Method v1.1 · Sep 26, 2026

How we measure

A Pickgap score is a standardized test, not what every customer sees. It asks the same local questions the same way each time, so scores can be compared over time and between businesses. This page is the whole method, and every report links to the version it used.

What we measure

Pick rate: the share of AI answers that name the business, out of 250 answers in a full scan. We report it overall and for each AI surface, always with its noise range and its sample, such as “12% ±4 · 30 of 250 answers”. It is not a rank. AI surfaces don't keep a fixed list, so we count how often a business is named across many answers instead.

The questions

Each business gets 25 questions a customer in its city would type. They cover 6 question types, and every business is asked all of them:

• Templates per trade. Dentists, lawyers, HVAC companies, plumbers and med spas have their own question sets, written the way their customers ask. Other trades use general templates filled with their services.
• Duplicates removed. Before a scan runs, questions that ask the same thing in different words (“best dentist in Tampa” and “top rated dentist in Tampa”) are found by meaning and replaced, so no more than 10% of a business's questions overlap.
• Service-area businesses with no public address aren't asked about walk-ins or office visits.
• The same questions every time. A change in the score is a change in the answers, not in the test.

The AI surfaces and the searcher's location

We ask 5 AI surfaces through their official APIs or a Google-results service, with web search on, as a searcher in the business's city. Where a surface accepts a location, we send the city as the searcher's approximate location:

Google doesn't show an AI Overview for every search. We record whether one appeared at all (“AI Overview shown in 14 of 50 searches”) separately from whether it named the business. A search with no AI Overview counts as an answer that named nobody, because that is what the searcher saw.

Runs per question

AI answers vary from one ask to the next, so a full scan asks each question 2 times on each surface: 25 questions × 5 surfaces × 2 runs = 250 answers. A weekly check asks the first 5 questions once on each surface (25 answers), a prospect audit the first 20 (100 answers), and the free check the first 5 (25 answers). The smaller the sample, the wider the noise range.

Finding the business in an answer

A parser lists every business an answer names, in order. The business being measured counts as named when any of these match:

• Its name, compared on the distinctive part: legal forms (LLC, PLLC, DDS), trade words (“dental”, “law firm”) and the city are set aside, and the rest must be at least 85% alike. “Bright Smile Dental” matches “Bright Smile Dentistry of Tampa”; “Smile Dental” doesn't.
• Other names it goes by, added on the business as aliases.
• Its website, when the answer links to its domain.
• Its phone number, when it appears next to the mention.

A business listed twice in one answer counts once, and its position is where it first appears. We lean conservative: a false match would inflate the score, which is worse than a miss. Other businesses in the answer are its competitors; spellings of the same competitor are grouped by the same name rule.

The noise range

Every pick rate comes with a range, written as ±. We resample the questions 1,000 times, with replacement, and take the middle 90% of the pick rates. Questions are the unit we resample because the answers to one question move together. With few questions, or a pick rate of 0% or 100%, resampling alone would claim more certainty than the sample has, so the range is widened to at least the 90% Wilson interval. The resampling is seeded, so the same answers always give the same range.

Movement labels

When we compare a scan with the previous scan of the same kind, we label the change:

• Up or Down, with the points (“Up 8 points”), only when the change is bigger than the combined noise of both scans.
• No real change, with the combined noise (“No real change (±6)”), when it isn't. The two ranges add up the way independent errors do: the square root of the sum of their squares.

Alerts follow the same rule: we email only when the pick rate moves at least 10 points and the move is outside the noise.

The other numbers

• Pickgap: the points between the business and the competitor named most often.
• Share of voice: the business's mentions against the top 5 competitors'.
• Cited sources: the sites the AI surfaces linked to when they answered, and how often the business was named alongside each.
• The answers: the full text of every answer is kept with the scan, and your team can read and search it in the app, so every number can be checked against the answers behind it.

When a surface fails

Each call is retried up to 3 times. If a surface still doesn't answer, the scan completes with the others, as long as at least 2 surfaces answered, and the report says which answers are missing and how many the score uses. Failed answers are left out of the pick rate, not counted as misses.

What it can't tell you

• What one person sees. The apps personalize answers by exact location, history and account, and change their models often. A standardized test, not what every customer sees: results in the apps can differ.
• Why an AI chose a business. The fix list points to what the picked businesses have in common (reviews, listings, website content), not to a guaranteed cause.
• Future answers. AI surfaces change; that is why tracking asks the same questions again every week and every month.

The fix list

Fixes come from rules, not a language model, and each shows the evidence behind it: review counts and recency against the top 3 competitors (from Google), sites the AI surfaces cite that don't name the business, what its website says about its city and services, and whether its details match across sources.

Versions

This is method v1.1, from Sep 26, 2026. When we change how questions are written, asked, matched or scored, the version changes and reports say which one they used. Scans from before September 26, 2026 used v1.0: 25 questions on ChatGPT, Gemini and Perplexity (150 answers) with a 95% range.

Questions about a result? Email hello@pickgap.com. See also the sample report.