Research · proprietary benchmark

ChatGPT Real Estate Visibility Benchmark

Our primary dataset: a locked set of 64 buyer and seller questions, each asked four separate times against ChatGPT (OpenAI), with every answer preserved and classified. Two anonymized US markets have been captured so far.

Request your analysisView the methodology
Provider
ChatGPT (OpenAI)
Capture date
August 2026
Markets captured
Greenville, SC · Jersey City, NJ (both anonymized)
Version
v1.0 · published 14 September 2026

Benchmark overview

The benchmark exists to answer one question with evidence rather than anecdote: when someone asks an AI assistant who to hire in a given real estate market, who does it actually name, and what is it reading when it names them?

Each market edition uses the same structure — 64 questions, four repetitions each, a maximum of 256 valid captured answers, one provider, one capture window. The structure is fixed so editions can be compared with each other.

Measured
Counted in a captured dataset with a published denominator.
Directional
A pattern observed in the data; not a precise or repeatable figure.
Example
An illustration of the format. Not a verified outcome.
Pending verification
Supplied or observed, but not yet traced to a primary source.
Not tested
Outside the scope of the measurement that was run.
Not observed
Looked for in the captured set and not found there.
Private
Held in the underlying dataset and not published.
Anonymized
Real measurement, identities withheld without publication permission.

Why repeated AI answers vary

The same question asked twice does not reliably return the same answer. Responses are generated probabilistically, the retrieval step can surface different pages between runs, and the underlying models and indexes change over time. A single screenshot is therefore not a measurement of anything.

Repetition is how that variance is handled here. Asking each question four separate times turns a one-off answer into a frequency with a denominator — a team recommended in 4 of 256 answers is a different statement from a team that happened to appear once.

This does not remove uncertainty. Four repetitions is enough to distinguish a consistently recommended team from an absent one; it is not enough to resolve small differences between two similar teams.

DirectionalRepetition reduces, but does not eliminate, run-to-run variance.

Prompt taxonomy

Hire-intent questions
“Who should I hire to sell my condo in this market?” Direct requests for a recommendation.
Best-of questions
“Who are the best agents in this city?” Requests for a ranked or shortlisted set.
Neighborhood questions
Questions scoped to a named neighborhood or building rather than the whole market.
Transaction-type questions
Buyer-side, seller-side, first-time buyer, investor and luxury framings asked separately.
Record and event questions
Questions about specific notable transactions in the market.

The question set is written for each market and locked before capture begins. Questions are not added or edited mid-window, so an edition cannot be steered toward a flattering result.

Recommendation-share methodology

Unit
One captured answer.
Numerator
Captured answers in which the team is recommended — put forward as someone to hire or work with.
Denominator
Valid captured answers in that market edition (maximum 256).
Mentioned, not recommended
Named in the answer without being put forward. Counted separately and never merged into the recommendation figure.
Absent
Not named anywhere in the answer.
Multiple teams per answer
One answer can recommend several teams, so shares across teams may sum above 100%.
Exclusion rules
An answer is excluded when the provider returned an error, refused the question, or returned no market-specific content. Excluded answers are removed from the denominator, and the valid-answer count is published.

Citation-share methodology

Unit
One citation instance, not one answer.
Captured
Every URL cited in a captured answer is recorded against that answer.
Multi-source rule
A single answer can cite several sources, so citations are counted per citation, not per answer.
Indexed frequency
Indexed source-appearance frequency (most frequent class = 100). Numerator: Citations belonging to that source class within the captured answers. Denominator: Citations belonging to the most frequent class (listing portals).
Retrieved vs mentioned vs cited
We can only observe what the answer cites. What the provider retrieved but did not cite is not visible to us and is never reported.

Source-class taxonomy

Indexed frequency (listing portals = 100) · Jersey City, NJ · residential (anonymized) · ChatGPT (OpenAI) · captured August 2026 · 256 valid captured answers
Source classIndexed frequency
Listing portals100
Brokerage sites44
Local publications22
Community forums12
Video7
The team's own site0

Indexed source-appearance frequency (most frequent class = 100). Citations belonging to that source class within the captured answers. Citations belonging to the most frequent class (listing portals). A single answer can cite several sources, so citations are counted per citation, not per answer. Observed association only; no causal claim is made.

These figures come from the Jersey City edition only. Source-class figures for the Greenville edition are held in the dataset and not published.

Market comparison

The two editions are reported separately and never averaged. They used the same structure but different question sets, different markets and different observed teams, so a combined figure would not mean anything.

Published figures by market edition

  • Observed team recommended (Jersey City edition)

    4 / 256

    Measured
    Numerator
    4 answers recommending the observed team
    Denominator
    256 valid captured answers
    Market
    Jersey City, NJ · residential (anonymized)
    Date
    August 2026
    Provider / model
    ChatGPT (OpenAI) — OpenAI-powered search response; exact model metadata unavailable for this benchmark
    Source
    Recommended First captured answer corpus
    Limitation
    One market, one provider, one capture window.

