How this is made
In short: ByTheRank's research system was designed by our editor, Maximus Milazzo, together with AI agents, and is run by a team of AI research agents (Anthropic's Claude) plus automated software. Here is what it does:
- What we collect. Official test results: the figures every certified product must report to the U.S. EPA's ENERGY STAR program, measured under U.S. government or industry test procedures, plus official U.S. electricity prices. No customer reviews, no paid placements, and no hands-on tests of our own.
- How we score. Each page says which numbers matter for its question and how much each counts. Software scores every product on those numbers and adds them up. The weights and the full table are on the page, so you can check the result, or download the table to re-rank it for your own priorities.
- How we check. A separate AI agent that did not build the page recomputes the winners from the raw data and must pass it against our published rules before a new or redesigned page goes live. Routine data refreshes are re-checked by software and go back to that agent if the top 3 change on more than a third of the pages.
- How often we refresh. Software downloads the data again every 30 days by default and rebuilds and re-checks every page; each page shows its Updated and Next refresh dates.
- Who is accountable. Our editor, Maximus Milazzo. Corrections: [email protected].
Technical details (sources, formulas, quality gate, data files)
The split between AI agents and software
Frontier-model research agents decide what is worth ranking, find and vet datasets, design each ranking's method and audit the results. Deterministic code on our own servers collects the data, computes every metric and score, rebuilds the pages and runs the checks. Every number on a page comes from code applied to public data; routine refreshes are not rewritten by AI.
1. Research: which questions deserve a page
A research agent reviews the questions people search for (search autocomplete suggestions) and checks whether a public, licensed dataset can answer them better than existing pages. A council of independent agents proposes, critiques and scores candidate pages; a page is built only if it states the question it answers, the evidence, the decision it helps with, and a number computed here rather than copied.
2. Sources
U.S. EPA ENERGY STAR certified-product lists (data.energystar.gov), which report each model's values as measured under the applicable U.S. Department of Energy or industry test procedure; U.S. Energy Information Administration electricity prices (Electric Power Monthly). Each page names its dataset, record ids and retrieval date.
3. Formulas
- Score: each criterion is scored against the best product on the page (best = 100%; lower-is-better values as best divided by value), multiplied by its declared weight; weights sum to 100% (how we rank).
- Five-year electricity cost: certified annual energy use (kWh/yr) × the EIA U.S. average residential price (date on the page) × 5.
- Duplicate listings: certification lists often carry one machine several times, sometimes under different brands. Listings with identical certified values on every reported field are counted as one model, and each page states both counts (distinct models and listings) and exactly what was merged.
- The cut: every model tied with the last one on the page's selection number is listed, so the cut never picks among equals.
- Why #1 is #1: every ranking shows the #1 vs #2 margin per criterion, the deciding number with its source and date, any equal scores (they share a rank and the page states the order rule) and the smallest weight change that would put #2 first, or that none would. All of it is computed from the scores.
4. Quality gate
Before a new or changed set of pages is published, a separate quality agent re-derives the winners and key figures from the raw data and issues a written PASS/FAIL verdict against a published rule set (usefulness, sourcing, no near-duplicate pages, structured data matching visible text, every #1 defensible for the page's question, and a justified order: each page shows why its #1 is #1). New page sets, and a sample of changed ones, also get a blind audit by a further agent on our strongest model. Automated checks refuse any page with fewer than 3 products, a missing value, weights that do not sum to 100%, an undated price or a missing source; a refresh that changes the top 3 on more than a third of the pages goes back to the quality agent.
5. AI disclosure
Introductions and FAQ answers are written with AI models (Anthropic's Claude) from templates filled with the computed numbers. No AI model picks a winner, sets a score or supplies a figure.
6. Open data
Every ranking table is available as CSV and JSON from its page and from all open data. Commissions never change a score (disclosure).