A caller asks your AI receptionist whether you are open Saturday. It says yes. Your Google Business Profile says closed. Your website footer says 9 to 2. Your booking tool will not show a Saturday slot at all. One question, four answers, and the caller either shows up to a locked door or books with whoever answered clearly.
In the audits I run for small businesses, the same fact usually lives in four to seven places, and none of them is marked as the one that wins. Nobody planned it that way. A receptionist updated the booking tool. A contractor updated the website. The Google listing was set up in 2022 by someone who no longer works there. Every individual update was reasonable. The result is a business that answers the same question three different ways depending on where you ask it.
AI turns small inconsistencies into lost revenue
Before AI, a stale listing cost you an occasional confused caller, and a human at the front desk absorbed the difference. They knew the real hours, so the wrong footer rarely mattered. AI removes that buffer twice over. Your own AI receptionist or chat assistant states whatever version it was given, confidently, on every call. And the assistants your customers use (ChatGPT, Google’s AI results, voice assistants in cars and kitchens) read your public surfaces and repeat whichever version they decide to trust. There is no human in either loop to say “that is out of date.”
Here is the version that costs money, as an illustrative example with no client data in it. An HVAC company in Aurora, Illinois. Does it serve Naperville? The website says 40 mile radius, so yes. The Google Business Profile lists 12 Kane County towns, so no. The booking tool rejects ZIP 60540. The voice agent was told “we serve the western suburbs,” so it says yes to calls from Indiana.
An AI assistant asked that question tends to lean on the structured listing data, which says no. So you lose a search you would have won, while your own agent books a truck roll to Gary. Both failures trace to the same root: four surfaces, four versions, and no rule anywhere about which one is true.
The fix is a decision, not a tool
Nothing about this requires new software. Somebody with authority decides the actual rule, writes it down once, and publishes it to every surface in one pass. The hard part is the deciding. “Do we serve Naperville?” has a real answer, but in most businesses nobody has been forced to give it, so every surface holds somebody’s best guess from a different year.
The artifact is deliberately boring: one page, or one spreadsheet tab, titled something like Business Facts. It has an owner whose name is on it, a last-reviewed date, and the official version of every fact you publish anywhere. It is not a brand document and not a wiki. It is the answer key.
The seven facts to pin down first
You do not need to document everything. Seven facts cause most of the damage, because they are the ones customers ask about and AI answers about.
1. Hours, including the exceptions.
Not just the weekly grid. Holiday rules, seasonal changes, lunch closures, and the difference between phone hours and door hours. If Saturday is appointment-only, write that as the rule, because “open Saturday” and “bookable Saturday” are different answers to the same caller.
2. Service area, as one definition.
Pick one shape: a radius, a county list, or a ZIP list, and retire the others. A radius on the website and a town list on Google will always disagree at the edges, and the edges are exactly where the marginal jobs live.
3. What you do, and what you refuse.
The refusals matter as much as the offerings. A dental office that does not place implants, or an HVAC company that does not service oil furnaces, should say so in the facts document, so no agent, human or AI, books work you will have to cancel.
4. The prices you are willing to state.
Decide which numbers are public (service call fee, cleaning price, consultation cost) and which are quoted per job. An AI agent that invents a price is a liability; an agent that says “the visit fee is $89 and the tech quotes the repair on site” closes calls.
5. Booking rules.
Do new patients get same-day slots? Is a deposit required? How far out can someone book? These rules live silently inside your booking tool’s configuration. Write them down as sentences, because your voice agent and your front desk need to say them out loud.
6. Phone, escalation, and the after-hours path.
Which number is the main line, who gets the emergency at 2am, and what counts as an emergency. If the answer is “burst pipe yes, dripping faucet no,” that sentence belongs in the document, not in one technician’s head.
7. Payment and insurance.
Which insurers you accept, which financing you offer, whether you take checks. For practices, this is often the first question a caller asks and the last thing anyone updates after a network change.
Publish in one pass, then give it one owner
Once the document exists, the publishing is a checklist, not a project. In one sitting, update the website, the Google Business Profile, the booking tool’s settings, the voice agent and chat assistant configuration, the major directories, and your social bios, all from the same page. The one-pass part matters: spreading it over three weeks recreates the drift you are trying to kill.
Then the process rule: facts change in the document first, and whoever changes it updates every surface in the same sitting. Add a quarterly review to catch what slipped. This is the least glamorous work in the building, which is exactly why it is nobody’s job until you make it someone’s job by name.
Why facts do not belong in your voice agent’s prompt
When businesses deploy a voice or chat agent, the tempting shortcut is to paste the whole facts document into the system prompt. I build these agents, and I am telling you not to. A prompt is just another copy, the fifth place the fact lives, and it starts drifting the day someone updates hours anywhere else. Worse, models paraphrase: a precise rule compressed into prose comes back out as “we are open until 6 most days,” which is a new fact you never approved.
The pattern that holds up is lookup, not memory. The agent reads hours and open slots from the live calendar, checks the service area against a lookup table, and quotes prices from the same data your website renders. The prompt carries behavior: tone, what to do when unsure, when to hand off to a human. The facts stay in systems that have one authoritative value at any moment.
This is how the AI receptionist we run for My Smile Miami works: it books against the practice’s real calendar rather than a prompt’s memory of the schedule, so it can only offer slots that actually exist. In its first month it booked 93 patients, roughly $27K in booked appointment revenue, and the accuracy is most of why. An agent that answers from live systems is trustworthy on the boring questions, and the boring questions are most of the calls.
Where this breaks down
Honest limits, because there are several. The document goes stale the moment it loses its owner; this is a process commitment, not a weekend project, and if nobody’s name is on it you will be back here in a year. Some surfaces cannot express your real rule: Google’s hours grid has no way to say “emergencies only after 6pm,” so you publish the conservative version and let the phone handle nuance. Multi-location businesses multiply everything by the number of locations, and franchise rules can override your preferences entirely.
And automation has its place, but second place. Wiring your website, listings, and agents to read from one live data source is a real project we build, and it is only worth building after the decisions exist. Automating the publishing of undecided facts just distributes the confusion faster.
If you want an outside pass at this, the free AI audit starts exactly here: we find where your facts disagree, which surfaces AI is actually reading, and which conflict is costing you calls. But you do not need us for the first step. Open a blank page, title it Business Facts, and answer the Saturday question once, for good.