What A.I. Gets Wrong About Bike Shopping, According to Someone Who Tests Them for a Living - bicycling.com

9 min read

Artificial intelligence is coming for bike shopping.

Not in the dramatic sci-fi way, where a robot rolls into your garage and replaces your favorite torque wrench with a subscription plan. I mean in the ordinary, already-happening way: Search engines now serve A.I. answers, browsers serve buying suggestions, and riders are asking ChatGPT, Gemini, Perplexity, and whatever comes next to help choose their next bike.

That is not automatically bad. Bike shopping is now complicated enough to make a person want to lie down in a dark room. Aero versus weight. Race versus endurance. Tire clearance. Wheel depth. Integrated cockpits. Bottom bracket standards. Claimed weights. Claimed aero savings. Claimed everything.

So I tried an experiment. I asked artificial intelligence a question I’ve spent most of my career trying to answer with my legs, hands, eyes, backside, and occasional poor judgment.

What is the best road bike?

I gave it some guidelines to make the search more in line with my preferences. The bike had to be aerodynamic, light, stiff, sharp-handling, comfortable, and able to clear modern, fast, wider road tires of at least 35mm. It had to make sense for an amateur rider on real roads, riding around 30 kilometers per hour, not a WorldTour rider humming along at 50 kilometers per hour with a mechanic following in a team car.

Then I pushed the search wider. I asked it to look beyond the traditional cycling media and consider Weight Weenies, Reddit, YouTube, German lab testing, independent wind tunnel results, foreign-language forums, social media, Chinese and Taiwanese sources, and smaller brands. I wanted the obvious giants and the less obvious contenders: Specialized, Trek, Giant, Canyon, Cervélo, ENVE, Allied, Bridge, FiftyOne, Parlee, Fairlight, Moots, Litespeed, Mason, Argonaut, BlackHeart, X-Lab, Tavelo, SEKA, Winspace, and more.

The result was messy.

Also, at first, extremely conventional.

A.I. didn't initially identify the bikes that best fit my parameters. It began with the popular race bikes that are widely discussed online. It only considered the bike that ultimately became its top choice after I explicitly asked it to do so.

The main lesson is that A.I. can help with bike shopping, but only if you understand its capabilities and limits, and know when to stop relying on it and trust your own judgment.

Below are the drawbacks of using A.I. to shop for a bike, how you can avoid them, and ways A.I. can assist you in making a better new bike choice.

A.I. Defaults to the Easy Answer

The first thing I learned is that A.I. does not automatically understand where road bikes are headed.

Left mostly alone, it did what a lot of people do when asked to name the best road bike: It drifted toward the easy answers. The biggest brands. The most visible models. The halo bikes that dominate launch coverage, pro racing, search results, YouTube thumbnails, and forum arguments.

That is not useless. The Specialized Tarmac SL8, Trek Madone, Canyon Aeroad, Giant Propel, Cervélo S5, Cannondale SuperSix Evo, and bikes like them are famous for a reason. They are exceptional machines.

But “easy answer” is not the same as “best answer.”

A.I. did not surface up-and-coming Asian brands like X-Lab, Tavelo, SEKA, or Winspace. It did not immediately dig into smaller builders with compelling modern road-bike ideas, like Parlee, FiftyOne, Bridge, or BlackHeart. And it did not recommend the Canyon Endurace CFR, the bike that ultimately made the strongest case as the one that best answered my question, until I prompted it to consider that specific model.

That matters because A.I. can make popularity look like authority.

It was not malicious. It was not stupid, exactly. It was doing what the internet trains these tools to do: find the answers surrounded by the most noise.

The more interesting answers appeared only after I pushed it toward the way many amateur riders actually ride: fast pavement, bad pavement, long days, dirt connectors, and speeds closer to 30 km/h than 50.

A.I. did not discover the trend. I had to point it at the trend.

That is probably the most important lesson for anyone using A.I. to shop for a bike: The answer is only as good as the question. Ask a generic question and you will get a generic answer. Ask a sharper question, built around how and where you actually ride, and A.I. gets more useful.

But the expertise has to come from somewhere. And that is from humans.

Your Prompts Can Bully the Result

Every time I gave it a new bike to consider, A.I. tended to take that bike more seriously. Every time I reminded the bot that I was only interested in bikes that fit 35mm and wider tires, tire clearance became more important. Every time I asked about smaller brands, the list shifted toward smaller brands.

A.I. is programmed to give you responses that it thinks you will like. But that means a rider can accidentally prompt their way into the wrong bike.

Ask for “the fastest road bike,” and A.I. will probably hand you an aero race bike that might be miserable on your roads.

Ask for “the most comfortable bike,” and it might drift toward endurance bikes that feel too muted if you like sharp handling.

Ask for “a road bike with big tire clearance,” and suddenly you start seeing gravel bikes sneaking into the room and pretending they were invited.

The better question is not, “What is the best bike?”

The better question is, “What is the best bike for the way I actually ride?”

That means being honest about your roads, your speed, your flexibility, your fit, your tire preferences, your maintenance tolerance, your budget, and your vanity.

Especially vanity.

We all have it. Pretending otherwise is how people end up on bikes they respect but do not love.

A.I. Believes Marketing Copy Too Easily

Some bike companies are very good at describing their bikes.

Sometimes they are describing real things. Sometimes they are describing aspirations. Sometimes they are describing the bike they wish they had built. And sometimes they are describing the emotional state they hope you enter before clicking add to cart.

A.I. has trouble telling the difference.

