Skip to content

How the score is measured

Engine version 2026-11-01.1. Page updated 2026-10-21

The AI Visibility Score measures, from 0 to 100, how often AI engines recommend a local business when people ask them where to go, how high they place it, and how well they speak of it. Citation computes it in code from real answers that are stored once and never change.

When someone asks an AI engine for the best tacos in the neighborhood, your business either comes up or it doesn’t. Citation measures which. A score that grades your livelihood should show its work.

So this page shows all of it. Every answer in a report is a real, stored engine response: measured, never simulated. Every number is computed by code from those stored answers, not estimated by an AI. And we measured the instrument’s own error bars, so a report can tell a real move from noise.

How a report is made

How a report is made, as one instrument. An isometric drawing of a bench instrument. A question card on a stand is wired to three engine cabinets; their answers go into an archive of drawers; a reading head over a tray of answer slips classifies them; a calculator computes the score; a plate shows the five grade bands.OpenAIGeminiClaudeABCDF

Tap a number to see that step on a real report.

Every answer is a real, stored response

An audit asks AI engines the questions a real customer would ask when looking for a business like yours, then records who each engine recommends.

The question

Who has the best

Direct question, intent weight 1.00

OpenAI (web search)

San Francisco boasts a rich seafood heritage, offering a variety of exceptional dining experiences. 5 (4533 reviews) 399 The Embarcadero, San Francisco, CA 94105 Situated along the Embarcadero, Waterbar offers stunning bay views alongside a diverse seafood menu.

Named, in order

  1. 2Scoma's Restaurant
  2. 4Sotto Mare
  3. Waterbar Restaurant, position not stored
  4. Angler, position not stored
  5. Tadich Grill, position not stored

Not named: Sam's Grill & Seafood Restaurant

Adds to the score0.00×(1 + sentiment) / 2×1.00=0.000

Sam's Grill & Seafood Restaurant on this engine, all questions: 9.7 of 100

Gemini (Google Search grounding)

Because San Francisco’s seafood scene is so deeply rooted in history—stretching from the Gold Rush era to modern Michelin-starred innovators—the "best" seafood restaurant depends on the exact dining experience you are looking for.

Named, in order

  1. 2Sotto Mare
  2. 7Scoma's Restaurant
  3. Swan Oyster Depot, position not stored
  4. Hog Island Oyster Co., position not stored
  5. Thanh Long, position not stored
  6. Tadich Grill, position not stored
  7. The Anchovy Bar, position not stored

Not named: Sam's Grill & Seafood Restaurant

Adds to the score0.00×(1 + sentiment) / 2×1.00=0.000

Sam's Grill & Seafood Restaurant on this engine, all questions: 47.0 of 100

Claude (web search)

"Best" is subjective and varies by source, but a few names come up consistently across rankings as San Francisco's top seafood spots: Swan Oyster Depot (Polk Gulch) is the most frequently praised, iconic pick.

Named, in order

  1. 2Sotto Mare
  2. 3Sam's Grill & Seafood Restaurant
  3. Swan Oyster Depot, position not stored
  4. Anchor Oyster Bar, position not stored
  5. Hog Island Oyster Co., position not stored
  6. Dominique Crenn's Marina restaurant, position not stored
  7. Waterbar, position not stored

Adds to the score0.50×(1 + 0.70) / 2×1.00=0.425

Sam's Grill & Seafood Restaurant on this engine, all questions: 59.8 of 100

Read every stored answer on the sample report
One question from the sample report for Sam's Grill & Seafood Restaurant: what each engine answered, who it named, and what that answer adds to the score.
Stored
Every answer verbatim; a report reads only that record. Never fabricated, simulated or paraphrased.
Labels
The API and its grounding: OpenAI (web search) is the OpenAI API with web search on, not the ChatGPT app.
Not the app
Apps add personalization and memory; the API is the closest reproducible proxy.
Claude
Its probe runs web search on the Sonnet model.

