What AI actually reads when it answers about your business

There is a widespread assumption that AI assistants "know" things, the way a person knows things, and that the job is to get into their memory.

That is not how the answer about your business gets made.

Retrieval, not memory

When someone asks an assistant about a specific local business, a current price, or which provider to use, the useful answer cannot come from training data. Training data has a cutoff, and your prices changed after it.

So modern assistants search. They retrieve pages, read them, and write an answer grounded in what they found — the pattern the KDD 2024 paper formalised as a generative engine: retrieve from an index, then synthesise with a language model.

This has an immediate consequence that most businesses have not internalised:

There is no API that installs your business into ChatGPT's memory. What exists is the web it reads at the moment it is asked.

Why most small businesses come back empty

Run the test yourself and the failures cluster into four shapes.

1. The facts are inside an image

A photographed price board, a scanned menu, a PDF flyer. A human reads it instantly. A retrieval system often gets nothing usable.

2. The facts are written as marketing

"A luxurious balayage experience tailored to you" contains no price, no duration, no availability. It is pleasant and it is unquotable. Compare: "Balayage starts at $150 and takes about three hours."

3. The question is never actually answered anywhere

Customers ask "do you have parking?", "are you open Sunday?", "do you do this for kids?". Most sites never state these in a form that can be lifted into an answer.

4. The facts contradict each other

The website says one price, the booking page another, a directory a third. Given conflicting sources, an assistant will state one of them confidently — possibly the wrong one, to your customer, before they walk in.

What the research suggests you do

The KDD findings translate into unglamorous work:

• Write in question-and-answer form. Use the customer's phrasing as the heading. "How much is X?" beats "Our services."
• Put real numbers in. Prices, durations, hours, capacities. Statistics were the single strongest lever in the study.
• Say where the fact comes from when it matters — a certification, an award, a source. Citations measurably helped.
• Write plainly. Readability gains were smaller than statistics but still real, and they cost nothing.
• Mark it up. Structured data tells a machine which text is the question and which is the answer, rather than making it guess.
• Keep one source of truth. Contradictions are worse than gaps.

The part that is genuinely uncomfortable

You can do all of this and still not appear. The assistant decides. Its retrieval behaviour changes, its model changes, and none of that is announced.

What you control is whether there is something accurate to find. That is a real thing to control — it is the difference between a coin flip and no chance at all — but it is not a guarantee, and you should treat any vendor who offers one accordingly.

Frequently asked