AI for cycling performance data

Power-meter data (FTP, TSS, IF, NP), HR, and HRV across multiple seasons — usually fragmented across Strava, TrainingPeaks, and Garmin. Most peer-reviewed work on cycling sits in three buckets: mechanistic studies (small samples, tightly controlled), observational cohorts (large samples, noisy variables), and consumer-device validation papers (mixed quality, often vendor-funded). When you read AI-generated summaries on AI for cycling, treat the first two as signal and the third as buyer-beware. The 3-Layer method makes you triage these before they enter your personal ledger.

Consumer devices that surface a "Cycling" score almost always combine a small set of raw signals — accelerometry, optical heart rate, skin temperature, sometimes ECG — into a proprietary index. The score is opinionated, the raw stream is not. The Ledger layer of the method exports the raw stream so AI can analyze the underlying variables instead of the marketing score. That is where most insight lives.

Cross-validation studies (Stanford, ETH Zürich, and several EU centres in 2023–2025) consistently show that wearables are most reliable for trend direction and least reliable for absolute values — especially night-to-night cycling. Use the data the way it is actually accurate: deltas over weeks, not single-night verdicts. AI is well-suited to this kind of rolling-window analysis; humans staring at one number are not.

FTP tests come and go. Most riders never read why their best season was their best. The most under-discussed confounders are time-of-month variation, recent travel, alcohol with a 48–72 hour tail, ambient temperature, and any acute infection — all of which shift baseline values by more than most behaviour changes do. A good AI ledger tags these as covariates before drawing conclusions; a bad one quietly attributes the swing to whatever supplement you started that week.

Good evidence on cycling: pre-registered protocols, declared funding, raw data available, effect sizes reported with confidence intervals, replication in an independent cohort. Hype: single n-of-1 anecdotes generalised on social media, supplement-funded reviews, AI summaries that cite nothing. Have sourced AI summarise current evidence on polarized vs threshold training, recovery, and durability. Asking AI to mark every claim with "primary study", "review", or "opinion" before you act on it is one of the most useful prompts you can run.

Three shifts matter. First, long-context models can now read 60–90 days of your raw export in a single pass and find correlations no app dashboard surfaces. Second, sourced-search models (with citations) collapse the literature-review step from days to minutes — provided you verify the citations. Third, agentic workflows can run the same daily check-in you would otherwise skip. Build a focused 12-week block grounded in your own response to volume and intensity. The judgement layer — what to test, what to ignore, when to stop — is the part that stays with you.