  • Observed team named at all (Jersey City edition)

    8 / 256

    Measured
    Numerator
    8 answers naming the observed team
    Denominator
    256 valid captured answers
    Market
    Jersey City, NJ · residential (anonymized)
    Date
    August 2026
    Provider / model
    ChatGPT (OpenAI) — OpenAI-powered search response; exact model metadata unavailable for this benchmark
    Source
    Recommended First captured answer corpus
    Limitation
    Naming is not recommendation; the two are counted separately.

  • Most-recommended competitor (Jersey City edition)

    116 / 256

    Measured
    Numerator
    116 answers recommending the leading observed competitor
    Denominator
    256 valid captured answers
    Market
    Jersey City, NJ · residential (anonymized)
    Date
    August 2026
    Provider / model
    ChatGPT (OpenAI) — OpenAI-powered search response; exact model metadata unavailable for this benchmark
    Source
    Recommended First captured answer corpus
    Limitation
    Being recommended more often is an observation, not an explanation.

  • Observed team recommended (Greenville edition)

    0 / 256

    Measured
    Numerator
    0 answers recommending the observed team
    Denominator
    256 valid captured answers
    Market
    Greenville, SC · residential (anonymized)
    Date
    August 2026
    Provider / model
    ChatGPT (OpenAI) — OpenAI-powered search response; exact model metadata unavailable for this benchmark
    Source
    Recommended First captured answer corpus
    Limitation
    Zero in this window does not mean permanently absent.

  • Lower-producing competitor recommended (Greenville edition)

    17 / 256

    Measured
    Numerator
    17 answers recommending the observed competitor
    Denominator
    256 valid captured answers
    Market
    Greenville, SC · residential (anonymized)
    Date
    August 2026
    Provider / model
    ChatGPT (OpenAI) — OpenAI-powered search response; exact model metadata unavailable for this benchmark
    Source
    Recommended First captured answer corpus
    Limitation
    Production and recommendation are different measurements of different things.

  • Recommendation share by prompt category

    Not published

    Private
    Numerator
    Not published
    Denominator
    Not published
    Market
    Both editions
    Date
    August 2026
    Provider / model
    ChatGPT (OpenAI)
    Source
    Held in the dataset
    Limitation
    Per-category denominators are small; publication is pending a larger sample.

  • Additional US markets

    In progress

    Not tested
    Numerator
    Not published
    Denominator
    Not published
    Market
    Not yet selected for publication
    Date
    Not published
    Provider / model
    Not published
    Source
    Not published
    Limitation
    Nothing is claimed about markets that have not been measured.

Editions are reported separately. Greenville and Jersey City figures are never combined or averaged.

Authority versus AI visibility gap

Greenville, SC · residential (anonymized) · ChatGPT (OpenAI) · captured August 2026 · 256 valid captured answers · production from RealTrends 2025 closed volume
Observed party2025 closed volumeMarket positionAnswers recommending
The team (you)$78.9MTop-5 team by volume0 / 256
Competitor$33.0MOutside the top 20 by volume17 / 256

Greenville edition only. Production figures are verified closed volume; recommendation figures are counted from captured answers. The two are different measurements placed side by side, not a cause and an effect.

DirectionalAcross both editions, recommendation frequency did not track verified production. That is a pattern in two markets, not a law.

Raw-data and evidence policy

Raw answers are preserved
Full answer text, capture timestamp, provider label and every cited URL are stored for each captured answer.
Answers are not published in bulk
The preserved corpus is held privately and shared for inspection on request, so it cannot be scraped or re-published out of context.
Entities are anonymized by default
No team is named without written publication permission. None is currently on file.
No downloadable data file is offered yet
There is no public CSV or API for this dataset. When one exists it will be linked from this section.
Corrections are versioned
Any change to a published figure gets a new methodology version and an entry in the update history below.

To inspect the preserved answers behind any figure on this page, write to francisco@recommendedfirst.com. Requests are answered with the captured answers themselves, not with a summary.

Update history

DateVersionChange
14 September 2026v1.0Benchmark page published. Greenville and Jersey City editions reported separately under methodology v1.0.

Every future change to a published figure will be added here with a new version number.

Limitations

What this cannot prove

  • One provider only: ChatGPT (OpenAI). Exact model metadata is unavailable for these captures.
  • Two markets. Nothing here describes any market that was not measured.
  • One capture window, August 2026. Results change as models, indexes and sources change.
  • Four repetitions per question resolves presence and absence, not fine differences between similar teams.
  • Nothing here reveals how any provider ranks or selects sources internally.
  • No causal claim is made between any source, any intervention and any recommendation.

How to cite this benchmark

How to cite this page

Recommended First, ChatGPT Real Estate Visibility Benchmark, Jersey City, NJ, captured August 2026, v1.0.

For the Greenville edition, cite: Recommended First, ChatGPT Real Estate Visibility Benchmark, Greenville, SC, captured August 2026, v1.0. Please cite a single market edition rather than the benchmark as a whole, and link to this page so the conditions travel with the figure.

Keep reading

See who AI recommends in your market — and where you stand.

One client per market. Start by telling me where you work.

Request your analysisView a sample AI visibility audit