Athletica.ai Under The Bonnet: Science-Led Or Marketing-Led?

If you are searching for an athletica ai review, you have probably already read the landing page. It says science, it says polarised, it says adaptive. What it doesn’t say is whether the sessions that land in your calendar match any of that. This post deals with that gap. I trace the claim back to its literature, then show you how to check the delivered sessions against it with your own files.

A caveat first. The numbers in the worked example below come from a representative four-week block I’ve built to show the method. They are not a lab result and not a verdict on every athlete’s plan. Run the audit on your own export before you believe me or Athletica.

What “polarised” actually means in the source papers

Most of the cycling-facing polarised talk traces to Stephen Seiler’s work. His observational research (Seiler & Kjerland, 2006, on elite Norwegian endurance athletes) found roughly 75-80% of sessions at low intensity, a small slice at high intensity, and very little in the middle. The key detail is the three-zone model, anchored to physiological thresholds:

• Zone 1: below the first lactate or ventilatory threshold (LT1/VT1).
• Zone 2: between the two thresholds.
• Zone 3: above the second threshold (LT2/VT2).

Two things get lost when this is turned into product copy. First, “80/20” is a rough description of session counts in elite athletes, not a prescription proven to beat everything else. Controlled trials are mixed. Stöggl & Sperlich (2014) found polarised training beat threshold, high-volume and high-intensity-interval blocks in well-trained athletes over nine weeks. Others, like Neal et al. (2013) in cyclists, found gains but with small samples. Second, the middle zone in Seiler’s model is a physiological zone. It is not the same as the Coggan “tempo” or “sweet spot” bands you see in your head unit, because those are percentages of FTP, not thresholds.

So when a platform says it is polarised, there are three testable questions:

1. How does it map your power to three zones?
2. What fraction of time lands in each?
3. Does the middle zone stay small, or does it quietly grow?

Tracing Athletica’s claim

Athletica’s public material describes training built on the polarised model, with intensity distribution adapting to your block and your level. That is a reasonable position to hold. The honest reading of the literature is that it supports “mostly easy, some very hard, limited middle” for athletes with enough hours. Nothing in it supports applying that rigidly to someone on six hours a week, where pyramidal distributions often show up in the data and perform fine.

What I’d want from a platform making this claim is simple. Show the zone mapping. Show the planned distribution per block. Let me export the plan so I can count. If any of those are opaque, “science-led” is an assertion, not a method.

The audit method

You need a plan export (the planned workouts as .zwo, .fit or calendar entries) and your thresholds. In intervals.icu you can sync planned workouts and read the zone times directly. In TrainingPeaks-style tools you’ll get it from workout structure.

Step one is the mapping. Without a lab test, use a practical proxy:

The cut points are debatable. If your VT1 is closer to 75% of FTP, shift the line. Whatever you pick, pick it before you look at the plan, so you aren’t fitting the boundary to the result.

Step two is to sum planned seconds per zone across the block, not per session. A session with a 20-minute threshold block inside a 90-minute ride is mostly Z1 by time but still contributes to the middle. This is where “polarised on a per-session basis” claims fall apart.

Step three is Treff’s polarisation index (Treff et al., 2019): PI = log10((f1 / f2) × f3 × 100), where f1 to f3 are the fractions of time in each zone. A distribution counts as polarised only if f1 > f3 > f2 and PI is above 2.0.

Worked example: a four-week block

Take a 10-hour-a-week athlete with FTP 285 W. The block below is the kind of output you’d expect from a “sweet spot plus VO2” plan, built to illustrate the arithmetic.

Week  Total   Z1 (<228W)   Z2 (228-299W)   Z3 (>299W)
1     10:00   7:00         2:25            0:35
2     10:30   7:20         2:35            0:35
3     9:45    6:50         2:15            0:40
4     7:00    5:20         1:20            0:20
Block 37:15   26:30        8:35            2:10
Share         71.1%        23.0%           5.8%

Now the maths. f1 = 0.711, f2 = 0.230, f3 = 0.058.

PI = log10((0.711 / 0.230) × 0.058 × 100) = log10(3.09 × 5.8) = log10(17.9) ≈ 1.25.

That is pyramidal, and not close to polarised. Z2 time is nearly four times the high-intensity time. Marketing says 80/20; the arithmetic says 71/23/6.

Compare with what a genuinely polarised version of the same 37 hours looks like: 77% / 8% / 15%.

PI = log10((0.77 / 0.08) × 0.15 × 100) = log10(9.63 × 15) = log10(144.4) ≈ 2.16.

The gap between 1.25 and 2.16 is the whole story. Same athlete, same hours, completely different training.

Where the drift comes from

If you run this and find a plan like the first one, the usual causes are boring:

• Zone labels. Platforms with their own five- or seven-zone scheme often put “endurance” up to 75-80% of FTP and then label anything from 80% as “tempo”, so a polarised intent still arrives with a thick middle.
• Threshold work doing double duty. Two 20-minute efforts at 95% are marketed as “hard” days, but under the three-zone model they sit in Z2, not Z3.
• Adaptive adjustment. When the system trims a failed VO2 session, it tends to swap in something moderate. Over a block, that nudges time into the middle.
• Time-crunched weeks. At 6 to 7 hours, a polarised split gives you perhaps 45 minutes of true Z3 in a week. Many tools can’t resist adding “useful” middle work.

What I’d conclude, and what I wouldn’t

My position: treat Athletica’s polarised label as a hypothesis until your own export confirms it. The literature it invokes is real, but the literature is narrower than the copy. Seiler describes what elite athletes do at high volume. He did not hand a template to a masters rider with a nine-to-five.

That doesn’t make the platform bad. A pyramidal plan with a PI of 1.3 can be a perfectly good plan for a 7-hour athlete. The problem is only the label. If the delivered distribution doesn’t match the stated model, you can’t attribute your progress, or your stall, to the model.

For how this compares with the other tools I’ve put through the same audit, the AI Coaching Platforms, Tested pillar has the side-by-side distributions.

Run it yourself in ten minutes

1. Export four weeks of planned workouts.
2. Set your three-zone cut points before you look.
3. Sum planned seconds per zone across the block.
4. Compute the PI. Anything under 2.0, or f2 above f3, is not polarised by Treff’s definition.
5. Repeat on completed rides. Planned and executed distributions often differ by 5 to 10 points in the middle zone, because people ride the easy days too hard.

Step five matters most. If your actual Z2 share is 25% while the plan said 8%, the platform isn’t the problem. You are, and no amount of AI will fix that without a power file to prove it.

Paste your block totals into a chat model and ask it to compute PI, but give it the formula yourself. Left to guess, it will quote “80/20” back at you and call it polarised.