Our method

How we rank, which prices we read and how often, when we say "wait", and how we make money. What is not built yet says so, as planned.

As of November 14, 2026. The figures on this page come from the same data as the ranking and the prices, and change with them. The text was last updated on November 13, 2026.

• 92 models in our ranking tested by Just A Dad Approved, results read between September 9 and October 17, 2026
• 111 models in our ranking tested by Vacuum Wars, results read on September 29, 2026
• 5 test sources whose picks we show beside the score: RTINGS (list as of September 25, 2026), TechRadar (list as of September 25, 2026), Tom's Guide (list as of September 25, 2026), Vacuum Wars (lists from September 17 to September 26, 2026), Wirecutter (list as of August 4, 2026)
• 27 lists from those sources: overall rankings and picks for one use
• 138 models in our ranking, 90 models with a price today
• 6 stores that ship to the United States

1. The ranking: our own score

We rank with a score of our own, out of 10. It rests for 60 percent on test-site results: measurements made the same way for every robot. For 25 percent it rests on owners’ ratings at the stores that ship to the United States, and for 15 percent on the facts the maker states. Price does not count: it says nothing about how a robot cleans, and a score that follows the price shifts with every discount. The price comes after that, on the budget pages, in the best value sort, in the Best value label and in the alternatives.

The test-site results now come from Just A Dad Approved and Vacuum Wars: for each robot, how much dirt it lifts deep out of carpet, how it picks up pet hair, how well it mops, avoids obstacles and navigates. We take those measurements ability by ability, never a test site’s overall verdict, which often weighs in owners’ ratings and features that we weigh ourselves already. Every figure is on the product page, with a link to the test and the day we read it.

Only robots a test site has tested are ranked. We still follow the robots no test site has tested yet: they are listed under "Other models", with their price but without a place, so a model without a test is never our best pick. Only models sold by US stores count.

Beside the score we show the picks of 5 testers: RTINGS, TechRadar, Tom's Guide, Vacuum Wars, and Wirecutter. For example "best for pet hair" or a place in their top list, with a link and the date of the list: the day that version of the list was theirs. Those picks do not count in the score, since most are one line without a measurement. The advisor does use their picks for pets and carpet to order robots that fit your answers equally well. We copy no text.

One robot, two names

A test counts for the robot the test site tested, in every market that sells it, under the name that market uses: the robot Tom's Guide and Vacuum Wars test as the Dreame X60 Max Ultra Complete is sold in Europe as the X60 Pro Ultra Complete, and counts in both editions. Where the maker does not link the two names itself, we lay the specifications side by side. The product page gives the other name under "Also sold or tested as". Where the two markets sell different bundles, a pick counts per market: Wirecutter's top pick, the Roborock Q7 M5+ (the Q7 M5 with a self-emptying station), counts for the Q7 M5+ here and for the Q7 M5 in the Netherlands, where only the M5 is sold. A pick for a model no US store sells does not count here.

2. The score: how the parts come about

Test-site results (60 percent). We put every measurement on 0 to 10 with fixed anchors: in the carpet test, 50 percent of the dirt picked up is a 0 and 100 percent a 10, and in a verdict in stars 0 stars is a 0 and 5 stars a 10. The five abilities weigh as follows: deep clean on carpet 25, pet hair 20, mopping 20, obstacle avoidance 15, and navigation 20 percent. A robot counts as tested when abilities worth at least 70 percent of that weight are measured. An ability no test site measured counts at the average of the robots with the same station type, so a robot is not lifted by what was never tested. The fewer test sites stand behind a result, the more we pull it toward the average of its station type: with one test site it keeps 31 percent of its distance from that average, with two 47 and with three or more 57 percent. So one lucky result does not lift a model to the top. When more test sites test the same robots, we put them on one scale through the robots they both tested, so a strict tester does not sink its robots.

Owners’ ratings (25 percent). One figure per platform. Ratings at Amazon do not count. A model sold in several colours counts once, with the version that has the most ratings. The maker’s own web store does not count. We work with an average in which 150 ratings at the market’s mean count first, so a handful of five-star ratings does not beat thousands of good ones. Three stars is a 0, five stars a 10. Ratings a store has not shown for two weeks no longer count. A robot without ratings counts at the market’s mean where a market has ratings. The stores we read in the US give us none yet, so here the part is left out for every robot alike.

