A GeoGenius study of how ChatGPT, Grok, Google AI, and Claude cite sources for real-estate queries. 720 probes, 4,455 citations, 1,002 unique domains — and only 1.9% of sources are cited by all four engines.
9 patterns repeated cleanly across markets, engines, and regions.
Only 19 of 1,002 cited domains (1.9%) are used by all 4 engines. The rest is engine-specific. This sample does not establish a universal ranking formula.
37.7% of ChatGPT citations are google.com — and 100% of those are Google Maps URLs (Business Profile pages), not Google search. Even more surprising: ChatGPT is the only engine that cites google.com. Grok, Google AI, and Claude all sit at 0%. This does not prove that Business Profile completeness causes inclusion or exclusion.
In rural markets ChatGPT’s GBP dependence jumps to 45% vs 33% in urban. When local website content thins, ChatGPT leans harder on Google Business Profile. This association does not isolate the effect of a listing change.
ChatGPT → Google. Grok → Zillow. Gemini → FastExpert. Claude → US News. Three of four are stable across all density tiers.
Despite being a top-10 source overall (#7), Realtor.com doesn’t lead for any individual AI engine. Brand recall ≠ AI citation.
The ONLY aggregators every AI engine cites are Zillow, Redfin, and Compass. Not Yelp, not Realtor.com, not HomeLight, not US News.
Every one of ChatGPT’s 359 Google citations is a Maps lookup for a named person or team, which let us recover exactly who it surfaced: 288 distinct agents and teams across 23 of the 36 markets. These URL-derived names do not reveal the model’s internal selection or verification process.
Individual agent websites are the LARGEST single citation source (36.8%). Combined with regional + national brokerage sites, agent-or-broker-controlled surfaces are 48% of all AI citations — bigger than every aggregator combined. This editorial classification describes the sample, not the effect of improving any one surface.
ActiveRain, LinkedIn, Medium, Inman, Nextdoor, X and TikTok were cited ZERO times out of 4,455. Reddit got 79. The engines cite consumer conversation and third-party evaluation — not agents publishing for other agents. And even that reaches only half the field: ChatGPT and Claude cited community sources 0 times.
Click any pin to see the top-cited sources for that market across all four AI engines.
A problem with your API key prevents the map from rendering correctly. Please make sure the value of the APIProvider.apiKey prop is correct. Check the error-message in the console for further details.
Aggregated across all 720 probes. Google leads, but only barely — the long tail is enormous.
4,455 total citations across 1,002 unique source domains.
Pick an engine to see its top 10 sources. Notice how little overlap there is between any two of them.
Out of 1,002 unique sources, only these 19 appear in every engine's citation pool. Here is the part nobody expects: only four are platforms or media. The other fifteen are individual agents' and teams' own websites.
Critical insight: The only aggregator platforms cited by every engine are Zillow, Redfin, and Compass. Realtor.com, FastExpert, Yelp, HomeLight, US News — none of them. Optimize the universal three first; engine-specific platforms second.
The bigger insight: the remaining fifteen aren't platforms you can join — they're agents who earned their way onto this list on a domain they own. Every one of them dominates exactly one market; not a single agent site appeared in two. There is no national winner here. And the most-cited individual agent in the entire study isn't in New York or Los Angeles — they're in Savannah. Burlington, Vermont placed two firms on this list; Boise placed two. National reach didn't decide it. Local density of evidence did.
Every one of ChatGPT's 359 Google citations is a maps/search/ lookup for a specific person — meaning ChatGPT had already chosen an agent and went to Google Maps to verify them. The names survived inside the URLs.
Interpretation: these citations link to Google Maps searches. They do not prove a hidden business-verification step or that an incomplete listing prevents a recommendation.
Smaller markets produced 108 named agents versus 87 in major urban markets in this sample. The observation alone does not establish why the counts differ.
Method note: this recovery is possible for ChatGPT only. Grok, Gemini and Claude cite the agent's own site or an aggregator, which doesn't carry the name in a parseable position. Entity names are taken verbatim from the citation URLs; brokerage suffixes were collapsed so "Jane Doe" and "Jane Doe – XYZ Realty" count once.
The "rural GBP tax" effect: as markets get smaller, ChatGPT leans harder on Google Business Profile, and the other engines lean on a narrower set of authoritative national sources.
We categorized all 1,002 cited domains into 11 buckets. The headline: individual agent websites are the biggest single category — bigger than every aggregator combined.
All four AI engines combined.
Agent-or-broker-controlled surfaces (highlighted in green): 48.08% of all citations. That includes individual agent websites (36.81%), regional brokerage/team sites (7.72%), and national brokerage chains (3.55%). Bigger than aggregators (24.71%).
