How PickScore measures

The one thing to know first

PickScore measures whether an AI assistant named a business in its answer. It does not measure whether the assistant recommended, picked, preferred or ranked that business. An answer that mentions a business in passing, or even unfavourably, still counts as naming it. Every page that shows a result says so.

The presence check (methodology p1)

The presence check is lightweight. It makes at most two requests to an AI assistant with live web search, and produces four signals in words. It runs for every check, as part of the free eligibility check, and nothing from it is shown before payment: its four signals appear only in the paid presence check. It never produces a number, a score, the name of another business, a source, a rank or an order.

• AI Recognition: the assistant is asked about the business by name, type and place. Detected, not detected or unclear says whether its answer found information about that business at that place.
• Official Website: whether that same answer stated a website of the business's own. A social profile or a directory listing does not count. We do not verify who owns the website.
• Public Identity: strong, moderate, weak or unclear says how consistently the answer could establish the name, address and contact details.
• Sample Customer-Choice Signal: one question of the form “What is the best (type of business) in (place)?” is put to the assistant, and the signal says whether the answer named the business. The question comes from a fixed template and is never written by a model. The answer is shared by every business of that type in that place and may be up to 30 days old.
• Full PickScore: always “not measured” in the presence check, because it has not been.

What the presence check cannot tell you

It does not measure how often AI names or recommends a business. One sample observation does not establish recommendation performance: being named once is not being recommended, and not being named once is not being ignored. A business location is measured afresh at most once every 30 days; inside that time a new check of it reuses the saved result, and the presence check shows the date it was checked and the date from which a new check measures it again.

The Full PickScore (methodology m2)

The paid measurement starts only after a payment has been verified. It does not use anything measured for the presence check.

• The questions. For each market (a type of business, in a city or area, in a country) a model writes eight short, typical customer questions for an AI assistant; they are written for the market and are not taken from real customer searches. They contain no business name. They are written once per market and used for every business in that market, so businesses in the same market are measured on the same questions. The question set is part of the paid report and is not published.
• The systems. Each question is put, on its own, to each measured system with live web search, asked as a customer in the business's city would ask it: the assistant is told that city and country as the customer's approximate location, the way consumer apps share a phone's approximate location, and is asked to answer in the language of the question. The country is also given to the search tool where the tool supports it. Since 26 September 2026 two systems are measured: Claude Sonnet 5, a model made by Anthropic, through the Anthropic API with live web search, which is the primary system (its answers decide whether there is a PickScore at all), and GPT-4.1 mini through the OpenAI API with live web search. These are the vendors' APIs, not the consumer products: ChatGPT and Claude.ai themselves are not measured. The exact model is recorded with every answer: every answer row in a report says which system and which model gave it, and a report says which systems it was measured on. Other assistants, such as Gemini, Copilot or Perplexity, are not queried, and they may answer differently.
• What counts as named. A second pass lists the businesses each answer named. A question counts for the business when its name, or the other name the owner gave, appears as whole words in the answer's text or in that list. Where the automatic match is uncertain — an ordinary one-word name, or another business whose name is close to it — nothing is counted automatically: a person reads the answer and decides before the report is released. This is naming, not recommendation.
• The score. The PickScore is the number of answers that named the business, out of the number of answers received across the measured systems; a count per system is shown beside it. A question that could not be answered is retried up to three times. If fewer than six questions end up answered, no score is released: the owner is told plainly, and we either complete the measurement or refund the payment. When some questions stayed unanswered, the report says how many, so five answers are never presented as eight.
• Competitor evidence. For each question the report lists the other businesses the answer named. It does not claim those businesses were recommended, only that they were named.
• Technical checks. The website, domain and email checks in the report are run directly over the internet, not by AI. So is a read of the website's own robots.txt file, which shows whether it asks AI crawlers to stay out of the whole site; the report says only what the file asks, never whether an assistant reads the site. When a check could not run, the report says “not tested” rather than “no issues found”.
• The guide and the plan. The advice sections are researched with live web search and written by a model, then read by a person before release. Prices and options in them were researched on the date shown and can change.
• The order of the action cards. Cards are ordered by priority as the plan judged it (high, medium, low); among equals, what we proved comes before what we inferred (a defect one of our own checks found, then one the measurement shows, then one inferred from the named competitors); among those, the quicker fix first; otherwise the order the plan gave. No card estimates a gain in customers or revenue, and marking a card completed moves nothing: only a new measurement can.

How fresh the evidence is

• Every answer carries the date it was asked, the model that answered and the methodology version.
• An answer to a market's question may be reused for another business in the same market for up to 7 days. After that it is asked again.
• Research behind the guide sections is reused within a market for up to 14 days.
• The presence check's signals are reused for 30 days per business, and the sample answer for 30 days per market.
• A report is a reading taken on its date. AI answers vary from one run to the next and change over time; a report does not show a trend.

What PickScore does not measure

• Whether an assistant recommends, prefers or ranks a business, or the order in which businesses appear in an answer.
• Any assistant other than the one named above, and any voice assistant.
• Rankings in Google or any other search engine, website traffic, calls, bookings or sales.
• Why an assistant answered as it did. The report gives researched, likely reasons and what to fix; it cannot see inside the assistant, and no one can guarantee how an assistant will answer.

Review and delivery

Every full measurement and complete report (tiers 2 and 3) is read by a person before it is released, and is delivered at a private link within 24 hours of payment, not instantly. The presence check (tier 1) has no review step and is released the moment payment is verified. Report pages are private and are not meant to be indexed; the public sample is made from a real check with every business name removed.

Change log

• 26 September 2026: m2 introduced. Each question is now asked as a customer in the business's city: the assistant is told that city and country as the customer's approximate location, as consumer apps do, and is asked to answer in the language of the question. Under m1 only the country was given, to the search tool, which the assistant itself does not see; asked a question that said "near me", it asked back where the customer was instead of searching, and that question could not be answered. Nothing measured under m1 is reused under m2. p1 is unchanged.
• 24 September 2026: wording corrected to match the service. The four p1 signals are shown only in the paid presence check (tier 1); the free step is an eligibility check whose result is not shown. p1 and m1 are unchanged: the signals, the 30-day reuse of a saved result and eight questions per market are as before.
• 26 September 2026: a second system measures beside the first: GPT-4.1 mini through the OpenAI API, on the same questions, with a count per system beside the overall one. The primary system alone decides whether there is a PickScore; the other's unanswered questions are shown as partial coverage. Reports measured before this date were measured on one system and say so.
• 24 September 2026: the measured model is named, here and under every answer in a report (Claude Sonnet 5, Anthropic API, live web search); each answer row also shows the date it was asked and, on a released report, the captured answer; the executive summary opens with three lines built from the measurement itself; every defect is one action card with its evidence, a written order (below) and the customer's own progress on it.
• 21 September 2026: p1 introduced (the free AI presence preview replaced the free score). m1 unchanged: eight questions per market, naming counted as whole-word matches, answers reused for up to 7 days.