The best agents in this market are not the ones AI names.

Two agents, the same sales band,
three thousand appearances apart.

Ranks eleven and twelve by closed volume both sold between $75M and $100M over the same period. To within a band, they did the same amount of business.

One was named once. The other was named 2,034 times.

Production window January 2024 to August 2026. Visibility window 27 July to 25 August 2026.

Ranked by sales on the left.
Ranked by AI on the right.

Same twenty-five people, ordered two ways. Each thread follows one agent from where their production puts them to where the assistants put them.

Presence and prominence
are different things.

Producing at the top does make you more likely to be named at all. An agent in the top twenty-five is roughly twelve times more likely to appear than an agent drawn from the market at large, and nineteen of these twenty-five appeared at least once. On the question of whether you exist to an assistant, production matters.

On the question of how often, it stops mattering almost entirely. The agent ranked twenty-second appears more often than everyone ranked second through ninth. The agent ranked seventh, selling alongside people who appear hundreds of times, appears nine.

Production puts an agent roughly in the right neighbourhood of the visibility ordering, then has almost nothing to say about the distances inside it.

Six of these agents were never named once. Each closed between $60M and $100M. They are not marginal operators. They are among the most productive people in a county of thousands of licensees, and to an assistant answering "who should I hire" they do not exist.

We measured ourselves against
a stricter standard than we sell.

Citelight's own extractor decides whether an agent was named by reading bold and list formatting across three assistants, and it skips AI Overview entirely.

Every figure on this page instead comes from a direct scan of the full stored answer text on all four assistants, which is the standard a sceptic would apply. Where the two disagree, we report the number that is harder for us.

Measuring with our own extractor produced a headline we had to withdraw. That correction is logged at the foot of this page.

All twenty-five, both orders.

Bands rather than exact volumes: publishing rank beside exact volume would identify individuals to anyone with MLS access, which in this market is everyone.

What we did, precisely.

What counts as being named

An agent is named if their name appears in the answer text of at least one assistant response to a market query in the window, on any of the four assistants, in any formatting. Frequency is reported alongside, because presence and prominence are different claims. Deliberately not a citation requirement, not a majority-of-runs requirement, and not our own extractor's rule.

Name matching

First-name stem plus surname, so diminutives are caught. Not cosmetic: under the corrected match, rank four moves from 19 appearances to 463 and rank ten from 71 to 691. Both counts are retained so the correction is auditable.

Production data

MLS closed sales, 1 January 2024 to 20 August 2026, Broward County, ranked by total closed volume, reported as bands to preserve anonymity.

The assistants changed underneath us

Perplexity and AI Overview were constant. ChatGPT changed model on 3 August, at the window boundary. Gemini changed twice inside it. "We asked Gemini" describes two different models within this window and four across the wider corpus, so any movement over time is confounded with model change and must not be read as a change in the market.

Query set

Market queries only. More than thirty distinct queries surfaced at least one name, led by head terms like "best real estate agents in Fort Lauderdale FL". Identity probes of the form "tell me about <name>" were excluded: they can only return the name we supplied, and including them would have manufactured this study's own headline.

The nine-times figure is a floor

It compares the share of top-25 agents named with the share named across the market. Both are computed under our narrower extractor, because the market-wide denominator has only ever been calculated that way. Under the broad definition used everywhere else on this page, the true share is higher in both groups. We publish the floor rather than estimate the ceiling.

Why there is no correlation coefficient

We computed one and withheld it. A Spearman coefficient on this data moves by about 0.03 when the definition of "named" changes and by roughly 0.3 when the population changes, ranging from 0.24 to 0.56 on the same records. It was also computed against cast counts, which store surface forms, so one agent with 1,523 mentions can sit across seven rows and be correlated as a fragmented measure. A single coefficient would imply a precision this data does not support, and it is the number most likely to be quoted back without its caveats.

What this doesn't show.

• A 45.6% name-match rate between assistant output and MLS records. This is the study's ceiling. Names we could not match to a licence are not counted, in either direction.
• Production is Broward-wide; the queries are Fort Lauderdale-specific. An agent whose volume comes mainly from Palm Beach reads as high-production and low-visibility here for reasons that have nothing to do with AI.
• Two of four assistants changed model inside the window, so comparisons across time are confounded with model change.
• No causal claim. Production and visibility move together to a degree. Nothing here shows that either produces the other.
• One market, one window. Nothing establishes that this shape holds elsewhere or at another time.
• Appearances are not weighted by demand. An appearance in a rare neighbourhood query counts the same as one in a head term, though they are worth very different amounts.
• Bands compress real differences. Two agents in the same band may be far apart in actual volume.
• A staleness hypothesis was tested and not published. The idea that visibility outlives market activity rested on a single case and was contradicted by the other 185 examined.

What we changed, and why.

• 4 September 2026 · corpus figures corrected from the whole table to the evidence. The corpus figures on this page and on the methodology page described our database, not our evidence: they counted every row, of which 68.2 per cent came from the test and stress accounts we ran before and beside the real cohort. Corrected: 35,925 citations (was 113,097), 65 readings (was 1,018, a count of per-platform result rows, never of readings), six weeks (was thirteen), 369 agents discovered (was 1,079), 171 corroborated (was 391). Two findings moved with them. The share of producing agents ever named is 4.9 per cent, not 6.2, and a top-25 producer is about twelve times more likely to be named than a randomly chosen producer, not nine. The test readings had added lower-ranked producers and unmatched names while the top producers appeared in real and test readings alike, so removing them raised the enrichment. The top-ten, top-25 and top-100 rates and the rank correlation hold. On the thirteen weeks: no count reproduces it. It is the calendar span from our first test reading in the nineteenth week of the year to our first real reading in the thirty-first, which is the window before evidence existed. We published the length of our setup as the depth of our data. The corrected figure is six weeks, the thirty-first to the thirty-sixth. Corroborated now means a name seen in three or more separate real readings, read from the per-reading record rather than a stored counter, because the counter counted stress readings and a stress account’s reading is not the evidence that sentence implies. From today these figures are served live from the same source the product reads, and every one is recomputed independently against it.
• 28 Aug 2026 · Headline withdrawn and replaced. An earlier version reported that 11 of the top 25 were never named. That count came from our own extractor, which reads only bold and list formatting and skips AI Overview. Under a direct scan of full answer text, five of those eleven do appear. The correct count is six, and the finding is now stated as the appearance gap rather than as a count of absences.
• 28 Aug 2026 · A claim was cut. "All ten highest-producing agents are named" is arithmetically true and misleading: rank seven appears nine times, which is presence without prominence.
• 28 Aug 2026 · Production boundary moved from Fort Lauderdale city to Broward County, after a boundary test removed 57% of a separate claim.
• 3 Sep 2026 · Rewritten. Same data, same findings, no figure changed. The AI ranking column and the crossing diagram are new views of the table that was already here.