As markets get smaller, individual agent sites decline and aggregators + GBP take over.
| Category | Major urban | Mid-size suburban | Small rural |
|---|---|---|---|
| Individual agent websites | 45.0% | 34.5% | 30.7% |
| Aggregators | 20.6% | 24.6% | 29.0% |
| Review/ranking sites | 13.9% | 11.7% | 9.5% |
| Google Business Profile | 6.1% | 7.8% | 10.3% |
| Regional brokerage / team | 6.1% | 9.0% | 8.1% |
| National brokerage | 2.7% | 3.4% | 4.5% |
| Community / forum | 1.2% | 2.5% | 2.5% |
| Other non-aggregator | 3.0% | 6.0% | 3.8% |
In urban markets, agent sites are 45% of citations. In rural markets that drops to 31% — and aggregators (+8.4 points), GBP (+4.2 points), and national brokerages (+1.8 points) absorb the difference. The fix for rural visibility isn’t just "build a site" — it’s a complete GBP plus a presence on the platforms AI defaults to.
Same 4,455 citations, sliced four ways. The differences are dramatic.
Almost never cites aggregators. Citations come from agent-owned web presence — direct sites, brokerage pages, and Google Business Profile.
Spreads citations across categories more evenly than any other engine. Reddit + Quora drive forum share.
Leans hardest on aggregators (Zillow, Realtor.com, FastExpert) and surfaces national brokerage chains more than any other engine.
Leans hardest on third-party review and ranking sites (Yelp, Expertise.com, US News). Less agent-direct than ChatGPT but more curated than Google AI.
Green bars are agent-or-broker-controlled surfaces. Across every engine, those categories combined exceed any single non-agent category — which is the underlying signal behind finding #08.
Start with a free website-readiness diagnostic. Fresh account-based recommendation checks are separate observations; this historical study is not a list of domains that cite your business.
No credit card · Website diagnostic · Recommendation checks require an account
Open data. Reproducible. We welcome scrutiny and follow-up research.
| Engine | Attempted | Completed | Failed |
|---|---|---|---|
| ChatGPT | 180 | 180 | 0 |
| Grok | 180 | 179 | 1 |
| Google AI | 180 | 175 | 5 |
| Claude | 180 | 88 | 92 |
Sample: 36 U.S. markets stratified across three density tiers (12 each) and four regions (Northeast, South, Midwest, West).
Queries: 5 standardized natural-language queries per market — generic discovery, seller intent, buyer intent, luxury specialty, and local search.
Engines & models: ChatGPT (OpenAI Responses API, gpt-4o-2024-08-06 + web_search), Grok (xAI Responses API, grok-4-fast-non-reasoning + web_search), Google AI (Vertex AI, gemini-2.5-flash with Google Search grounding), Claude (Anthropic Messages API, claude-sonnet-4-6 + web_search).
Total volume: 622 completed and 98 failed attempts out of 720 probes (36 × 5 × 4), 4,455 citations, 1,002 unique source domains. Deduplicated to exactly one probe per (market, engine, query) tuple.
Success rates: ChatGPT 100% · Grok 99% · Google AI 97% · Claude 49%. Claude's lower success rate is an artifact of our own API provisioning during the run, not model behavior: 84 of its 92 failed probes returned an account credit-balance error and 7 more hit a rate limit. On the probes Claude did complete, it returned citations 87 of 88 times. We disclose this because it means Claude's citation pool (556) is smaller than the other engines' and its shares carry greater uncertainty; no confidence intervals are supplied here.
Aggregation: All citation URLs were normalized to their registrable domain (e.g., www.zillow.com/agents/... → zillow.com) before aggregation.
Reproducibility: Every figure in the headline findings, engine, market, density, region and crossover sections is computed directly from the probe-level and citation-level exports — no figure is entered by hand. The named-agent counts are derived from ChatGPT's Maps citation URLs by a script committed alongside the data. The one exception is the 11-category source classification below: grouping 1,002 domains into categories requires judgment calls no script can make on its own, so treat those percentages as a considered editorial grouping of the same citation set rather than a purely mechanical result.
Versioned public data: download the sanitized probe table, citation-domain counts, summary and model versions, and data dictionary, reuse terms and limitations. Version ai-citations-2026-v1 omits raw provider errors, full URLs and named-agent details. Its reduced tables reproduce headline counts, not the named-agent or editorial classification analyses.
Report generated May 14, 2026. This is the source report generation date, not the observation interval. The saved exports lack per-observation timestamps; exact collection dates are unknown in this release.
A quick note on cookies
We use cookies for analytics — to understand which pages help real estate agents and to improve the product. No ads, no selling your data. You can change your mind anytime in our privacy policy.