What are you feeding your AI?

A proptech data loop reality check | Hannah Parker, Inoki

I promise this is not another AI-written waffle piece about how AI will fix or kill the world (maybe depending on how many "life-ending" Steven Bartlett wines you had last night), but an updated observation on what is actually happening in the proptech and marketplace space right now, specifically around AI features, product evolution, and whether any of it is building something that lasts.

I've spent over 15 years working in the portal and proptech space, and the last year has possibly seen more change than the previous ten combined. What's notable is that the businesses actually moving forward are not necessarily the ones shipping the most AI features. They're the ones staying true to their core reason to exist. Investors are simultaneously panicking about AI spend (hello new item on the P&L) and demanding "how do we add more AI?" But as with any technology change, the smarter question should be "where in the journey could AI actually help?"

After watching demo's and testing out features from real estate agent platforms, marketplaces and new entrants, I am fully with you, the dopamine hit of a well-built AI experience is very real. But after maybe the hundredth flashing gradient button or whooshing ai star icon, what is actually adding lasting value?

The flywheel is still a thing

Whether you are a portal, a real estate agent, a housebuilder, proptech or a niche CRM, you are all taking bites from the same layer cake within the exact same flywheel. Together we enable research, discovery, shortlisting, viewings, valuations, marketing, finance, negotiations and legalities until horrah we reach a transaction point.

The problem is that most of these touch points operate as silos. Data generated at one stage rarely feeds back into the stage before it. Businesses keep investing heavily at their respective funnel peaks without ever really knowing what happens when the lead leaves that part of the wheel. That is not a problem a magic AI button can fix no matter how 'whooshy' it is (technical term). It comes back to the less glamorous world of data infrastructure. AI will not fix the cycle if the data loop is broken to begin with.

Your vector database is your ‘clean eating’ champion

If you operate a consumer-facing interface that generates leads, you are probably deep in building richer personalisation and testing out AI embedding models right now. The goal being that user behaviour gets converted into vectors, those vectors get matched to properties (nearest-neighbour), the model surfaces recommendations, and in theory the whole system gets smarter over time.

But based on what I have seen in the market so far, there are some great intentions, but significant gaps in this data diet.

Take a portal or real estate agent’s website. There is a wealth of behavioural data there: clicks, saves, search history, time on listing. The model builds what looks like a confident picture of a buyer: three-bed semis in Bristol, south-facing garden, under £450k. That looks like strong intent signalling, but it is only half the story.

Back in March I spoke at the Property Portal Watch conference in Bangkok about the kind vs. wicked data problem in property. A kind data environment gives you fast, clear, complete feedback loops: the model sees the outcome of its predictions and improves. This is where AI has moved fastest; tech, law, science, finance etc. Property data however is inherently wicked: not only is it ambiguous, but the journey is long, fragmented across platforms, the outcome is often months away, and that outcome is invisible to whoever started the search.

Unless you are feeding the model outcome-related data alongside your intent signals, the model's intelligence is artificially capped. You don't get smarter, and that is exactly where it gets expensive. The model keeps predicting with the same confidence, but it never learns whether it was right. It's only ground truth is engagement proxies (clicks, saves, time on listing), never the one thing that actually matters: did that family find their home? You pay for the same compute either way, for an intelligence that has hit its ceiling, and you keep spending on the campaigns that drove those leads without knowing which ones actually convert. As long as the human outcome stays invisible to the model, the spend doesn't get more efficient, it just repeats itself. If you are investing at the top of the funnel without analytics feeding back from the conversion point, it is like making a sandwich with no bottom slice: all the good fillings keep falling out.

The five layers worth thinking about

When I work with property businesses on data strategy, I map it across five layers: raw listing data at the base, enriched property attributes above that, third-party datasets covering neighbourhood, planning, risk and transport signals, behavioural data from users, and finally proprietary outcome data at the top. Most portals have the bottom two reasonably well covered. The third is improving, but the top two, particularly that proprietary outcome layer, are where the real AI super powers get built. Today, it is the layer most businesses are skipping (or hiding?).

