Method

How we research robot vacuums

Q: Why don't you average the lab scores?

Because the labs aren't measuring the same thing. RTINGS judges by eye how much spilled debris a robot leaves behind, and it presses synthetic pet hair into low carpet. Vacuum Wars weighs the sand a robot digs out of medium carpet at full power, and it tests thresholds one added strip at a time on a single step. Consumer Reports uses human hair for its robot hair test. TechGearLab uses real pet hair and weighs what comes out of the carpet. PCMag deliberately runs robots in their default mode. The Hook Up tests on deep, high-pile carpet. Just A Dad Approved doesn't measure at all and scores the hassle of months of use at home instead. A robot can come out on top in one of these tests and trail in another, and both results can be honest. Averaging them would produce a number nobody ever measured. So each figure stays attached to the test behind it, and the page tells you which test looks most like your floors. When two sources disagree, you see both. Most testers simply repeat the pascal figure printed on the box. Vacuum Wars is the one tester that measures suction at the brush, where the robot actually meets the floor. Across the twelve robots on file with both a claim and a measurement, a bigger number on the box did not mean more suction at the floor. Pascals have become the horsepower of the robot aisle: printed large and checked rarely. We take what a lab measured and ignore where it placed a robot, because some leaderboards blend in shopping data that has nothing to do with how well a robot cleans. The gaps get reported too. No tester measures more than one unit of a model, so nobody knows how much one robot varies from the next off the same line. No lab measures reliability or customer support model by model either. When a page warns about a dock that quits after a year, that warning comes from owners, and the page says so.

Swept doesn't test robot vacuums. It reads every maker, lab, reviewer and owner on each exact US model, keeps every figure with the test that produced it, and tells you plainly what that adds up to.

By David King, EditorUpdated November 5, 2026

Nobody at Swept has ever run a robot vacuum across a rug, and we would rather tell you that in the first line than bury it in the footer. What we do instead is read everything. That means the maker's box and US manual, every independent lab that published a test, the reviewers who lived with a robot for months, and the owners who wrote about the morning it wedged itself under the couch for the fifth time. We set the promise against what was measured, one exact US model at a time, and then we tell you what we would buy.

Who writes it, and why a jet engine builder

David King edits Swept and answers for every page on it. Before he edited anything, he spent his career on factory floors. His last post was Production Leader at Rolls-Royce, where he oversaw the assembly of jet engine sections bound for Boeing and Airbus airliners. Before Rolls-Royce, WEC Group employed him as a Production Manager. His full working history is on his LinkedIn profile (opens in a new tab), and there is more about him on our About page.

He is not a robot vacuum expert, and he doesn't pretend to be. The engine line gave him a habit instead: ask what the tolerance is, how it was measured, whether you can trace the number back to its source, and what the root cause is when the part fails. On a production floor, a part is only as good as the measurement that passed it. I read a robot vacuum box the same way, and it's remarkable how many big numbers arrive with no gauge behind them.

What we never do

Do you test the robots yourselves?

No, and it's better that you hear it from us. We have no testing facility of our own, we don't take robots from makers, and placement on a page isn't for sale. Every experience on our pages belongs to whoever had it. That could be a lab that measured something, a reviewer who watched it happen, a manual that states it, or owners who keep reporting the same fault. When a page says a robot gets over the step into the kitchen, it also tells you whose test showed it. Our opinions are our own. The evidence always belongs to someone else, and we say whose it is.

How the research is done

The research is AI-assisted. AI models read makers' product pages and US manuals, independent labs' published tests, reviewers' write-ups and videos, and owners' reports on Reddit and YouTube. They work like a careful research assistant. They follow each lead back to where it started, rather than trusting a blog that quotes a blog that quotes a press release. We keep the full text of every web source we rely on, so a figure can be checked against the words it came from long after the original page has changed. Reddit threads read before November 2, 2026 are kept as summaries, and we count them separately because they aren't kept in full.

As of November 4, 2026, the file holds 27 exact models and 1,273 individual facts. Of those, 579 are measurements, 161 are things someone saw happen, and 533 are makers' own claims. Behind those facts sit 497 sources whose complete text we have saved, along with 148 Reddit threads stored only as summaries.

