Short answer
A poisoned SERP is a Google results page that looks like a normal one, but most of the organic results on it have nothing to do with the keyword. There is no captcha and no error, so a rank tracker saves the page, and the report says your site dropped. Since late September 2026 a single rank check can no longer be read as a position.
I still track rankings, with three changes. The data comes from A-Parser with its Bypass poisoned SERP option and residential proxies from the country and city of the search. An empty or cut-off page counts as a failed measurement. And any keyword whose top 10 suddenly shares fewer than half of its domains with the last check is measured again before I believe it. In my test of 12 November 2026 this setup gave 83% good measurements against a real browser, with a mean position error of 2 places inside the browser's top 5. Below is how far off it still is.
What a poisoned SERP is, and what is known about it
Three public sources described the problem between October and November 2026. I read all three on 12 November 2026; the numbers in the table are theirs, not mine.
| Source | What it reports | Limit |
|---|---|---|
| SE Ranking study, published 3 November 2026 | On 10,000 US desktop keywords, the share with unrelated results jumped from 0.09% on 16 September to 57.3% on 17 September. In 96.5% of the cases the unrelated results started at position 9 or lower, never in the top 5 | US and desktop only; they call it "unrelated results" |
| DataForSEO post, published 23 October 2026 | A poisoned page has fewer than half of the usual sites for the keyword. 20 to 30% of the pages it collected were affected from 26 September; after two retries and a history check 2 to 3% remain; noticeably fewer since 10 October. All large providers are affected | The provider's own data, no independent check |
| A-Parser article, 6 November 2026 | A new option, Bypass poisoned SERP, in the settings of its Google parser, with the vendor's test on two US pages through residential proxies in New York | One test by the vendor; how the option works is not described |
Is it deliberate? Nobody knows: neither study claims it, Google has said nothing, and I make no claim either. I can only describe what it did to my numbers.
How it showed up in my own tracking
My setup had two parts: Search Console for sites I can verify, and a rank tracker for everything else. Search Console is the closest thing to ground truth, but it gives an average over all devices and countries for the period, not the position a person sees in one search. And it exists only for verified sites. For PBN projects, where connecting Search Console would leave a footprint, and for competitors, I had only the tracker.
Then its numbers started to jump: on one keyword my site stood third. Over three weeks, from 21 September to 11 October, the tracker found it in 14 of 21 checks and missed it in 7. The misses were not a drop. The page itself had changed: three versions of the top 6, with almost no sites in common.
Before you blame the SERP, rule out the boring cause. If the tracker shows a different URL of your site from one check to the next, that is two of your pages competing for the keyword, and the fix is in the article on keyword cannibalization, not in the data source. In my case the URL was the same; the competitors around it were not.
Three data sources tested against a browser
My reference was my own browser in the same city as the proxies, in the local language: what a real searcher sees. All tests ran on Estonian Google, mobile.
- Bright Data SERP API, 11 October. On that keyword it put my site third in 5 runs out of 5, the same as my browser.
- 30 keywords, two runs each, on Bright Data and DataForSEO, 15 October. Two runs of the same source shared only about half of the top 10: 0.51 on one and 0.52 on the other. Between the two sources the overlap was 0.28. At 5 parallel threads Bright Data lost 68% of the calls; it can probably be tuned, but I found no setting that made it cheaper or steadier for this job.
- Search Console as the judge, 26 October. 31 keywords of a verified site, four rounds 10 minutes apart. Where Search Console had the site in the top 30, DataForSEO found it in 38% of cases, Bright Data in 21% and A-Parser with my first settings in 13%. A-Parser returned only 51% valid measurements then, blocked by captchas, against 97% for DataForSEO. The number one result stayed the same across all four rounds for 68% of keywords in A-Parser and 33% in DataForSEO.
