AI Business Check
Methodology
What we report
We report what AI assistants said about a business, check those statements against evidence, and point to fixes on sources the owner controls. There are no rankings, no composite grade and no placement promises.
1. The reference: facts the owner confirms
We find the business with Google Places (when available) and read the website: the home, contact, services, pricing and about pages it links to, and any schema.org LocalBusiness data. The owner sees what each source says, side by side, and confirms or corrects the phone, hours, address, website, services, service area and pricing notes. That snapshot is the reference; public sources are kept separately, each with the date we read it. Ratings and review counts are always stored with their platform and date.
2. The questions
Discovery questions never include the business name, a competitor, or any fact about the business. Questions about the business use its name and location only, never the answer. The location is written into every question, so every assistant sees the same words.
- Discovery
best {service} in {city} - Discovery
{service} near {area} open now - Discovery
affordable {service} in {city} - Discovery
who should I call for {job} in {city} - About the business
Tell me about {business} in {city}. - About the business
What are {business}'s hours, phone number, and address? The business is in {city}. - About the business
Is {business} in {city} currently open for business? - Careful buyer
Who is {business} in {city} not a good fit for? - Careful buyer
What isn't included in {business}'s service or price? The business is in {city}. - Careful buyer
What costs are easy to overlook when hiring {business} in {city}? - Careful buyer
When would someone be better off choosing another {service} in {city} instead of {business}? - Careful buyer
What recurring concerns, if any, are supported by available reviews of {business} in {city}? Cite the evidence and distinguish allegations from established facts.
The careful-buyer question C4 adds “instead of {business}” so the question is about this business.
3. The assistants and how we call them
- OpenAI Responses API with web search:
gpt-5.4-mini - Gemini API with Google Search:
gemini-2.5-flash - Perplexity Agent API (Sonar) with web search:
perplexity/sonar - Claude API with web search:
claude-haiku-5-5(off unless enabled in our configuration)
Every call is a fresh, single-turn request through the provider’s official API with its web search tool on, at the lowest temperature the model accepts (some reasoning models accept none; we record that). Citations come from the provider’s search metadata, never from URLs we find in the answer text. Each answer is stored with the exact model the provider reports, the tool configuration, whether a search actually ran, its timestamp and its cost. A provider error or a failed search is recorded as an error, never as “not mentioned.”
4. Sampling
- Free check: OpenAI Responses API with web search and Gemini API with Google Search; 3 discovery and 2 business questions; one answer each; at most one targeted retry, only for a critical or ambiguous result. Hard cost cap $1.50.
- Verified Business Check: OpenAI Responses API with web search, Gemini API with Google Search, Perplexity Agent API (Sonar) with web search; all 12 questions; two observations of each, at least 45 minutes apart; one more round of discovery questions only when the business appears in some answers but not all. Hard cost cap $6. Reviewed by a person before delivery.
5. Extraction
A low-cost model (claude-haiku-5-5) turns each answer into strict, schema-validated JSON: the businesses named, the facts stated, exact quotes, cited URLs, claims made without a source, and whether the question was actually answered. It extracts; it does not judge. Every quote is checked against the answer text, and anything the answer does not contain is dropped.
6. Fixed rules and labels
Phone numbers are normalized to ten digits; hours are parsed per day and compared with a 15-minute tolerance; addresses are compared on street number, street name and ZIP; websites on their domain; open or closed status against the owner’s confirmation and, for “open now,” against the confirmed hours at the time of the answer. Each finding gets exactly one label:
- Confirmed error
- Contradicts a fact you confirmed.
- Sources conflict
- Public sources disagree with each other; waiting on your confirmation.
- Unsupported claim
- We found no evidence either way.
- Confirmed fact
- Matches a fact you confirmed.
- Subjective / opinion
- An opinion, not a fact that can be checked.
- Not checked
- Outside what this check verifies, or worded in a way our rules could not read.
“Not found” is never “false.” An assistant declining to guess a price is not an error. Claims about complaints, pricing practices or quality are never labeled confirmed errors by software; a person reviews each one before any customer sees it.
Critical means a wrong phone number, wrong hours, reported closed when open, a wrong address, or a service wrongly reported unavailable.
7. The three measurements
- AI Discovery: “Named in X of Y completed discovery answers.” Only discovery answers that completed and answered the question are in Y. Assistants are never compared as if they were the same product.
- AI Accuracy: the number of findings with each label.
- AI Trust Signals: concerns an assistant tied to a source, claims with no source, and answers that said the evidence was insufficient.
They are shown separately and never combined into a score.
8. Competitor gap
The businesses named most often in completed discovery answers, compared on public, checkable facts only: service pages found, a published service area, published pricing, rating and review count with platform and date, and discovery mentions. Anything we did not look up says “not checked.” These are observed differences, not proven causes.
9. Re-runs
Each check stores its benchmark: the exact questions, location, assistants, models, settings, number of runs, scoring rules version, and the confirmed-facts snapshot. A recheck reuses the identical benchmark, so before and after are measured the same way.
10. Abuse limits
The free check verifies your email with a one-time code before any paid call, allows one free check per business every 30 days, limits checks per connection per day, and stops for the day at a global spend cap.
Limitations
- Results are what these assistants returned through their official APIs on the test date.
- Answers vary from run to run, even with the same question.
- Consumer apps (the ChatGPT app, Google AI Mode, Perplexity’s site) may answer differently from the APIs we tested.