Method

How the measurement is done.

This page exists so you can check the work. It is the method behind the Miami Shortlist Index and behind every monthly report SignalTide sends a client.

01

The prompt set is fixed before anything is measured

Twelve booking intents, each asked three ways — thirty-six prompts per property, written the way a traveler actually asks: by neighbourhood, by stay type, by occasion, by constraint. Wording moves an answer more than almost anything else, which is why each intent is phrased three ways. Once a property's set is written it does not change, because a set that moves cannot be compared month to month.

02

The competitor set is named, not implied

Leadership names the properties they consider competitors before the first run. Results are reported against that set. This stops the report from quietly redefining success when the numbers are unflattering.

03

Five engines, two languages, three runs — 1,080 recorded answers

Thirty-six prompts across ChatGPT, Gemini, Perplexity, Claude and Google AI Overviews, in English and Spanish, repeated three times each. That is 1,080 answers per property, every one queried through the provider's own API with web search enabled, logged out, with no personalization or account history, and every one dated with the full engine response saved to the evidence file you receive. API answers are close to, but not identical to, what the consumer chat apps return, and this is stated rather than glossed over. Spanish is not optional in this market. A single run is mostly noise; repetition is where the noise shows itself.

04

Results are reported as frequency, never as a position

A property is reported as appearing in, say, eleven of thirty runs for a question — not as finishing fourth. This is the single most important rule here, and the reason is below.

05

Descriptions are checked against what is true

Being named is not the same as being named correctly. Each answer is compared against the property's actual amenities, location and policies, and errors are logged as findings.

Why frequency, and not a ranking

AI answers are not stable. Ask the same question twice and you can get two different lists. Independent 2026 research found only about 20% of brands appear across five consecutive AI runs, and that citation overlap between ChatGPT, Perplexity, Claude and Gemini is roughly 11%.

A single-run ranking is therefore not a finding. It is a coin toss, published. Reporting how often a property appears across repeated dated runs is the only honest way to describe placement, and it is the only way a client can re-run the test themselves and get the same answer we did.

Source: 2026 AI search visibility research. Reference.

Likely source, not proven cause

SignalTide can show that an assistant said something wrong, identify the listing most likely to have produced it, and document the citation the assistant gave. It cannot prove that one listing caused one answer: these systems do not disclose that. Findings are reported as likely source, documented citation, or conflicting information — never as confirmed causation. The same applies in reverse: placement can be measured, but no booking is attributed to it.

In practice that means every finding is filed as one of three things, and the label is part of the finding:

  • Likely source. The assistant said something wrong and one listing is the probable origin — an old OTA description, a stale directory entry. Named as likely, not as proven.
  • Documented citation. The assistant named its source. That one is checkable, and it is the strongest kind of finding on offer.
  • Conflicting information. Two sources disagree and either could be what the assistant read. Reported as a conflict to resolve, not as a diagnosis.

A finding that cannot be placed in one of those three is not reported as a finding.

What this method cannot tell you

It can prove that placement moved. It cannot prove that a booking came from it. Travelers who find a property through an assistant do not arrive carrying a label saying so, and anyone claiming a clean attribution chain from an AI answer to a reservation is selling something.

So the monthly report shows placement and movement, and sets the property’s direct-versus-OTA share beside it. The two are reported next to each other and the reader draws their own line. No ranking position is promised, and no number of bookings is promised.