Original research · ArribaIA
Which Dénia restaurant the AI recommends
We are sending thousands of real queries to ChatGPT and other assistants asking where to eat in Dénia, save every full response, and extract which restaurants get mentioned, in what position, and with what attributes. It's not a survey or a list of our own favorites: it's a measurement with a public methodology.
The experiment in numbers
The first data cut, verified
These figures are from the first data cut, closed and verified. This isn't the study's cumulative total: collection keeps moving, and we'll publish later cuts as they close.
Measurement window
September 11-13, 2026
Queries included
451
Mentions extracted
1,278
Distinct restaurants
498
With capturable response text
236 of 451 (see the data quality notice)
Data quality notice
Of the 451 calls in this cut, 236 returned capturable text. The gap isn't spread randomly: it's concentrated in three specific day, time-slot, and assistant combinations (detail below) totaling 217 calls, of which only 2 returned text. Those calls completed their run and used tokens, but the text couldn't be recovered due to a capture-pipeline gap, not because the assistant recommended nothing. Every figure on this page counts them toward the total queries, but they can't contribute mentions.
What we're studying
How Dénia restaurants show up in AI responses
The question this study answers is specific: when someone asks ChatGPT or another AI assistant where to eat in Dénia, which restaurants does it cite, in what order, and with what information.
To answer it, we automate sending realistic queries (the kind a person would actually type while looking for a restaurant, not loose keywords), save each assistant's full response, and use a second model as a referee to extract the restaurants mentioned, their position, and their attributes.
What this study is not
- It is not a customer survey or a popularity vote.
- It is not an editorial ranking based on the ArribaIA team's own taste.
- It doesn't measure food or service quality: it measures visibility and recommendation frequency in the responses of the AI assistants analyzed, under this experiment's specific conditions.
How it's done
The experiment's design, summarized
The full design, with every variable justified and its limits stated, lives in the methodology. This is the summary for anyone who only needs the general frame.
Automated queries
A system generates and sends realistic questions to each assistant ('where can I get rice for dinner in Dénia tonight?', for example), not keyword lists.
Crossed variables
Each query combines cuisine, search intent, language, budget, and Dénia zone, to cover how people actually ask, not a single repeated script.
Time slots
Queries are spread across several time slots of the day, Madrid time, to see whether the time of the query changes the response.
Extraction with a referee model
A second model reads each response and extracts which restaurants are mentioned, at what position, with what attributes, and with what tone. This extraction can contain errors, and we explain that in the methodology.
Normalization
Extracted names are normalized deterministically so the same restaurant doesn't get counted twice over a spelling variation.
Articles and analysis
What we've published from the data
Each piece analyzes a different dimension of the same dataset: ranking, the full data, and six per-variable analyses. All of them are published from the first data cut.
Dénia restaurants most mentioned by AI
Dénia restaurants ranked by mention count and average position in the responses of the AI assistants analyzed. A visibility ranking, not a quality one.
ReadExplore the full dataset
Explore the full Dénia restaurants GEO study dataset: queries by variable, ranking, metric definitions, and a JSON dataset download.
ReadHow we measure GEO ranking for Dénia restaurants
Full design of the Dénia restaurants GEO study: variables, models, extraction process, definitions, and limitations. Data collection is continuous; the ranking publishes once we close a verified cut of the data.
ReadIs showing up in the API the same as showing up in the ChatGPT interface?
Why measuring via API isn't exactly the same as measuring with an AI assistant's consumer interface, with each provider's own documentation as evidence.
ReadWhat changes by the query's cuisine type
How AI visibility changes by the query's cuisine: which categories concentrate mentions on a handful of restaurants and which spread them across many.
ReadWhat the time slots show, and don't show
What this cut's three time slots show, and why we can't separate the effect of time of day from the effect of day of week with this specific design.
ReadWhat changes by what the person needs
How AI recommendations change based on what the person actually needs: immediate booking, future booking, discovery, or a situational query.
ReadWhat changes by the query's budget
What changes when a query states an explicit budget versus when it doesn't, and which restaurants dominate each spend level.
ReadHow much the assistants agree with each other
How much DeepSeek and OpenAI (with web search) agree on which Dénia restaurants to recommend, and why comparing OpenAI without search isn't reliable in this cut.
ReadWhat we actually can compare, and what we can't
What changes when OpenAI answers with the web search tool on: real citations, verifiable sources, and restaurant coverage, versus a without-search sample too small to compare.
ReadWhy it matters
Why this matters to a Dénia restaurant
Asking ChatGPT where to eat, instead of opening a map or a list of reviews, isn't unusual behavior anymore. If a restaurant doesn't show up in those answers, or shows up with wrong information, it loses a recommendation channel that no longer depends solely on Google.
This study doesn't promise a formula for ranking better in ChatGPT: it documents, with a public methodology, which restaurants show up today, how often, and under what conditions, so anyone, restaurants included, can review the data and draw their own conclusions.
Generative search doesn't replace traditional search, it adds a different layer: a written answer instead of a list of links. Understanding that difference is the starting point of the discipline ArribaIA calls GEO.
Limitations we acknowledge upfront
- Any figure we publish is a snapshot of one specific moment: collection keeps moving after that snapshot, so the full dataset is always bigger than whatever was last published.
- Assistants can answer the same question differently at a different moment.
- Automated restaurant extraction can make mistakes; we document how we mitigate them and which we accept.
- A provider's API isn't necessarily identical to what a person sees using its web interface, which we cover in detail in GEO UI vs API.
Methodology and transparency
You can check exactly how we measure this
We publish the full experimental design before we have a closed ranking: variables, collection cadence, models, extraction process, and limitations. If anything in the final ranking doesn't match this, it's our mistake and we correct it.
Go to the methodologyFrequently asked questions
This usually answers questions before you write to us
Is this a ranking of the best restaurants in Dénia?
No. It measures how often and in what position each restaurant shows up in the responses of the AI assistants analyzed, under this experiment's conditions. A restaurant can be excellent and not show up, or show up and not be the best fit for everyone.
Did you survey customers or collect reviews for this study?
No. The entire dataset comes from automated queries to AI assistants and the extraction of their responses. There's no human opinion, survey, or vote in the measurement process.
My restaurant doesn't show up, does that mean it has bad GEO?
Not necessarily. Showing up or not in this specific experiment depends on many variables (what was asked, in what language, with what budget) and on how each assistant builds its answer. It's a signal, not a complete diagnosis.
Can I ask you to analyze my restaurant or my sector?
Yes. The contact form has a specific option for that. This study's architecture is built so it can be repeated for another sector or another city.
Why do some pages say data is missing?
Restaurants, press, and researchers
If your restaurant shows up in the study, if it doesn't and you want to understand why, or if you work in research and want to collaborate, write to us through the study's form.