The same questions for everyone

Every audit runs a fixed matrix of query templates, filled in deterministically from the business profile (category, neighborhood, city and state), the way the generic matrix asks “best pizza restaurant in Temescal, Oakland”. The kind of business is Google’s main category for the listing. When that category is broad, such as Restaurant, Store or Cafe, the audit uses the first more specific kind Google also lists, and the owner can pick another of Google’s kinds before the audit runs. For a few kinds whose Google name misleads, the questions can use plainer words from a short list we keep, such as boba shop for Google’s Tea house, and the businesses it is compared with are then found by those words anywhere within 10 miles, not only within the usual 2. The report says which one it asked as. We never ask the engines about you by name. Every question is one a customer would ask without knowing you exist.

Stored rows from the report for Sam's Grill & Seafood Restaurant. Open the report

Direct, Direct questionBest in the area, Who has the best, Top rated, One to recommend×1.0
Near, Near meWorth going to nearby, A good one nearby, Close by, any suggestions×0.9
Occasion, OccasionSpecial occasion, First-time visitor, This weekend×0.8
Attribute, AttributeGreat service, good reviews, Affordable, high quality×0.7
The shipped matrix: 12 templates across four intent families, named by the labels the report uses.
Same profile
Same queries, so scores in a category compare.
Verticals
Specialized verticals swap in their own matrix of the same shape.
Public
The structure and the weights. The verbatim query strings stay private.
Why private
Public probes could be gamed word for word; search engines keep theirs private too.

Three matrices, one shape

A generic matrix would ask nonsense of a law firm (“best lawyer for a special occasion”), so verticals whose customers ask differently get a matrix of their own.

Every matrix shares one intent shape and the same weights, so the score denominator is identical across verticals and a repair shop’s B means the same thing as a taqueria’s B. One illustrative query is shown per matrix; the rest stay private.
Local businesses (generic)Direct 4 at 1.0, Near 3 at 0.9, Occasion 3 at 0.8, Attribute 2 at 0.7

Every business that is not a law firm or an auto repair shop: restaurants, cafes, shops, services.

Occasion here means: A moment a customer is planning for, like a date night or a group dinner.

“best pizza restaurant in Temescal, Oakland”

Estate planning attorneysDirect 4 at 1.0, Near 3 at 0.9, Occasion 3 at 0.8, Attribute 2 at 0.7

Law firms are typed as a bare lawyer by Google Places, so the practice area is set at enrollment and selects this matrix.

Occasion here means: A situation that sends someone looking: setting up a living trust, probate after a death, writing a will.

“best estate planning attorney in Oakland, CA”

Auto repair shopsDirect 4 at 1.0, Near 3 at 0.9, Occasion 3 at 0.8, Attribute 2 at 0.7

Google Places types the business as car repair, so no one has to tag it.

Occasion here means: A situation an owner types in a hurry: a check-engine light, a brake noise, a pre-purchase inspection.

“best auto repair shop in Temescal, Oakland”

Shape
Every matrix keeps the same four families, counts and weights; only each family's meaning changes.
Trades
Home-services trades (plumbers, electricians, roofers) are not audited yet; a report never starts with questions meant for another kind of business.
New verticals
Published here, shape, weights and one example, before the first audit runs on them; the rest stays private.

How the score is computed

For each engine, every query contributes to a score from 0 to 100. Three quantities determine each query’s contribution:

Rank weight, by the position an engine names you at

  • Position 11.00
  • Position 20.70
  • Position 30.50
  • Position 40.30
  • Position 50.30
  • Position 6 or lower0.15
  • Not named0.00
DFDCBA
The report's grade object for the sample report, Sam's Grill & Seafood Restaurant: the letter, the five bands and a tick at the stored score.
rank_weight
Where the engine placed you, weighted as the bars above show.
sentiment
How the answer characterizes the business, from −1 to +1, rescaled so a neutral mention still earns half credit.
intent_weight

The weight of the query’s intent family in the matrix: how much that kind of question matters to a real customer.

Direct1.00
Near0.90
Occasion0.80
Attribute0.70
score = 100 ×Σ( rank_weight× (1 + sentiment) / 2× intent_weight )Σ intent_weight
Absence
A question that does not name you stays in the denominator, so absence costs points.
Headline
The simple mean of the per-engine scores, over the engines that completed.
Failed engine
Excluded, never averaged in as a zero.
Grades
Map directly from the blended score:
GradeBlended score
A80 and above
B60 to under 80
C40 to under 60
D20 to under 40
F0 to under 20

The AI classifies; code computes

Run the same stored answers through the pipeline twice and you get the same score twice.

