Why this matters to you
Travelers are increasingly asking ChatGPT, Gemini, and other AI assistants to recommend hotels. These AI models don’t just search — they choose. They read your hotel’s data, compare it to competitors, and recommend one.
If you run an independent hotel, there’s a problem: the AI already has a preference, and it’s not based on your rooms, your ratings, or your reviews. It’s based on the brand name — and branded hotels win.
Unlike Google Ads, you can’t buy your way in. The fix requires understanding how the bias works — and that’s what we set out to measure.
AI doesn’t just search. It chooses — and it has a preference.
Key findings at a glance
01
AI picks branded hotels 60% of the time
Even when only 2 out of 5 options are branded. That’s a 20-point lift over the 40% ‘fair’ baseline — just from being a recognized name.
02
It’s the brand name itself — not hotel quality
When we hid the brand, branded selection dropped by 11 percentage points. Same hotels, same data, same ratings. The drop proves the name does the work.
03
Brand bias is strongest when hotels are similar
In competitive markets with comparable attributes, the bias jumps from 11 to 16 points. AI uses brand as a tiebreaker — and independents lose the tie.
04
Fake brand names have zero effect
We tested 100 made-up brands. Zero impact. The AI only responds to brands it learned from training data — so brand equity in the AI era is web presence.
What this means for your hotel
AI is already in the loop between traveler and booking. When a guest asks “best hotel near the convention center,” the AI returns one answer — not a ranked list of ten. If you’re an independent hotel, you’re statistically less likely to be that answer, even when you deserve to be.
The good news: brand bias is a data problem, and data problems have solutions. You can’t change what the AI was trained on, but you can change what it reads about you going forward — and you can give it enough structured data about your property that brand familiarity stops being the easiest tiebreaker.
What to do about it
1. Flood your digital footprint.
Get your hotel mentioned on travel blogs, local guides, review platforms, and industry sites. Every high-quality mention trains the next generation of AI models to “know” your hotel. Branded hotels have decades of web mentions built up — your job is to close that gap faster than they widen it. Start by claiming and optimizing profiles on all major OTAs, review sites, and local directories.
2. Structure your unique selling points as data.
Add structured data (schema.org markup) to your website for your specific amenities, nearby landmarks, transit times, and unique features. Don’t just say “great location” — list exact landmarks and walking distances. Our separate feature-importance research suggests location context is one of the strongest factors AI uses to pick hotels. If that data is missing from your site, the AI can’t use it.
3. Differentiate where the AI can see it.
Identify 2–3 features that make your hotel genuinely different from nearby branded competitors — a rooftop bar, a specific conference center walkability, locally sourced breakfast — and make sure these are prominent and clearly presented in your site data. When AI sees a real difference, it’s less likely to fall back on the brand name as a tiebreaker.
The evidence
Finding 1 — Branded hotels win 60% of the time
We gave an AI model a list of 5 hotels — 2 branded and 3 independent — and asked it to pick the best one. If the AI had no brand preference, it would pick a branded hotel about 40% of the time (2 out of 5). Instead, it picked branded hotels 60% of the time.
+50%
lift in recommendation rate just from having a recognizable name. That’s like a coin that lands ‘branded’ 6 out of 10 times.
Branded hotel selection rate — real brands (Google Hotels cohort)
37 real Google Hotels query sets · 5 hotels per set (2 branded + 3 independent)
The 5.86-point gap between QAS-adjusted (54.41%) and observed (60.27%) is what quality can’t explain. The brand name adds something on top of quality. Binomial p < 0.001; QAS-adjusted p = 0.025.
Branded hotel selection rate — synthetic control (made-up brands)
100 synthetic query sets · same 2-branded + 3-independent structure
With fake brand names, the observed rate falls *below* the fair baseline (p = 0.024) — no brand lift at all. This is the control that isolates real-brand recognition as the driver.
But maybe branded hotels are just nicer?
A fair question. To rule that out, we recalculated the expected rate using what researchers call a Quality-Adjusted Selection (QAS) baseline — adjusting for each hotel’s actual attributes like price, star rating, review scores, and amenities. If branded hotels really were just objectively better, the observed rate should match the QAS-adjusted expectation. It didn’t. The QAS expected rate was 54.4%; branded hotels were still picked 60% of the time.
Statistical test · Finding 1
Tested across 37 real Google Hotels query sets. We used a binomial test — a standard method for checking whether an observed rate differs meaningfully from an expected one. Result: p < 0.001, meaning less than a 1-in-1,000 chance this came from randomness. The QAS-adjusted gap also reached significance (p = 0.025).
Finding 2 — Hiding the brand name drops branded selection by 11 points
We ran the exact same test twice. Same hotels. Same data. Same ordering. The only difference: in round two, we hid the brand name from the AI. When the AI could no longer see “Hilton” or “Holiday Inn,” it picked branded hotels 11 percentage points less often (49% instead of 60%).
That’s the causal effect of the name alone — nothing else changed between the two rounds. And in mixed sets where the AI had a genuine choice between branded and independent, the drop was even bigger: 16 points.
