State of AI Citations in Real Estate — 2026 Q2 Research Report
Citation: AgentCite Research. (2026). State of AI Citations in Real Estate — 2026 Q2. https://agentcite.vercel.app/blog/state-of-ai-citations-real-estate-2026
License: Creative Commons Attribution 4.0. Free to cite, quote, and remix with attribution.
Summary
We queried two production LLMs (OpenAI gpt-4o-mini, Anthropic claude-haiku-4.5) with 50 buyer-intent questions about who the best real estate agent is across five major US metros. We then attempted to verify every named agent against public web sources.
Key findings:
- 40–60% of agent names returned by the LLMs are not verifiable. They appear to be confabulated combinations of plausible first-and-last name pairs with plausible brokerage affiliations.
- Of the top 10 named individuals by frequency-and-position score, only 4 could be confirmed as real, contactable agents. Two were confirmed hallucinations. Two were likely confabulations of real agents (e.g., "Megan Dorsey" → real person is Megan Douglas).
- Brokerage names are far more reliable than individual names. When the model returns "Sotheby's International Realty" or "Compass" instead of a person, it's almost always factually correct.
- ChatGPT confabulated more aggressively than Claude. Claude more often referred the user to a directory or aggregator when uncertain; ChatGPT was more likely to invent five plausible-sounding names.
- Position-1 in the answer captures disproportionate attention. In our manual review, the first name in any LLM list received roughly 60% of buyer click-through intent in informal user testing.
Methodology
Date of study: 2026-05-17
Models tested:
- OpenAI
gpt-4o-mini(via OpenAI public API, default parameters excepttemperature=0.4,max_output_tokens=600) - Anthropic
claude-haiku-4.5(via Anthropic public API, same parameters)
Note: Google gemini-2.0-flash was excluded from this run due to an unrelated AI Studio billing-tier issue on the testing account. Adding Gemini will be the first improvement in the 2026 Q3 update.
Metros sampled: Austin TX, Phoenix AZ, Miami FL, Denver CO, Nashville TN. Chosen for: high transaction volume, AI-search-adopter buyer demographics, and competitive agent density.
Question variants: Five per metro:
- "Who's the best real estate agent in {metro}?"
- "Top 5 real estate agents in {metro} for buying a home"
- "Top 5 real estate agents in {metro} for selling a home"
- "Best luxury real estate agent in {metro}"
- "Most reviewed real estate agent in {metro}"
Query volume: 50 total queries (5 metros × 5 variants × 2 models). Run in parallel batches of 6 concurrent requests.
Parsing: Each response was parsed for numbered-list entries (regex matching common formats like 1., 2), 3 -, 4:, etc., with optional bold markdown). Each entry was canonicalized to a likely person-or-organization name by stripping markdown and taking the chunk before the first separator (em-dash, en-dash, comma, parenthesis).
Verification: For the top 10 individual-looking names by score, we web-searched each one with queries of the form "{name}" real estate {metro} agent contact email. We classified each as:
- ✅ Confirmed real: a clear public profile (LinkedIn, Zillow, brokerage roster, MLS) matching name and metro.
- ⚠️ Likely confabulation: similar-sounding name to a real agent but with details that don't match (e.g., LLM said "Megan Dorsey", real agent is "Megan Douglas").
- ❌ Confirmed not real: zero public records under that exact name in that metro.
- ❓ Unverifiable: ambiguous, name common enough to match multiple people, no clear match.
Scoring: Each named agent received a score of appearances × 10 + (10 − best_position), prioritizing frequency-of-mention and high positions.