A language model

  • Reads each stored answer: whether a business is present, at what rank, and with what sentiment
  • Lists the other businesses an answer names
  • Pulls the stated facts out of the fact answers
  • Writes the plain-language prose in the report

Code

  • The query matrix
  • The intent weights
  • The score
  • The grade mapping
  • The gap detection triggers

What counts as a gap

A report names your top three gaps. Each one is detected by a fixed rule over stored data, never by an AI deciding what looks wrong.

  1. 1

    AI answers cite third‑party sources, not your site

    11 of 36 answers

    One fix closes both: add your website to your Google Business Profile.

The rule that put this card on the report:

citation_source
Your site is cited in fewer than 33% of the answers that cite anything (domains counted per answer). With no website on file at Google it always fires, and when a cited domain matches the business name the card says the site is cited but not on file.

Stored rows from the report for Sam's Grill & Seafood Restaurant. Open the report

places_profile
The stored Google profile lacks a field engines draw on: the website, the neighborhood, or both. No threshold.
review_delta
A named competitor has at least 2× your reviews or a rating 0.3 stars higher, same Places snapshot; the headline names which, the evidence quotes both.
accuracy
At least one fact graded wrong against your Google listing; it then ranks first. Not stated and unclear never fire it.
generic_category
Your Google main category is broad (such as Restaurant, Store or Cafe) and the audit asked about a more specific kind. It states the two categories only; it does not touch the score.
Prose
The AI only explains a gap code found; each gap quotes its stored evidence.
Not shown
A gap that does not fire. A leading rating is said, never made a deficit.

Five facts, graded in code

Being recommended is not enough if the answer then gets your details wrong, so every audit checks up to five facts against your Google listing.

AddressCorrect

Google lists
600 Guerrero St, San Francisco
Stored rows from the report for Tartine Bakery. Open the report
Questions
One per fact your Google listing holds, up to five: hours, phone, address, kind of business, website.
Question text
Only your business name and town, never the phone or hours, so it cannot hand over the answer.
Engines
Free: Gemini (Google Search grounding). Pro: OpenAI (web search), Gemini (Google Search grounding) and Claude (web search). Stored and cached like every answer.
Reading
A language model reads each answer once: stated or not, its exact words, how sure. It never sees Google's value.
Found
A claim counts only if found in the stored answer, as one stretch or, for a joined list, every number in order and most words. A number the answer never gave fails.
Listing
Your Google listing frozen at audit time, refreshed when more than seven days old.
Hours
Every named day within 15 minutes of Google, opening and closing; regular hours only.
Schedules
Any contradicted day is wrong; a partial schedule with no contradiction is correct.
Phone
Any number matching the listing's last ten digits is correct; wrong only if none match; a second location is not wrong.
Address
Street number and a shared street name; suite, city and zip ignored.
Website
The registered domain; the page itself for social and booking hosts; the full host name for shared site builders.
Category
A shared word with Google's category is correct; anything else unclear; never wrong in this version.
Not stated
Not given, said not to know, or sent the reader elsewhere. Counts against the engine: an assistant that cannot state a fact cannot act on it.
Unclear
Under 0.7 sure (a floor we chose, not from a study), claim not found, or nothing to compare; shown, left out of the count.
Headline
Correct facts out of every fact except the unclear ones.
Score
Not part of the AI Visibility Score; it never moves the grade.
Version
Grader 2026-09-30.1, stamped on every verdict. Every figure is a code rule or your own count.

Agent Reach, checked in code

Agent Reach checks whether an AI agent can open a business's website and read it. It is never a score and never moves the AI Visibility Score.

Reachable: Partial.

The homepage has no LocalBusiness structured data, so agents have to guess your hours and address.

Fix: Add LocalBusiness structured data with the same name, street address, phone and hours as your Google listing, so agents can read your hours and address correctly.

Not part of your AI Visibility Score.

