How we measure whether AI picks you

We publish one number about your business, and this page explains where it comes from, what it can and cannot tell you, and where it is uncertain. If any of it stops being true, this page changes first.

The two things on this site

The free check on the front page is one day. 5 questions go to 3 assistants, 15 answers come back over a couple of minutes, and you read every one of them. That is a sample of a single day, and the results page says so on the number itself.

The weekly check repeats the sampling every week for as long as you want it, and what you pay decides how far down the question list below it goes. That is where a rate you can lean on comes from. One day cannot tell you whether a fix worked; four weeks of days can.

What we ask

We write down the questions a real customer would type when they are looking for a business like yours in your town. Not “PicksYou reviews” and not your own name: the questions someone asks when they do not know you exist yet.

There are two lists, because there are two ways a customer looks for a business. Someone who wants a dentist names a town; someone who wants a thing online names nothing but the thing. So a business people look for in a town is asked the 25 questions in the first list, and a business that sells online is asked the 25 in the second. No business is ever asked from both.

A plan is a cut of whichever list a business is on. The free check asks the first 5, Starter carries on down the same list to 10, and Growth asks all of them. The last column says which plans reach each question.

For a business people look for in a town

Your trade and your town fill the two slots, so a photographer in Lyon is asked about photographers and never about dentists.

For a business that sells online

There is no town slot in any of these and no country either. The buyer of an online shop or a remote service is anywhere, so the questions are the ones somebody types when they want the thing itself: the best, the most reliable, the cheapest to order from, the one for a beginner. Only your trade fills a slot.

Either list only ever grows downwards. Move from one plan to another and the earlier questions are still the earlier questions, so your history stays comparable instead of restarting.

Twenty trades are on the form, from dentists and plumbers to photographers, restaurants, gyms and driving schools, and “something else” takes the word you type and puts it in the same slot. Each language has its own list, written by a person, not run through a translator at the moment of asking.

The exact question we sent is stored beside the answer and printed above it, so you never have to take our word for what we asked.

We ask those questions through the official APIs the assistants publish. We do not log into a consumer app and copy what it says, and we never pretend to be a person using one. Everything we get back is stored exactly as it arrived, so any number on your dashboard can be traced to the answer it came from.

How often

The free check runs once, when you ask for it. The weekly check runs on the same day every week, and each question goes to every assistant we cover. On the higher plans each question is asked more than once, because assistants do not give the same answer twice and one reading of a noisy thing is not a measurement.

A week is an ISO week counted in UTC, so the boundary does not move when clocks change or when a server is in a different country from you.

Appearance rate, and why it comes with a margin

Your appearance rate is the share of sampled answers that named your business. If we made forty checks last week and your name came up in fourteen of them, that is 14 of 40.

The “plus or minus” is the part most reports leave out, and it is the honest part. Forty checks is a sample, not a census. If we had run forty different checks the same week, the count would have come out somewhat different by luck alone. The margin says how much wobble is normal at that sample size. Ours is a Wilson interval, which is the standard way to do this for small samples; the practical thing it buys you is that it never reports a rate below zero or above one hundred, and it never claims certainty at the edges. Nought out of ten checks does not mean “never”, it means “we did not see it in ten tries”.

The headline uses a rolling four week window, so one strange week does not swing your number, and a real improvement takes a few weeks to show. That cuts both ways and we would rather it did.

When a match is not obvious (a shortened name, a spelling with the accents dropped, a business whose name is two ordinary words) our matcher hands it to a person instead of deciding. Until that person confirms it, it does not count as an appearance. So the number can be behind the truth, but never ahead of it.

Why we never show a position in a list

Nobody can give you a position in an AI answer, and anyone who sells you one is selling a number they invented. There is no list inside ChatGPT to be number three of. There is one answer, written fresh, that either mentions you or does not, and that changes between users, phones, days and phrasings.

So we count. Out of the questions we asked, how many named you. That is a fact you can check yourself by reading the stored answers, which is exactly why it is the number we publish.

The keywords page: a Google position, read once a week

If you track phrases in your account, we read Google's own list of results for each phrase once a week, on Monday, through a search data provider (DataForSEO), to the first 50 results, on desktop, in your own market. If your website is in that list, the page prints where it was, with the day it was read and the provider that read it beside the figure. Every figure on that page carries that sentence, because a position without a date and a source is not a fact anybody can check.

This is a different kind of fact from the rest of this page, and that is the whole reason it is allowed. Google publishes an ordered list of results, so a place on that list on that day is a fact about the list. An assistant writes one answer, fresh, with no list behind it, which is why the section above says we never show a position for one. A position in Google and a count of answers that named you are two facts about two different things, and neither explains the other.

There is no margin on it and no arrow beside it. One reading of one list is one observation, not a sample, so there is nothing to put an interval around. Two Mondays that differ tell you the list moved, and below the first few results it moves by itself week to week, so the page prints last Monday's figure beside this Monday's as two facts and draws no direction between them. Nothing on it says you rose or fell.