Customer ratings come from the stores themselves. We are not able to check whether each one comes from a real buyer. We trust the stores to do so, and we do not change the reviews. They count for 25 percent of our score, as one figure per platform.

Features (15 percent). Points for what the maker states: a station that empties the bin (10), washes the mop (10) and dries it warm (6), a roller mop (12) or spinning pads (10), lidar (10), a camera that recognises obstacles (12), under 3.5 inches tall or a 1.2-inch threshold (4 each), and a few smaller ones. We count only the facts we know, so a missing figure is no penalty, and only from 60 points of known facts.

When a part is missing, it is left out and the others count for more: the score is the weighted mean of the parts there are. On an equal score, the higher test-site part comes first, then more ratings. The score is worked out again after every price run, since the ratings change every day. Beside the score we say how sure it is: high with two or more test sites, or with one test site and at least 300 ratings, otherwise medium. What never counts: the price, our commission, payment for a place, a store’s best-seller list, the suction the maker states and a test site’s overall verdict.

Example: today’s number 1

The Roborock Qrevo Edge 3 Pro leads our ranking today. This is how its score comes about:

1. Test-site results, what the test site measured: Deep clean on carpet 8.8 (25%), Pet hair 9.1 (20%), Mopping 7.3 (20%), Obstacle avoidance 7.8 (15%), Navigation 6.7 (20%). The weighted mean: 7.98.
2. 2 test sites tested it, so the test-site part keeps 47% of its distance from its station type’s average of 6.84: 6.84 + 0.47 × (7.98 − 6.84) = 7.38.
3. Owners: no ratings at the stores that ship to the United States, so this part is left out.
4. Features: 87 of the 89 points for the facts the maker states: 9.78.
5. The score: (0.60 × 7.38 + 0.15 × 9.78) / 0.75 = 7.86, printed as 7.9.

Air purifiers

An air purifier’s score has two parts, not three. There is no test-site part: no test site that measures purifiers lets its results be stored and reused, so the registers stand in for it. Owners’ ratings (40 percent) count by the rules above. Features (60 percent) are points for what the maker and the registers state, 98 in all: a CADR certified by AHAM (10), clean air per watt on ENERGY STAR’s measure (15 from 5.6 cubic feet a minute per watt, 10 from 4.4, 5 from 3), a HEPA filter of a stated class (15, or 5 for a HEPA-type filter), 25 decibels or less on the lowest setting and 50 or less on the highest (8 each), a carbon filter (6), an auto mode with a laser particle sensor (8, with another sensor 4), no ioniser or one that switches off (8), CARB certification (5), a filter that lasts a year (5), an app (4), Matter (2) and a warranty of three years or more (4). We count only from 55 points of known facts. A fact we do not know counts at the share of known points the market’s ranked purifiers earn, so a model that states fewer facts is not lifted by it.

Only purifiers with a known CADR, the clean air they deliver, and enough known facts for a features part are ranked. The others get a provisional score and no place. The confidence is high with a CADR certified to AHAM AC-1, or to the Department of Energy’s test that adopts it, and at least 300 ratings, otherwise medium. The price never counts here either.

Headphones and earbuds

A pair of headphones’ score has two parts, not three. There is no test-site part yet: the test sites that measure headphones do not let their results be stored and reused, or are not counted yet. Owners’ ratings (40 percent) count by the rules above, and a model without ratings counts at the market’s mean. Features (60 percent) are points for what the maker states, 100 in all: adaptive or hybrid noise cancelling (15, with a microphone only inside or only outside 8), a transparency mode (5), the play time on one charge (15 from 8 hours for earbuds and 30 hours for headphones, 8 from 6 and 20 hours, with noise cancelling on where the maker states it), fast charging in 15 minutes or less (5), connecting to two devices at once (10), a codec beyond SBC and AAC (8), LE Audio or Auracast (5), an equaliser in the app (7), a Hi-Res certification (3), water resistance from IPX4 (8), wear detection (5), wireless charging (4), USB-C (4) and a warranty of two years or more (6). We count only from 55 points of known facts. A fact we do not know counts at the share of known points the market’s ranked models earn, so a model that states fewer facts is not lifted by it.

