reports what real searchers saw. If impressions and clicks for that page held steady on the day your tracker says it vanished, the tracker read a bad page. Nozzle says the same thing: if Search Console still shows impressions and clicks, nothing happened to the site.
Did it last? Give it three days. A real ranking change stays. Four in five poisoned spells in Nozzle's data were a single check.
Does the page look like that keyword's usual neighborhood? Compare against the sites that normally rank over the past few weeks, not just yesterday, because yesterday might have been a decoy too. A results page suddenly full of social posts and forums, or one that stops at page three or four, is a hint. None of those signs is reliable on its own.
Did anything ship on your side? A theme change, a migration, new redirects, a robots edit, a product feed problem. If yes, investigate that first, because a real drop after a real change is still the most expensive thing to ignore.
If Search Console is steady, the drop reversed, and nothing shipped, write it down and move on.
How the work moves
Before you react to a ranking alert
- Check Search Console
Same page, same dates. Steady impressions and clicks mean the tracker probably read a bad page.
- Wait three days
Most poisoned spells in Nozzle's data were a single check and reversed within three days.
- Compare with history
Look at the keyword's usual page-one sites over recent weeks, not just yesterday.
- Check your own changes
Migrations, redirects, robots edits, theme or feed changes come first if anything shipped.
- Then decide
Only a drop that lasts and shows up in first-party data earns a rewrite or a rollback.
If your own software reads Google, it can be confidently wrong too
This goes beyond SEO reports. Plenty of small businesses have quietly built something that reads Google results on a schedule: a competitor price check, a weekly visibility report, a lead list scraped from local searches, an AI assistant that looks things up on the open web before it answers a customer.
All of those share the weakness Nozzle describes. The fetch succeeds. The page parses. The data looks clean. And it's wrong, so it gets saved, charted, summarized, and maybe acted on. A failed request at least shows up as a failure. A decoy shows up as a result.
If you own a tool like that, a few changes make it far less gullible:
- Compare each pull with its own history. If most of a keyword's usual sites disappear at once, flag the row as suspect instead of storing it as fact.
- Cross-check against first-party data. Search Console and your own analytics come from real visits. Use them as the referee.
- Prefer sanctioned sources. Official APIs and licensed data feeds cost more than a scraper. They also don't hand you a quiet decoy.
- Let uncertain data wait. An agent or automation should not change prices, send an alert to a client, or rewrite a listing on a single pull that disagrees with everything else.
It's the same habit we argued for in treating tool returns as data until you prove they are safe. A web page an agent fetched is evidence, not truth.
What not to do this week
Don't rewrite your category pages because of one red row. Don't roll back a launch that Search Console says is doing fine. Don't fire anyone, and don't buy a second rank tracker to settle the argument, since it collects pages the same way and may be getting the same decoys.
Do go back and annotate late September in whatever reports you keep. Nozzle's worst days were September 27 to 29, and any month-over-month comparison that leans on them will look worse than reality. If a ranking number feeds a client report or a bonus, say so in the report rather than letting it stand.
And keep a definition next to the chart. A ranking position from a scraper and a click count from Search Console are different measurements of different things, which is the point of defining the metric before you dashboard it.
Put this into practice
Take the last ranking alert that made your stomach drop. Open Search Console for the same page and the same dates. If the two disagree, you just learned how much to trust that alert going forward.
Then list every place your business reads Google automatically: reports, price checks, lead lists, assistants. For each one, write down what happens when it gets a believable wrong page. If the honest answer is "we'd act on it," add a history check and a first-party cross-check before the next run. That is a small, bounded piece of business process automation work, and if the reader is an AI assistant, it belongs in how you build AI applications and agents. If you want a second pair of eyes on one of those pipelines, tell us what it reads and what it does next.
Related work: Spin:Market.
We build alongside the people doing the work.
Bring us the problem you’re thinking about.