How ChatGPT assembles a home search and sale

The research and guidance module contains 52 answers across six questions: eight each on neighborhoods, flood/insurance, valuations, places to live and independent guidance, plus 12 on first-time-buyer guides. The requested help ranges from narrowing a location to understanding a transaction. A named agent is often unnecessary to satisfy it.

Neighborhood research combines evidence with different meanings

The neighborhood question starts with Zillow four times, Redfin twice, Niche once and NeighborhoodScout once. The final focused toolkit has a different ordering: Redfin leads seven and Niche one. Five answers change that first position within the response.

Redfin, GreatSchools, Walk Score and FEMA each receive substantive roles in all eight answers. They address market information, schools, transport and risk rather than compete for one job. Zillow appears substantively in five, not all eight. Every answer calls for checking school boundaries with the district and using local police information; seven explicitly advise visiting in person.

Three answers offer illustrative weighted decision matrices. Their weights differ, although each totals 100%. These are model-generated planning examples, not observed preferences or Gist's recommended neighborhood-scoring method. One says a severe flood or insurance problem can remain a dealbreaker despite a high overall score. The answers themselves therefore resist treating a composite rating as complete evidence.

Flood risk, insurance availability and price are separate checks

FEMA is first-listed in all eight flood/insurance answers. All eight also retain First Street/Risk Factor/Flood Factor as complementary modeled exposure, local or seller history, and property-specific insurance quotes. RE22-E01 explicitly separates risk, availability and affordability. No answer treats the official map designation as sufficient.

The source roles matter. The official designation, a model of potential exposure, local flood history and a quote for actual coverage answer different questions. Two answers put NFIP/FloodSmart before First Street in the numbered menu; that is a program/information role rather than another independent insurer.

The limits are not identical in every response. Seven discuss CLUE, while five explicitly limit access to individual NFIP claim histories. Four tie the checks to a due-diligence or contingency period. The report records those qualifications as dated model statements and does not adopt their insurance or legal details as current guidance.

The citations add a separate observation: content.naic.org appears in all eight saved citation lists even though the ranked advice begins with FEMA. One answer does not name NAIC in its prose. Conversely, modeled-risk advice appears in all eight while help.firststreet.org is cited in three. A citation is neither a complete inventory of recommended resources nor a sentence-level attribution.

Exhibit 6. Six research tasks, six different kinds of help

Scroll sideways to read all table columns.

Full-response observations within the exact questions above. Resource features, risk/legal statements, estimate accuracy and current counseling arrangements are not independently verified by this study.

An estimate supplies context; a price strategy remains a different output

Zillow is first in all eight home-value answers. All eight combine Zillow, Redfin and Realtor.com with a local comparative market analysis, or CMA. Two begin the closing ordered workflow with Redfin's comparable-sales tools rather than Zillow's estimate. Every answer warns against using an automated average or range as the final asking price.

RE23-E03 says:

“Use Zillow, Redfin, Realtor.com, and Chase together, but don't set your asking price based solely on their average.”

Chase appears in seven answers. The exception develops Redfin Owner Estimate, which is a feature of Redfin rather than an additional independent provider. Three answers explicitly discuss shared data or error dependence. RE23-E07 later calls the estimates “independent” despite an earlier shared-data warning; the report retains that terminology tension rather than inferring independent corroboration from separate logos.

The numerical accuracy claims also vary. One answer gives a Redfin off-market error rate of 7.28%; three give 7.46%. Those are model-stated figures, not accuracy measured by Gist or verified current documentation. All eight use hypothetical property examples rather than sales observed in this study. The research can show that the advice distinguishes estimates from pricing, but cannot conclude that the cited website or a local CMA is empirically more accurate.

A guide publisher and the first item to read can differ

CFPB is the first publisher in 11 of the 12 guide answers. The closing reading shortlist starts with Freddie Mac in seven, CFPB in three, Fannie Mae in one and NerdWallet in one. Eight answers change the first publisher between the opening and closing lists.

The exception is substantive: RE24-E03 puts NerdWallet's step-by-step guide first and keeps it first in the final three. Other answers use NerdWallet documents for tips or mistakes and Realtor.com for searching, offers, agents or inspections. Commercial publishers therefore remain part of the advice; an “only public resources” story would be false.

CFPB and HUD each receive developed resource recommendations in all 12 answers, Freddie Mac in nine, Fannie Mae in five, NerdWallet in four and Realtor.com in one. Different documents from the same publisher are not counted as separate organizations. All 12 distinguish borrowing approval from personal affordability and discuss ownership costs beyond a mortgage payment. The report does not independently validate the outputs' course, program, legal-timing or article-update claims.

