AI in classifieds: building the portal is cheap now, getting the listings isn't

The argument about AI in classifieds is mostly happening from the outside. McKinsey, OC&C and Apax have all written about whether the big marketplaces are under threat, and most of it is about ChatGPT and the other assistants sitting between the buyer and the portal. OC&C puts generative search at up to a billion prompts, growing more than 20% month on month, although they don’t publish how they got there. Apax comes back the other way and says more than 70% of traffic to the leading marketplaces arrives direct and more than half through apps. That’s an investor’s claim, but it rings true to me.

I’ve been looking at it from the challenger’s side. I started with realestate.com.au in 2010 and I’ve done pretty much everything in real estate platforms since, and this year I built one of my own with coding agents. The code turned out to be the cheap part, and the hard part is the same as it’s always been.

I rebuilt a big chunk of a property portal

I’m not going to pretend I’ve replicated everything a property portal does. They’re sprawling businesses that cover a huge range of small pieces, and a lot of it builds on depth that takes years, things like valuation and local requirements like energy certificates. What I have done is take a substantial chunk of the core business and replicate it. And not an MVP that’s on Vercel and falls over the moment you throw serious traffic at it. A scalable, production-ready application, with the alerting and the rest of what keeps it up.

The first pass took about two days, and rounding it out took a week or two. I built it on Cloudflare at the edge, which meant it scaled without me doing much about it. Agents have been merging their own work to production on it since January, and I’ve written about how that works separately.

None of this came from nowhere. I had ten-odd years of knowing what these businesses actually do, and I spent a lot of time on the spec. You’re not trying to replicate somebody’s exact code, you’re trying to replicate the product. Along the way I hit the same problems REA hit years ago. Ingest was hitting the site so hard it was taking it down, and the fix was to split it off and put a queue in between, which is exactly where a whole bunch of companies ended up. It felt like speed-running ten years of portal engineering.

The run-cost gap is real

I priced out what mine would cost to run at the volumes of an established portal, and it comes out at roughly an order of magnitude less than what a portal like that pays. I’ve designed the platform to be extremely cheap to run, and at national volumes that adds up to a very big gap.

I keep telling CIOs and CTOs that they need to be really aware they’re losing a moat here. The cost of building and running a portal at national scale used to keep a lot of people out, and it’s collapsed on both sides. In a paid vendor advertising market, where agents are already unhappy about what they’re being charged, that difference is room a challenger can use.

Portals have always been easy to write

The catch is an old one. Every two or three years in Australia, some agent group would get together and start a portal. The ability to write a portal is nothing new. And then they’d discover that the absence of visits was the biggest problem. It’s a distribution problem. It’s always been a distribution problem.

I’ve watched a lot of those upstarts get crushed by Domain and REA. It’s the same with Slack. You can build a Slack clone in no time flat now, but open source alternatives to Slack have existed for a very, very long time. Building the thing was never what kept the challengers out.

AI makes the cheap part cheaper. It doesn’t touch the expensive part, which is attention: getting the traffic, getting the people, getting them to come back.

The listings problem

On a classifieds site the distribution problem has a second half, which is supply. Nobody visits a portal with no listings on it, and nobody lists on a portal with no visitors.

The big problem with entering any market is getting the listings. Even if you get the buy-in of an agency, even if the agency says yes, you can have my listings, often they don’t control them quite as much as all that. The listings live in a CRM you have to integrate with, and the CRM is often owned by the major portal, which isn’t going to add you.

The feed itself is less of an obstacle than people think. REA Group used, and still uses, XML over FTP. These days it isn’t even a real FTP server, it’s a fake one that picks the files up and processes them. They decided that reworking every real estate agent’s software interface into the portal was too much work and too error-prone when there was an existing standard and format. Developers, myself included, have a bit of an allergic reaction to FTPing files around, but it’s a good decision. You have a contract, they have a format, and you don’t have to integrate each vendor’s software one at a time. What keeps a challenger out is who controls the pipe, not the format.

The way I’ve gone at it on my portal is to not depend on the pipe at all. I built ingestion that can take any agent website, whatever it looks like, and pull out the listings as long as they’re reasonably complete. All I need is a list of agents, and I have every listing for an entire country. You hit each agent website every week to keep it fresh, and that gives you SEO juice as well, which means you can go to the agent and actually make a deal. On top of that I’ve built agent products that don’t need traffic volume to matter, like an AI assistant they can embed on their own site that answers questions about their listings and the agency.

That’s a way in, not a guarantee. It gets you supply, but you still have to pay the traffic bill and earn the visits.

What this means for incumbents

I was evaluating this as a threat as well. If anyone can come along and build platforms, then REA Group and the other big portals have a harder life with all these pop-ups. But it’s still not that easy. You need deep, deep knowledge of all the things a portal does, and that’s before you even get to distribution, to getting people to actually use it. Some of the engineers inside these companies could do something similar, because they have that knowledge. Most people starting from scratch don’t.

The moat for incumbents is the direct traffic, the app installs and the relationships with agents, built up over a very long time. If Apax is anywhere near right about direct and app traffic, that’s what protects a portal from ChatGPT as well. The change I can actually point to is that the code and the hosting stopped being a barrier.

The risk I’d watch is on the inside. Portals like REA have historically had a lot of their revenue come from one place, and that makes it hard to focus on anything else. When the core product is that dominant, a cheap challenger in some corner of the market doesn’t look like a priority.

If you’re running an established portal, the build and run cost is the part of your moat that’s leaking. The traffic and the agent relationships are what you need to protect, and traffic in particular is the part a challenger can’t build with agents.

If you want a second pair of eyes on where your platform sits on that, that’s the kind of thing an architecture audit covers.