Document · policy
How a score is produced, what it is built from, and when we refuse to produce one.
A FilterSpecs score is 100 points across four components, shown out of 10. It is computed from the product record and nothing else: we do not test products, we verify certifications against the certifier's own database, and a product we cannot verify does not get a score.
Every point comes from one of these. There is no discretionary component, and no revenue term — see how we make money.
| Component | Points | What it measures |
|---|---|---|
| Certified coverage | 40 | The share of the contaminants that matter for this product type which the product is actually listed for, weighted by how much each one matters. |
| Claim integrity | 20 | Starts full and is deducted against: a marketing claim with no listing behind it costs more than one backed by a published lab report, and more again when the contaminant is a high-priority one. |
| Cost of ownership | 25 | Five-year cost per gallon against the median for the same product type. Purchase price alone decides nothing. |
| Practical fit | 15 | Warranty length, cartridge lock-in, whether a third-party cartridge voids the listing, required clearance, and install difficulty. |
A product with no certification and no manufacturer claim gets no score, and neither does one whose running cost we cannot compute. We show the gap instead. A score built on absent data reads exactly like a score built on verified data, which is the reason not to build one.
Where a component's data is missing but the product still qualifies for a score, the remaining components are scaled to 100 rather than the missing one being counted as zero. A product is not penalised for a fact we have not found; it simply is not credited for one either.
Running costs depend on how much water a household puts through a system. Rather than guess per product, we publish one assumption per install point and apply it everywhere. Change the assumption and every cost figure on the site recalculates at the next build.
A cartridge is replaced when either its stated life in months or its stated capacity in gallons runs out at that volume, whichever comes first. That is usually sooner than the months figure on the box.
Some manufacturers publish no replacement interval at all. We work one out rather than leaving the running cost blank, in this order.
A capacity rating in gallons measures how long the filtration stages hold their reduction claim, so it stands in for a missing interval on a sediment prefilter, carbon prefilter, carbon block, post filter and nothing else. A reverse osmosis membrane, a UV lamp, softener resin and a remineralizer are not replaced on that cycle. Letting them inherit a filtration capacity is not conservative, it is simply wrong: it assumes the membrane is changed as often as the sediment prefilter, and produces a running cost several times too high.
So where a component's class is not one a capacity governs and the brand publishes no life for it, that component is left out of the running cost entirely and the page says which. We do not substitute a typical life. A stated omission is honest and checkable; an invented number is neither, and it would look exactly like a sourced one.
A derived interval costs the product 3 points of its cost of ownership score, and a figure that had to leave a component out costs a further 2. That is not a punishment for being cheap to run; it is the price of not telling buyers when to change the filter, which is a real cost of owning the thing. It also closes a hole: without it, a brand that published nothing would drop the component entirely and have the other three scaled up, so silence could raise a score. The second deduction exists for the same reason: an incomplete figure is the brand's doing, and it should not be free.
The direction errs deliberately. A certified capacity is the point at which the system was last shown to still meet the standard, not the point the manufacturer would tell you to change the filter, so a derived interval runs shorter than a published one and the cost comes out high. A brand cannot improve its figure by staying quiet.
A certification only earns coverage points when it is relevant to what the product is. An NSF/ANSI 372 lead-free listing says the fittings are lead-free; it does not say the system reduces lead. The standards counted per product type:
| Product type | Standards |
|---|---|
| Under-sink reverse osmosis | NSF/ANSI 58, NSF/ANSI 53, NSF/ANSI 401, NSF/ANSI 42, NSF/ANSI 372 |
Every factual field on a product record carries the source it came from, and those sources are listed at the foot of each page.
Sources rank in this order: the certifier's own listing database, then a regulator, then a manufacturer's performance data sheet, then the product manual, then a spec sheet, then the brand's website, then a retailer's listing. A retailer listing never establishes a specification on its own.
A value we cannot verify is recorded as missing, not estimated. Missing is always better than wrong.
We publish a figure, show the one we did not publish with its source, and say how we chose. We used to publish neither and print "sources disagree", and that was worse: it threw away a number a certifier prints and left the page saying less than either source on its own. A disagreement between two certifiers about the same model is worth reporting — but it is worth more next to a number than instead of one.
Which figure we publish is decided in this order.
A field is blank only when no source publishes a value at all.
Least favourable is not always the smaller number, so the direction is fixed per field type rather than decided case by case:
| Field | We publish the |
|---|---|
| Production rate | lower figure |
| Flow rate | lower figure |
| Capacity in gallons | lower figure |
| Percent reduction | lower figure |
| Filter life | lower figure |
| Maximum permissible effluent | higher figure |
| Influent challenge | lower figure |
| Price | higher figure |
A maximum permissible effluent is the most contamination a product may leave in the water and still pass the test. A higher limit is therefore the weaker guarantee, not the stronger one, so where two sheets disagree we publish the higher figure. Publishing the stricter of the two would overstate every product certified under the looser one, across the whole site at once. Price works the same way for the opposite reason: the higher of two disputed prices is the one that does not flatter the running cost.
Ranked lists are ordered by score. Commission rate is not an input, and cannot be: the code that scores a product is handed a copy of the record with the commercial block removed, so it does not have the number to be influenced by.
Saying a product holds no certification is a strong claim, and for most of this year we were not entitled to make it as often as we did. Certifier search engines are not complete. On 20 September 2026 we found IAPMO's certificate endpoint serving four certificates as active through 2029 while its own search returned nothing for them under any search term, including their file numbers. The listings were live; the search could not see them.
So we now distinguish three things an empty result can mean, and word them differently on the page.
The middle claim is weaker than what most sites imply when they call something uncertified. It is also, as far as we know, stronger than what any of them actually did: searching all three databases and publishing the date is not common practice.
Every search we run is preceded by a control query against a product known to be listed. If the control comes back empty the database is not answering correctly, and we record nothing rather than record an absence — a broken search and a genuine absence look identical, and only one of them is a fact.
If a figure on this site is wrong, tell us and we will fix it and say what changed. Contact.
Every change to this methodology, and every change to the commercial disclosure, is listed here with the date it took effect. A methodology that can be revised without a record is not a methodology.