NFL consensus power rankings

How the composite is built. Each source is truncated to its top 32 and scored 32 points for first down to 1 for 32nd; a team a source does not rank earns nothing from it. Every source is then weighted by its tier share, split within the tier. Tier shares renormalize over the tiers that published this week, so a tier going dark never breaks the ranking, and a tier carrying only one source is capped at 35% because one outlet is not a tier. Ties break on how many sources ranked the team, then its single best rank, then alphabetically. The tiers are ordered by how little bias they carry, not by how famous they are. A betting market is the least biased signal available: it is real money, continuously repriced by people who lose that money when they are wrong, and it has no reason to favour a brand, a conference or a television window. An analytics model is next, objective and results-derived, but it is one method with one set of assumptions, and different models genuinely disagree. Human rankings come last because their known failure modes all point the same way: voters anchor on where a team started the season, they reward reputation and blue-blood brands, they are slow to drop a name team that keeps losing, and they see a fraction of the games they rank. That is not a knock on the AP poll, which is a fine measure of what people believe. It is simply the reason it does not lead here. Spread is the gap between a team’s best and worst rank among the sources that rank them; the count under a team name says how many of the 5 scoring sources that is. A source whose data is more than 30 days old is excluded from scoring and shown struck through.