Poisoned SERPs: your rank tracker was never the search results
This week Semrush published a statement about what the industry has taken to calling poisoned SERPs. Google, it says, has changed how it responds to some automated search requests, so that results pages can look normal but differ from what a typical user would see. It calls this an industry-wide issue and says it detects clearly invalid results, excludes them from its database and flags doubtful ones for re-collection. It doesn't say how many results are affected, and it notes that Google hasn't commented. Ahrefs said much the same a few days earlier.
The reaction has been alarm. Mine was a shrug. Rank tracking has been a poor witness for years, and it's worth being clear why before deciding how worried to be.
What's being reported
Three accounts are worth knowing about, and none of them has been independently audited.
Nozzle reported that 8.0% of the results pages it collected since 13 September were poisoned, against 0.17% before that date. It counted a page as poisoned when fewer than half of a keyword's usual top-ten sites turned up and then came back on a later collection. Ahrefs said that between 5% and 20% of results served to SEO data providers have been altered, depending on the provider, and that it's filtering for abnormal results and cleaning stored data. Semrush's example was a search in Brazil for a specific baby stroller that returned a dental supply store's equipment page, because both pages use the Portuguese word "carrinho". It describes two forms: a page that is plainly off from the top, and a page that looks fine until you scroll and find loosely related results further down.
The vendors measure it differently. On the day Nozzle measured around 27%, another rank tracker reported 1.4%, using a different baseline. And Google hasn't commented. The suggestion that the target is large-scale data collection for AI systems is industry speculation, not a statement.
What was already wrong
A rank tracker has never reported the search results. It reports one set of them, from one location, on one device, as seen by one automated request that doesn't behave like a person. Real results shift with who's asking, where they are and what they searched yesterday, so for plenty of queries there is no single "position three" to report. A tracker reduces all that to one tidy number, which is why it looks authoritative.
What's different
It's quiet, and it looks plausible. A broken tool announces itself. A page of results that is wrong in the lower half, or on page one for one keyword in twenty, doesn't. You'd only notice if you already knew what the answer should look like. Most people reading a rankings report don't, and the report presents the drop with the same confidence as everything else.
What to lean on first
Data your own site collected. Search Console shows the impressions and clicks Google actually recorded for your pages, and GA4 shows what the people who arrived then did. Neither is perfect (average position is an average, and some queries are withheld), but neither is built from requests Google can choose to answer differently.
Semrush's own advice to customers who spot unexpected ranking changes, especially across unrelated keywords, is to compare them with their own data, such as Search Console. That leads to a few plain checks before anyone panics about a drop. Did the same keywords fall in the tool and in Search Console? Did unrelated keywords all move on the same day? Did the tool post a notice about its own data? If a position has fallen twenty places overnight but impressions and clicks for that query haven't moved, suspect the data before the site.
Nobody's number is the truth
It doesn't much matter whether a figure comes from Google, Semrush, Ahrefs or a tool you built yourself. Each is a view put together by someone with their own method and their own reasons. Google's is no exception. Search Console reports through Google's definitions: positions are averages, some queries are withheld, and it decides what counts as an impression. It's a better witness than a scraper, but it's still a witness. The poisoning story is just the first time the disagreement between the witnesses has been loud enough to notice.
The one number nobody else gets to define for you is what the business actually received: enquiries, orders, revenue. Everything else is a clue about why.
Rankings are the heron, not the river. They're the thing that stands out and gets reported, while the water underneath, visibility turning into visits, visits into enquiries and enquiries into sales, is what actually moves anything. A poisoned tracker just means the heron has been repainted.
The same trap, with AI answers
It's worth keeping in mind as everyone rushes to put a value on being cited in AI Overviews and AI answers. Here's what can and can't be counted today. Some platforms report citations, meaning how often the site is cited in their AI answers, but a citation is visibility, not a visit. Google's AI Overviews and AI Mode clicks are folded into ordinary web search in Search Console, so they can't be separated out. Visits and sales that arrive from AI assistants can be counted in GA4, which is the one measurable value, and on the ecommerce site I've been tracking it's small and patchy from month to month.
So any figure for what a citation is worth is a guess with nothing to check it against. It also isn't the same thing as the site having recovered. On that site the AI citations came back faster after a migration than the Google clicks did, which tells you something about visibility and almost nothing about revenue. I'd report both, keep them separate, and be wary of anyone who sells the two as one number.
Use the trackers as direction
They're still useful for competitor movement, rough visibility and keeping an eye on a keyword set. They just shouldn't be the number that decides anything. If a report is built on one of them alone, ask what it would say if it were wrong, and whether anyone would notice.
Which is where I'd start with any site. I'd rather know how a page is doing from the people who actually landed on it, and what they did next, than from a robot that Google may have decided to mislead.
Sources: PPC Land on Nozzle's findings; PPC Land on Ahrefs' statement; Semrush, "Poisoned SERPs: An update on Google ranking data in Semrush", 27 October 2026.