AI football betting prompts built around xG and the fair price
Football is the hardest sport to prompt well: three outcomes, low scoring, and the most efficient market in betting. A prompt that ignores the draw or trusts a 4-0 result over the xG behind it will lose slowly and confidently.
What a football prompt has to get right
The first job of a football prompt is arithmetic, not opinion: strip the bookmaker margin out of the 1X2 prices so the model starts from a fair baseline instead of an inflated one. Everything after that is an adjustment — form measured in xG rather than points, the home-away split of each side, who is missing, and how many days of rest each team had.
The second job is to keep the draw honest. Roughly a quarter of matches in the big European leagues end level, and models left to their own instincts systematically under-price that. Both prompts below force a three-number probability set that sums to 100%, so an under-weighted draw becomes visible immediately instead of hiding inside a confident "home win".
The third job is scale. Football is by far the largest slate on the board — well over a thousand fixtures land in a fortnight — so a prompt that only works when you have read the team news is a prompt you will use twice. Both versions below are written to run on whatever you can paste in from a match page, and to say so when that is not enough.
De-vig before you predict
Convert 1X2 odds to implied probabilities, remove the overround, and treat the result as the baseline. A prompt that starts from raw odds is starting from a number that already sums to more than 100%.
xG over results
Five-match samples of goals are almost noise. xG for and against, shot volume and shot quality tell you whether a run of wins is real. Value lives where the table lags the underlying performance.
Home-away split, not season averages
Many sides are a different team away from home — deeper block, fewer shots, more draws. Feed the split explicitly, otherwise the model averages the two into something that describes neither.
Rotation, injuries and congestion
A midweek European tie three days earlier, a suspended centre-back, a keeper change. These move the price more than most narratives, and they are the factors a model cannot guess — you have to supply them.
Football prompts v1 and v2 — and how they differ
The same model, two instruction sets, two different betting personalities. Run both on the same matches; that comparison is the only thing that settles the argument.
| Version | Focus | Style | Best for |
|---|---|---|---|
| v1 | Fair 1X2 baseline, then small adjustments | Disciplined | 1X2, consistency |
| v2 | Shot quality and goals markets | Goals-hunting | Totals and BTTS value |
You are a disciplined football betting analyst. Match: {home} vs {away}, {league}, {date}. Market 1X2: {odds}.
Step 1: convert the 1X2 odds to implied probabilities and remove the bookmaker margin to get a fair baseline.
Step 2: adjust that baseline only for verifiable factors — form measured in xG for/against (last 5), confirmed injuries and suspensions, home form for the home side and away form for the away side, days of rest and travel. Do not adjust for narratives or motivation.
Keep the draw honest: it is roughly 25% in most top leagues.
Output exactly:
1) Home / draw / away probabilities summing to 100%
2) Main pick + confidence 1-10
3) Best value market (1X2 / Over-Under 2.5 / BTTS) and the reason
4) Most likely correct score
5) One-line reasoning
If your fair price matches the offered price, answer "no bet".
You are an attacking-metrics football analyst. For {home} vs {away} ({league}, {date}):
Base your read on xG for and against, shot volume and shot quality, set-piece threat and how high each defensive line plays — not on results. Lean into goals markets when both attacks create real chances, and away from them when either side suppresses shot quality.
Compare every conclusion with the posted line {odds} and flag where the market disagrees with the underlying numbers.
Output exactly:
1) Predicted score
2) Over/Under pick with the line you are using
3) BTTS yes/no
4) 1X2 pick + confidence 1-10
5) The single decisive factor
If the xG samples are too small to separate the sides, say so and answer "no bet".
Placeholders in braces are filled automatically when you run a prompt from a match in the AI Lab. Pasting into your own chat window works too — just replace them by hand.
What to feed the model, and what a usable answer looks like
Feed it this
- League, matchweek and kick-off date — plus cup context if the match is a dead rubber.
- Last five matches per side with xG for and against, not just scorelines.
- Home form for the home team, away form for the away team — separately.
- Confirmed absences: injuries, suspensions, and any keeper or centre-back change.
- Days of rest since the last match and travel distance.
- The line: 1X2, Over-Under 2.5 and BTTS. Weather too, if it is extreme.
