Free snapshot

Methodology

How the Dopbot Index is produced, what the numbers mean, and where they stop being reliable.

What the index measures

The index measures how often AI assistants recommend a brand when a shopper in the United States, Mexico or Canada asks a buying question. It is a measure of visibility in AI answers. It is not a measure of sales, product quality or market share.

The questions

Each category has 5 prompts, written to read like what shoppers type into an AI assistant. There is one prompt for each intent: best overall, a budget with a price ceiling, premium, a specific need, and an alternative to the category's best-known brand. The prompts for a category are listed on its page in the index.

Markets

The index covers three markets. Each market has its own list of categories, chosen for the shoppers of that country, and its own pages. A category in one market is not comparable with a similarly named category in another, because the questions, the prices and the brands on sale are different.

  • United States (/index). Questions in English with prices in US dollars. Web search is set to the United States. Updated monthly, on the first Monday of each month.
  • México (/mx/index). Questions in Mexican Spanish with prices in Mexican pesos. Web search is set to Mexico. The index pages are in Spanish; this methodology page is in English. Updated monthly, on the second Monday of each month.
  • Canada (/ca/index). Questions in English with Canadian wording and prices in Canadian dollars. Web search is set to Canada. Updated monthly, on the third Monday of each month.

Three things differ between markets: the country location sent with each web search, which shapes which sites and stores the search returns, the language and wording of the questions, and when the market is sampled. Nothing else changes: the engines, the number of samples per prompt and the counting rules are the same in every market. Free snapshots and the paid tracker use the United States.

Update schedule. Each market is updated once a month, on its own Monday, so the three markets are never collected in the same week. A market runs in the ISO week whose Monday (UTC) is its Monday of that calendar month: the United States on the first Monday (a Monday on the 1st to the 7th), Mexico on the second (8th to the 14th) and Canada on the third (15th to the 21st). A week whose Monday is the fourth or fifth Monday of its month has no run. The date of the Monday decides, so a week that starts in one month belongs to that month, and the gap between two updates of a market is four or five weeks. Each market's pages show that market's latest published update, so between updates you see the previous update, not an empty page. The change column compares with the market's previous update (normally one month earlier), not with the calendar week before. A market's first update runs as soon as the market is added, and after that it follows this monthly schedule (a market's scheduled update is skipped only when its first update already ran earlier in the same calendar month, which happens only right after launch).

The prompts stay the same from one update to the next so that the numbers are comparable. When we do change a prompt, the change since the previous update for that category is less reliable. In the alternative prompt, the named brand is usually mentioned but not recommended, so it is normally not counted as a recommendation.

How we sample

  • Engines. Every prompt is sent to 2 engines:
    • OpenAI, through its developer API with the web search tool turned on.
    • Perplexity, through its Sonar developer API, which answers from live web search.
  • Web search is on. Search is enabled on every request so that answers can reflect current products, prices and reviews.
  • Country location. Requests set the user's approximate location to the country of the market: the United States, Mexico or Canada. The prompts of each market are written for shoppers there (see Markets above).
  • Samples per prompt. Each prompt is asked 1 time per engine in each update. That gives 10 answers per category per update (5 prompts, 2 engines, 1 sample each).
  • Cadence. Each market's index runs once a month: the United States on the first Monday of the month, Mexico on the second and Canada on the third (see Markets above). A run starts on Monday (UTC). A week is labelled by its Monday. Weeks whose Monday is the fourth or fifth Monday of the month have no run. Results are published when all answers for the run have been collected and processed.
  • No personalization. Requests carry no account, search history or memory.

How brands are extracted and counted

  1. An AI model reads each answer and lists the brands it names, in order of first appearance, with the product when one is named. It also marks whether the answer presents the brand as a recommendation or only mentions it, for example as a comparison or a warning. Retailers and marketplaces are stores, not brands, so they are left out. At most 15 brands are kept per answer.
  2. Brand names are normalised so that capitalization, punctuation, accents and endings such as "Inc." or "Co." do not create separate brands. We merge known spelling variants by hand.
  3. A brand is counted at most once per answer, however many times the answer names it. Only recommendations count.
  4. Share is the number of answers that recommend the brand divided by all answers sampled for the category in that update. The columns for each engine use the same formula on that engine's answers only.
  5. Average position is the average of the brand's position, the order of first mention among the brands in an answer, across the answers that recommend it. 1 means it was named first.
  6. Cited sources. Every web source an engine reports for an answer is recorded and reduced to its domain. The sources table counts citations per domain.
  7. Change is the difference in percentage points between this update's share and the previous update's. "New" means the brand was not recommended in the previous update. The page names the week of the previous update.

What the numbers mean

A brand with a 40% share was recommended in 40% of the answers we sampled for that category in that update. That is its visibility for these questions at that time, and nothing more.

What the numbers do not mean

  • They are not sales, market share or customer satisfaction.
  • They are not our opinion of a product, and a ranking is not an endorsement.
  • They do not predict what any one person will see. AI answers vary from run to run, even for the same question.
  • They are coarse. With 10 answers per category, one answer moves a brand's share by about 10 points. Treat gaps of a few points as ties and look at trends across updates.
  • They cover the categories we track and the brands the assistants chose to name. A brand that is missing was not recommended in the answers we sampled.

API answers versus consumer apps

We sample the engines through their developer APIs because that lets us ask the same questions on a fixed schedule and record the sources. API answers approximate what people see in the consumer apps from the same companies, but they are not identical to them. The apps may use different model versions, system instructions, search tools and personalization, and they know more about the user. Use the index as a consistent yardstick that moves when AI recommendations move, not as a copy of any one app.

Independence

Dopbot is independent. It is not affiliated with, endorsed by or sponsored by OpenAI, Perplexity or any other AI provider, or by any brand that appears in the index. Brands cannot pay to be included or ranked, and the paid tracker does not change the index.

Report an error

If a brand name looks merged, split or misspelled, or a number looks wrong, email hello@dopbot.com with the category, the week and what you expected to see. We check every report and correct wrong names and mappings. We do not remove or reorder brands on request.

Get a free snapshot