What has been released, and does it help close the loop?

I know we all want to skip to the review parts, so let's take a drive by at what has shipped over the last twelve months and who is building something defensible. I prefer to keep these evals/ reviews pretty simple, asking: does it improve the experience for the person buying or selling a home, does it create the foundations of a smarter learning loop, or does it just make the business feel more ‘lit’?

*For everyone born before 1990 translate lit as ‘cool’. Gen Z language is a whole other discussion!

• Top of class: Zillow's AI Mode

A strong example of AI going well beyond search. The architecture matters here: this is a multi-agent system, not a single model. In plain terms, that means separate models handle distinct tasks (search, memory, recommendation) and orchestrate between themselves. This is how you get coherent, cross-domain reasoning without trying to cram everything into one massive prompt. Zillow has also introduced cross-session memory (personal fav), meaning the system builds a relationship with the buyer over time rather than starting cold with every visit. That is a very different competitive ambition to most portals and one with significant implications for agents and #2/3 portals, who might have thought they were competing just on listings.

Try it; www.zillow.com

• One to watch: REA’s realAssist

This feature builds on the success of their realEstimate valuation tool and focuses on the seller, a point in the flywheel that is chronically underserved in the chase for buyer leads. But what makes REA worth watching closely right now is not any single feature: it is the data strategy bubbling away underneath all of them. The Planitar/iGUIDE acquisition (AI-driven 3D spatial capture and virtual tours) is yet another well-placed piece in a stack they have been assembling deliberately for years. The Neighbourlytics acquisition, the Ray White data partnership, the Jitty investment, the OpenAI collab. Each one hits the same question “what data do we need to make AI work for property?” That is a fundamentally different question to "what AI feature should we build next?" which, let's be honest, is taking up an awful lot of transcribed meeting note server space right now. This strategy could make REA a formidable long-term opponent for anyone in their market.

Try it; www.realestate.com.au

• Biggest potential: Rightmove's conversational AI search (built on Google Gemini)

I've picked this out not because it is the most polished release, but because it is where the green shoots of something bigger are showing. The natural language query tool is getting closer to how buyers actually think. Pushed to answer more personalised or ambiguous requests, it still struggles to translate intent beyond standard filters, but as a starting point it is a meaningful step.

The renovation cost calculator is a fun addition: it does not deliver a personalised wow moment, but as a mechanism for capturing non-standard buyer intent data, it is far more strategically interesting than it looks.

Rightmove's real asset here is scale. As one of the most visited websites in the UK, they sit on decades of proprietary behavioural history, rich lead data from calls and messages, and deep area and property insight. Combine those layers with a properly closed feedback loop from lead to outcome, and you have the foundation for something that could fundamentally shift the UK property search experience. The potential is enormous. The execution is the question.

Try it; www.rightmove.co.uk

• The close-loop player: iovox’s AI Agent & AI Insights

Call tracking has historically been treated as table stakes: a compliance checkbox for lead attribution. What the latest iovox AI suite does is upgrade it from a reporting tool into a data loop mechanism. Call transcriptions are processed to extract structured intent signals, lead quality indicators and outcome data, all of which feeds back into top-of-funnel product and marketing decisions. This matters because voice and conversation data represents at least half of the leads a portal or agency generates, and in most businesses today that data evaporates the moment the call ends. When you close that conversational data loop, a few things shift: attribution becomes grounded in actual outcomes rather than activity proxies, campaign spend follows real conversion signals, and the AI features being invested in start to compound rather than plateau. Worth watching how portals in particular use this enhanced data flow as part of their broader AI roadmaps.

Try it; www.iovox.com

• The partnership play: Realtor.com's RealAssist AI

Another Google stack partnership (Gemini plus Google Maps grounding) ‘RealAssist AI’ by Realtor is possibly the most direct challenge to Zillow's AI Mode we've seen from another portal. This is currently in beta, but what we can gather so far is that it offers cross-session memory, natural language search, neighbourhood context: all from the off. The Google Maps integration means a buyer can ask "somewhere I can get to Midtown in under 30 minutes" and the system treats that as a core search dimension, not a filter bolted on at the end.