A phrase we read that found nothing says "not in the first 50" and never a nought: your site was somewhere below what we paid to read, or nowhere, and we cannot tell those apart. A Monday we could not read at all says so, with the reason, and is never printed as a phrase that was not found.

The AI Overview column is a yes or no about that day: whether Google put an AI Overview on the page for that phrase when we looked. It is not a statement about what the overview said, and nothing on that page claims the overview named you.

How many phrases you can track is part of the plan: Starter tracks 10 and Growth tracks 50. Stopping a phrase frees the slot straight away and keeps every Monday it was already read, so the record does not shrink when you change your mind, and starting it again continues the same history. The spreadsheet you can download holds one row a phrase a Monday, including the phrases you stopped.

What the usual range is, and is not

Every count we publish carries a range, and that range is one specific piece of arithmetic: a Wilson interval on a binomial proportion. The model behind it treats each sampled answer as one draw that either named you or did not, with the same chance of naming you on every draw, and the draws independent of each other. That is the assumption, said plainly. It is the assumption an interval of this kind always makes, and most reports that print a margin never tell you they made it.

So the range says this and nothing more: if the world stayed as it was during the window we sampled, and we asked the same questions of the same assistants again, here is the band of rates we could not tell apart from the one we measured. It is not a forecast of next week. Assistants change their models without telling anyone, your site changes, your competitors change, and none of that is inside the arithmetic.

The headline's range and the calculator's illustration are two different objects. The headline is a range around answers we really sampled, over a rolling four week window. The calculator inside your account holds that measured rate still and asks a hypothetical question: how wide would the range be if a week were made of more answers of the same kind. Nothing in that answer was sampled, which is why the page is titled an illustration, why it states its assumptions on itself, and why a business with nothing checked yet gets an example rate it picks by hand instead of a band drawn over numbers nobody measured.

One thing the illustration cannot do, and does not claim: asking new questions, or adding an assistant, does not make the same measurement more precise. It measures something different and reports a rate over a different set of answers. A wider week is worth having for what it covers, not because it sharpens the number you already had.

And one thing we have not done. The honest test of an interval is a coverage check: run the method many times over data whose true rate you know, and count how often the interval contains it. We have not run that check on our own samples yet, so we do not tell you our intervals have been verified in practice, and we would rather say that here than let the word Wilson do the work. It is an open measurement task on our list, and this page will carry what we found once we have run it.

Which assistants we cover

So a free check today is 5 questions across 3 assistants: 15 answers. Every assistant we add to it adds 5 answers to that, this page says so before your number moves, and the older weeks stay what they were rather than being quietly recalculated. An assistant that asks only on a plan is not in that count and never has been: the free check is the same three for everybody.

Which businesses an answer named

After an answer arrives we read the business names out of it, with a second call to the same OpenAI plumbing, and store the list with the answer. That is where “AI named” on your results page comes from.

We claim nothing about those businesses beyond the plain fact that this answer named them. Not that they are better than you, not that they beat you at anything, not that they paid anyone. If your name is missing from the list, the honest reading is that this one answer did not mention you.

What a sampled answer is, and is not

That is the whole reason we sample repeatedly and report a rate with a margin instead of showing you one flattering screenshot. One answer is an anecdote. Forty answers a week, over four weeks, is something you can make a decision on.

What else we check, and what each check can tell you

Beside the answers we take six readings about your business. Each one is labelled on your page with where it came from: measured means we fetched it ourselves, vendor means a search data provider looked it up for us, and research means a published bar we cite rather than something we measured about you.

Google Maps listing (vendor)

Whether Maps returns a listing matching your name in your town today. It cannot tell you whether the listing is correct or complete, only whether we found one and how confident the name match was.

Google reviews (research)

The review count on that listing, held against a bar of 150 reviews. The bar comes from published work on what the businesses assistants name tend to have; it is a comparison, not a measurement of your quality, and it is only as good as the listing we matched.

Local best of pages (vendor, then measured)

Whether pages like “best plumbers in your town” exist for your trade, and whether they name you. We fetch up to three of them ourselves and skip any page whose robots.txt tells us not to. It cannot tell you why an editor left you out.

AI crawler access (measured)

Your robots.txt, read by us, for the crawlers the assistants use. It tells you whether they are allowed in. It cannot tell you whether they have actually been, and a page can still be missing for other reasons.

Bing index (vendor)

How many pages of your site Bing has indexed, asked as a public site query. ChatGPT search reads Bing, so an empty answer here is worth knowing. It is not a Google index check and it is not your own Bing Webmaster data.

Business details in your site (measured)

Whether your homepage carries structured business details with a name, an address and a phone number that an assistant can read without guessing. We read the homepage only, so a detail buried on another page will look missing here.

When a reading cannot be taken, the page says “we could not check this”. That is never a pass. If you have no website, three of the six cannot run at all, and the page names which three instead of quietly scoring them as fine.

