Been using TR since the beginning (2013). I shied away from it in 23 & 24 just because I needed a change. I started using it again this fall. I had just come from taking 6 weeks off to clear my mind and let some nagging things heal. I started with a ramp test and it was low, 220. Started with Mid Volume base 2. Hit every workout. AI detected 227. Thought it was low so I did an FTP test on Zwift, the banners at every minute really seem to push me. Tested at 253. Manually set my TR FTP to 253 for Mid Volume base 3, hit every workout with no issues. AI detection gave me 1 watt to 254. Did another Zwift Ramp test and tested 273. Seems in my case, AI detection is off. I would think though if you are able to complete every workout with no issues in a 4 week block you would at least get a 2-3% gain?
Verification ride time 
Go knock out 2 Ă— 20 at a wattage of you choosing. If you do it at an RPE of 8, you have a good number.
ETA: The ramp test tells you less than nothing
Very true…I hated the ramp test when it first came out. I was a fan of the 2x8 minute tests. I should try going back to those.
I have always preferred doing an actual ftp test over guessing with a ramp test or ai.
Go ride at what you think your ftp is, you will know very quickly if it is accurate or not 
. Plus this way you get an accurate tte. That is also quite helpful in training.
Finishing a 4 week block doesn’t necessarily mean you will see any gain in FTP, especially not 2-3%. You’re riding the renewb gains right now and that can skew things a bit.
Many people say that they test high on a ramp test. If you are able to do it and work at that level then it may be fine for you. But this is also going to depend on what kind of workouts you are doing and in the TR world how deep into the workout levels you are.
For me a ramp test is going to be better than a 2x8 and a 20 minute test is going to be better than those two. The other option is a long format Kolie Moore test. That last one is the only one I care to do now. If you get your FTP dialed in then AI detect can be pretty good. Then obviously something like a 2x20 for confirmation mentioned above.
Since you are on a come back depending on the history of data that TR has for you AI may be less accurate until you build up some rides.
The other thing to consider is that when you bump your FTP your progress levels will drop, there are benefits to just ignoring FTP bumps for a bit and training through. I ran but ignored an AI FTP bump recently for this very reason.
It’s worked well for me in terms of keeping my training on track. Is it an accurate FTP? I don’t know but it works for my training which is all I care about.
I’d suggest to add context, consistency and results for a better perspective.
Using the optic of results then TrainerRoad AI works.
Bring in consistency and the ability to continue to train without burn out is another consideration. Burn out, either physically or mentally is a limiter and the ability to train consistently and avoid injury or illness provides the optimised route
Of course we are all unique and younger riders have a higher tolerance to training load and require less recovery.
For me TrainerRoad is the best system for getting me fitter within a certain time frame than any other platform.
I’ve found them all to be different so trusting one source’s AI over another’s is unrealistic for me (and I don’t just mean the TR one). I’ve switched to the Kolie Moore test and find it to give me a much more accurate number that gives me confidence in my training. Also, it’s a great workout!
Ive found the GP Lama version of the MAP Test to be the most accurate. Plus it gives me a good (in my case) number for my VO2max as well as my FTP and MAP.
Nah. The 2 × 8s aren’t much better. Better just to warm up, do 5 minutes at sweetspot, ramp up to what feels like FTP and keep going until your RPE suddenly shoots up. That way you’ll have an FTP number and a TTE (on the assumption that your effort lasted at least 35 minutes or so)
I’m guessing you’re luckily right in the middle of the bell curve. 
“Below is how you can perform this test yourself using Zwift to calculate your MAP and an estimate of your FTP. This is an excellent task to complete prior to doing your 20 minute FTP test.”
As above, the article itself suggests only using the resulting estimated FTP number as an input to another test to further estimate your FTP.
Side note: MAP nunbers derived from ranp/step tests are very sensitive to the starting wattage, step length and step size of the protocol.
I only use AIFTP. Haven’t tested since it was introduced.
I do think AIFTP and the other “AI” features that TR makes claims about function better on data from TR workouts than from other workouts / data
As most of my riding (at this time of year in particular) is on TR, coupled with me choosing to ride without any data sources outside, I am happy to accept AIFTP.
FWIW, in RPE / HR / fatigue response compared to pre-AIFTP days, I think it is pretty accurate (for me).
The more data I relied upon from outside TR, the more likely I think I would be to need to test - if only to verify a couple of times that AIFTP was keeping up with the non TR data.
It can only be as good as the data going in, so if you’re doing lots of maximal efforts it’s probably pretty close. As an aside, finishing workouts in a base block doesn’t mean FTP is set right - if everything is too hard by a factor of 10% but the workouts are tempo/sweet spot you’ll still be able to get through them. When I’ve made this mistake in the past the issue shows horribly when it’s time to start proper vo2 max and I can barely make it half way through the first interval!
Choose a system and stick to it. Same discussion
go round about powermeters and/or trainers.
Consistency is king, if you always measure the same way, you will see how you will perform relative to that benchmark. That’s all you need to know.
But purely anekdotal, the ramptest, 20 minute test and the pure and the “go for an hour” test, all were somewhere around where the ai detection puts me. The ramptest a little bit higher than the others, due to the anaerobic component. But all within acceptable margins.
If things are too hard or too easy, you can always adjust intensity. The path to improvement is 95% a highway that’s the same for everyone, the last 5% ask for a tweak here and there. That’s where you do you.
I think it’s as accurate as it needs to be to set training.
Threshold workouts with a PL of between 5 or 6 and are decent “tests”.
That said I think I’m lucky in that Kolie Moore, ramp test and now FT detection all line up pretty well for me.
I got feisty and I lost to AI FTP detection recently.
I’ve been using the feature since it was released and was feeling special and unique after the off season. Detected my FTP, said “ no way, man! “ and proceeded to get almost the exact same result with the ramp. Been on TR with a power meter for years.
If AI detection didn’t work as advertised then we would hear about it on these forums… we love complaining 
AIFTP is pretty similar to the ramp test for me. I only use AI now and it seems to consistently set training targets accurately.
That being said, my aerobic system is garbage so my SS and threshold PLs are always lower than my
VO2 and there’s no way I could hold my AIFTP for an hour.
What is the Zwift FTP test like?
I just hammer out the 20 minute test, and usually TR’s FTP detection is close. Besides that, intervals.icu will update you threshold, if you are signed up for training peaks you will get an update there too, and if you run a Garmin head unit, it will also update your ftp.
I’ve considered doing the KM ftp test, but I’ve resisted because I like the bonehead simplicity of the 20’ test and it’s what I have done for the longest time. Why change it? Afterall, it’s only used in setting training zones, and the zone itself are not exact either.
Where does TP do an update for your FTP? I’ve never seen that.