Brand visible vs. brand masked — all sets
Identical hotels, only the brand line differs · Google Hotels cohort (top) vs. synthetic control (bottom)
Real brands: 11.4 pp drop when the name is hidden (McNemar χ² = 33.6, p < 0.001). Synthetic brands: 0.1 pp — statistical noise. The real-vs-fake contrast is the causal proof.
Brand visible vs. brand masked — mixed sets only
Sets where branded and independent hotels genuinely compete — the AI has to actively choose
In competitive markets, real-brand bias grows to 16.3 pp (McNemar χ² = 30.7, p < 0.001). Synthetic brands stay flat — the brand becomes a tiebreaker only when the AI actually recognizes the name.
Statistical test · Finding 2
We used McNemar’s paired test — the right statistical tool when you run the same experiment twice on the same items and want to know if flipping one variable changed the outcome. All-sets: χ² = 33.6, p < 0.001. Mixed-sets: χ² = 30.7, p < 0.001. Both results are far beyond the threshold for statistical significance.
Finding 3 — Fake brand names have zero effect
As a control, we created 100 sets of synthetic hotels with made-up brand names and ran the same visible-vs-masked test on these.
The result: zero difference. Showing or hiding fake brand names produced a 0.1 percentage point delta — effectively pure noise. The AI isn’t reacting to the presence of a “Brand:” label. It’s reacting specifically to brands it recognizes from its training data — the countless web pages where real brands dominate the conversation.
Real brands vs. fake brands
Visible → masked delta · higher number means more bias
The AI responds to brands it knows, not to the label itself. This confirms the mechanism: bias comes from exposure in training data.
Your brand equity in the AI era is literally how often your hotel appears across the internet.
Finding 4 — Brand bias is a tiebreaker for similar hotels
When we restricted the analysis to “mixed sets” — where the AI had to genuinely choose between a branded hotel and an independent with comparable attributes — the brand effect jumped from 11 points to 16 points. When the AI can’t easily differentiate hotels on ratings, price, or location, it falls back on the one thing that feels “safe”: the brand name.
This is the worst-case scenario for independent hotels competing in busy markets. The more similar your hotel looks to a nearby branded competitor in the AI’s eyes, the stronger the brand bias works against you.
The solution flows directly from this: the more clearly differentiated your data is, the less room the AI has to default to brand recognition.
Frequently asked questions
If you’re an independent hotel or a small operator competing in markets with Marriotts, Hiltons, or Holiday Inns, yes. This research tested scenarios where independents and branded hotels compete for the same guests. If you run a branded hotel, this research explains one reason AI already favors you — and why you shouldn’t take that advantage for granted.
We tested GPT-5.4 for this specific brand-visible-vs-masked experiment. Our cross-model research (separate chapter) shows different AI models have different overall selection patterns — Claude, for example, tends to make more independent-friendly choices than GPT. But all models are trained on internet data where brands dominate, so some level of brand familiarity bias is expected across all of them.
Yes. The most impactful immediate step is ensuring your hotel has comprehensive, structured data on your website and across all platforms. The second is building your digital footprint — every travel blog mention, every review site profile, every local guide listing builds your ‘brand equity’ in AI training data. These aren’t quick fixes, but they compound over time.
Traditional SEO gets you ranked in a list of 10 links — the traveler still decides. AI recommendations pick *one hotel* for you. The stakes are higher because there’s no ‘page 2’ in AI — you’re either the recommendation or you’re invisible. The tactics overlap with SEO (structured data, content presence), but the goal is different: you’re optimizing to be **chosen**, not just found.
Yes, but for different reasons. The brand boost varies by brand — Holiday Inn was the most common brand in our dataset, so smaller or less-indexed brands benefit less. And if AI brand bias becomes well-known, platforms may start correcting for it, meaning the current advantage could shrink. The long-term play is earning the recommendation through data quality, not just name recognition.
How we ran the experiment
2,740
Total trials
137
Query sets
5
Hotels / set
2×
Paired conditions
We tested GPT-5.4 across 137 hotel selection scenarios — 37 using real Google Hotels API data and 100 using synthetic controls. Each scenario contained 5 hotels: 2 branded and 3 independent. Each set was tested multiple times with different orderings to rule out positional effects.
The core design was a paired visible-vs-masked comparison. We ran every query twice — once with brand names visible, once with the brand line stripped from each hotel card. Same hotels, same attributes, same orderings. Any difference in selection had to be caused by the brand name alone.
Statistical tests included binomial tests against the 40% baseline, McNemar’s paired test for causal brand effect, and quality-adjusted baselines (QAS) to rule out the possibility that branded hotels simply have better attributes.
Limitations. We tested GPT-5.4 only; other models may behave differently. Hotel sets were fixed at 5 options with 2 branded and 3 independent — real AI searches may present more options and different brand ratios. Brand composition in our real-data sample was skewed toward brands like Holiday Inn, Radisson, and Hyatt, which may not reflect every local market.
Want to know if AI is biased against your hotel?
Huxo’s AI Visibility Report shows you exactly how you stack up against branded competitors in your specific market — which AI models recommend you, which skip you, and what to fix first.