Exactly how this index is produced

What we do

We ask publicly available AI assistants the questions a customer would ask them, through each provider's official API, and we store the answer verbatim. We then use a separate structured model call to read that stored answer and list the businesses it named, in order, with a verbatim snippet for each. Every business name published on this site came out of a model's own answer to one of the prompts below.

We do not scrape websites, we do not buy data, and we do not have a relationship with any AI provider. The only external requests this system makes are API calls to the engines listed below.

Engines and models

Temperature is set to 0 wherever the provider supports it. The exact model string returned by the provider is stored on every run, so a historical row keeps saying what actually ran even after a provider deprecates the model.

Cadence

Tracked questions are re-run on a six-hour cycle, oldest first, subject to a hard daily spend cap. Each run is stored as a new row rather than overwriting the last, so the history is a record rather than a snapshot. Every result block on this site shows the engine, the model, and the run date.

Extraction

Business names are pulled from the raw answer by a structured model call, never by pattern-matching. Published snippets are verbatim substrings of the stored answer, capped at 200 characters; if a snippet cannot be matched back to the stored text, it is not published.

The badge

A business is badge-eligible when it was named in at least two of the last three runs of the same question on the same engine. One hit is noise. Eligibility is computed only from the public seed set of questions — never from a question added at a customer's request — so a badge cannot be manufactured or bought. The badge image states the engine, the question, the ratio, and the run date, so a screenshot carries its own expiry. Below the threshold, the badge renders a neutral "not currently named" variant rather than failing.

What this cannot tell you

• AI answers are non-deterministic. The same prompt can return different businesses on different days, from the same model, at temperature 0. A single run is evidence of one answer, not of a ranking.
• Answers are personalized in the real world. We query the API without location signals, history, or a logged-in account. A customer asking the same question in a consumer app may see something different.
• Being named is not an endorsement of quality, by the engine or by us. We publish only whether a business was named, in what position, and the sentence it appeared in. We never publish a model's characterisation of a business's quality, and we never generate commentary of our own.
• Absence is not a verdict. A business not named may be excellent, new, or simply unclear to a model. That is the gap this index measures — nothing more.

Corrections and removal

Removal is unconditional and immediate: any request naming a business suppresses it on submission, with no verification. We accept that a competitor could abuse this, because the alternative — us deciding whose removal is legitimate — is worse. Getting back in does require proof of ownership, via an email at the business's own domain or a DNS TXT record. Evidence rows are retained; only the public surface changes. See the opt-out page.

How the prompts are written

We currently track 25 questions across 1 metro. Every one is built from the same two templates, so results stay comparable between categories:

Who are the best {category}s in {metro}? Please recommend specific
businesses or practitioners, ranked, and cite sources if you can.

What are the best {category}s in {metro}? Please recommend specific
businesses, ranked, and cite sources if you can.

The first form is used where the buyer hires a person or practice, the second where they hire a company. Nothing else is added — no system prompt, no location metadata, no examples, no follow-up turns.

The exact text of any prompt is published on the page that shows its results, under "The question we asked", so every individual claim on this site can be checked against the precise input that produced it. We publish it there rather than as one combined list here, because the point is that each claim is verifiable — not to hand over the whole question set as a single file.

Browse every question we track →