The Question Buyers Are Starting to Ask
Sometime in the past eighteen months, a certain type of buyer began arriving at viewings having already spoken to an AI. Not to an agent, not to a friend with a holiday apartment in Puerto Banús — to a chatbot. They had asked it about median prices in La Zagaleta, about the difference between a frontline golf villa in Nueva Andalucía and a hillside plot in Benahavís, about what a service charge looks like in a gated community on the Golden Mile. Some of the answers they received were reasonable approximations. Others were fabricated with complete fluency and no apparent awareness of the problem.
This piece is an honest attempt to describe what is actually happening when buyers use general-purpose language models to research property on the Costa del Sol, and where that diverges from what a tool built specifically for this market can do. I am writing it partly because the question of *ai real estate marbella* is now a genuine search term — people are looking for a comparison, and most of what they find is either promotional or technically shallow. I want to be precise about the mechanics, because precision is what this kind of decision requires.
How Generalist LLMs Handle Property Queries
ChatGPT, Claude, Gemini and their peers are trained on large, undifferentiated corpora of text. They are very good at synthesising patterns from that text and producing fluent responses. What they cannot do, structurally, is tell you the current asking price for a specific villa in Cascada de Camoján, because that information is not in their training data in any reliable form, and even if it were, training data has a cutoff date. The listing may have sold. The price may have been revised three times. The property may have moved from a public portal to an off-market arrangement.
When a buyer asks a generalist model for a specific property recommendation in Sierra Blanca, something predictable tends to happen: the model generates a response that sounds like a property description. It may include approximate square metreages, vague references to panoramic views, and a price that sits within a plausible range for the zone. None of this is sourced. The model has no access to a live feed. It is, in the clinical terminology, hallucinating — producing output that is statistically coherent but factually unverifiable. The buyer, unless they already know the market well enough not to need the tool in the first place, has no obvious way to detect this.
This is not a criticism of the technology in the abstract. Generalist models are genuinely useful for certain research tasks: understanding the legal structure of a Spanish property purchase, comparing mortgage frameworks, learning what a nota simple contains. These are questions where the answer is relatively stable and well-documented. Asking for specific listings is a different category of question entirely.
What a Grounded Tool Does Differently
A property-specific AI tool operates on a different architecture. Instead of generating answers from a static training corpus, it retrieves information from a live, structured database and uses the language model only to interpret and communicate what it finds. The database is the source of truth; the model is the interface.
At Muse Selection, the working catalogue draws from three data feeds — Inmobalia, Resales-Online, and Zoddak — deduplicated to approximately 670 active residences at any given moment. When a buyer asks the AI Concierge on museselection.es a specific question — four bedrooms, sea view, under €4 million, within ten minutes of Marbella centre — the tool queries that catalogue. Every response it gives is anchored to a real reference. If a property has sold, it is no longer in the feed. If the price has changed, the feed reflects that change.
This is the structural distinction that matters. The generalist model is fluent but ungrounded. The dedicated tool is narrower in scope but factually accountable. In a market where a misread price point of €500,000 is a rounding error rather than a disaster, accountability is not a marginal concern.
The Off-Market Problem
There is a further layer that no AI tool, however well-engineered, currently handles well, and it is worth being straightforward about it. A significant portion of the residences that change hands in the premium zones of the Costa del Sol — La Zagaleta, El Madroñal, the quieter roads above the Golden Mile — are never listed publicly. They move through networks of introductions, through agents who have maintained relationships with the same families for years, through conversations that happen in person rather than on portals.
At Muse Selection, roughly 300 residences sit in this category at any given time. They are shown by introduction only. No feed carries them. No AI tool, including our own, can surface them through a standard query, because they are not in any structured data layer that a tool can access. This is not a failure of the technology; it is a reflection of how a meaningful share of this market actually works.
What this means practically is that a buyer who relies entirely on any AI tool — generalist or dedicated — to scope the market is seeing, at most, the publicly listed portion of it. In the sub-€3 million segment, that portion is substantial and the tool is genuinely useful. As budgets move upward and the properties become rarer and more discreet, the gap between what is queryable and what is available widens. The tool becomes a first filter, not a complete picture.
Accuracy as a Question of Architecture, Not Brand
It is worth spending a moment on why this distinction — grounded versus ungrounded — is architectural rather than a matter of one product being more carefully built than another. A generalist language model cannot be made accurate about live property listings by improving its training data, because training data is by definition historical. You can give a model access to the web via a search plugin, and some implementations do this, but web search retrieves whatever is publicly indexed at that moment, which is a different thing from a curated, deduplicated, professionally maintained feed.
The feeds that serious property platforms maintain are valuable precisely because they are not the public web. They are structured, they are regularly reconciled against reality, and the people maintaining them have professional reasons to keep them accurate. When a property sells, someone updates the feed because their business depends on the feed being correct. That incentive structure does not exist in the same way for a general-purpose AI company whose product is a conversation interface.
This is also why the question of which AI tool to use for property research in Marbella is not primarily a question about AI. It is a question about data. The AI layer — the natural language interface, the ability to ask questions in plain language and receive coherent answers — is, at this point, relatively commoditised. What differentiates tools is the quality and currency of the data they are permitted to query.
What Buyers Should Actually Do
The most sensible approach I have observed among buyers who navigate this market well is to treat AI tools as a first orientation layer, not as a source of transactional information. Use a generalist model to understand the legal framework of a Spanish purchase, the tax implications of residency, the general character of different zones — Sotogrande's relative distance from Marbella, the practical difference between a Benahavís address and a Marbella one for everyday life. These are questions where the model's broad training is an asset.
For specific listings, pricing benchmarks, and zone-level inventory, use a tool that is explicitly connected to a live data source and can tell you where its information comes from. Ask it to cite a reference. If it cannot, treat the response as orientation rather than fact.
And for the segment of the market that is genuinely off-market — which in the €3 million-and-above range is a non-trivial share of what is actually available — accept that no tool replaces the relationship. The introduction still matters. The conversation that happens in an office on Avenida Arias Maldonado, or over lunch, or at a viewing arranged quietly through someone who has been in this market for years, surfaces inventory that no feed will ever carry.
AI changes the efficiency of the queryable layer. It does not change the nature of the layer beneath it.
---
The buyers who arrive best-prepared are not those who have spent the most time with any particular tool. They are the ones who understand, before they start, which questions a tool can reliably answer and which questions require a different kind of attention. That distinction — between what is computable and what is not — is not a weakness of current AI. It is simply an accurate description of where the technology sits, and where the market sits, at this particular moment.