Measure what AI answer engines say about a business — then fix it.
AI Business Facts is a working software product operated by Quantumatic LLC. It asks the major AI answer engines — ChatGPT, Gemini, Claude and Perplexity — the questions real customers ask about a business, compares every answer against verified source-of-truth facts, and reports two things: how accurate those answers are and how often the business appears at all.
When the answers are wrong, the tool shows exactly which field is wrong on which engine, and helps publish corrected, machine-readable facts from the business's own website so the engines have something better to read next time.
What the tool does
Four jobs, all built on the same measurement engine. Everything below is available to a signed-up account today unless it is explicitly marked otherwise in what's live and what isn't.
Fact accuracy audit
We ask each engine a fixed set of factual questions — hours, phone, address, website, services, price range, and industry-specific fields — and score every answer against the business's verified facts. You get a per-field, per-engine breakdown of what is right, what is wrong, and what the engine refused to answer.
Appearance rate
Separately from accuracy, we run the "who should I use for X near me" style prompts a customer would actually type, across engines and repeated samples, and measure how often the business is named at all. Low accuracy on answers you never appear in is a different problem from being absent — so we never blend the two.
AI exposure on paid keywords
Upload — or, once connected, sync — the keywords you already pay for, and we check which of them now trigger an AI Overview or AI-generated answer above the ads, on Google and Bing. That tells you where paid clicks are being intercepted by an answer you are not part of.
Publish and prove the fix
Corrected facts are published as machine-readable structured data (JSON-LD, a business.json endpoint and an llms.txt file) from the business's own domain via our WordPress plugin, search engines are notified, and every change is stored as a dated receipt so you can prove the fix shipped.
How a measurement actually runs
- Establish the truth. Before we ask any model anything, we build a source of truth: the business's Google Places listing, a scrape of its own website, and the facts the owner or agency has confirmed in their account. Nothing is scored against a guess.
- Query the engines through their official APIs, with web search switched on. Questions go to OpenAI, Google Gemini, Anthropic and Perplexity over their paid, documented developer APIs, every one of them asked to search the live web — because the product question is "what does AI tell a customer right now," not what a model memorised. We never script, scrape or automate the consumer chat interfaces of any provider.
- Record whether it actually searched. Asking an engine to browse doesn't mean it did — any of them may answer from memory with the search tool sitting unused. So we read each provider's own retrieval receipt (citations, grounding metadata, search-tool results) and mark every answer as grounded or not. If a grounded call fails, we never quietly re-ask a non-browsing model and score that instead: the answer is recorded as missing, with the error. Every report shows how many of its answers were confirmed live-retrieved.
- Score field by field. Each answer is compared to the truth value with a matcher built for that field — phone numbers normalised to digits, URLs normalised by host and path, hours parsed and compared by day. "I don't have that information" scores zero rather than being quietly skipped.
- Repeat and report the uncertainty. Answers move between runs, so a single run is not a measurement. Results are reported with the sample size behind them and an interval — never as a bare point estimate that implies more precision than the data supports.
- Watch for drift. Scheduled re-runs track the same measurements over time and flag regressions, so a correction that silently stops holding becomes an alert rather than a surprise.
How we report numbers
These are product rules, not marketing copy. They constrain what the software is allowed to display.
- Accuracy and appearance rate are the headline metrics. We do not headline a "rank" or a "share of voice" number against competitors. AI answers are not a ranked list, and presenting them as one would be a made-up precision.
- Every number carries its sample size and interval. A percentage is shown with the number of samples behind it and a confidence interval, and with the run's grounding rate — how many of those answers we could confirm came from a live web search rather than the model's memory.
- We'd rather report nothing than report a guess. When the engines don't return an answer we can check, that field shows "could not verify" and is left out of the score entirely — it isn't counted as the business getting something wrong. A verified correction and a lost-visibility alert both require grounded evidence before we'll put them in front of you.
- Official provider APIs only. Every model call is a billed API call under the provider's developer terms. No consumer-UI automation, ever.
- No autonomous writes to your accounts. Nothing is published, changed or sent on your behalf without a human approving that specific change, and every write is logged with a timestamp.
