The method
How Prop Lab Organizes NFL Prop Research
Prop Lab turns NFL player prop research into a clearer weekly read. We put the current market alongside the matchup, a player’s role, availability, and game conditions, so you can see what is driving a prop, and where the uncertainty still is.
By Prop Lab
The problem
What Makes This Hard
Football is hard to predict. Some of that is the fun of it, and all of it makes this a hard problem to work on. A yardage prop can turn on one broken tackle or one red-zone call. Those are real events. They're also not repeatable skill, and a method that treats them as signal will confidently learn the wrong thing.
The season doesn't give you much to work with either. Seventeen games is short, roles move around inside it, and any one player-week is a tiny sample. There's an enormous amount of data and most of it is noise.
Then there's the number itself. A posted line already carries the market's read of the matchup, the injury report, the weather, and the player's role, priced by people who do this full time. Any honest method has to start there.
So the job is narrower than predicting the game. Some weeks it's finding context the market hasn't fully absorbed. Plenty of weeks the analysis just lands where the price already is, and that's a result too. Agreement means the number held up under a second look, which is worth knowing before you take either side of it. The work is being honest about which of those two you're looking at.
The AI part
Why Asking a Chatbot Isn't the Same Thing
Ask a general AI assistant like ChatGPT or Claude whether Patrick Mahomes throws for more than 230.5 yards against the Chargers this week and you'll get an answer. It'll be well written, and it'll sound like it thought about it. You'd do about as well flipping a coin.
The model isn't doing anything wrong. It's answering the question it was handed with what it has, and what it has is thinner than it looks. Its training data has a cutoff, so anything it already knows about Mahomes comes from a snapshot that's months old. To cover that, it searches the web, which sounds like the fix until you think about what a search returns. The pages that rank for a question like that are the pages built to rank for it. Picks posts, aggregators, previews written on Tuesday. What comes back is a well-organized summary of what other people published about the game, and that isn't the same as analyzing the game.
Two things happen next that are easy to miss. The first is that an assistant won't tell you what it couldn't find. If snap counts and route participation weren't in the pages it read, it doesn't stop and say the most important inputs are missing. It answers with what it has, in the same tone it would use if it had everything.
The second is that how good the answer sounds has nothing to do with how good the evidence was. Writing fluently is the thing these systems are best at. A model will say a player should comfortably clear his number with a strong case behind it or with almost nothing behind it, and both versions read the same. The phrasing of your question shapes it too. Ask whether the over looks good and you'll tend to get a case for the over. Ask about the under and you'll often get a case for the under. Same player, same game, same number.
That's where the actual work is. Prop Lab hands its AI layer the same structured set of market, matchup, availability, and usage data for every prop, in the same shape, and asks the same questions in the same order. What counts as enough evidence is set by the system, not by how a question got worded. When an input is missing it gets labeled instead of glossed over. When the conviction isn't there the label is Pass, and no amount of well-written reasoning gets to override that.
None of that guarantees a read is right. It means every prop was held to one standard, and you can see what the standard is. The principles below are it.
Why I built Prop LabHow we think about it
The Principles Behind the Read
These principles explain how we read a prop: start with the market, model the football context, and make uncertainty clear. For the plain-English version of the same research process, see the four-question framework in Learn.
Start with the market
Respect the consensus before looking for the context it hasn’t fully absorbed.
The market is where we start, not what we fight
A posted line already encodes a lot of information. We treat sportsbook consensus as a prior and look for the context it hasn’t fully absorbed, instead of assuming the market is wrong by default.
Model the football
Build the read from repeatable opportunity, full-field availability, and only the information that existed before kickoff.
We model opportunity, not just production
Recent yardage can mislead: a big game on few touches isn’t repeatable. We model how much of the offense actually flows to a player, including downfield and red-zone opportunity. Opportunity is more stable than output, and it’s what a line has to be beaten with.
Availability is about more than the player
A prop’s outcome depends on who else is on the field. Availability is modeled across other players, including teammates and opponents, not just the named player in the market.
No hindsight
Features for a given week are built only from weeks before it. A later week’s data never leaks backward into an earlier read. That’s the fastest way for a model to fool itself by grading its own homework.
State uncertainty honestly
Make conviction explicit, preserve a real pass, and show where the information stops.
Pass is a real answer, and it’s enforced
Below a confidence floor, a prop isn’t a Lean. It is a Pass. That rule lives in the system instead of in anyone’s judgment, so a weak read can’t be talked into a recommendation.
Confidence is an output, not a vibe
Every recommendation carries an explicit confidence band and score. Conviction is something the system has to state, not something the reader has to infer.
We would rather show a gap than fill one
When market consensus is only partial, or a source is unavailable, we label it instead of quietly guessing. A visible gap is more useful than a confident number built on missing data.
