Why this matters to you
When a traveler asks an AI for a hotel recommendation, the AI reads every attribute on your listing — price, rating, location, amenities — and silently decides which ones are worth paying attention to.
If you don’t know which signals actually drive the decision, you end up optimizing the wrong things. Spending weeks getting your wifi labeled “premium” buys almost nothing. Fixing one missing landmark in your location description can reshape which hotels you even compete against.
We ran the experiment that answers this directly: remove one feature at a time, see how often the AI’s pick changes, and rank features by how much each one matters.
Location context reshaped the entire competitive set. Wifi didn’t move the dial.
Key findings at a glance
01
Location context is the dominant signal
Removing the ‘nearby places’ field flipped the AI’s pick in 45% of trials and collapsed rank correlation to 0.28 — a fundamental reshuffle.
02
Price stabilizes more than it selects
price_per_night is the only feature whose removal actually reduces position instability (STSR from 0.368 → 0.277). The model uses price as an ordering axis, not a deal-breaker.
03
Amenities are tiebreakers, not drivers
No single amenity crosses 23% shift rate. Most have near-zero ability to overturn a unanimous pick (odds ratio < 0.15). Individually weak, collectively present.
04
Review count behaves like an amenity — not a rating
Despite being a credibility signal humans weigh heavily, review_count patterns with wifi and parking (20% shift, ρ = 0.71) — not with overall_rating.
What this means for your hotel
Your effort should follow the AI’s attention, not your intuition. Three of the 13 attributes — nearby places, overall rating, and location rating — account for roughly twice the selection influence of the other ten combined. Everything below star rating is a tiebreaker.
The practical consequence: if your listing has a sparse location description, you’re competing against a smaller, worse version of yourself. If the data the AI reads doesn’t name the landmarks, transit stops, and walking times around you, the AI fills that gap with someone else’s hotel.
What to do about it
1. Treat your location data as the primary battleground.
The “nearby places” signal (landmarks, transit, commercial districts) was the single strongest lever in our test — a 45% shift rate, 3× the weakest feature. Make sure your website, your OTA profiles, and your schema.org markup name every significant landmark within walking or short-driving distance, with actual distances or times.
2. Keep pricing visible and stable.
Price doesn’t dominate which hotel gets picked, but its presence stabilizes the decision. A hotel whose rate is missing or ambiguous forces the AI to reason with fewer anchors — and that’s when position bias and brand bias get louder. Visible, consistent pricing is worth more than chasing a lower nightly rate by a few dollars.
3. Don’t over-invest in amenity noise.
Getting “free wifi” or “parking” surfaced cleanly on your listing is table stakes. But chasing long amenity lists and hoping the AI rewards comprehensiveness is a poor use of effort. The data says amenities sway already-uncertain picks — they don’t create them.
The evidence
Finding 1 — Location context dominates
“Shift rate” is the percentage of trials in which removing a single feature changed which hotel the AI picked. A higher shift rate means the feature carried more decision weight.
Feature shift rate — how often removing a feature changed the pick
GPT-5.4 · 61 Google Hotels query sets · 16 permutations · 976 trials per feature
All 13 knockouts clear Fisher’s exact p < 0.001 with Cohen’s h > 0.8 — meaning every feature has *some* measurable effect. The ranking above is the honest order of that effect.
0.28
Spearman rank correlation after removing nearby_places. Closer to 1.0 means the knockout only nudged the top pick. A value of 0.28 means the full ranking collapsed — the model considered a fundamentally different set of hotels once location context was gone.
Finding 2 — Price is a stabilizer, not a selector
Most features change which hotel gets picked without changing whether the AI’s pick is consistent across orderings. Price is the exception. When we removed price, the model’s position instability (STSR — how often the pick flips when you reshuffle the list) dropped sharply.
STSR change after removing each feature — negative is more stable
Baseline STSR = 0.368 (lower is more stable); removing price is the strongest stabilizer
No single knockout clears Mann-Whitney U p < 0.05, so the price stabilization is directional, not conclusive at current sample size. But price is the only feature that consistently reduces permutation instability — it anchors the decision rather than driving it.
Why this matters operationally
If the AI is more consistent when price is visible, then missing prices amplify all the other biases — positional, brand, and otherwise. Keeping accurate nightly rates in front of the model is a cheap way to make everything else more stable.
Finding 3 — Amenities are weak individually, collective as tiebreakers
We also measured “conditional shift”: splitting trials into unanimous sets (where the baseline pick was the same across all 16 orderings) and split sets (where orderings disagreed). If a feature is a real decision driver, removing it should overturn even unanimous picks. If it’s a tiebreaker, it only matters when the baseline is already wobbling.
