AI Recommends a Different Hotel Every Time You Ask. That's the Point.

So here's a fun experiment. Ask ChatGPT for the best boutique hotel in, say, Portland. Write down the answer. Close the window. Open a new one. Ask the exact same question. There's a 45% chance you get a different hotel at the top. Not a different order... a different hotel entirely. Only about 60% of the properties even show up the second time. Now go ask Gemini the same question. Then Copilot. Then Google's AI Overview. According to an ongoing study from Kollective covering 13,500 AI-generated answers and roughly 9,600 distinct properties across 100 destinations, those platforms agree on the single top hotel about 4% of the time. Four percent. Seven out of ten hotel names appear on just one platform and nowhere else.

Let that sink in for a second if you're an operator who just paid a vendor to "optimize your AI visibility."

Look, I get the instinct. For 20 years, the game was Google. You had a ranking. You could track it. You could hire someone to improve it. SEO was knowable... tedious and annoying, sure, but knowable. There was a list, your hotel was on it somewhere, and the job was to move up. So when AI search started replacing traditional search (and it is... the trajectory is obvious), the natural reaction was to treat it the same way. Find the ranking, optimize for it, measure progress. Except there is no ranking. These systems are probabilistic. The output changes based on phrasing, session history, prior queries, model updates, and what amounts to a weighted dice roll inside the architecture. You're not position 7 trying to get to position 3. You're a set of data points that may or may not get surfaced depending on variables you can't see and the model itself can't fully explain.

This matters because vendors are already selling "AI ranking optimization" to hotels the way they sold SEO packages in 2010. I talked to a hotel group last month that was quoted $2,800 a month for an "AI visibility platform" that essentially monitored how often ChatGPT mentioned their properties. That's it. Monitoring a number that changes 45% of the time between identical queries. The vendor couldn't explain the mechanism for improvement because there isn't one in the traditional sense. What you can actually do... and this is the unsexy part that doesn't sell $2,800/month contracts... is make sure your property data is rich, structured, and machine-readable. Semantic schema. Detailed, accurate, consistent information about your hotel that AI systems can parse without hallucinating. That's not optimization. That's hygiene. It's the equivalent of making sure your phone number is correct on Google Maps. Essential, but nobody should be charging you a premium subscription for it.

The Kollective study also surfaces something interesting that deserves attention: there's a "winner-take-most" pattern where certain properties consistently appear across platforms while the rest rotate in and out. That's not ranking. That's data density. The hotels showing up everywhere tend to have the richest online footprint... reviews, structured data, content depth, OTA presence, media mentions. They're not gaming an algorithm. They're just more thoroughly documented than the competition. Which, if you think about it, is the same reason some hotels always ranked well on TripAdvisor. It wasn't magic. It was volume and consistency of guest-generated content.

Here's what actually changes your Monday morning. Stop thinking about AI ranking and start thinking about AI readiness. Is your property data structured correctly across every platform? Are your room descriptions, amenity lists, and location details consistent and detailed enough for a system to confidently recommend you? Because the study's real finding isn't that AI recommendations are broken. It's that they're working exactly as designed... pulling from available data and making probabilistic selections. If your data is thin, you're invisible. Not ranked low. Invisible. There's a difference, and the difference is that no amount of "optimization" fixes a data problem. You fix a data problem by actually having better data.