After A-Parser added its option in early November, I took the keyword for which I had written down the real top 30 from my browser on 10 November. With the option, 9 of the browser's top 10 were in A-Parser's top 10, and all 5 of the top 5. DataForSEO had 4 of 10 and none of the top 5. One keyword and one browser is a hint, not proof, so on 12 November I ran a wider comparison: 6 keywords, 8 runs each, 48 measurements per source. A measurement counted as good if it did not fail and at least 7 of its top-10 domains matched the browser.
| Setup, 12 November 2026 | Good measurements of 48 | Top-10 overlap with the browser | Mean position error in the browser's top 5 |
|---|---|---|---|
| A-Parser, proxy pool of the same country | 40 (83%) | 0.68 | 2.0 places |
| A-Parser, pool of IPs from mixed countries | 25 (52%), about 70% once it settled | 0.57 | 2.1 places |
| Hosted SERP APIs, two providers, three configurations, default settings | 15 to 44% | 0.20 to 0.49 | 2.7 to 5.7 places |
The hosted APIs ran on default settings, without deep tuning, and these products may have changed since. DataForSEO itself reports fewer poisoned pages since 10 October. Its post calls the problem industry-wide and tells clients to treat a one-day drop as unconfirmed until two readings agree. I agree; I just needed a measurement I could act on the same day.
The setup that worked
The tool is A-Parser, which I run on my own Enterprise licence, and its SE::Google parser. Five settings made the difference:
- Bypass poisoned SERP (
smartSessionWarmupin the API). Inside the parser it is a session warm-up. It did not change which results came back, but it made runs finish: with the option 3 runs of 3 completed, without it 2 of 3, and one of them returned nothing for 8 keywords. At 5 pages neither run without the option finished. - Residential proxies from the same country and city as your reference. A worldwide pool does not work: Google redirects to the domain of the proxy's country, and the parser reports a mismatch. A pool of mixed countries gave 52% good measurements against 83%, and one site in one keyword moved between places 1 and 4 from one IP to the next. Without proxies, and with datacenter proxies, A-Parser returned 0 results in my tests of 11 October.
- Language, country, Google domain and city set the same as in the browser.
- Device: mobile. All my projects are tracked on mobile, and a mobile page differs from a desktop one.
- Three pages, not five. Past position 30 the parser and my browser stopped agreeing.
What did not help: the full browser engine gave 0 successful queries, an experimental split option gave 0 results, and the single main browser, the duplicate filter and the results-language parameter all stayed inside the noise. Three runs per variant is a tendency, not proof.
The A-Parser recipe in the files below packs this into a ready addTask body and a Python wrapper (3.8 or newer, standard library only). The wrapper sends the task, waits, parses the result, retries failed keywords and prints one status per keyword: FOUND, NOT_FOUND or FAILED. A page counts as failed when it has no organic results, or fewer than 8 per page while Google reports 200 or more results. Its --dry, --parse and --selftest modes need no network; the self-test runs 33 checks against a mock A-Parser. The wrapper itself has not been run against a live A-Parser yet, only the integration in my tracker that it was taken from.
The README lists the traps that cost me hours. \n and \t in the result format must be typed as text, a backslash and a letter: real newline characters made all 6 keywords of a live test come back empty. Result variables need the $p1. prefix. A task without configPreset is rejected. An option given as a [name, value] pair is accepted by the API and silently ignored. And A-Parser tells proxies apart by address:port, so two pools with the same ports merge into one set.
With my proxy pools the warm-up passed for 7 countries (Estonia, Norway, Sweden, Finland, Lithuania, Slovakia, Bulgaria) and failed for 3: the United Kingdom, the United States and Hong Kong. On 12 November, 5 US keywords at 3 pages gave 0 successes and 47 failed warm-ups. The vendor reports a passing US test, so this is most likely a property of one pool, not of the country. Test your pool on 10 keywords before you buy traffic.
How wrong a position still is
Against my browser, on one keyword, with the option and a same-country pool, page 1 shared 70 to 90% of its domains, page 2 50 to 75%, page 3 10 to 44%, and pages 4 and 5 almost nothing. The exact position matched in 7 to 16% of cases. Google does no better against itself: two searches in my browser, one right after another, gave the same exact position 7% of the time.