The evidence
What Actually Moves a Prop Line
The posted line is a starting point, not a diagnosis. Opponent tendencies, expected game script, venue conditions, player availability, and changing roles all shape the football case. Sportsbooks also respond to how a market is being bet, which can move the number without changing Prop Lab's underlying inputs.
Explore NFL player propsSignal 01
Opponent Strengths & Tendencies
A defense is more than its overall ranking. Look at how it handles this player’s position and the kinds of plays that create opportunity for him.
A matchup starts with the player’s actual job, not a defense’s overall ranking. We consider how that unit handles the relevant position, the kinds of plays it tends to allow, the pressure it can create, and whether it gives up explosive or red-zone opportunities. A defense that erases outside receivers can present a very different problem for a slot receiver.
Game Script & Tempo
Game flow sets the range of plausible volume. Pace, score pressure, and likely play calling can add chances or take them away before the game even starts.
The same player can have a very different opportunity set in a fast, competitive game than in one where his team is protecting a lead. A team expected to trail may need to throw more, while a faster pace creates more plays for everyone. We use the likely flow of the game to frame the volume a prop could realistically receive.
Weather
Weather isn’t a blanket adjustment. We check whether wind, rain, or cold can actually reach the field, and how that changes each offense’s choices.
Weather matters when it can change how a game is played. Wind can make deep passing less attractive, rain can affect ball handling, and cold can alter play calling. We read those conditions alongside the venue and each team’s tendencies, so a dome isn’t adjusted for weather that never reaches the field. Most weeks it changes nothing. We’d rather say that than find a wind story in every game.
Injuries
Injuries change more than a player’s status. Absences can redistribute touches, protection, and defensive attention across the entire roster.
Availability is rarely a clean yes-or-no question, and a prop depends on more than the named player. When a receiver is ruled out, targets can shift to teammates. When a lineman is missing, pressure can change a quarterback’s options. We account for those roster effects instead of treating an injury report as a single-player update.
Changing Roles & Usage
Roles can change before a season average catches up. We look at whether a player is gaining work, losing it, or sharing it differently before relying on a recent box score.
Recent production doesn’t always show where a workload is heading. A player can be gaining routes, carries, targets, or high-value opportunities before the box score catches up, or losing them while recent results still look strong. We compare current usage with the role a player has established to judge the opportunity in front of him, not just the last result. This is the hardest one to get right. A role change is obvious in hindsight and ambiguous while it’s happening, and we’re sometimes late to it.
Market Pressure & Sharp Action
A posted line is also a price. Sportsbooks can adjust it when betting demand builds on one side or when respected bettors take a position.
Sportsbooks can move a line when one side attracts enough betting demand to create lopsided exposure, or when respected bettors take a strong position. Those moves often arrive alongside news and football context, but they don’t automatically explain a player’s expected outcome. Prop Lab doesn’t use betting demand or sharp action as a model input. This is context for how and why a posted market can move during the week.
Reading the Signals
How to Read the Output
Each read starts with a structured set of market, player, matchup, and game context. The AI weighs those inputs together, checks where they agree or conflict, and applies the same standard to every supported prop. That keeps the output consistent and objective instead of letting one random signal drive the answer.
What comes out is a side, over or under, paired with one of four confidence levels instead of a raw probability. That pairing is a deliberate choice. A number like 63.4 percent implies a precision the underlying data doesn't support, and false precision is its own kind of dishonesty. Four levels say what we can actually defend: how much separation the evidence created, and how much the signals agreed.
So read a label as relative strength, not as odds. It describes the state of the evidence, not the chance a player clears his number. Alongside it, every read carries an explicit confidence band and score, so you can see how firm the read is without having to infer it from tone.
Resulting read
Strong Take
The strongest conviction the AI analyst puts behind a side. Multiple signals point the same way, the evidence separates clearly from the market, and there’s less uncertainty left than usual. It’s the setup that deserves your closest look. It is not a proven betting edge and still isn’t a guarantee.
Take The analyst is confident in the side, but it isn’t a clean setup. The evidence clears our bar with real uncertainty left over, so treat it as a read worth investigating. It’s not a promise about the result.
Lean There’s a side the analyst prefers, but the separation is smaller or the signals are mixed. Worth keeping on your radar as new information arrives. There isn’t enough separation yet to call it a clear play.
Pass We still name a side, because there’s almost always one we prefer. What’s missing is the conviction to act on it at this price. Either the evidence doesn’t separate from the line, or important signals disagree. We’d rather show you an honest Pass than force a recommendation.
What we don't claim
Football is inherently tough to predict and many factors influence the outcome of a game (which is why we all love to watch it). Player availability, roles and usage, weather, market lines, and other factors can change at a moment's notice without warning, and any data source can be incomplete or delayed. For those reasons, our methodology and research framework doesn’t guarantee an outcome or replace independent judgment.
League sources
Methodology FAQ
See the full board
Put the research to work on every slate
Create a free account to explore the full board with model signals, matchup context, injuries, weather, and line movement in one place.