Methodology
How these numbers were produced, and what we refuse to do
A comparison site has exactly one asset: that its numbers can be trusted. Everything below exists to make that checkable by a reader rather than something you have to take on faith.
Where the data comes from
Every specification is read from the manufacturer’s own product listing — the brand ’s own website, never a marketplace listing, a competitor’s review, a press release or another comparison site. We read 7 manufacturers directly. Where a brand publishes a specification in a table, an image caption, an embedded product payload or a specification sub-page, all of those are fair game; where a brand simply does not publish a figure, we do not fill the gap.
Prices are vendor list prices captured at collection time. They move constantly, and the price we hold is a reference point rather than a live quote — always check the vendor page before buying.
How a value gets in
- Collection. Each product page is fetched once and cached, so a re-run reproduces the same dataset rather than silently drifting.
- Spec-sheet isolation. Product pages routinely contain other products’ specifications — comparison modules, “related products” carousels, and on some vendors even the site navigation. We locate the vendor’s own specification table as the densest contiguous run of label/value rows and extract from that, discarding the rest.
- Normalisation. Twelve vendors word the same specification a dozen different ways. Values are converted to one unit each — watt-hours, watts, pounds, minutes, months — and a field is only emitted when a label and a unit pattern both match.
- Evidence capture. Every stored value keeps a verbatim quote from the source page. This is not a note for our benefit; it is what makes step 5 possible.
- Verification. Each value’s quote is programmatically re-checked against the source text. If the quote cannot be found, the value is discarded, regardless of how plausible it looks. A fabricated number cannot survive this, because a fabricated number cannot also appear in the vendor’s own page.
- Physical plausibility. Surviving values are checked against physics. A complete LiFePO4 station delivers roughly 30–60 watt-hours per pound; a weight implying more than 70 Wh/lb belongs to a different product on the same page. That failure mode is real, we hit it, and it is now detected automatically.
- Conflict reporting. Two independent extraction methods are run over every page. Where they disagree the disagreement is recorded rather than averaged away, so systematic problems stay visible.
Current dataset
| Measure | Value |
|---|---|
| Models tracked | 76 |
| Brands tracked | 7 |
| Individually verified spec values | 840 |
| Average fields per model | 11.1 of 19 |
| Models with a captured list price | 73 of 76 |
The fields we try to track
Not every vendor publishes every field. A missing value is displayed as “—” everywhere on the site and the model is excluded from any ranking that needs it. We never substitute a typical value, a predecessor’s value, or another model’s value.
| Field | Meaning |
|---|---|
capacityWh | Usable battery capacity in watt-hours |
acOutputW | Continuous AC output in watts |
acSurgeW | Peak / surge AC output |
powerLiftingW | Boost or "power lifting" wattage |
chemistry | Battery chemistry (LiFePO4, NMC, LTO, lead-acid) |
cycleLife | Rated charge cycles to 80% capacity |
weight | Net weight, normalised to pounds |
rechargeAcMinutes | Time to full from an AC wall outlet |
solarInputW | Maximum solar input wattage |
acOutlets | Count of AC outlets |
usbCPorts | Count of USB-C ports |
usbAPorts | Count of USB-A ports |
carPort | 12V car port / cigarette outlet |
upsSwitchoverMs | UPS / EPS transfer time in milliseconds |
warrantyMonths | Stated warranty length, normalised to months |
maxExpandableWh | Maximum capacity with add-on batteries |
noiseDb | Stated noise level |
ipRating | Ingress protection rating |
operatingTempF | Low end of the operating temperature range |
Derived metrics we compute
Vendors publish specs; they do not publish the ratios that make them comparable. Every derived figure on this site is a function of verified values only, and disappears when an input is missing:
- Cost per watt-hour — list price ÷ capacity. The fairest single value metric in the category.
- Watt-hours per pound — capacity ÷ weight. The portability measure, and the one buyers most often wish they had checked.
- Inverter headroom — continuous AC watts per 100 Wh. High values mean a big load on a small battery.
- AC recharge rate — capacity ÷ vendor charge time, expressed in watts. Makes charge times comparable across different battery sizes.
- Lifetime throughput — capacity × rated cycles. Converts a chemistry claim into a single comparable figure.
What we will not do
- We do not invent reviews, ratings or testimonials. No invented people, no fictional user quotes, no made-up star ratings. Fabricated consumer reviews breach the US FTC rule on consumer reviews (16 CFR 465) and Google’s spam policies, and they are simple fraud. Our editorial analysis is written by us and labelled as ours.
- We do not claim hands-on testing. No “we measured”, no invented lab figures, no implied first-hand experience we do not have.
- We do not estimate missing specifications. A gap in the table is information: it tells you what the vendor is not willing to publish.
- We do not copy specs from other review sites. If a number is not on the manufacturer’s page, we do not have it.
- We do not present our composites as measurements. The home-backup readiness score and composite rankings are our judgement; their inputs and weights are published so you can disagree with them precisely.
Known limitations
Stated plainly, because a methodology page that only lists strengths is marketing.
- Vendor figures are vendor claims. Cycle life, noise and surge ratings are measured by manufacturers under conditions they choose. We verify that a figure was published, not that it is true.
- Coverage is not uniform. Larger brands publish more, so they rank in more dimensions. Coverage of the dataset is 76 models from 7 brands — a fraction of the category, chosen to be the models buyers actually consider.
- Three models still lack a captured price. DJI sells its Power line through a storefront that blocks requests from our collectors, so we show no price rather than an old or approximate one.
- Seven models lack vendor imagery. Where a vendor page exposes no usable product image, we show a labelled placeholder instead of stock photography that might misrepresent the product.
- No acoustic or electrical measurements. Noise, inverter waveform quality, thermal behaviour under sustained load and real-world efficiency are not knowable from a spec sheet. Where they matter, we say so and link to what the vendor claims rather than pretending to know.
Corrections
If a figure here disagrees with the vendor page, the vendor page is right and we want to know. Email [email protected] with the model and the field, and it will be corrected at the source rather than patched on the page — the dataset is regenerated from vendor pages, so a fix propagates everywhere at once.