Only models whose form factor we know (in the ear, on the ear, over the ear or open) and with enough known facts for a features part are ranked. The others get a provisional score and no place. The confidence is high with at least 300 ratings, otherwise medium. The price never counts here either.

Home batteries

A home battery’s score has two parts, not three. There is no test-site part: no lab tests home batteries across brands yet, and we run no tests of our own. Owners’ ratings (30 percent) count by the rules above, and a model without ratings counts at the market’s mean. Features (70 percent) are points for what the maker states, 100 in all: the capacity you buy, the usable figure where the maker states it and the nominal one otherwise (20, none at 1 kWh or less, all from 8 kWh, in a straight line between), the round-trip efficiency on the AC side (15, none at 80 percent or less, all from 95 percent, in a straight line between, and nothing for a figure measured on the DC side or not said where), the standby draw (5, all at 1 W or less, none from 10 W, in a straight line between), the warranty’s years (10, none at two years, all from ten, in a straight line between, and 5 less when it holds only while the battery stays online), the capacity the warranty promises at its end (5, at 70 percent or more), the charge cycles the maker rates it for (10, 8 in a straight line from 3,000 to 8,000 cycles, and 2 more when it states that at least 70 percent of the capacity is left after them), certifications (15, 10 for a battery-safety code, IEC 62619 or VDE-AR-E 2510-50, or UL 9540 with UL 9540A, and 5 for a grid code), an electricity meter it reads (4), a dynamic electricity tariff (3), a local interface without the maker’s cloud (3), charging in the cold, by heating its own cells or charging from 0 °C or below (5), and backup power in an outage (5). Power and room to grow score nothing: they are what a home needs, not a quality, and the filters serve them. We count only from 55 points of known facts. The efficiency, the standby draw and the capacity at the warranty’s end count for every model: when the maker does not state one, the model earns no points for it. Any other fact we do not know counts at the share of known points the market’s ranked models earn, so a model that states fewer facts is not lifted by it.

Only models whose kind we know, other than a bare battery module, and whose maker states the nominal capacity, the warranty’s years, the rated charge cycles and the AC output are ranked. A battery module is left out because its efficiency and standby draw are those of the inverter it is paired with. The others get a provisional score and no place. The confidence is high with at least 300 ratings, otherwise medium. The price never counts here either.

3. Prices: read, checked and kept

At every store we follow we read the price and whether the model is in stock, and where the source gives them the seller and the delivery time too. That comes from the product data of a brand's own web store, and for Amazon.com from Keepa, a service that tracks prices at Amazon. We read and keep Amazon’s prices, but do not show them: Amazon allows its prices on a site like ours only through its own API, which we do not use yet. So Amazon.com appears on a product page as a link, and the price is on Amazon itself. We read all 6 stores every day. The table says it per store. Every price shows when we last saw it.

We show a price as the store shows it, even when it is half or double what we saw the week before: a cut or a rise can be the store’s own mistake, and we report what the store shows. Judging whether a price or a discount is genuine is work still to come. We keep every observation, so the price history and the 90-day low come from our own data.

4. The real price: shipped, and compared like for like

• Shipping to the United States counts, at every store.
• At the same price, a store in your own country comes first, then a store selling its own stock before a marketplace seller, then the shorter delivery time, then free returns and the longer return window. Commission never counts.
• Prices are shelf prices, before sales tax. Sales tax depends on your state and the store adds it at checkout, so no single figure would be right for every reader.
• Delivery time is part of the deal: tomorrow or in five days makes a difference.
• Planned: showing sales by the store itself and sales by a marketplace seller apart.
• Planned: showing refurbished and used offers as a separate category, never mixed with new.

5. Fake discounts

US federal rules set no fixed period for a "was" price. The FTC's guides only ask that it be a price the item was really offered at, openly and for a reasonably substantial time. We hold every store to the European rule instead: a discount is real only when the sale price is below the lowest price that store charged in the 30 days before the sale. Because we keep the list price next to every observed price, we can check and show this: claimed "was $1,299". The lowest price we saw in the last 30 days was $1,099.