The feature I am most looking forward to playing with though is the visualisation tool: buyers can see how a property might look at different times of day, in different seasons, or with potential exterior modifications. That sounds like another ‘shiny thing’ feature, but it is actually a clever aspiration data capture mechanism, the portal can start to understand what buyers are imagining, not just what they are clicking. It echoes the logic behind Rightmove's renovation cost calculator, but goes a step further into the emotional layer of the decision (big fan).

As I write this though, I can't help but question whether building on Google's infrastructure is the best bet? Is it a long-term competitive advantage or a structural dependency? Will going faster today cost you tomorrow? Zillow has built proprietary models but Realtor.com has moved faster by assembling someone else's stack (and looks like they have executed extremely well). Both can have success here, but it will be worth watching whether the learning loop compounds into something Realtor.com owns over time, or stays anchored to Google's roadmap. 

Take a look; Realtor

• The functional play: Scout24’s HeyImmo

Similar intent to Zillow on the relationship and conversational search front but 'HeyImmo' positions itself as "a sparring partner that guides you through the whole platform experience." Where Zillow has committed to a full multi-agent architecture, HeyImmo is currently closer to structured function calling, the AI invokes specific tools to answer queries rather than reasoning freely across domains. That is not a criticism btw, it feels like a very sensible starting point. If the learning loop is built properly for both qualitative and quantitative signals (pre and post lead generation), and if the model evaluations (aka the benchmarks used to measure whether AI responses are actually improving) are performing, this could evolve into something more powerful. Scout24 is already expanding dialogue-based search into the core platform, suggesting the ‘HeyImmo’ approach is graduating quickly from experiment to mainstream in their product suite. 

The business made some ambitious claims at their capital markets day in May (apparently there was much more shared in the room). My between the lines read was this will hit on a multi-agent approach across all user touchpoints and automating the full transaction workflow. If delivered, that could change portal and agent relationships in a material way.

Try it; www.immobilienscout24.de

• The non-tech play: Hemnet's "sell first, pay later"

A useful reminder that the best product decisions do not always involve a model or a dataset. For those less familiar: Hemnet operates a consumer-direct monetisation model in Sweden, where sellers pay for listing visibility rather than agents carrying the cost. "Sell first, pay later" defers that payment until the property transacts, removing a friction point for sellers in a market that has been under significant pressure. Sweden's housing market saw one of the sharpest corrections in Europe over 2022 to 2024, and ouch, Hemnet's stock price felt it. This is sharp commercial UX designed to stimulate seller-side activity while the market recovers, no AI required. Smart product strategy sometimes means knowing which problems to attack that bypasses the technology toolkit and puts a pin in features FOMO.

Try it; www.hemnet.se

Final thoughts

The businesses I am watching closely right now are not those with the flashiest press releases (though I am here for those too!), but those who are quietly building smart closed-loop ecosystems and keeping the user at the absolute centre of their decisions. The family searching for a home does not know or care what a vector database or multi-agent strategy is. They care whether they are making a good decision, whether the information in front of them can be trusted, whether someone picks up when they call, and whether the process feels designed around them rather than forced upon them.

So perhaps the stop and check ourselves questions we should be asking today are more… what’s that first outcome data point we are not currently capturing, what’s that one thing that we really should be building on first?

You do not need to solve the whole wicked data problem at once. You just need to start making it a little kinder.

---

Who is Hannah Parker? 

Founder of Inoki, a proptech specialist consultancy. 

“In additional to an unheathly obsession with product releases and data strategy across portals and the proptech sector, I also work with a range of businesses from startups to portal unicorns and investors, supporting in product and growth strategies. Based in Bordeaux, France, I work with clients from Australia to London.”

Sounds interesting? Get in touch and let’s see what ideas we can create together. 

hello@inoki.agency

hannah@inoki.agency

inoki.agency | linkedin.com/in/hannahfparker