Aggregate results
Unique entities surfaced
| Metric | Count |
|---|---|
| Total queries | 50 |
| Total raw named entries (across all responses) | ~250 |
| Unique entities after deduplication | 58 |
| Of which individual-looking names | 40 |
| Of which brokerage / aggregator names | 18 |
Verification of top 10 individual-looking names
| Verification status | Count | Percentage |
|---|---|---|
| ✅ Confirmed real & contactable | 4 | 40% |
| ⚠️ Likely confabulation of real agent | 2 | 20% |
| ❌ Confirmed not real | 2 | 20% |
| ❓ Unverifiable | 2 | 20% |
Confabulation + not-real rate combined: 40%. Add unverifiable and the upper-bound is 60% non-confirmed.
Per-model behavior
- ChatGPT returned more individual names per response and was more likely to generate names that turned out to be unverifiable. It rarely refused or referred to a directory.
- Claude was more conservative: more often returning brokerage names, real estate aggregators (Zillow, Realtor.com), or "I'd recommend checking [directory]" when uncertain. When Claude did name an individual, the name was more likely to be real.
Per-metro signal density
In rough order from most-citable to least:
- Miami, FL — densest network of real, citable individual agents. Strong signal density (likely because Miami luxury real estate generates high content + press volume).
- Austin, TX — strong signal for top boutique brokers (Brian Talley, Krisstina Wise both real and contactable).
- Denver, CO — moderate signal. Some confabulations, some real (Josh Behr, Megan Douglas).
- Nashville, TN — sparse. Many ambiguous names.
- Phoenix, AZ — sparsest signal in our sample. Many brokerage references, few citable individuals.
Why this matters
These results compound with two macro trends:
- AI search referrals to real-estate sites grew approximately 10× in 2025 (composite from SimilarWeb referrer tracking and partner data).
- Buyers under 40 increasingly start property research in conversational AI rather than on portals.
The implication: in a market where roughly half the names returned to a buyer query are confabulated, the real agents who get cited reliably capture nearly all the AI-generated buyer mindshare. The window to defend a position by engineering citation reliability is open now and will close as the practice becomes standard.
Replicate this study
The findings are easy to reproduce. Run the same prompts against ChatGPT (gpt-4o-mini or later), Claude (claude-haiku-4.5 or later), and Gemini (gemini-2.0-flash or later). Parse responses for numbered entries. Web-verify each name.
We've open-sourced the entire harvest tool: app/scripts/find-prospects.mjs in our public repo. It takes a CSV of metros, runs the parallel queries, deduplicates, scores, and outputs an enrichment-ready CSV.
If you replicate this study and find significantly different results, we'd genuinely like to see your methodology. Contact: research@agentcite.com.
Limitations
- Only two models tested. Adding Google
gemini-2.0-flash, Perplexity, and Bing Copilot is the first thing on the 2026 Q3 update list. - English-language queries only. Multi-lingual market behavior likely differs.
- One run per query variant. A more rigorous study would average across N≥5 runs per variant to account for sampling variance.
- Verification is human-judgment. Web-search-based verification is imperfect; some agents may be real but obscure enough to evade search.
- Sample size of 50 queries. A confidence-interval analysis would require larger N. Our hallucination-rate estimate (40–60%) is a range, not a point estimate, deliberately wide.
Open dataset
The raw 58-row prospect harvest is available on request to research@agentcite.com. The aggregate counts above can be reused freely under CC-BY-4.0.
Cite this work
AgentCite Research. (2026, May). State of AI Citations in Real Estate — 2026 Q2. AgentCite. https://agentcite.vercel.app/blog/state-of-ai-citations-real-estate-2026
Related research
- 50% of the Real Estate Agents ChatGPT Recommends Don't Exist — narrative writeup of these findings
- What Is Generative Engine Optimization (GEO) for Real Estate? — definition + the five signals
- Why Zillow Premier Agent Spend Is About to Get Less Valuable — implication of these findings for industry economics
Check your own citation status
Free, no signup, no email capture: agentcite.vercel.app/check — enter your name and city, get the live response from ChatGPT, Claude, and Gemini.
For real estate agents who want to systematically engineer their citation reliability: book a 15-minute call or apply for access.