Results · AI test · Method
AI search test: how we check what assistants say about a manufacturer
We ask AI assistants the questions a buyer's engineer would ask about a company, and check every answer against the company's confirmed record. This page sets out the method, and how to run a rough first pass on your own plant. Results appear here only with the client's written permission, each with its date and the assistants we asked.
Two tests, kept apart
This page describes the AI assistant test. We ask named assistants about one company on a stated date, and score each answer against what the company confirmed in its Blueprint.
It is separate from the AI Knowledge Test, which grades a website's own content against ten buyer questions and asks no assistant anything. The two measure different things, so we never add their results together. Both methods are on our methodology page.
Method
Four steps, run the same way every time so a re-test can be compared with the first run.
-
Step 1, OP 10
Questions
The questions a buyer's engineer would ask before a call: limits, specs, certifications and fit for the application.
-
Step 2, OP 20
Assistants
Named assistants, each asked the same wording, on a stated date.
-
Step 3, OP 30
Scoring
Each answer marked correct, partial, wrong or 'I don't know', against the company's own confirmed record.
-
Step 4, OP 40
Checks
Every contradiction checked twice, and every count reconciled before it's published.
Why assistants guess
An AI assistant answers from what it can read. When a website doesn't state a limit, a spec or a certification, the assistant fills the gap with what's typical for that kind of company, and says it with the same confidence as a fact.
A confident wrong answer costs more than 'I don't know'. An engineer who hears the wrong line speed or the wrong certification moves on to the next name, and the company never learns the question was asked. Nothing in its analytics shows the call that didn't come.
The fix starts on the website. Each wrong answer traces back to a sentence that isn't there, or a number locked in a PDF or a photo.
What the fix looks like
We put the missing limits, specs and proof on the pages buyers and assistants read. Each fact goes up as plain text, with its unit, after the client confirms it.
Then we ask the same questions of the same assistants on a stated date, and set what changed beside the first run.
Run a first pass on your own plant
You can run a rough version of this test on your own company in about an hour.
-
Step 1, OP 10
Write ten questions
The ones your buyers' engineers ask before a call: biggest part, tightest tolerance, materials, certifications, lead time.
-
Step 2, OP 20
Ask two or three assistants
Use the same wording in each, and write down the date.
-
Step 3, OP 30
Score against your floor
Correct, partial, wrong or 'I don't know', judged against what your plant actually runs.
-
Step 4, OP 40
Trace each miss
Find the page where the right answer should be. Usually it isn't there yet.
See the lesson
The Hoist Says 15 Tons
What happens
- 01An engineer asks an AI assistant who can machine a big casting
- 02The assistant answers with confident numbers
- 03The hoist plate on the floor says otherwise
- 04The shop never hears the answer was given
Questions about the test
Can you run this test on our company?
Yes. It runs inside a program, after the Blueprint, because we need your confirmed record to score the answers against. We scope it on the review call after your free Buyer's-Eye Audit.
Will fixing our pages make AI assistants recommend us?
We can't promise that, and nobody can. What we can do is put your facts where assistants read them, then ask the same questions on a stated date and show you what changed.
Find out what AI says about your plant
Start with the free Buyer's-Eye Audit, and we'll scope it on the review call after it. Every program starts with the paid Buyers Blueprint, which gives us the record to check the answers against.