The method, as planned
How we would measure what buyers are told
Nobody can open an assistant and read off where you rank, because there is no ranking to read. What you can do is ask the questions your buyers ask, many times over, and write down what comes back. This page walks through how we’d do that and where it stops being reliable. None of it is built yet.
What’s a long-lasting perfume under $100?
Competitor A
Smoky amber and cedar. 50 ml, $85, lasts through the evening.
Competitor B
Sweeter floral with rose. 75 ml, $96, easy to find online.
Competitor C
Lighter citrus for daytime. 50 ml, $70, sold at major retailers.
Your brand
Not named in this sample. A likely reason: the page never states notes, sizes or where to buy.
Choose the questions
- What happens
- We start from how your buyers phrase a decision, using sales calls, support emails, reviews and the searches you already see. We keep a few dozen questions, grouped by stage: discovering options, comparing two, and checking that a choice is safe.
- What you would see
- A question list you could explain to a colleague, tagged by stage, market and language. Questions written the way people actually write them in each language.
- Where it gets shaky
- No assistant tells outsiders how often a question is asked. Your list reflects your buyers and says nothing about market volume.
Ask each one repeatedly
- What happens
- Every question is asked many times, per assistant, market and language, because the same question gets different answers. We would record which route was used: the consumer app that buyers see, or the developer interface that is easier to run at scale.
- What you would see
- A count of runs next to every result, and the route used. For example: “named in 3 of 10 runs.”
- Where it gets shaky
- Ten runs is a small sample. Nobody has yet published how many runs are enough, and we would say so rather than guess.
Keep the evidence
- What happens
- For each answer we would store the full text, the date, the assistant, the settings, the brands named, the claims made about each brand and the sources shown.
- What you would see
- Raw answers you can read yourself, with wrong facts flagged: an old price, a closed branch, a feature you do not have.
- Where it gets shaky
- Answers can be personal to the person asking, shaped by their history or location. A sample cannot reproduce that.
Read the gaps
- What happens
- We would separate four situations that get lumped together as “visibility”: not named at all, named but described wrongly, named without a reason, and named with a source linked.
- What you would see
- For each question, which of the four you are in, next to the same view for the competitors named most.
- Where it gets shaky
- This shows what was visible, not why a system chose it. Providers do not expose that.
Choose the actions
- What happens
- Gaps become a short list ranked by strength of evidence and effort. Many items would be ordinary: correct a fact, complete a product page, allow the right crawler, earn a mention on a site buyers already read.
- What you would see
- Each action with its evidence, who would do it, and how confident we are.
- Where it gets shaky
- Some gaps have no fix inside your control. The list would say so.
Check again
- What happens
- After a change, the same questions are asked again. We compare ranges across runs and report any difference smaller than normal run-to-run variation as noise.
- What you would see
- Before and after, with run counts. Traffic and inquiries shown alongside, with a plain note on how weak the link can be.
- Where it gets shaky
- A visit from an assistant is often not tagged as one, and a buyer may ask an assistant and then search your name. Cause is hard to prove.
Where this stops
It can’t see private answers, count demand, prove cause, rank you or guarantee anything. The full version, with sources, is on the methodology page.
You can do a rough version this afternoon
Before any tool, ten questions and a notebook will tell you a lot. The buyer guide has the exact routine, including how to avoid fooling yourself.
The questions look different in every industry
Hotels
“A quiet boutique hotel near the old town”
Asked as a feeling and a place. Answers hinge on location facts and trust.
HotelsB2B software
“Best CRM for a small contracting firm”
Asked as a fit for a workflow. Answers hinge on use cases and integrations.
B2B softwareClinics
“Dentist for veneers in Chicago”
Asked as a service in a place. Answers need accurate services, and clinicians should review them.
Clinics
Questions people ask
Is this running anywhere today?
No. This is the method we would build. The concept preview uses sample data written by hand.
Why not show one score?
A single number hides the run count, the route and the spread. We would rather show a range and let you see the raw answers. If we ever show a score, we will publish how it is made.
Do you need access to my website or analytics?
Sampling assistants does not. Matching answers to your own pages would involve reading your public pages. Anything connected to your analytics would be a separate, optional decision, and nothing like that exists today.
Does this work in other languages?
It is meant to. Whether answers differ between languages in ways that matter is something we want to test with partners, so we do not claim it yet.
Try it now, by hand
Ten minutes with an AI assistant will tell you a lot
Open ChatGPT or any AI assistant in another tab, then answer seven yes-or-no questions about your business. Every No or Not sure adds a step to your to-do list. Your answers stay in this browser.
Question 1 of 7
Open two AI assistants and ask the question your buyers ask most. Is your business mentioned?
For example: “What’s the best [what you sell] in [your city]?”
Bring the question you’d most hate to be missing from.
We’ll talk through how that question would be sampled, and where the method would be weakest in your market.
Talk to us about your buyers