6. Buy now or wait

On every product page we set today's price against the prices of the last 90 days. For each day the lowest price with shipping at the stores the page shows counts. A price holds until we see a new one, and a day on which the store gave no price, sold out or not readable, ends it. From 14 days with a price we give a verdict, and the rules apply in this order:

1. Lowest price: no day of the period was cheaper than today.
2. Good price: on at most 10 percent of the days with a price it was cheaper than today, and today's price sits in the lowest third of the range between the period's lowest and highest daily price, and the period's lowest daily price is no more than 10 percent below today's. We name those days and the lowest price.
3. Waiting may pay off: it was cheaper on more days, and the lowest daily price of the period was at least 5 percent below today's. We name the day and that price, so you can check it.
4. Usual price: it was cheaper on more days, but always by less than 5 percent.

The verdict and the label beside the price come from the same rule, and the price chart draws the same lowest daily price, so you can read it off. When the prices do not yet span 90 days, the verdict says "since we started measuring" with the first day instead of "90 days". Next to the verdict are the lowest and highest price of the period, how often the price dropped and when it last did. Sometimes the honest answer is "wait". That costs us a commission today and earns a reader who comes back.

7. Alternatives

Under the price on a product page are at most 4 alternatives, worked out again every day from the ranking and today's prices. Only models a store sells today count, and each shows how much more or less it costs than this model:

• Better for about the same price: the best-ranked model of the same station type that stands higher and costs at most 10 percent more today.
• Better for a bit more: the best-ranked model of the same station type that stands higher and costs more than this model today, by at most 30 percent, as a second choice beside the first.
• Nearly as good for much less: a model from the 3 places below that is at least 15 percent cheaper today, the highest placed first.
• One station type up or down: the best-ranked model with more station (from no station to self-emptying station to all-in-one station) for at most 10 percent more, and the best-ranked one with less station that is cheaper.
• When that makes fewer than two, the best-ranked lower model of the same station type that is cheaper today is added.
• When the model is on sale nowhere today, we show the best-ranked models of the same station type that are, at most 10 percent above this model's last price, and the station types above and below it.

Which store pays a commission plays no part: the rules are the same for every store.

Best value

On the lists and the product pages some models carry the label "Best value". We line up every ranked model on sale today by its lowest price, with prices from the same choice of shops as the page. The cheapest model gets the label, and after it every model that scores higher than all the cheaper ones. So each step up in price buys a higher score. A higher score counts only from 0.5 points (out of 10) above the previous label, so a gap that reads as equal on the page earns no new label. A model that costs more and does not score higher than a cheaper one does not get the label. The label does not say a model is cheap: even the dearest model can carry it, when no cheaper model scores as high.

The label comes from the whole ranking, on a filtered list too, so a model left out by the filter cannot hand it to a weaker one. The "Best value" sort puts these models first and then the rest, each part by score per dollar.

8. How we make money

Best at Best is free. Links to stores can be affiliate links: buy through one and we may earn a commission from the store or its affiliate network. The price you pay does not change. The commission never influences the ranking, the alternatives or the advice. Those come from the same data for every store. We sell nothing ourselves and take no money for a place in the ranking or for a review. A product page says so next to the buttons to the stores as well.

Today we earn only from links to MOVA (brand store): 1 of the 6 stores we follow. Buy through such an affiliate link and we receive a commission from the store or from its affiliate network. The other stores pay us nothing, and the ranking, the lowest price and the advice come from the same data for every store.

A brand’s own web store can have an affiliate link too (MOVA (brand store)). That maker then pays a commission on a sale, as any other store does, and never for a place in the ranking or for a review.

9. What we do not do

• No paid tests (Consumer Reports) as a source: they are licensed and cannot be checked freely.
• No average of other sites' final scores: from a test site we use its measurements per ability, never its overall verdict, and owners' ratings count by how many there are. No aggregators of aggregators.
• No copied review text. From testers only their measurements, positions and labels, with attribution.
• No reselling of products. We are a guide, not a store.