Stored rows from the report for Tartine Bakery. Open the report

Reads
The website on the Google listing, once per audit, HTTPS only, obeying its robots.txt, at most 3 fetches: robots.txt, the homepage and one page that should not exist. Code only, no language model.
AI crawlers allowed
None of GPTBot, OAI-SearchBot, ClaudeBot, Google-Extended, PerplexityBot is blocked in robots.txt, and the homepage shows bots no challenge page.
Readable without JavaScript
Before any script runs, the homepage text has the name, the city and hours, menu, services or phone, in at least 300 characters.
Hours and address readable
LocalBusiness structured data matching Google's name, street number and postal code, phone and hours, so agents read them correctly.
Real 404 for missing pages
A page that does not exist answers 404, not a normal page or a redirect home.
Page identity tags
A title, a meta description, a canonical link and an og:title.
Result
One line, the worst result, with its evidence and fix. Fail: the first check fails, or two of the other four do. Partial: any partial, or one of the other four fails. Otherwise pass.
No website
Nothing is fetched, and the line says so.
Version
Checks 2026-10-08.1, stamped on every stored result.

How we keep it honest

The noise floor was measured, not chosen

AI engine answers vary between runs even when nothing about a business changes.

10-point bar
Same day5.70 points
Across days6.70 points
Audit to audit9.46 points
Three ways of repeating an identical audit (within one day, across days, and between full audits), each drawn as two standard deviations in score points, the same units as the 10-point bar. Run nf-2026-07: 17 clean repeat audits, completed 2026-07-18.
Measured
Repeated identical audits across several days, before any move counted as signal.
The bar
Under 10 points on the blended score is noise: two audit-to-audit standard deviations, rounded up.
In your report
Score history reads a move under 10 points since your last audit as steady; only a larger one is a real gain or drop.

We ask one question twice

Every audit asks its first question a second time, up to 200 repeats a day across all audits, and reports what came back.

OpenAI named 2 businesses the first time. 1 of them came back when we asked the same question again. Your business appeared both times.

OpenAI (web search)

Gemini named 2 businesses the first time. 2 of them came back when we asked the same question again. Your business appeared both times.

Gemini (Google Search grounding)

Claude named 5 businesses the first time. 5 of them came back when we asked the same question again. Your business appeared both times.

Claude (web search)

In Citation's own repeat runs the score moves about 5 points between identical asks, which is why nothing under 10 points is reported as news.

Stored rows from the report for Tartine Bakery. Open the report

Engines
Free: Gemini (Google Search grounding). Pro: OpenAI (web search), Gemini (Google Search grounding) and Claude (web search).
Same ask
Same question text, engine and UTC day, under its own cache key so it is never served the first answer.
Two days
If the asks would span two UTC days, nothing is stored.
Compared
The businesses each answer named: intersection over union (Jaccard), computed in code from classifier-matched names.
Names
Normalized as everywhere else, so a respelling is not a different business.
Shown
Counts, not the ratio: one pair of answers is not a rate.
Never scored
Not in the score, leaderboard, gaps or answer count, so the reading can never move the grade.
Movement
Run nf-2026-07: audit-to-audit standard deviation 4.73 points, about 5, which sets the 10-point bar.
Average
None published yet; it will be computed from stored pairs once there are enough.

An audit never changes

An audit is an immutable snapshot. The engine answers, mentions, scores, and gaps are stored once, and a report URL renders from those stored values forever.

Read later
Nothing is re-generated when a report is opened.
Stamped
Every audit carries the engine version at the top of this page.
Versions
A change to probe models, extraction, the analyzer or the scoring math bumps it; versions are never compared, and score history starts a new line.

What an audit costs

Real answers cost real money. Every probe is a paid API call with web search turned on.

One cold audit, 36 scored answers, priced from the token counts the providers reported on real audits (measured 2026-08-26, before the fact questions and the repeated question were added). A same-day repeat of a query on the same engine is served from the stored answer and costs nothing.
OpenAI (web search)$0.05
Gemini (Google Search grounding)$0.49
Claude (web search)$0.681
Analyzer (classification)$0.027
Gap prose$0.001
Per cold audit$1.249
Why published
We publish the bill so the limits on the free tool read as arithmetic rather than a sales tactic.
Limits
Free: 5 audits a day per visitor, 150 a day for everyone, and a daily spending breaker that pauses new audits before the budget is gone. Pro: 10 a day and a weekly re-audit of one business.

