Methodology

Applies to every page under /chatgpt-rankings/ and /chatgpt-agency-rankings/

Where the data comes from

Every week we ask ChatGPT a fixed set of eight buying-intent questions per category (the exact prompts are listed on each category page) and record the complete answers. Rankings are computed from which brands those answers name and in what order — nothing is hand-picked, and no brand can pay to appear.

The numbers

Every number on a category page pools the same window: the five most recent weekly readings (fewer while a category is young). A reading is one weekly collection of answers, so the window holds roughly forty answers — a single week's ~8 answers is too small a sample to rank on. The leaderboard column states the span it covers.

The exceptions are deliberate: the trend chart plots the individual weekly readings (it is the history view), and the verbatim answers and quotes are the newest reading's — they are quoted content, not statistics.

The quadrant

The visibility-vs-prominence chart plots the two windowed numbers above — no composite scores. Each faded circle is the same window ending one reading earlier. Its midlines arefixed, not category-relative: visibility at 50% of answers, prominence at position 3.0. Fixed lines mean a brand only changes quadrant when its own numbers change, and quadrants are comparable across categories. The chart renders only when at least five brands have a measurable average position.

Gates: what we refuse to publish

Pages are built only from windows with real data: a category page fails its build rather than render from zero answers, a newest reading more than eight days old, an implausible brand count, or nonsensical shares. Per-prompt leaderboards are suppressed below four answers. Quotes are verbatim sentences from real answers — if no complete sentence names the brand, no quote is shown.

Cadence and history

Data collects weekly; completed weeks are frozen and never recomputed, so history is stable. Pages state their “as of” date. ChatGPT’s recommendations legitimately change — including when OpenAI updates the underlying model — so week-to-week movement reflects what a ChatGPT user would actually have seen.

Agency city pages

Each city page tracks 8 prompts, one per industry (the exact prompts are listed on each city page), asked with the city's country as the location. The leaderboard pools all 8; the industry section shows each prompt's own top three once it has at least four answers. Every number uses the same window and gates as the category pages. Variant spellings of one agency's name are grouped under a single name, and the AI assistants the prompts name (ChatGPT, Gemini, Claude) are not counted as agencies.

A city page is published only once its data is real: at least 6 agencies each named in 2 or more answers across the window, and at least 5 of the 8 industries with a board. A published page stays up while it keeps at least 4 such agencies, and is removed below that. Cities without enough agencies ChatGPT recommends have no page.

Questions about the method? Talk to MentionBird.