So I measured noise separately. On 12 November I ran 30 neutral Estonian keywords (weather, recipes, public services, transport) through A-Parser four times, and the same set twice through DataForSEO, plus 30 English keywords twice through DataForSEO for the United States. This measures how a tool agrees with itself, not accuracy: there is no browser reference on these 60 keywords.
| Two runs of the same source | A-Parser, Estonia | DataForSEO, Estonia | DataForSEO, United States |
|---|---|---|---|
| Mean top-10 overlap | 0.76 | 0.60 | 0.48 |
| Pairs with 8 or more of 10 in common | 60% | 39% | 33% |
| Pairs with 4 or fewer of 10 in common | 7% | 26% | 50% |
| Shift of a top-3 site, median / 90th percentile | 0 / 4 | 0 / 12 | 0 / 14 |
| Shift at positions 4 to 10, median / 90th percentile | 1 / 7 | 3 / 10 | 1 / 10 |
| Shift at positions 11 to 20, median / 90th percentile | 2 / 5 | 1 / 6 | 1 / 7 |
Two limits matter here. DataForSEO had two runs per market sent back to back, so its numbers rest on 23 and 30 pairs of runs. And part of its US results came back as Google redirect links without a readable domain; without those rows the US overlap reads 0.59 instead of 0.48. Between A-Parser and DataForSEO on the same Estonian keywords the overlap was 0.37. Which of them looks more like what a person sees, this data does not say.
The nastiest error is the outlier. On 58 keywords of one project on 11 November, 32 shared 8 to 10 of 10 domains with a fresh measurement and 14 shared 4 or fewer. My rough estimate: 10 to 15% of single measurements show the other version of the page. One site appeared second for a keyword; in 22 of the next 24 measurements it was not second, my browser agreed, and its real place was around 20th. One bad measurement in ten hardly shows in a report of a thousand keywords, but it matters for the one keyword you are deciding on.
When a position change is a drop, and when it is noise
The "plus or minus 2" for the top 3 is the mean error rounded, not an interval: the 90th percentile of the run-to-run shift there is 4 places. From these numbers I made a short decision table; the full version is the rank error bars cheat sheet, where every number carries its date, sample size and limit.
| What I see | What I do | Why |
|---|---|---|
| One keyword drops hard on one day, one measurement | Do not report a drop; measure again with the same settings | 10 to 15% of single measurements show the other page |
| The cell is empty or the page is cut off | Treat it as a failed measurement and retry only that keyword | An empty page is not a lost position |
| A move within the band's median: top 3 up to 2, positions 4 to 10 by 1, 11 to 20 by 2 | Ignore it | That is ordinary run-to-run noise |
| A move between the median and the 90th percentile | Wait for a second measurement before writing it down | 90th percentiles: 4, 7 and 5 places |
| A site vanishes from positions 21 to 30 | Not a drop on one measurement | Only 10 to 44% of page-3 domains matched the browser |
| Many keywords flagged on the same day | Suspect the data source first and rerun the next day | Third-party episodes reached 20 to 30% of pages |
The same logic applies when you judge whether SEO work is paying off. A move of two places inside the top 10 is inside the noise, so in the month-by-month view of SEO results I look at the direction over several checks, not at one day.
The re-check rule in my tracker
The tracker I use for client and own projects is my own; the source of positions is switchable per project, DataForSEO or A-Parser, and nothing breaks if I switch back. Two rules came out of these tests.
Rule one: an empty or cut-off page is a failure, not "site not found". On the first large runs, 18% of 613 keywords came back empty on one project, and 206 of 487 failed on another, after one retry. Now the tracker retries failed keywords by itself, up to three rounds, and starts a new round only if the previous one recovered at least 15%. After that, 1 to 3 failures remained per run of about 500 keywords. In the table, cells have three states: found, not found and failed, and failed is red, so a failed measurement cannot be mistaken for a lost position.
Rule two: re-measure the jumps. After a run, the tracker compares each keyword's top 10 with the previous check from the same source. If fewer than half of the domains match, the page probably flipped, and the keyword is measured once more. The second measurement is kept only if it is closer to the previous check than to the first one. If the site really dropped, both new measurements agree with each other, so the new result stays. If the first one was a flipped page, the second looks like the old check and replaces it. Re-measures are capped at 40% of the keywords of a check, so a real Google update cannot blow up the proxy bill; in practice they add about 25% traffic.
