Nobody Knows Who’s Good Yet. Here’s Why Week 4 Matters.
Two games can tell us something. They cannot tell us enough. Early in the season, huge parts of college football have not crossed paths yet, which makes national rankings much shakier than the number beside a team’s name suggests. Our own data shows when that starts to change.
Every September, the same thing happens. A team crushes two overmatched opponents, jumps in the polls, and suddenly everyone wants to know whether it is a playoff team. Another team wins ugly twice and gets written off. After only a couple of games, we talk about the rankings as if the country has already been sorted out.
It has not. The problem is not just that two games are a small sample. The bigger problem is that most teams have not played enough connected competition yet. Michigan may have played two teams, Georgia may have played two completely different teams, and Oregon may be sitting in another part of the schedule altogether. There may be almost no game evidence connecting those teams to one another.
That means early-season college football is not really one national comparison yet. It is a bunch of smaller groups trying to be ranked on the same list. Our live network pageshows exactly how quickly those groups begin to connect.
After two weeks, college football is basically 39 mini-leagues
Think of every game as a bridge between two teams. Once enough bridges exist, you can trace a path from almost any team in the country to any other team through shared opponents and opponents of opponents. That is what allows an opponent-adjusted rating to compare teams nationally.
Early in September, those bridges barely exist. We pulled the published 2026 FBS schedule from CFBD and counted how many separate groups of teams are actually connected after each week.
| Through week | Separate groups | Largest group |
|---|---|---|
| 0 (openers) | 130 | 2 of 138 teams |
| 1 | 87 | 6 of 138 teams |
| 2 | 39 | 17 of 138 teams |
| 3 | 3 | 129 of 138 teams |
| 4 | 1 | 138 of 138 teams |
Through two weeks, the sport is still split into 39 separate groups. The biggest one contains only 17 of 138 teams. If Team A is the best team in one group and Team B is the best team in another, there may still be no path of game results connecting them. We are trying to decide which is better without having much common evidence.
By Week 3, almost the entire country has connected. By Week 4, all 138 teams are part of one network. For the first time, the season itself gives us a chain of results connecting everyone.
Week 4 does not suddenly make every ranking correct. Four games are still four games. It is simply the first point this season when ranking the whole country is based on one connected body of evidence instead of dozens of separate islands.
That is how an 0-2 team ended up No. 4
Two games into 2026, one version of our own ratings produced something obviously wrong: Sam Houston was 0-2 and ranked fourth in the country.
The reason was not that Sam Houston had secretly played like a top-four team. Sam Houston, Troy and Tulsa were sitting inside a small pocket of connected teams that also included Oregon and Indiana. With so few games linking that pocket to the rest of the country, strength from the top of the group could spill into teams that had done very little to earn it.
In plain English: the computer had not seen enough football yet.
the overall rating, Adj. Net, is designed to pull extreme early results back toward average until more evidence arrives. That helps prevent wild rankings like Sam Houston at No. 4. But no model can create information that has not happened on the field yet.
Our own history says two games are not enough
Instead of only criticizing the AP Poll, we tested the rating model against itself. We went back through eleven full seasons of our historical ratings from 2014 through 2025, excluding 2020, and asked a simple question: how similar are the rankings after about two games to the rankings at the end of the season?
The easiest number to understand is the top 10. After roughly two games, only 4.1 of the teams in our top 10, on average, are still there at the end of the season. After roughly four games, that rises to 5.8 of 10.
The full rankings tell the same story. Our rank similarity score improves from 0.71 after about two games to 0.83 after about four. Early rankings clearly contain useful information, but they become noticeably more stable once teams have played more football and the national schedule is connected.
The AP Poll starts the season guessing too
Every ranking system has the same basic September problem: there is not enough current-season football to work with yet. Human polls fill that gap with preseason expectations. Computer models use some combination of previous seasons, recruiting, returning production or conservative early-season adjustments.
Those preseason expectations are far from perfect. Across the twelve seasons of the College Football Playoff era from 2014 through 2025, a study of 300 preseason-ranked AP teams found that only 57% finished the season in the final Top 25 and just 32% finished in the final top 10.
So when a preseason No. 4 team starts 2-0, its Week 2 ranking is not suddenly based on two games alone. A lot of what we believed in August is still baked into that number. Sometimes that prior is useful. Sometimes it is wrong. Either way, two games have not fully replaced it yet.
Last year’s team tells us less than it used to
The obvious solution to having too little new data is to lean on last year. But modern college football has made that harder too. Rosters change faster than they used to, so the team wearing the same logo in September may look very different from the one that finished the previous season.
| Season | Avg. returning production | Teams under 50% returning |
|---|---|---|
| 2014 | 59.5% | 30% |
| 2016 | 63.8% | 23% |
| 2018 | 60.4% | 32% |
| 2020 | 61.8% | 29% |
| 2021* | 69.4% | 19% |
| 2022 | 55.6% | 39% |
| 2023 | 55.2% | 45% |
| 2024 | 47.6% | 56% |
| 2025 | 39.9% | 65% |
| 2026 | 44.0% | 59% |
*2021 is inflated by the NCAA’s blanket COVID eligibility waiver, which allowed many seniors to return for an additional season.
In 2014, only 30% of FBS teams returned less than half of their production from the previous year. In 2025, that number reached 65%. The average team returned just 39.9% of its production.
That creates a bad combination for early rankings: we do not have enough games from this season yet, and the information from last season is becoming less reliable too.
Start taking them more seriously around Week 4.
Not because Week 4 magically reveals who is good. It does not. But by then, every FBS team is finally connected through the schedule, and our historical testing shows the rankings become meaningfully more stable around the same point.
That is the real takeaway: early rankings should come with less confidence. Watch the games. Argue about the top 10. Have fun with it. Just understand that in Weeks 1 and 2, everyone is working with an incomplete picture, including us.
Network figures are PRIME Football’s own analysis of the published 2026 FBS schedule (via CFBD). Rank-similarity figures use Spearman correlation across eleven historical seasons (2014-2025, excluding 2020, which has no published data) of the site’s own rating history. Because site week-numbering shifted slightly across seasons, each season is aligned by median games played (≈2 and ≈4) rather than raw week label. Returning-production figures are CFBD’s /player/returning data (national average and per-team usage share, 2014-2026; 2013 has no published data). AP Poll historical figures via RotoWire’s 12-year preseason-poll study.