How AI Rankia Data measures AI answers

The site reads precomputed snapshots, one per country and tree, rebuilt after every monthly run. The numbers below are the current one, built on Oct 29, 2026.

Categories = nodes with answers or search demand. Niches = sampled leaves (at least one AI answer). Brands are distinct per tree and country, so the total counts a brand once per country. The snapshot also holds 4,211 distinct stores and 16,415 distinct prompts.

Product categories. Every sampled niche gets the same two prompts in every country: a short product search, run on Google AI Mode, and a buying question, run on ChatGPT. Some niches also carry a Google AI Overviews answer.

Industries. Each index asks its own set of questions, the same ones the published index uses, on several engines.

One answer per prompt and engine. We keep the latest completed run (or, failing that, the latest empty answer), so re-runs never weigh a prompt twice. Shopping uses the latest completed capture of that prompt and engine. Ads count every ad seen on any run of that prompt and engine, deduplicated by advertiser and title. Failed runs are ignored; empty answers count in the sample.

The sample is deliberately visible: every AI metric on the site carries n = answers and the date of its last answer, like this: n = 37.4K answers·16.4K prompts·Oct 29, 2026. A single niche rests on a handful of answers; a top-level category on hundreds.

Engine names are normalised: Google AI Mode, ChatGPT (with search), Google AI Overviews, Gemini, Perplexity, Claude and Grok; anything else is grouped as “other”. Which engines answer each tree today:

Brands named in the text of the answers. A brand counts once per answer, however many times it is repeated.

count

answers the brand appears in

byModel

the same share computed inside each engine's answers

avgRank

mean position of the brand in the answer (1 = named first)

In product categories, brand visibility and cited sources come from Google AI Mode and AI Overviews answers for now (see caveats). Left out of every ranking: the index's own target brand, placeholders («Unknown», «Genérico», «N/A»…), media, social and search sites, the services we use to capture answers, and names that the answer analysis tags as a source, guide, institution or website in at least 75% of their mentions and that never appear as a product maker. Stores never enter brand visibility: they are counted under shopping.

Bands follow the US Department of Justice: below 1,500 is fragmented, 1,500–2,500 moderate, above 2,500 concentrated. We also report the leader's share (top 1) and the top 3. Concentration is “not enough data” with fewer than 3 brands or fewer than 3 AI Mode / AI Overviews answers, so single niches (one AI Mode answer) are not rated.

Only real product cards count: every ChatGPT shopping card, and Google AI Mode items with a price, a Google Shopping link or a store that shows prices elsewhere in that country. Video, social and editorial tiles are not products. Items are deduplicated per answer by normalised title and store. Products are never read from the answer text.

products

distinct normalised product titles

stores · product brands

share of appearances (one per product card; a store named in the text with no card adds 1); count = answers

median price

in the country currency (MXN, EUR, USD); prices whose currency cannot be read are skipped, and AI Mode cards are priced in the searcher's country

examples

up to 12 products, round-robin over engines, one per store first, best documented first; images only over https (today only ChatGPT cards carry them)

Stores vs brands. A name is a store when it is a known retailer (Amazon, Mercado Libre, Walmart, Liverpool, Coppel, Costco, El Corte Inglés, MediaMarkt, PcComponentes, Target, Best Buy…) or when it appears as a carousel retailer more than twice as often as a product maker. So Liverpool is a store, while Samsung or Tramontina, which also sell direct, stay brands. Retailer variants merge (Amazon.com.mx, Amazon MX…). ChatGPT cards without a brand get one when the title starts with a brand already known in that country.

Only for engines where ads are captured: ChatGPT today. Google ads, when they join, count only inside the AI Overview block and the ad right after it, never ordinary search ads. The ad count is the number of ads seen; advertisers are ranked by ads, with the domain taken from the ad (deep-link hosts resolve to the advertiser).

The hosts cited in Google AI Mode and AI Overviews answers (the citation's domain, else the URL host, without «www.»), one per URL per answer.

This snapshot reads 161,820 citations across both trees and the three countries.

Google searches / month

Google Keyword Planner monthly searches for the search chosen for each niche, 12-month average (Oct 2025 – Sep 2026)

AI asks / month

an estimate: Google volume × the share of commercial searches each AI engine already answers (table below)

trend

(sum of the last 3 months ÷ sum of the first 3 months of the window − 1) × 100

peak month

the month with the most searches

CPC

average cost per click from Keyword Planner, in MXN; converted for Spain and the United States with the European Central Bank reference rate of the build day

competition

Keyword Planner's LOW / MEDIUM / HIGH (parents take the most common value of their niches)

Niches without a chosen search have no demand. A niche that repeats another niche's search keeps 0 volume, so the search is counted once. Parents add up their niches: Google searches, AI asks, the per-engine estimate and the monthly series (month by month); CPC is the volume-weighted mean; trend and peak month are recomputed on the summed series.

Example: Pianos (Mexico) has 301,000 Google searches a month × 41.5% ≈ 124,915 estimated AI asks a month.

A parent category adds the raw countersof every answer below it and computes its shares again; it never averages its children's percentages. So a parent's brand share is the share of all its answers, and small niches don't weigh as much as big ones.

sample.answers

AI answers under the node, empty answers included

sample.queries

distinct prompts

sample.niches

sampled leaves (at least one answer)

demand percentile

percentile of AI asks / month among the categories of the same level in the same country (0–100)

openness

100 − the top-3 brands' share (50 when concentration can't be measured)

ad pressure

the highest ad rate across engines

The score ranks each category against the scored categories of its own level in the same country, so it is relative: high from 65 (better than about two thirds of its peers), medium from 45, low below. Nodes with fewer than 2 answers are not scored.

The prompts run again every month and the snapshots are rebuilt right after, so each month adds a new reading and history accumulates. The current snapshot (Oct 29, 2026) holds answers from Apr 11, 2026 to Oct 29, 2026. Every page shows the date of the last answer behind each number.

What you should know about this snapshot, in our own words:

• Product categories: brand visibility and cited sources come from Google AI Mode and AI Overview answers only. ChatGPT brand mentions and ChatGPT citations are left out for now: the first ChatGPT run of these prompts went through an API path instead of the real ChatGPT app. The real-app monthly run of 27-Nov-2026 replaces it. ChatGPT already counts for native shopping cards and ads.
• Shopping captures are still pending for about 9% of answers; shopping figures cover the captures already closed.
• Spain: Google AI Mode rarely shows native catalogs (about 12% of short searches).
• Demand: Google Keyword Planner monthly searches, Oct-2025 to Sep-2026 (12-month average). AI asks per month are estimated from Google volume × the share of commercial searches each AI engine already answers.
• CPCs for Spain and the United States are converted from MXN with the European Central Bank reference rate of 2026-10-28.
• Product categories: every sampled niche gets the same two prompts in every country, a short product search (Google AI Mode) and a buying question (ChatGPT). Native shopping figures count only real product cards (price, store or catalog link), never products read from the answer text.
• Answers are refreshed monthly; the next refresh is around 27-Nov-2026.

AI Rankia Data only uses prompts owned by AI Rankia's own index. A client's prompts, brands and workspaces never enter it: every row is checked against the database when the snapshot is built, indexes that belong to other workspaces are left out, and the build stops if a single node fails the check. The web app never queries the database: it reads the finished snapshots.