That became obvious when it evaluated bikes with especially clean product stories: light, aero, comfortable, fast, race-inspired, endurance-ready, performance-focused, compliance-optimized, wind-tunnel-developed, real-world-tuned.

You have read those words before.

So have the bots.

The problem is that those claims are made in a vacuum. A bike can be “aero” and still not feel especially fast under you. A bike can be “comfortable” and still feel dull. A bike can be “race-inspired” in the same way a crossover SUV is “rally-inspired.”

During an early round of recommendations, it got too excited about a number: tire clearance. Bigger number; better bike, the bot thought. And that is how the recommendations for a road bike ended up with gravel bikes barging into the argument.

But a road bike that clears 35mm tires and a gravel bike that clears 45mm tires are not separated by one centimeter of rubber alone. They are separated by geometry, handling, stiffness targets, rider position, wheelbase, stack, reach, trail, and purpose.

Manufacturer weights, aero gains, stiffness claims, and comfort claims should all be treated as claimed unless independently verified. And even when the numbers are real, they still need interpretation.

One watt in a wind tunnel may matter less to you than the ability to run the right tire at the right pressure. A slightly heavier frame may be irrelevant if the bike rides better, fits better, or saves you from a proprietary cockpit nightmare. A bike that looks slower on paper may be faster for you because you can stay comfortable, confident, and aero on it for three hours instead of 30 minutes.

A.I. likes clean answers because clean answers are easy to produce.

Bikes are not clean answers.

Bikes are compromises.

It Cannot Feel the Bike

A.I. can tell you that, on paper, a bike is light, aero, stiff, and has clearance for 35mm tires.

It cannot tell you whether that bike will make you want to stay out for an extra hour.

A great bike is not always the right bike. I have tested bikes that were objectively excellent and emotionally dead. I have tested bikes that made little sense on paper and still made me invent errands just so I could ride them again.

This is where human insight matters. A.I. does not know how a front end loads in a fast corner. It does not know whether a bike feels dead in a sprint. It does not know whether a compliant bike feels suitably damped or mushy. It does not know if the cockpit shape annoys your hands after an hour. It does not know if a bike’s story makes you feel something.

And yes, that last part matters.

No one rides a bike blind. The badge, the paint, the history, the shop that sells it, the friend who loves it, the pro who won on it, the builder who crafted it, the shape of the fork crown, the way it looks leaning against your garage wall. All of that affects how we experience a bike.

That does not make us fools.

It makes us human.

Use A.I. for the Shortlist, Not the Decision

The best way to use A.I. for bike shopping is not to ask it to crown a winner. That gives it too much unearned authority.

Use it to build a shortlist. Use it to surface models you forgot. Use it to compare claimed tire clearances, weights, geometries, and intended use. Use it to ask, “What am I missing?”

Then stop.

Go read real reviews. Look for actual ride impressions, not just launch-day spec regurgitation. Measure your current bike and double-check the geometry charts. Verify the claimed weight. Confirm whether the bike’s maximum tire clearance is realistic with the tires and rims you plan to use. Think about replacement parts. Think about traveling with the bike (if that is something you do). Think about whether you can live with the cockpit. Consider whether the bike has complex integration and your local shop’s labor rates. Talk to the dealers and owners of the bike you’re considering. Ride the bike if you can.

And listen to people who have actually ridden a lot of bikes.

Not because we are magical bike shamans, but because testing teaches you where the spec sheet lies by omission.

If fit is even slightly in doubt, start with Bicycling’s Guide to the Perfect Bike Fit. A great bike in the wrong position is still the wrong bike.

So, What Did the Experiment Find?

After all the arguing, correcting, and re-prompting, the bot decided the best answer for my specific question was the Canyon Endurace CFR.

But that is not the first, or even the fifth, bike that it said best answered my question.

A.I. totally excluded the Canyon for quite a while and only considered it after I prompted it. Only then did it start comparing the Endurace CFR against the specific criteria: road-focused handling, aerodynamic efficiency, light weight, comfort, and clearance for wide and fast modern road slicks.

That is important. The eventual winner was not something A.I. immediately and brilliantly unearthed from the depths of the internet. It was a bike that the human had to bring into the room.

Once it was there, though, the Canyon made the fewest excuses. It was road-first, light enough, aero enough, aggressive enough, comfortable enough, and cleared modern road tires without wandering into gravel-bike cosplay.

I will note that I have not ridden this bike, so I don’t yet know whether how it looks on paper matches up with how it performs in the real world.

On some level, the Canyon was not the answer I wanted. I wanted something more romantic. Something from a small builder, maybe. Something titanium, perhaps. Something with more story and less direct-to-consumer efficiency.

But the exact winner is not the point.

The more useful discovery was this: A.I. did not replace my real-world experience, an honest ranking of my needs and shortcomings, and the need to use my own brain to know what questions the bot had missed.

The machine gathered. I questioned. The machine ranked. I argued. The machine overreached. I corrected it. The machine missed the eventual winner. I added it. Then I judged the result.

That is probably the healthiest way to use A.I. when shopping for a bike: not as the expert, but as a fast, tireless assistant who can gather information but still needs a human who knows when a gravel bike has wandered into a road-bike conversation.

A.I. can help you get to better questions faster.

It might help you find bikes you might not have considered.

It can help you compare the claims.

But it cannot replace judgment. It cannot replace experience. It cannot replace the feeling of clicking into a bike, standing on the pedals, and knowing within the first few minutes whether the thing underneath you is alive or merely expensive.

So use A.I. if you want.

But do not let it buy your next bike for you.

That part still requires legs, hands, eyes, a backside, and a human brain.