Housing advertising and audience targeting

In May 2024 HUD issued guidance addressing the application of the Fair Housing Act to the advertising of housing opportunities through online platforms that use targeted ads, noting that platforms often use AI to target ads to specific groups and that such targeting must not discriminate against protected classes [1].

The reason this is worth its own page is that engineers building housing products usually do not think of themselves as being in the advertising business, and several common features are advertising in the relevant sense.

Three places this shows up

Explicit targeting. A campaign configured to reach a defined audience. This is the obvious case and the one most teams already treat carefully, since selecting audiences by characteristics related to protected classes is understood to be a problem.

Delivery optimisation. This is the case that surprises teams. Nobody sets a discriminatory targeting criterion. The system is told to maximise engagement or minimise cost per enquiry, and it learns from historical response data which users are most likely to respond. If historical response patterns differ across groups, and they generally do, the optimiser will deliver the ad disproportionately to some groups without any targeting criterion having been set.

The system is doing exactly what it was asked to do. The requested audience was neutral. The delivered audience was not.

A concrete version of this: a "lookalike" or "similar audience" feature, common in ad platforms, that takes a seed list of people who previously engaged with a listing or a campaign and finds other users who resemble them on whatever signals the platform uses internally, which the advertiser generally cannot see or control. If the seed list itself reflects a skewed pattern of past engagement, the lookalike expansion propagates and can amplify that skew, because the platform's similarity model is doing exactly what it is designed to do: finding more people like the ones already in the list. The advertiser set no discriminatory criterion and often cannot even inspect what the platform's similarity model used to build the expanded audience. That opacity is itself a reason to measure the delivered outcome rather than only reviewing the campaign's stated targeting, since the stated targeting can be completely neutral while the actual mechanism behind it is not visible at all.

Ranking and recommendation. A search results page ordering listings for a user. A feed suggesting properties. A notification deciding which new listing is worth alerting someone about. Each of these decides which housing opportunities a person learns about, and each is typically built by a team that has never been part of a fair housing conversation, because it is filed under relevance rather than advertising.

That third category is where we would expect most unexamined exposure to be in a typical proptech product, precisely because nobody has framed it as the same problem. It is worth asking directly which team owns ranking or the recommendation feed and whether fair housing has ever come up in that team's own reviews, since the honest answer in most organisations is that it has not, and that gap is the finding rather than a footnote to it.

Measure the delivered audience

The single most useful engineering consequence of all this: the property to measure is the audience that actually received the impression, not the audience that was requested.

That is a data model requirement before it is an analysis one. The system has to record, for each listing or campaign, the composition of who was actually shown it, in a form that can be assessed against the composition of the eligible population. Most ad and feed systems record delivery in aggregate for billing and performance and do not retain what is needed to answer this question later.

The comparison that matters is between the delivered audience and a reasonable baseline of who could have been shown it. A delivered audience that differs sharply from the eligible population is the signal, and it is a signal regardless of whether any targeting criterion explains it. In fact the cases with no explanatory criterion are the interesting ones, because they are the ones nobody would have found by reviewing the campaign configuration.

Constrain the optimiser

If an optimiser learns to deliver a housing ad unevenly, the engineering options are the ones available for any constrained optimisation problem.

Remove features that carry protected-class signal from the optimisation, understanding from why removing the protected field is not a defence that this is necessary and not sufficient. Constrain the optimiser explicitly, so delivery composition is bounded relative to the eligible population rather than following engagement without limit. Or take housing listings out of the general optimisation path entirely and deliver them under a different, simpler policy.

That last option is worth considering seriously rather than dismissing as a worse option. Housing is a category where the cost of getting delivery wrong is high and the marginal benefit of aggressive optimisation is modest. A simpler delivery mechanism for housing content, applied uniformly, is defensible, explainable, and cheap to verify. Sophisticated optimisation of housing ad delivery gains a small amount of efficiency and adds a category of risk that is hard to measure and harder to explain.

Ranking is the same problem

If your product ranks listings for users, the same analysis applies, and it applies to a feature nobody flagged.

The questions are identical. Does the ranking model use features correlated with protected characteristics? Does the resulting exposure of listings differ systematically across user groups? Are certain listings systematically shown to a narrower or different audience than others, and does that pattern align with anything it should not?

The practical starting point is simply to ask, for a sample of listings, who saw them. Most teams cannot answer that question today, and finding out that you cannot answer it is itself the first finding.

A specific mechanism worth checking for in a ranking model: features derived from historical engagement with a listing, such as click-through rate or time spent viewing, fed back into the ranking of future listings. This is a completely standard technique for improving relevance and it creates a feedback loop. If a listing received less engagement in the past because it was shown to fewer people in a particular group, and the ranking model treats low historical engagement as a signal that the listing is less relevant, it will show that listing to fewer people going forward, which produces even less engagement, which reinforces the signal further. Nothing about the model's objective mentions a protected characteristic, and the loop can still produce a pattern where certain listings, or certain groups of viewers, are shown a narrowing set of options over time. Breaking the loop generally means periodically exploring outside what the model currently ranks highest, rather than only exploiting what it already believes works, so that a listing is not permanently disadvantaged by an early measurement that itself reflected uneven exposure.

The vendor question

Many housing products advertise through platforms they do not control. The targeting, optimisation, and delivery happen inside somebody else's system.

That does not remove the need to understand what is happening, and the practical position is to use the controls the platform provides for housing-related advertising, which major platforms have introduced, and to retain records of the audience configuration used for each campaign. What you can measure is limited by what the platform reports, which is an argument for keeping your own records of what you asked for, and for treating platform-provided housing ad controls as a minimum rather than a complete answer.

There is also a practical distinction worth drawing between running a housing campaign on a general-purpose ad platform and running it through a specialised listing syndication channel, such as a multiple listing service feed or a rental listing aggregator. The general-purpose platform's optimisation and audience tools were built for advertising broadly and then adapted for housing content, which is why the housing-specific controls exist as an overlay rather than as the platform's original design. A specialised housing channel is more likely to have been built around the constraint from the start, which does not make it automatically safer, but does mean the questions to ask a specialised vendor are different from the questions to ask a general platform: ask the specialised vendor how listings are ordered and to whom, and ask the general platform what its housing-specific controls actually restrict, since "we have housing ad controls" describes a feature category rather than a specific guarantee about what is and is not restricted.

If you want help working out what your product currently decides about who sees what, get in touch.