- No connector is a single point of failure. Every place we offer an API connection, the manual path — CSV upload, typed facts — keeps working. If a connection breaks or you never grant one, the product still functions.
The Google data the tool uses
Some of what the product does depends on Google APIs. Here is the complete list, in plain terms. The privacy policy names the exact scopes, fields, storage locations, retention periods, and how to revoke and delete each one.
| Google API | What we do with it | Status |
|---|---|---|
| Google Sign-In openid, email |
Optional way to sign in to an existing account. We receive your Google account ID and email address, and use them only to find the matching account and start a session. We never post anything to your Google account. | Live |
| Google Ads API adwords |
If you connect your Google Ads account, we read keyword and search-term performance — keyword text, impressions, clicks, cost, conversions and conversion value — so the AI-exposure report can run against the keywords you actually spend on instead of a CSV you export by hand. adwords is Google's only Ads scope and there is no read-only version of it to request; what we do with it is read. We never create, edit, pause or budget a campaign, and we never write conversions back, and every change we ever make on your behalf needs your approval and is logged. |
In review |
| Google Business Profile API business.manage |
If you connect your Business Profile, we read the profile fields AI answers are grounded in — name, address, phone, hours, categories, attributes, website, review count — and compare them to what the engines say. Corrections are written back only for the specific fields you approve, one approval per change. | In review |
| Google Places API | Looks up publicly listed business information (name, address, hours, phone, rating) to build the source of truth for an audit. Public listing data — not data from your Google account. | Live |
| reCAPTCHA v3 | Protects the free audit form from automated abuse. | Live |
Connections marked "In review" are not enabled yet. The connector code and the manual equivalents are built, and the OAuth applications are going through Google's verification and Google Ads API review. Until that clears, the AI-exposure report runs from a keyword report you export from Google Ads and upload yourself, and Business Profile facts are read from public listing data. We would rather tell you that than show a button that does not work.
Our use of information received from Google APIs adheres to the Google API Services User Data Policy, including the Limited Use requirements. Google user data is never used to train AI models — ours or anyone else's.
What's live and what isn't
An honest capability list. If it says "in development," it is not something you can use today, and we don't bill for it.
| Capability | Status |
|---|---|
| Free AI visibility audit (ChatGPT + Gemini, public) | Live |
| Fact-accuracy audit across four engines, per field and per engine | Live |
| Appearance-rate measurement with repeated sampling and intervals | Live |
| Cross-engine consistency view (where engines disagree with each other) | Live |
| AI Overview tracking on Google and Bing for tracked keywords | Live |
| AI-exposure report from an uploaded Google Ads keyword report | Live |
Structured-fact publishing via the WordPress plugin (JSON-LD, business.json, llms.txt) + search-engine notification | Live |
| Scheduled re-audits, history, drift alerts and client-ready PDF summaries | Live |
| Agency workspace: many client businesses, per-client access, optional white-label | Live |
| Google Ads direct connection (replaces the manual upload) | In development |
| Google Business Profile read + approval-gated write-back | In development |
| Microsoft Advertising direct connection | In development |
| Shopify plugin (WordPress is supported today) | In development |
Who it's for
A business that wants to be found
You run one business and customers increasingly ask an AI assistant instead of opening a map. You want to know whether the answer they get is right, and to fix it when it isn't. Sign up directly, connect your site, and work through the corrections the audit surfaces.
An agency with a book of clients
You manage many businesses and need a repeatable, defensible measurement to put in a monthly report — with the sample sizes and intervals visible, so a client can challenge a number and get a real answer. One workspace, per-client dashboards, optional white-label branding.
What this tool is not
- Not a rank tracker. There is no position #1 in a generated answer. We measure whether you are named and whether the facts are right, not where you "rank."
- Not a guarantee. No one controls what a model says. We measure it, we improve the machine-readable inputs it can read, and we re-measure. Anyone promising a guaranteed placement in AI answers is selling something we won't.
- Not a content spinner. We do not mass-generate pages to game an engine. The corrections we publish are the business's own verified facts, in a format machines can parse.
- Not an autonomous agent on your accounts. Connections are read-only by default; the one write path (Business Profile corrections) requires you to approve each change.
See what AI says about your business.
The free audit runs in about a minute and needs a business name and an email address. No card, no install.