Research can precede professional selection without replacing it

All eight place-comparison answers put location research before selecting or interviewing an agent; none selects a named individual. Niche and BestPlaces are substantive options in every answer. Niche leads five initial lists and BestPlaces three, while the final first-position split is four each. The order changes in one response rather than producing a uniform winner.

The research sources themselves have different status: official statistics, commercial ratings and resident discussion are not interchangeable evidence. Only two answers explicitly warn that relocation guides may be promotional, and only three explicitly advise visiting. Those useful qualifications should not be generalized to all eight.

Independent guidance introduces a further distinction between a gateway and a service. HUD leads all eight answers and is paired with CFPB in all eight. NeighborWorks and state housing-finance agencies also have roles in all eight, while particular nonprofit organizations appear as examples in four. These are not four unique final counseling selections.

Three step-by-step endings start with reading CFPB material before contacting a counselor, even though HUD heads the opening list. Every answer qualifies fees or service availability; a statement that all counseling is free would overstate the text. The stable advice combines education, finding an appropriate local service and checking eligibility. It satisfies a different need from an agent or brokerage recommendation.

Local shortlists change while agent-fit advice persists

Question group · ChatGPT · October 19, 2026 Pacific collection. Each n is saved answers to that exact question. RE01–30 abbreviate RE-B1-01–30.

Scroll sideways to read all table columns.

The local module asks the same first-time-homebuyer question for Los Angeles, Chicago, Atlanta and Dallas. The complete cells contain eight, eight, 12 and eight ChatGPT answers, respectively: 36 answers in total. All name people or teams as well as organizational routes. Every full response is an interview menu conditioned on fit, not an exclusive final hiring decision.

The city is explicit prompt text. The study did not set or test GPS/IP location, Maps ranking or a representative sample of each city's agents. Names, affiliations, experience claims and performance descriptions below are reported only as they appear in the saved outputs; they have not been independently verified.

In Los Angeles, Compass leads five initial lists and Redfin three. In Chicago, Baird & Warner leads seven and @properties one. Ebby Halliday heads all eight Dallas lists. Atlanta's initial entries are less uniform in type: five answers start with an individual, one with a team and six with a brokerage. Even the entity level being ranked varies across that cell.

The more stable instruction concerns the actual buyer agent. Every answer in each city emphasizes the person, neighborhood, budget or first-time-buyer fit. RE28-E06 says:

“An agent's recent transactions, responsiveness, references, and familiarity with your specific neighborhood matter more than brokerage branding.”

This is the saved model's evaluative advice, not independent evidence that the named candidates meet those criteria.

The first answer is a fragile roster

The expanded reading materially changes the account of named local recommendations. Several individuals featured in the first answer appear only there, even when a brokerage or route remains prominent in later answers.

Exhibit 7. Selected candidates from the first local answer, checked across the complete cell

Scroll sideways to read all table columns.

Purposive audit of selected names in the first response, not an exhaustive ranking of local agents. Counts mean substantive appearances in the repeated advice and do not verify identity, affiliation, quality or availability. People and teams remain different entity levels. Exact city questions appear above; supporting IDs are retained in the local reading codes.

The Chicago case prevents an exaggerated “all names are unstable” conclusion. Becky Sexson recurs in six of the seven later answers, for seven of eight overall. Other first-answer candidates in the same cell appear only twice. The observation is variation in candidate persistence, not a universal failure to repeat names.

Atlanta supplies the strongest first-answer limitation. The four selected Ansley-affiliated individuals in the initial answer each occur only in that response. Later answers introduce different individuals, teams and brokerage routes. A feature built solely around the initial four would make a momentary roster look like the study's settled account of local advice.

A brokerage heading and the interview list can have different orders

The ending adds a second distinction beyond recurrence. RE30-E06 opens with Ebby Halliday but closes its interview list with Kathy Beck/Keller Williams first, Angie Rodgers/Compass second and an Ebby specialist afterward. In Atlanta, RE29-E08 opens with Atlanta Intown but begins the final interview list with Mike Lacy. These are ordered conditional menus, not necessarily incompatible recommendations.

The Dallas answer itself says:

“This is an interview shortlist, not a verified ranking of the five best agents.”

It also explicitly leaves current availability, rates and recent first-time-buyer results unverified. Treating the initial brand or final first person as a validated local winner would override the answer's own boundary.