Good output has
- Three probabilities for home / draw / away summing to exactly 100%.
- A main pick plus confidence 1-10, with a low score allowed.
- The best value market of the three (1X2, Over-Under, BTTS) and why.
- A most-likely correct score, which exposes an incoherent probability set fast.
- One decisive factor in one line — no paragraph of hedging.
- A clear "no bet" when the fair price and the offered price agree.
Where football prompts usually go wrong
- Under-weighting the draw (it is around 25% in most top leagues).
- Reading one 4-0 as a step change instead of variance.
- Motivation narratives ("they need the win") replacing data.
- Totals picks that ignore how each side actually creates shots.
Odds, model context and market drift for each fixture are on the football matches with odds and AI picks board, so most of the input list above can be copied straight from the match page.
Prompting the three markets football actually offers
A prompt that only answers "who wins" throws away most of a football card. The three liquid markets reward different reasoning, and asking for all three in one answer is also the cheapest coherence check you have: a 1X2 read, a totals read and a correct score that contradict each other tell you the model is guessing.
Three-way, draw included
Demand three probabilities that sum to 100% and compare each with the de-vigged price. The draw is the honesty test — a model that prices it under 20% in a tight league match is not reasoning, it is picking a favourite.
Totals need shot creation, not results
Over-Under is a question about how each side generates and concedes chances: shot volume, shot quality, set-piece threat, defensive line height. Two teams that both create little produce unders regardless of how attacking their reputations are.
Both teams to score
BTTS is close to two independent scoring questions, so ask for each side's chance of scoring separately before the yes/no. It is also where a weak keeper or a missing centre-back moves the honest number most.
Handicaps and correct score
Ask for a most likely correct score even when you are not betting it: it exposes an incoherent probability set instantly. If the model says 55% home win and predicts 1-1, one of those two numbers is wrong.
Measure both versions before you trust either
Store both versions
Save v1 and v2 as separate prompts in the AI Lab so every run is attributed to a version instead of blurring together.
Run them on the same matches
Pick fixtures from the football board and lock both forecasts before start. Same slate, same information, no hindsight.
Judge on ROI, not hit-rate
A value prompt taking underdogs will always look worse on hit-rate and can still be the profitable one. Settlement and scoring are automatic once the match finishes.
The AI Lab starts on the $19 tier with one sport and five stored prompts, which is enough for a full v1-versus-v2 comparison in football. Open a free trial to run it on today's card, or read the prompt library overview for the shared structure behind every sport.
Football prompt questions
Why should a football prompt de-vig the odds first?
Because bookmaker prices include a margin, so implied probabilities sum to more than 100%. If the model treats them as fair it will systematically overestimate every outcome and see "value" where there is none. Removing the overround gives a baseline that is honest enough to argue with.
Where do I get xG numbers to paste in?
Any public source you already trust works, as long as you use the same source consistently — mixing providers introduces differences bigger than the effects you are trying to measure. On PropickAI match pages the model and market context are shown alongside the odds, which is usually enough for the disciplined v1 prompt.
Do these prompts work for lower leagues?
The v1 market-anchored prompt travels well, because the price carries most of the information. The xG-driven v2 needs data that often does not exist below the top divisions — in that case either supply what you have or stick to v1.
How should the prompt treat the draw?
As a real outcome with a real probability, not as a rounding error. Force three numbers that sum to 100% and compare the draw against the de-vigged market price. In tight, low-scoring leagues the draw is frequently the fairest price on the coupon, and a model that never picks it is telling you about its bias rather than about the match.
Should I run one prompt across every league, or write one per competition?
Start with one prompt and one league so the comparison is clean, then widen. League context (typical goals, home advantage, refereeing) shifts the reference points enough that a prompt tuned on the Premier League will misprice a low-scoring second division — which is exactly the kind of drift the AI Lab dashboard makes visible.
Prompts for the rest of the board
Find out which football prompt actually wins
Start the 5-day AI Lab trial without a card. Bring your own AI key, run v1 and v2 on today's football card, and let the dashboard settle it on a virtual $10,000 bank.
AI Lab $19/mo after the trial · cancel anytime · Full Access $49/mo for all sports and the official 7-AI consensus