Every version of the instrument

Three parts of the instrument carry their own version: the engine, the fact grader and the Agent Reach checks. The current engine version is 2026-11-01.1.

  1. 2026-07-18.1The engineD27

    Versions begin. The noise floor was measured (run nf-2026-07) and the movement threshold was set from it at 10 blended points. Every audit from here on carries its version.

    Not comparable because: Audits before this date carry no version and are never compared for alerts.

  2. 2026-07-29.1The engineD28

    Estate-planning attorneys got their own query matrix, place names in legal queries gained their state, and an engine that returned nothing scoreable is dropped from the blend instead of counting as zero.

    Not comparable because: A blended score with a dropped engine is a mean over fewer engines than before.

  3. 2026-08-25.1The engineD57

    Claude (web search) became the third engine. The blended score is now a mean over three engines.

    Not comparable because: A three-engine mean and a two-engine mean measure different things. This version was stamped only on internal audits; it was replaced the next day before any public report used it.

  4. 2026-08-26.1The engineD57, D58

    Every place name in the generic matrix now carries its parent (neighborhood with city, city with state), because location-blind engines answered for the wrong metro. Shipped together with the third engine so public reports saw one instrument change, not two.

    Not comparable because: Changed query text changes what the engines are asked, so answers before and after are different measurements.

  5. 2026-09-30.1The fact graderCurrentD90

    Fact grading begins. Hours agree when every day the answer names matches Google within 15 minutes per opening and closing time; a partial schedule with no contradiction counts as correct. Phone numbers agree on their last ten digits. Addresses agree on the street number and a shared street name, ignoring suite, city and zip. Websites agree on the registered domain, or on the page for social and booking hosts. Category counts as correct on a shared word and is never graded wrong in this version. A claim the extractor is less than 0.7 sure of, or that does not appear word for word in the answer, is shown as unclear and left out of the count.

  6. 2026-10-08.1Agent ReachCurrentD88, D94d (D94 addendum 1)

    The Reachable line begins. Code reads the website on the Google listing once per audit, over HTTPS only, obeying its robots.txt, with at most 3 fetches: robots.txt, the homepage and one page that should not exist. Five checks: AI bots allowed (GPTBot, OAI-SearchBot, ClaudeBot, Google-Extended and PerplexityBot not blocked, and no bot challenge); content without JavaScript (name, city, and hours, menu, services or phone in the raw page, with at least 300 characters of text); hours and address readable (LocalBusiness structured data whose name, street number and postal code, phone and hours match Google); a real 404; and basic identity (title, meta description, canonical link, og:title). The line shows the worst result, never a score.

  7. 2026-11-01.1The engineCurrentD118

    All three engines now answer with the model each consumer app gives its free users: GPT-6 Luna for OpenAI, Gemini 3.5 Flash-Lite for Gemini and Claude Sonnet 5.5 for Claude. Each question allows at most 3 web searches.

    Not comparable because: A different model gives different answers to the same question, so scores before and after are different measurements.

Engine entries
Each says why scores across it are not compared. A version cannot ship without its entry: a test fails if they disagree.
Models
Each engine answers with the model its consumer app gives free users: GPT-6 Luna (ChatGPT desktop app), Gemini 3.5 Flash-Lite and Claude Sonnet 5.5. Last checked against the apps on 2026-11-01.
Versus the apps
On 2026-10-31 we asked the ChatGPT, Gemini and Claude apps 5 of our questions and compared the businesses they named with our stored answers. Ours named 48% of the businesses the Gemini app named, 42% of the Claude app's and 23% of ChatGPT's. The apps also use the asker's location, the time of day, a maps tool and, in ChatGPT, the user's memory. An API probe has none of these, so Citation measures what the engines know from the web, not any one person's results.
Not a bump
Copy and design changes to this page. Only what is asked, how answers are read, or how the score is computed bumps a version.

See where you stand

The instrument this page describes will grade your business against the ones down the street: real questions, real answers, scored the way you just read. The report is free and yours to share.