One exact model at a time

Every fact is tied to the exact model it describes. It also records where it came from, when we read it, and whether it is a maker's claim, a measurement or something someone saw. This matters more than it sounds, because robot names are slippery. Makers reuse an old name on a new generation, and they sell the same robot with a different dock. The L40 Ultra and the L40 Ultra Gen 2 share most of a name and none of a test sheet. Reviewers also sometimes test a version built for shelves outside the US. We use only the US model's own results, and when nobody has measured the US model, the page says so.

A result for one model is never presented as a result for another. If we use a labeled sibling's figure, such as the same robot sold with a different dock, the sentence says it is the sibling's. A predecessor's results are never used at all. Where the exact model has no figure of its own, the page says "not measured". Borrowing another robot's homework would make our pages look more complete and would make them less true.

Why we never average the labs

Why don't you average the lab scores?

Because the labs aren't measuring the same thing. RTINGS (opens in a new tab) judges by eye how much spilled debris a robot leaves behind, and it presses synthetic pet hair into low carpet. Vacuum Wars (opens in a new tab) weighs the sand a robot digs out of medium carpet at full power, and it tests thresholds (opens in a new tab) one added strip at a time on a single step. Consumer Reports (opens in a new tab) uses human hair for its robot hair test. TechGearLab (opens in a new tab) uses real pet hair and weighs what comes out of the carpet. PCMag (opens in a new tab) deliberately runs robots in their default mode. The Hook Up (opens in a new tab) tests on deep, high-pile carpet. Just A Dad Approved (opens in a new tab) doesn't measure at all and scores the hassle of months of use at home instead.

A robot can come out on top in one of these tests and trail in another, and both results can be honest. Averaging them would produce a number nobody ever measured. So each figure stays attached to the test behind it, and the page tells you which test looks most like your floors. When two sources disagree, you see both.

Most testers simply repeat the pascal figure printed on the box. Vacuum Wars is the one tester that measures suction at the brush, where the robot actually meets the floor. Across the twelve robots on file with both a claim and a measurement, a bigger number on the box did not mean more suction at the floor. Pascals have become the horsepower of the robot aisle: printed large and checked rarely.

We take what a lab measured and ignore where it placed a robot, because some leaderboards blend in shopping data that has nothing to do with how well a robot cleans.

The gaps get reported too. No tester measures more than one unit of a model, so nobody knows how much one robot varies from the next off the same line. No lab measures reliability or customer support model by model either. When a page warns about a dock that quits after a year, that warning comes from owners, and the page says so.

Can I check the numbers myself?

Yes. Every review, roundup and comparison ends with "The evidence in full", where every lab's figure sits beside the method that produced it. The Home-Fit Finder lines up every figure for every robot on file, each one credited to the lab or maker behind it. Where nobody has measured something, it says "not measured" rather than leaving a blank.

How a page gets written and checked

Each page is drafted by an AI model working only from that page's evidence, with no web access while it writes, so a stray forum post can't slip in. Scripts then check every page before it can be published. They confirm there are no price figures and no claims of testing, that every number with a unit traces back to a fact on file for that exact model, and that every source link is one the page actually used. Every product section ends with a clear call (buy, skip or wait) and says who the robot suits. A skip always names a robot that would suit that reader better.

Why you see a price band, not a price

Prices change daily, so our pages show only the price band a robot sits in and leave the current price to Amazon. That way a page is never wrong by Tuesday.

Keeping pages current

Whenever we update the site, we check every product link for availability. A model that is no longer sold points to its replacement. Every page shows the date it was last updated, so you know how fresh the advice is.

How we make money

Who pays for this, and does it change the picks?

As an Amazon Associate I earn from qualifying purchases.

In practice, if you buy through one of our links, Amazon pays us a commission, and you pay exactly what you would have paid anyway. The commission never decides a call. A robot that doesn't suit you gets a skip, however much it would have paid us. No maker supplies us with robots, and none has a way to pay its way onto a page.

If a page tells you to buy something, you should be able to trace that advice to whoever measured it, on the exact robot you're about to order. If you can't, the fault is ours, not the robot's.

The evidence in full

9 sources