Domain Atlas / Housing & homelessness services
One engine, many rivals: a shared rent-setting model and the record it writes back
The United States and ten plaintiff states allege in United States v. RealPage, Inc. (M.D.N.C., filed 23 August 2024, amended 7 January 2025) that competing landlords contractually fed a single vendor nonpublic, competitively sensitive information — executed new-lease rents, renewal offers and rates, lease terms and occupancy signals — to train and run a common pricing algorithm that recommended rents back to all of them: the complaint alleges at least 80 percent of the commercial revenue-management software market for multifamily housing and data agreements reaching over 16 million units nationwide including units of landlords who were not customers, describes an 'Auto Accept' setting that implemented daily recommendations with no human review, a 'Governor' feature alleged to constrain price decreases more than increases, and vendor pricing advisors who monitored client acceptance and pushed property managers toward compliance, and reports a national acceptance rate of 40 to 50 percent for new leases across January 2017 to June 2023 against internal analysis finding nearly 60 percent of final floor-plan prices within 2.5 percent of the recommendation and more than 85 percent within 5 percent — allegations only, with no defendant having admitted wrongdoing and no liability adjudicated.[3]
What happened
RealPage, Inc. sells commercial revenue-management software to multifamily landlords: YieldStar and AI Revenue Management, plus Lease Rent Options, acquired from rival Rainmaker in 2017 in a transaction the Department of Justice complaint later cited in its monopolization count. The products generate daily unit-level rent and lease-term recommendations. What distinguishes them from ordinary pricing software, on the government's account, is the substrate: competing landlords contractually agreed to feed the vendor nonpublic, competitively sensitive information — executed new-lease rents, renewal offers and rates, lease terms, occupancy and future-occupancy signals — to train and run a common algorithm whose recommendations went back to all of them. The complaint alleges the vendor's data agreements gave it access to confidential lease-transaction data from over 16 million units nationwide, including units of landlords who did not use its pricing products at all; a sales representative is quoted pitching "we have over 16 million units of data." It alleges at least 80 percent of the commercial revenue-management software market for multifamily housing. On 23 August 2024 the Department and eight state attorneys general filed suit in the Middle District of North Carolina, pleading a Sherman Act Section 1 information-sharing and algorithmic-coordination scheme alongside a Section 2 monopolization count. On 7 January 2025 an amended complaint added six large landlords — Greystar, Blackstone's LivCor, Camden Property Trust, Cushman & Wakefield/Pinnacle, Willow Bridge and Cortland — and Illinois and Massachusetts joined, bringing the total to ten plaintiff states.
The mechanics the complaint describes are specific, and three of them do most of the work. "Auto Accept" implemented daily price recommendations with no human review and updated advertised rents automatically. The "Governor" feature required escalation or approval for out-of-band price changes and was alleged to be asymmetric, constraining price decreases more than increases. And the vendor employed pricing advisors assigned to client portfolios who monitored client acceptance of recommendations and pushed property managers back toward compliance; it also hosted user groups and forums where competing landlords' staff met and discussed pricing practices. Human discretion was nominally complete — any landlord could reject any recommendation — and the complaint quantifies the gap between nominal and effective: a national acceptance rate of 40 to 50 percent for new leases across January 2017 to June 2023, against internal analysis finding nearly 60 percent of final floor-plan prices within 2.5 percent of the recommendation and more than 85 percent within 5 percent. The record's own guidance is that raw acceptance understates influence while the spread metrics overstate independence, so deference belongs in a band rather than at a point. Scale figures are three different denominators and must not be mixed: 16 million-plus units describes the vendor's data reach including non-customers; roughly 4.5 million units across roughly 600 customers is a vendor-reported priced footprint whose attribution could not be confirmed in the source originally cited for it; and the White House Council of Economic Advisers separately estimated that at least 10 percent of all US rental units, and nearly one in four multifamily rental units, were priced with the software. Vendor performance claims — a marketed 3 to 7 percent revenue lift, and Greystar's report that its YieldStar buildings "outperformed their markets by 4.8%" in one downturn — are marketing and self-report, relayed by journalism, never independently measured.
Public scrutiny arrived from outside the industry and late. ProPublica's 15 October 2022 investigation by Heather Vogell documented the use of competitors' private lease data and found that in one Seattle ZIP code the ten biggest property managers ran 70 percent of apartments and all ten used the pricing software in at least some of their buildings — a concentration finding about the managers, not a claim that every one of those units was algorithm-priced. Senators Warren, Sanders and colleagues wrote to the company on 22 November 2022; ProPublica reported the Antitrust Division's investigation open the following day. In December 2024 the Council of Economic Advisers estimated that rental pricing algorithms cost US renters more than $3.8 billion in 2023, roughly $70 a month per unit in algorithm-priced buildings — model-based counterfactual estimates the council itself framed as a lower bound, because non-participating landlords also raised rents in response. These are contestable in method and are not measured overcharge. Private tenant class actions were consolidated in 2023 as In re RealPage Rental Software Antitrust Litigation (No. II), multidistrict litigation docket 3071, in the Middle District of Tennessee; on 21 November 2025 that court preliminarily approved 26 settlements involving 27 landlord defendants totalling $141.8 million for a class of renters of covered properties between 18 October 2018 and 21 November 2025, over objections from five state attorneys general, with a second batch of roughly $218 million announced in 2026.
The resolution is a sequence of consent decrees, none of them an adjudication. Cortland settled on 7 January 2025 (roughly 80,000 units across 13 states). The Department announced a proposed decree with Greystar on 8 August 2025 — no anticompetitive algorithms trained on competitors' competitively sensitive data, no vendor-hosted meetings of competing landlords, and a monitor if it uses uncertified third-party pricing algorithms; Greystar separately paid $7 million to nine states in a later multistate settlement announced 18 November 2025. On 24 November 2025 the Department filed its proposed decree with RealPage itself: a seven-year term, a court-appointed monitor for three years after approval with sweeping oversight, competitor data in models required to be at least 12 months old and not from active leases, no geographic modelling below state level for the existing demand and supply models, the Governor feature made symmetric, automatic-acceptance ranges made user-set and off by default, inspection rights, a written antitrust compliance program, and an obligation for the company to cooperate with the government's cases against the remaining landlord defendants — with no fine and no admission of liability. The Proposed Final Judgment and Competitive Impact Statement were published in the Federal Register on 5 December 2025 (90 FR 56286), opening the 60-day Tunney Act comment period; the notice published on 21 January 2026 is the LivCor decree, not this one. The stipulation was entered 26 March 2026. On 8 May 2026 the Department published its response to eight public comments, including the American Antitrust Institute's objection — argued in the institute's own commentary, whose "algorithmic cartel manager" phrase is the institute's and not the Department's — that the decree may not prevent the company coordinating common pricing rules. Final public-interest entry by the court remained pending as of July 2026. LivCor's decree followed in December 2025 and Willow Bridge's on 6 July 2026. Camden and Cushman & Wakefield/Pinnacle remained in litigation when the settlements were filed. In parallel and outside the courtroom, San Francisco enacted the first municipal ban in September 2024 (up to $1,000 per violation), Philadelphia followed in October 2024 (up to $2,000), and Minneapolis, Berkeley, Jersey City and Seattle followed in 2025; the company sued Berkeley on First Amendment grounds and initially fended off its ban. California enacted AB 325 on 6 October 2025 and New York enacted the first statewide ban on landlords using shared rent-setting software on 16 October 2025. The company denies wrongdoing, says its software was "purposely built to be legally compliant", notes acceptance rates under 50 percent, and attributes rent growth to housing undersupply — all vendor claims.
The sociotechnical reading
Almost every other cell in this Atlas has the same skeleton: a model scores a person, someone reads the score, and the argument is about whether they should have believed it. This one has no person-level decision at all. Renters are not scored, not notified, not able to appeal, and never touch the system — the affected population is entirely outside the loop, and the harm, if the government's account holds, is market-priced and diffuse rather than case-adjudicated. That single structural fact rearranges what governance can even mean here. There is no override to log because there is no adverse determination. There is no fairness metric to compute because there is no protected-attribute decision. The instruments that fit are the ones that act on structure: who may feed the model, how old what it reads may be, how coarse its geography may be, what it may write without a person, and who may talk to whom.
The mechanism worth learning from is a memory loop between rivals. Each operator's output became every operator's input: an accepted recommendation turned into an executed-lease record in the shared dataset, and the shared dataset shaped the next day's recommendation for all of them. Nothing in this Atlas's single-agency deployments has an analogue — those loops run inside one organisation's records, where at least one body owns both ends. Here the loop crosses the boundary between competitors, which means no single participant can see it, own it, or stop it, and every participant's honest local view is that they simply subscribed to good software. Two documented features give the loop a direction rather than mere noise. The price-movement band is alleged to have damped downward moves harder than upward ones, so what got executed and written back skewed one way. And the automatic-acceptance setting closed the loop with no person in it at all. A loop with a bias and no reconciliation step is a different object from a loop with error, and it is why the remedy reaches for the age of the data rather than the quality of the model.
The second structure worth naming is that the operator group closest to the model worked for the vendor, and its documented job ran the wrong way. Pricing advisors assigned to client portfolios measured how often recommendations were taken and pushed on-site staff back toward taking them. Everywhere else in this Atlas, the class nearest the model is the one expected to catch its errors; here that position was staffed by someone whose success was measured by adoption. Add the vendor-hosted forums where competing operators' pricing staff talked, and the peer pathways carry more of this network's behaviour than the recommendation edge does. That both channels are closed by name in the decrees is the strongest evidence available that both were carrying something — closing a channel is what regulators do to channels they think matter.
What the response teaches is the sharpest lesson of all: the regulators did not audit the model. They did not ask whether the recommended prices were right, and no independent evaluation of accuracy exists anywhere on this record — the only performance figures are vendor marketing. Instead they severed edges (no vendor-hosted meetings of competitors), staled the shared memory (a twelve-month floor, no active-lease data), coarsened its resolution (nothing below state level), re-symmetrised the damper (the band must treat increases and decreases alike), re-inserted the human (automatic acceptance off by default, ranges user-set), and grafted a monitor onto the vendor for three years. Read as governance rather than as law, that is a topology edit, and it is a demonstration that when the failure is coordination rather than error, accuracy is not the binding constraint. The honest boundary: none of this is adjudicated. Nobody admitted anything, no fine was paid on the federal decree, two landlord defendants were still litigating, and the decree itself was still awaiting the court's public-interest determination in mid-2026. The monitor has published no findings. Whether the remedy works is contested on the record by an antitrust institute and by five state attorneys general, and this Atlas does not resolve that dispute — it shows the pathways the decree edits, and marks the one instrument the decree does not contain: a reconciliation between what was executed and what was recommended, before the executed price becomes the record everyone trains on next.
The concepts used in this reading are defined in the Field Guide; the governance responses live in the Practice Library.