Affiliation adds another counting problem. The outputs explicitly present Ebby Halliday and Dave Perry-Miller as sharing a parent; the study does not independently confirm that statement. Multiple Compass-affiliated people or teams are not automatically multiple competing brokerages. Conversely, folding every person into a parent brand would erase which actual candidate the reader is told to interview.

For a multi-location brokerage or local professional, the research therefore calls for more than a city-level name count. The useful observations are the entity level, the person or team identified, the role in the proposed interview process, the recurrence across full responses and the qualifications surrounding the recommendation. The module establishes those observed patterns; actual professional fit requires evidence beyond the generated shortlist.

Sources: the question changes which organizations supply the recorded links

The source analysis covers the full 280-answer ChatGPT collection across all 30 questions. It measures whether a saved citation list includes a host at least once in an answer. A host is the address in the URL, with a leading “www.” removed; other subdomains remain separate. This is source participation, not a ranking of recommended agents or a claim that a page caused the advice.

Within that definition, Zillow's main host has the highest equal-question citation rate, followed by Redfin, Compass and HomeLight. Redfin appears somewhere in the source lists of more distinct questions than Zillow, even though Zillow's average within-question presence is higher. Frequency within a task and breadth across tasks describe different patterns.

Exhibit 8. Leading citation hosts under two weighting choices

Scroll sideways to read the full chart. Exact data and qualifications follow.

Scroll sideways to read all table columns.

The six leading normalized hosts under equal question weighting. That column averages the citation-presence rate within each of the 30 questions. Pooled weighting gives each saved answer equal weight, so questions with more repetitions contribute more. Other subdomains, including support.redfin.com, remain separate. Neither column is a recommendation rate or audience share.

The first two stay in that order under both weights. HomeLight rises from fourth to third when answers have equal weight; Compass moves from third to fourth. NAR's host also moves ahead of Realtor.com's. These changes do not make one formula wrong. They show why a panel with unequal repetition needs an explicit weighting rule.

The operational account was anchored on Redfin. This report does not use its tracked-brand scores, and it has not tested whether the account anchor affected the answers. The comparison therefore describes the archived collection, not an anchor-free experiment.

The whole-panel rate conceals different source settings

A source can be concentrated in the part of the panel where its function is directly relevant. HomeLight leads the agent-selection questions; NAR's host leads representation; CFPB's host leads research and guidance. Compass leads the four city questions. These are differences in the saved citations across different tasks, not causal effects of changing a few words.

Exhibit 9. The most frequent source host differs by question group

Scroll sideways to read the full chart. Exact data and qualifications follow.

Scroll sideways to read all table columns.

Editorial task groups partition the questionnaire. “Leading” means the highest citation-presence rate within that group, not the most suitable service. Exact questions precede their findings and appear together in the appendix. Full host tables, including all other sources, accompany the report.

The selling result illustrates concentration. Homecoin's leading group rate comes from citation presence in two of the five questions, not all five. In the research group, CFPB's host appears in three of six questions. A group average should not erase the tasks in which a source is absent.

Some domains also have related subdomains with substantial presence. support.redfin.com occurs in 30/280 answers across ten questions, separately from redfin.com. Adding the two numerators would double-count answers that contain both. An ownership-family analysis would require a declared mapping and a new union calculation; this report's whole-panel table intentionally uses hosts.

A repeated source list is not another independent endorsement

All 280 answer bodies are textually distinct. None of the 30 question cells uses a single unchanged unordered citation set throughout its repetitions. That does not establish independent retrieval. The agent-review question, RE09, has eight distinct answers but only four exact citation sets. Other cells reuse some sets while changing the prose.

Reading only the first response would also understate source breadth. Redfin's main host appears in six first-answer lists, but somewhere in 14 of the complete question cells. Zillow's corresponding breadth expands from nine to 12 questions. A first-answer count and complete-cell breadth have different opportunities to observe a source; they should not be placed on the same scale as if they were interchangeable rates.

Two useful bounded relationships survive the expanded audit. Zillow- and Realtor.com-family links appear in all eight RE09 agent-research answers. Houzeo-family links appear in all eight RE16 flat-fee MLS answers. The first concerns research tools; the second concerns a provider that can be both a discussed option and a recorded source. Neither relationship demonstrates sentence-level attribution, source neutrality or causal influence.

The contribution of the source analysis is therefore specific: it identifies where an organization's pages participate, how broadly, and how sensitive the comparison is to the panel's repetition. The semantic chapters establish what the answers tell readers to do. The two analyses should be read together without substituting one for the other.