Does it work? It is new. On its first live run of 12 November, a check of 9 keywords, 3 looked like jumps, 2 were re-measured (the third was over the cap), and both times the second measurement was taken. That is a smoke test. So I also replayed the rule on the four A-Parser runs of the neutral set: 44 of 624 cases (7%) looked like jumps, the rule took the second measurement in 21 and kept the first in 23. On those cases the overlap with a control run rose from 0.62 to 0.86. The cases overlap, since they come from the same 30 keywords, so read it as a direction.
The repeat-check script applies this rule to your own exports. It takes three CSV files with the columns keyword, position, domain: the last check you trust, the new check and, on the second run, the re-measured keywords. It prints which keywords flipped, which to measure again under the 40% ceiling and which measurement stands. It is one Python file without packages or network access, with a self-test of 22 checks. Two warnings from its README: the last check must come from the same tool, country, city and device, and the rule picks the measurement that is consistent with the last check, which is not the same as the true one.
What it costs
A-Parser does not report proxy traffic, so I read it from the proxy dashboard: 0.19 GB for about 6,050 requests to Google, or about 1.7 MB per keyword at 3 pages. My tracker's own estimate is 1.3 MB. Taking 1.0 to 1.7 MB per keyword, 1,000 keywords cost:
| Proxy price | 1.0 MB per keyword | 1.3 MB | 1.7 MB |
|---|---|---|---|
| $1.70 per GB (small pack) | $1.70 | $2.21 | $2.89 |
| $1.00 per GB (large pack) | $1.00 | $1.30 | $1.70 |
The prices are from my proxy provider's list of 12 November 2026. For scale, a hosted API queue at depth 30 cost about $1.50 per 1,000 keywords when I checked on 15 October and again on 12 November. Retries and re-measures add traffic only for the keywords involved, and the A-Parser licence comes on top; check which edition includes the HTTP API. So a check costs about the same as through an API; what I pay for is numbers closer to my browser on my keywords.
Check your own tracker in 15 minutes
This test needs no A-Parser:
- Pick 10 keywords where you know the real result. Search them in an incognito window with the right language and country, and write down the top 10.
- Run the same keywords through your tool three times.
- For each keyword, count how many top-10 domains match between runs. Fewer than half means a flipped page: do not trust that day.
- Compare the average with your browser. If page 1 shares 70% or more, the tool is good enough for trends.
To compare your numbers with mine, the repeatability dataset holds the raw measurements of 12 November: 4,642 result rows, 60 keywords, 237 measurements, one row per organic result with source, market, run, keyword, position and domain. The analyze.py script reproduces every number in the noise table above and runs on your own CSV with the same seven columns; use at least 2 runs, better 4. The data is licensed CC BY 4.0, the script MIT. All four files from the video are below this article.
Who this is not for, and what is not solved
This approach does not fit if you need positions deeper than 30, if you cannot get residential proxies for your country, or if you track thousands of keywords with nobody to watch the share of failed measurements: the proxy pool is the weak point. And if your current tracker passes the 15-minute test, do not switch.
What I have not solved:
- Coverage. With some proxy pools the warm-up fails on a captcha and almost every query comes back empty.
- Depth. Thirty positions, no more.
- The reference. One browser of one person, a handful of keywords, everything dated 12 November 2026. Google itself reorders the first three results between two searches.
- Languages and markets. I tested local-language Estonian keywords on mobile. Other languages, countries and desktop pages were not tested for accuracy.
- Terms of service. Automated queries are against Google's terms. I use this to monitor my own and client sites; how much risk you accept is your call.
- Permanence. DataForSEO says it tests new approaches every day. The fix may land in your own tool next month, and this method may become unnecessary.
One last check before you blame any tool: if a page disappears from every measurement and from your browser as well, the problem is not the SERP. Start with why Google is not indexing your pages. If you want the data source and the reports checked on your own project, I do that as part of a technical SEO audit.