Conditional shift — unanimous vs split baseline. Lower odds ratio (OR) = weaker effect on certain picks.
| Feature | Unanimous shift | Split shift | Odds ratio |
|---|---|---|---|
| nearby_places | 29.2% | 48.9% | 0.43 |
| kitchen_kitchenette | 16.7% | 24.4% | 0.62 |
| business_center | 14.6% | 23.2% | 0.56 |
| location_rating | 12.0% | 33.7% | 0.27 |
| overall_rating | 9.9% | 38.4% | 0.18 |
| star_rating | 9.4% | 24.0% | 0.33 |
| price_per_night | 8.3% | 33.7% | 0.18 |
| review_count | 4.7% | 24.1% | 0.15 |
| restaurant_onsite | 3.1% | 20.7% | 0.12 |
| parking | 2.6% | 19.9% | 0.11 |
| breakfast_included | 2.1% | 19.1% | 0.09 |
| gym_fitness | 2.1% | 20.7% | 0.08 |
| wifi | 2.1% | 19.4% | 0.09 |
The split is clean: nearby_places, kitchen, and business_center are the only features that can overturn unanimous picks at meaningful rates (OR ≥ 0.43). Every other amenity lives in the 0.08–0.15 range — they only surface when the AI was already undecided.
Finding 4 — Three tiers of feature influence
Aggregating shift rate, rank correlation, and odds ratio gives a clean three-tier classification of how GPT-5.4 actually uses hotel data.
Feature classification — three tiers by how the model uses each signal
| Tier | Features | Shift rate | Rank ρ | Role |
|---|---|---|---|---|
| **Primary drivers** | nearby_places, overall_rating | 33–45% | 0.28–0.54 | Reshape which hotels compete |
| **Structural anchors** | price_per_night, location_rating, star_rating | 21–29% | 0.63–0.70 | Set the decision frame without reshuffling |
| **Tiebreakers** | wifi, parking, breakfast, gym, restaurant, review_count, business_center, kitchen | 16–23% | 0.70–0.82 | Sway already-uncertain picks |
Two features do the thinking. Three set the frame. Eight break ties.
Statistical tests
Shift-rate significance uses Fisher’s exact test (all 13 features p < 0.001). Rank-correlation drops use Wilcoxon signed-rank (all p < 0.001). STSR-change uses Mann-Whitney U (no knockout clears p < 0.05, so the STSR effects are directional). Unanimous-vs-split odds ratios test whether a feature can overturn already-certain picks.
Frequently asked questions
No — location coordinates are a separate field. ‘nearby_places’ is the list of named landmarks, transit stops, and points of interest in the hotel’s description. The AI cares more about which landmarks appear with walking or driving times than it does about the raw address.
Every amenity is better than none, but past a basic set the returns are flat. Focus your differentiation effort on location context and quality signals (ratings, review count), not amenity padding.
Kitchens and business centers are the only amenities that meaningfully overturn unanimous picks (OR 0.62 and 0.56). Our interpretation: they signal trip type. A kitchen pivots the recommendation toward longer stays; a business center toward work travel. They’re not generic niceties — they change which hotel category fits the request.
This specific knockout experiment was run on GPT-5.4. Our cross-model chapter shows different models weight features somewhat differently — Gemini and Claude Opus have their own patterns. But the broad ranking (location context > ratings > price > amenities) is consistent with behavior we’ve seen across models.
Traditional SEO ranks pages in a list; AI picks one answer. In SEO, amenity keywords can meaningfully lift ranking. In AI selection, amenities are mostly inert — the model is picking a *hotel*, not matching keywords.
How we ran the experiment
13,664
Total trials
61
Query sets
16
Permutations per set
13
Features knocked out
We used a feature knockout (ablation) design. For each of 13 raw hotel features, we removed that feature from every hotel in every query set and re-ran the experiment, comparing selections to a full-feature baseline.
Query sets were drawn from real Google Hotels results — 10 hotels per set, real names, real ratings, real prices. Each set was shuffled into 16 different orderings to separate feature-knockout effects from positional effects. Total: 976 trials per feature (61 query sets × 16 permutations).
We measured four signals per feature: (1) shift rate — fraction of trials where the pick changed vs baseline, (2) STSR change — whether removing the feature made the model’s pick more or less stable across orderings, (3) Spearman rank correlation — whether the full ranking survived the knockout, and (4) conditional shift with unanimous-vs-split odds ratio — whether the feature could overturn already-certain picks.
Limitations. All trials are GPT-5.4, temperature 0, no reasoning tokens. Features were knocked out one at a time — we did not test combined knockouts. The 13 features reflect what Google Hotels exposes, which may not be exhaustive of what hotel websites provide.
Want to know which features are missing from your listing?
Huxo’s AI Visibility Report audits your hotel’s data the way GPT-5.4, Gemini, and Claude read it — and tells you which fields to fix first to show up in the recommendation.