Hospitality · AI search presence (GEO)
When travelers asked AI where to stay, the resort only showed up if they already knew its name
An audit platform that measures a website the way people now find hotels: search, accessibility, speed and mobile, plus a scored view of what six AI engines actually recommend, cite and say.
9 in 10
Answers that named the resort when the question already mentioned it or its area
< 1 in 5
Answers that named it when travelers asked generic questions, like the best resort for families
5×
More AI citations to the destination’s tourism site than to the resort’s own website
2×
Increase in the sister property’s AI readability after two targeted changes, same checks
01
The problem
Travelers are skipping the ten blue links. They ask an assistant for the best family resort, the best golf resort, a place for a weekend away. A hotel can rank well on Google and still be missing from those answers, and no standard SEO report will tell you.
The group ran two properties and one of them was being redesigned. It needed a single, trustworthy view of site health, with history. And before paying for any AI search work, it needed a measurable baseline, so that improvement could be proven instead of claimed.
02
Why it was hard
- AI answers are probabilistic and every engine behaves differently: different APIs, different search grounding, citations hidden behind redirects.
- Engines rarely use a brand’s registered name. Exact matching missed roughly one mention in eight.
- Automated mobile checks are mostly noise: a single site can produce thousands of layout “issues” that a person would never flag. A review done only by an LLM, on the other hand, places problems poorly and invents locations.
- Results had to stay comparable month after month while the pages, the tools and the site itself kept changing.
03
What we built
A multi-dimension audit platform with a client portal. A preflight decides what applies to each site, four engines measure in parallel, and every run is frozen into an immutable snapshot, so any two moments can be compared.
04
Where the innovation is
01
Ask AI the way customers do
Every engine runs with live web search turned on, so the answers reflect what a traveler sees today, not what a model memorized last year. All six are normalized to the same shape: was the brand named, what did the answer say, which sources did it cite.
02
Let the answers define the competition
A hand-picked competitor list produced a leaderboard of zeros. So the competitive set is derived from the brands the engines actually cite, with review sites and directories filtered out. The client sees who the AI recommends instead of them.
03
Store everything, recompute for free
Every answer and the exact sentence that mentions the brand are stored. Results can be recalculated and tone re-classified without paying for a single new query.
04
Code measures, the model reviews, code draws
A real browser renders each page at three device sizes and builds a map of every element. A vision model decides what is actually broken, but it can only point at elements on that map, so it cannot invent locations. Simple rules settle the obvious cases first, and only doubtful ones get a second opinion from a stronger model.
05
Comparable by design
The preflight decides which dimensions apply to each site, the holistic score only averages those, and snapshots never change after the fact. The whole workflow can be run by talking to an AI coding agent, but always through the same tested scripts.
05
What the AI engines said
Fifteen questions a traveler might ask, sent to six engines with web search on. The pattern was immediate: the resort was well represented when someone already knew it, and close to invisible when they did not.
Who the AI engines cite as a source
Citations across all answers, relative to the resort’s own website. Its pages were rarely the source the engines trusted.
15 traveler questions × 6 AI engines
Each cell is one live answer. Filled cells named the resort. Places and the brand are generalized.
View the data
| ChatGPT | Gemini | Claude | Perplexity | Copilot | Google AI Mode | |
|---|---|---|---|---|---|---|
| Is [the resort] a good resort? | ● | ● | ● | ● | ● | ● |
| Where to stay in [its community] | ● | ● | ● | ● | ● | ● |
| Beach resorts in [its town] | ● | ● | ● | ● | ● | ● |
| [The resort] reviews | ● | ● | ● | ● | · | ● |
| [The resort] vs [a competitor] | ● | ● | · | ● | ● | ● |
| [Its community] resort | · | ● | ● | ● | ● | · |
| Best resorts on the [region] coast | ● | ● | ● | ● | ● | · |
| Boutique beach resorts in [the destination] | ● | · | · | · | ● | · |
| Family beach and golf resort in [the destination] | ● | · | · | · | · | ● |
| Beach wedding venues in [the destination] | · | · | · | · | ● | · |
| Best golf resorts in [the destination] | · | · | · | · | · | · |
| Best family resorts in [the destination] | · | · | · | · | · | · |
| Resorts for corporate retreats and meetings | · | · | · | · | · | · |
| Weekend getaway resorts near [the capital] | · | · | · | · | · | · |
| Things to do in [its town] | · | · | · | · | · | · |
06
Results
- A clear, defensible GEO baseline: nearly every answer named the resort when the question already did, and fewer than one in five did otherwise. Five generic questions, including the one built around its signature amenity, returned nothing on any engine.
- A content plan tied to evidence: new FAQ answers mapped one by one to the questions with zero presence, existing ones rewritten, and structured data that contradicted the page copy flagged for correction.
- On the sister property, AI readability doubled after two targeted changes, structured data first and then llms.txt, measured with the same checks before and after.
- A single metadata fix lifted the boutique property’s search score by almost ten points overnight, again with identical checks.
- The mobile AI reviewer matched a human reviewer on every check of its pilot, with no invented references, at well under a dollar per full-site run.
- The platform is now the standard audit for a portfolio of client websites, backed by an automated test suite.
AI readability, sister property
Index, baseline = 100. Same checks each time: AI crawler access, structured data, server-rendered content, FAQ, entity signals, llms.txt.
Who decided each mobile layout check
Share of checks in one full-site run. Most never reach a model; only doubtful ones get a second opinion.
07
What we learned along the way
The zero we refused to publish
One engine reported the resort in none of its answers. That would have made a dramatic slide. Instead we traced it: the search API was rejecting the location parameter and the error was being recorded as “not mentioned”. Rerun correctly, every answer came back.
The score drop that was not real
Accessibility dipped on the redesign. A few heavy pages were timing out because too many browsers ran in parallel. Run one at a time, the score was unchanged. Concurrency was capped and runtime errors are now excluded instead of counted as failures.
The test site was quietly auditing production
The test environment’s sitemap listed live URLs, so part of the audit was measuring the wrong site. The crawler now rewrites every URL onto the host being audited.
Stack
- TypeScript
- Node.js
- Playwright
- Lighthouse
- React
- Tailwind CSS
- Vercel
- OpenAI
- Gemini
- Claude
- Perplexity
- Search APIs
- Vision models
- AI coding agents
Is AI recommending your business, or your competitor?
We run the same baseline for your brand: the questions your customers actually ask, which engines name you, who they cite instead, and what to fix first.
Book a discovery sessionClient identity, places and commercial terms are withheld under confidentiality agreements. Questions and brand names are generalized and some figures are shown as ratios or indexes; every one comes from a measured run. Delivered as the specialist technology partner of a production and development agency, which held the direct client relationship.