What the complete panel establishes

The national and local chapters describe a recurring separation between access and a decision the reader can rely on. Portals expose listings; matching services introduce candidates; MLS vendors provide distribution; financing services help bridge timing. The later checks concern the actual property, professional, agreement or obligation.

The pattern is specific to these questions and outputs. It is not evidence that every homebuyer follows the same journey or that the sequence generated transactions. It does provide a more complete account of where organizations participate than a pooled recommendation score.

Verification is distributed across the transaction

Freshness advice sends a portal user toward another source or a local agent. Agent-review advice combines transaction records and client reviews while questioning individual/team attribution. New-construction advice narrows a national builder to a community and actual construction team. The local shortlists still require interviews and verification of professional fit.

These checks do different work. Treating every named resource as an alternative provider would conceal why it appears. A company can participate as an initial destination, a benchmark or a verification resource, and those functions may coexist in one answer.

Control takes several forms

The privacy question favors browsing before introductions. Builder-independent representation prioritizes advocacy separate from the sales team. Flat-fee MLS answers distinguish exposure from professional transaction help. Buy-before-sell and preparation answers preserve repayment, fallback and timing obligations.

“More control” is consequently not one outcome to assign to one platform. It can mean choosing who receives contact details, choosing who represents the buyer, retaining work instead of outsourcing it, or deciding which financial obligation to accept. The report identifies those distinctions in the advice; it does not judge which arrangement is best for a particular household.

HomeLight demonstrates why role-level analysis matters even within one familiar organization. It appears in matching, relocation, a cash-buyer network and buy-before-sell contexts. The preparation cell sometimes explicitly distinguishes its equity/moving service from dedicated repair finance. Combining those appearances into one undifferentiated recommendation rate would lose the service the answer is actually discussing.

Repetition changes some findings and reinforces others

The first response often captures the broad division of roles. It is less reliable as a final roster or a complete account of qualifications. Homeward is absent from the first buy-before-sell answer but appears in 12/16. Several local individuals appear only in the initial response, while Chicago supplies a clear counterexample of persistence.

Exhibit 10. What survives the broader reading, and what needs a boundary

Scroll sideways to read all table columns.

All counts are within the question cells developed in the report or the complete ChatGPT source audit. No claim is made about a representative consumer population, causal wording effects or verified provider performance.

Implications for operators, publishers and research teams

A brokerage reviewing its visibility should distinguish its own services, partner introductions and the particular agent or team named. An agent-matching service should inspect whether it is positioned as the first shortlist generator, a proposal-comparison route or a later fallback. A seller-service company should retain the explanation of who performs the work and which obligations remain with the homeowner.

For publishers and information providers, the source tables identify where recorded links appear, while the semantic chapters show the surrounding advice. Public guidance, a provider's own program page and a commercial comparison resource are not automatically equivalent sources. Citation frequency alone neither proves independence nor explains why the answer recommended something.

For a research team, the expanded reading supplies a practical minimum: preserve the exact question, initial option, complete recommended action, entity role, material conditions and strongest exception. Use source presence as a separate observation, with its own unit and weighting rule. This study has not tested whether editing a page changes future answers or business outcomes.

The report's substantive conclusion is that ChatGPT's real-estate advice distributes a move across different organizations and checks. Familiar names can recur while the person, service, final priority or obligation changes. Understanding those differences is the research contribution; resolving the missing Google comparison remains further work.

Methods, definitions and limits

Panel and collection

The approved panel contains 30 analyst-authored, unbranded English questions for an intended United States context. Twenty-six concern home search, agent selection, representation, selling routes or research; four ask the same first-time-buyer question for Los Angeles, Chicago, Atlanta and Dallas. It is not a representative sample of search demand, a category census or a causal wording experiment. The questions deliberately ask for different services and kinds of help.

Gist's native Deep workflow collected the original batch on October 19, 2026 Pacific / October 20 UTC. This edition selects the 30 completed ChatGPT runs and all 280 nonempty completed answers, with 8–16 answers per question. Each record has a question ID, run identity, epoch index, response hash and saved citation list. “Epoch” is the archive's within-run answer index, not a separate consumer or survey respondent.

All 30 selected native runs show Done and Converged. The stopping display reports convergence under its brand-based procedure, with a 95% threshold; it does not establish semantic agreement, accurate advice, complete market coverage or statistical independence. Unequal repetition is part of the collection design and is retained in every cell denominator.

“ChatGPT” is the provider label in the saved exports. The exact underlying model/version, search configuration and complete generation parameters have not been established from those exports. Provider locale was not exposed in the recorded launch controls. A city named in a question is prompt context, not proof of GPS/IP localization or a test of Maps placement.

