How our ratings work

Nothing here is entered by hand. Eight upstream sources feed the site, each answering a different question.

Before any of it can be counted, each recorded deck has to be matched to an archetype. Where the source already publishes a name we recognise, we use it. In cEDH and Duel Commander the commander is the archetype, so the deck is matched on the commander it registers. Everything else is matched by comparing the list against the decks we already recognise in that format — and where nothing is a confident match, the deck is recorded with no archetype rather than being guessed at. That last case matters more than it sounds; see below.

An archetype's tier comes from its share of the decks we recorded in the format recently — how much of the field it is, not how well it did. Five tiers, from most to least of the field:

Only Challenge and Showcase results feed this figure, over a rolling recent window rather than all time, so a tier tracks the current metagame rather than the format's history.

In cEDH and Duel Commander the archetype is the commander, which splits the field far more finely — no single commander is ever a tenth of the metagame. Those two formats are graded on their own scale, so an S there means the same thing in context that an S means in Modern, and the two are not comparable as raw percentages. cEDH is also the one format where a tier reflects results as well as popularity: a commander that is played rarely but converts to the top tables unusually often can rate above its share of the field.

A card's badge in a competitive format answers how much of this format is built on this card — not how strong is this card in the abstract. A card inherits the prominence of the decks that play it, discounted by how heavily those decks commit to it: a full playset in a popular deck counts for much more than a single copy in the same list.

Cards are rated for several different reasons, and the reason is shown alongside the tier:

Sideboard cards are graded across the whole format rather than by their best single home, since a sideboard card's value is that many decks want a few copies — and discounted against maindeck play, because a sideboard card is only seen in some games of some matches. A sideboard rating never lowers a rating a card already earned from maindeck play.

Commander is a casual format with no meaningful tournament field, so its ratings come from how often decks that could play a card actually do, rather than from results. Cards that can be commanders are additionally rated on how often people build around them, and a card that scores well both ways is promoted and labelled as such. cEDH is the opposite case: those ratings come from recorded tournament decks, so they describe competitive play rather than deck registrations.

Tiers are per-format, so they cannot rank a Modern staple against a Pauper one — an S in Pauper and an S in Modern are two different claims. The power score exists for that comparison: one number per card, accumulated across every archetype in every format that plays it, with each appearance weighted by how much of the field plays that deck and discounted for sideboard slots.

What it rewards is worth being explicit about, because it is the whole design: a card that holds up popular decks outscores one that dominates a deck nobody plays. It follows that the power score will always understate a genuinely powerful card that has not been adopted yet, and will flatter an ordinary card sitting in a popular shell. It is a measure of a card's footprint on real metagames, not of its ceiling.

Where an event publishes round-by-round pairings, each round becomes one recorded result between two archetypes, and those results are pooled per archetype pair per format.

Draws count against both decks. A drawn match is a match that was not won, so draws sit in the denominator rather than being discarded — dropping them would inflate the win rate of decks that draw often. Two decks that frequently draw against each other will therefore both show below 50% in that matchup, which is the intended reading.

Every matchup figure is shown with the number of matches behind it. A win rate without its denominator is not a claim you can evaluate: 100% over two matches and 100% over two hundred are different statements, and they should not look alike. Both decks in a match must be identified for it to count at all, so matchup data thins out much faster than the metagame tables do.

Matchups are banded by how much evidence sits behind them. The bands are named for the sample, not for confidence — that is a deliberate choice, because no band in our data earns the word “confident”:

Below 5 matches no percentage is shown at all — just the record and the count. The cut points are ours, measured against our own distribution rather than borrowed: the last column is the share of cells in a 90-day, top-16 matchup matrix that fall in each band, pooled across formats. Narrower windows are thinner still; at 60 days almost no pairing in any format reaches 20 matches.

The band is encoded in more than one way on purpose — border style and weight, whether a percentage is printed at all, and the match count in text — so that none of it depends on distinguishing colours or noticing that a cell has been faded.

Where a matchup surface offers a time range, it filters on when the matches were actually played, so a narrower range genuinely describes a more recent metagame. It also describes a smaller one — narrowing the window is the fastest way to turn a solid sample into a thin one, so watch the match count as you change it.

One number for how few decks a format is. It is the Herfindahl–Hirschman index (HHI) — the standard measure of concentration, also known as Simpson's index — computed over one calendar month of tournament results:

             Σᵢ  nᵢ (nᵢ − 1)
   HHI  =  ——————————————
                 N (N − 1)

   nᵢ = decks of archetype i that month
   N   = total decks that month

Read plainly, that is the probability that two decks drawn at random from the month's field are the same archetype. We display its reciprocal, 1 / HHI, as the effective number of decks, because “this format is effectively 12 decks” is the same fact in units you already think in. Lower means more concentrated.

The (n − 1) and (N − 1) matter. The simpler form, “sum of squared shares”, is biased upward by roughly (1 − HHI) / N, which means a month with fewer tournaments in it reads as a more concentrated format. On a month holding 31 decks that error was 0.031 — about half the entire range Modern has moved through in eighteen months, manufactured purely by the sample size. The form above has no such bias, and models what is actually happening: drawing two decks from a finite field, without replacement.

Decks we could not classify are excluded. They are not one archetype, so counting them as one would inflate the index; counting them individually would deflate it. Every point reports what share of the field was classified, and that share is worth checking before comparing two distant months — our classification coverage has improved over time, and a coverage change can look like a concentration change.

Archetypes that have since died still count toward the months they were played in. A deck that was a tenth of the field last spring was a tenth of the field last spring, whatever it is doing now; dropping it would make every older month look artificially diverse and manufacture a trend.

Thin months are left blank rather than plotted. A month needs at least 200 classified decks across at least two events. Below that the reading swings by more than 20% on sample size alone, which for a chart people read against ban announcements is worse than showing nothing. A gap in the line is a month we do not have enough of, and the line does not bridge it.

Each source is checked daily against how recently it last delivered data, and an alert is raised if one goes quiet for longer than it should. Pages themselves are cached and rebuild at most once an hour, so a figure you see may be up to an hour behind the database.

These figures are a good picture of the recorded metagame. They are not the metagame. The differences worth knowing before you lean on one: