How AI Real Estate Agents Outperform Human Persuasion: Insights from Recent Research

Today, large language models (LLMs) have the capability to write texts, engage in conversations, and tackle intellectual challenges. However, their persuasive abilities are still not thoroughly understood. The art of persuasion is at the heart of numerous economic processes; it is estimated that up to 30% of the U.S. GDP is generated through advertising, lobbying, negotiations, and other forms of communication where “selling” an idea or product is crucial.

Nonetheless, the mass production of advertising content poses significant risks, such as the dissemination of misinformation or the manipulation of public opinion. This is particularly critical in fields where accuracy and reliability are paramount—such as real estate advertising. Clients take home purchases very seriously, necessitating descriptions that are not only emotive but also strictly factual.

The authors of the study “Grounded Persuasive Language Generation for Automated Marketing” propose using a large language model for the automated creation of marketing materials. Their approach consists of three key modules:

Justification Module. This focuses on breaking down the text into specific, measurable parameters of the object. Essentially, it mimics an expert’s logic and highlights market advantages (such as actual square footage, location, and finishing features).

Personalization Module. This enables the tailoring of the generated text to the tastes and interests of individual buyers. For some, the convenience of transportation links is paramount, while for others, the amount of sunlight in the rooms is more important.

Marketing Module. This ensures that the text retains factual accuracy while incorporating unique characteristics (those that are rare in the area or price range). Such details enhance the attractiveness of the offer.

The core idea is grounded in the theory of strategic communication: how to convey exactly the signals that will most effectively influence a real estate buyer’s decision, all while staying factual.

The researchers gathered over 50,000 listings from Zillow. Each listing contains multiple attributes, including size, cost, number of rooms, and so on.

Using the LLM, key features were automatically extracted from the texts. Based on this data, a hierarchical feature schema was built—a sort of “framework” that describes the variety of marketing advantages, from “proximity to the subway” to “prestigious neighborhood.”

Next, the researchers trained a simple neural network that translates “raw” factual attributes of the property (e.g., 80 square meters, 5-minute walk to school) into more abstract marketing features (“spacious layout,” “close to schools”).

Real individuals’ preferences were studied, noting the importance of certain attributes (like “clean entrance” or “view of the park”). This was utilized to allow the system to fine-tune the text for each specific client.

The authors tested how persuasive their model was compared to traditional real estate descriptions. To ensure objectivity, they employed an Elo rating system and conducted additional checks for inaccuracies.

It turned out that the AI realtor, based on this approach, received a higher rating compared to texts from professional realtors. This indicates a strong potential for LLMs in generating genuinely professional advertising offers with factually accurate descriptions.

As new modules were added—moving from justification to personalization and then to marketing—the quality and persuasiveness of the outcomes consistently improved.

Fact-checking revealed that the model seldom distorted data about properties. In contrast, human-written texts frequently contained inaccuracies, such as inconsistencies between stated square footage or room counts and actual parameters.

The authors also attempted to model user behavior using their AI system. However, while the model could partially replicate human assessments, there were instances where individual preferences were difficult to replicate. This suggests that while the simulation shows promise, it still has room for improvement.

Large language models have already learned the art of persuasion, and the real estate sector has served as an excellent testing ground for this. The researchers succeeded in creating an AI system capable of automatically generating “engaging” texts, which were rated higher than those produced by professional realtors.

This type of system can operate on large data volumes, generating content en masse. This has the potential to reduce costs associated with creating advertising and marketing texts, and allows for personalized offerings for each client.

It’s essential to remember that if such technology is misused, it can manipulate public opinion. If the system uses distorted information, there is a risk of losing trust in these types of services.

If you’re interested in AI topics, subscribe to my Telegram channel — where I regularly share insights on AI implementation in business, launching AI startups, and explain how all these AI wonders work.