Execution used a Redfin-anchored account. The report analyzes raw answer text and citations; it does not adopt the account's tracked-brand scores as category standings. The influence of the anchor is untested. Claims apply to this saved collection, not every ChatGPT interface or current output.

Complete-response reading

The previous foundation semantically reviewed the first answer in each question for each of two engines. Its files remain preserved. This edition reads and codes all 280 ChatGPT answers under one shared real-estate rubric, including the 30 previously read first answers. It therefore adds 250 previously unreviewed ChatGPT responses and rereads 30, without importing Perplexity evidence into any count.

One primary AI coder reads each assigned group, retaining the response identity/hash, salient entities, role, stance, conditions, conflicting signals, uncertainty and exact supporting passages. The report author develops the synthesis from those records and the saved source audit. This is not a blinded human coding study or evidence of measured intercoder reliability. Any independent recheck is disclosed with its actual scope in the review receipt.

The salient entity register is purposive: it preserves organizations and people important to the observed advice rather than claiming exhaustive extraction of every name. Every reported semantic fraction defines the feature and uses the complete stated question cell. No pooled all-category recommendation ranking is computed.

Roles and decision rules

A portal or directory is not interchangeable with an agent, team, brokerage, referral service, public counselor, cash buyer or financing/coordination service. One organization may occupy more than one role. Affiliations in a model answer are preserved as claims; related names are not automatically merged or counted as independent competitors.

“First-listed” means the first substantive main-list option, excluding introductory examples. An explicit overall label is a narrower textual observation than an exclusive whole-answer choice. A conditional recommendation depends on a stated criterion, and a shared menu retains more than one reasonable route. The closing action may differ from the initial ranking because the answer changes criterion. A contradiction requires incompatible advice or claims, rather than merely separate alternatives for different needs.

Citations are coded separately from recommendations. Source presence does not show that the page supports a particular sentence, caused a recommendation or generated attention or business. The study has not inspected each live page or independently evaluated its accuracy, ownership or commercial neutrality.

Source measures and sensitivity

The host audit covers every saved citation list. Hosts are lowercased and a leading “www.” is removed; other subdomains remain separate. A host counts once per answer regardless of its number of URLs. Bounded domain-family checks use the named principal domain and its subdomains. No other ownership is inferred.

Equal-question source rates first calculate presence within each question, then average the rates with equal question weight. Pooled-answer rates show how unequal repetitions change the comparison. Question breadth counts how many of the 30 cells contain the host at least once. First-answer breadth is separately labeled and is not equivalent to complete-cell breadth.

Text uniqueness uses the saved response hashes. Citation-set comparisons ignore order and duplicate exact URLs but retain query parameters. Distinct answers and changing source lists do not prove independent retrieval. The report reports descriptive fractions and weighting sensitivity, not population confidence intervals, market shares or causal effects.

Google coverage remains unresolved

No AI Overview answer enters this edition's findings or denominators. The original wave eventually had six completed and 24 failed runs, with empty records and visible missing-list or incomplete content among the wider archived evidence. A later 24-run recovery wave ended with all 24 failed; its terminal snapshot also contained changing record states and eleven native-counter/message-count discrepancies. Those waves are preserved separately rather than pooled into a synthetic sample.

Failed, empty or incomplete captures are not evidence that Google gave no useful answer or recommended no organization. Even substantive fragments from failed runs do not establish comparable completed coverage. The cause of the capture problems remains unresolved, and this report does not claim that a parser change elsewhere repaired the measurement worker.

Google's absence limits any comprehensive account of AI-mediated home search. It does not prevent a bounded report about the completed ChatGPT collection. Perplexity is also outside this edition; the earlier two-engine evidence and report remain preserved.

What this report does not verify

The report describes dated model statements about services, people, policies and risks. It does not establish current brokerage affiliations, agent performance, availability, pricing, privacy practices, financing terms, legal duties, insurance coverage or government-program eligibility. No consumer transaction, application, contact form or local service was tested.

Where the outputs make those claims, the research retains their meaning and uncertainty. Before using them as current consumer advice, a separate source and service verification would be required. The study also does not measure whether a mention or citation caused a visit, lead, booking, sale or revenue. These are boundaries on interpretation, not missing values to be coded as zero.

Full questionnaire and answer counts

Exact approved execution text, ChatGPT only. Each row has complete-response reading coverage. RE01–30 abbreviate the portable question IDs RE-B1-01–30.

Scroll sideways to read all table columns.