Nothing changes without your click

When we find something in the way, we write a draft: the lines for your robots.txt, a block of business details, a paragraph for your Google listing, a reply under one review. A draft is words on your screen and nothing else. Pressing the button writes one row with what was there before, what it says now, who pressed it and the minute, and only then does anything happen. Undoing it writes a second row that points at the first, and neither row can be edited afterwards, by you or by us.

There is no setting, key or role that lets anything on your site or your profile change without your click; the tests that prove it run on every release.

What leaves the product

There are 8 files you can download from your panel, and each one has a single grain: what one row of it is. Every file is UTF-8 with a byte order mark so a spreadsheet opens it as it was written, whatever letters your name is in, and every one is named the same way, after us, your business, what is in it and the day you took it.

The weeks

One row a sampled week, with the count, how many answers it came from, and the two ends of the usual range in whole counts. A week nobody sampled has no row, and a week we cannot draw a range for yet has two empty cells rather than a nought.

One week's answers

One row an answer of that week, with the assistant, the question we asked, whether it named you, and the moment it was asked, in UTC.

What was in the way

One row a blocker a check, with the day we read it, the sentence of evidence we read it off, and whether that came from us fetching something ourselves or from a search data provider.

Who arrived

One row a day and an assistant, counting the people who reached a page carrying your snippet. Somebody who read an answer and never came is in no row of it, because we never saw them; a day nobody arrived keeps its row, with a nought in it.

Named beside you

One row a business a week, with the count of answers that named it and the n that count came out of. It is made on the Growth plan; on Starter the same names are on your own page, week by week.

A closed month

One row a sampled week of that month, with the count, the n and the same cut by assistant the report prints. It is read from the frozen report, so it says today what it said on the morning the month closed, and a month that has not closed has no file.

What you approved

One row an approval, with what it touched, who pressed the button, what happened when it ran and the proof we fetched afterwards. An undo is a row of its own. What is not in it is the before and the after themselves: a robots.txt in a spreadsheet cell is a cell nobody can read, so both sides are printed in full on the approvals page and are in the account data export.

Your keywords

One row a phrase a Monday, with where your site was in Google's results that day, the page of yours that was found, whether Google put an AI Overview on it, when it was read, which provider read it, the market, how far down we read and the device. A Monday that found nothing has an empty cell rather than a nought, a Monday we could not read says so, and a phrase you stopped tracking keeps every Monday it was read.

What is in none of them is the assistants' own words, or the list of pages they cited. Showing you your own answers on your own screen is one thing; handing out a machine readable pile of another company's writing is a different thing with a different agreement behind it, so the answers stay where they are, kept word for word, on the week page every one of those rows came from.

Beside the files there is a read only API for the same figures, and a proof link you can send somebody who has no account: a page with the count, the n, the usual range and the answers, that you can revoke whenever you like.

Who else can see your business

A business lives in a workspace. Most workspaces hold one business and one person, the owner. An agency's workspace holds their clients, and the people who work on those clients are invited into it by email, each with a role and a list of the clients they may see. Nobody is charged for being in a workspace, and we never count the people in one: what an agency pays for is client slots.

There are three roles and this is all of them. A viewer reads and exports; an admin runs the work; the owner holds the money. Whichever role somebody has, they only ever have it for the clients they were given: an invitation can name every client of the workspace, including the ones it takes on later, or it can name a list, and a business added after that list was written is not on it until somebody puts it there.

One row of that table belongs to nobody in the workspace. Approving a change to a business's own website is never a role: the business itself approves it, from its own login or from a link we send that works once. An agency prepares the change and says what it would do; the person whose website it is presses the button. That is not a setting somebody can switch on, it is what the server does.

Ownership of a workspace moves in one act, by the owner, without asking us: the person it moves to becomes the one owner, the person handing it over becomes an admin, and every client, plan, question, receipt and invoice stays exactly where it was, under the names that made it.

When we email you, and how often

Four kinds of mail come out of an account, every one of them can be switched off from your account page, and every one of them carries an unsubscribe link that works in one press without signing in.

The weekly mail arrives when your weekly check has finished, with that week's count, the n it came out of, the usual range, and a link to read the answers. One a week, on the day your check runs.

The month report arrives in the first morning after a month closes: a PDF with the same figures as the page, frozen on the first of the month. You choose whether it comes monthly or not at all, and who else gets a copy. Anybody you add has to confirm by mail before we send them anything, and the refusal link in that mail is permanent: an address that has refused can never be added again, by you or by anybody.

An alert is sent only when one of these four things happened, and each one arrives with the counts on both sides of it so you can see the size of what moved.

An assistant stopped naming you for a question it used to name you for

It named you at least once in the four weeks before, and in none of the answers to that question in the four weeks since. The mail prints both windows with their own n.

The last four weeks are clearly different from the four before

The usual ranges of the two windows do not overlap at all, with at least two sampled weeks on each side. Two ranges that overlap are the ordinary swing of sampling, and we say nothing about those.

Something new is in the way

A check that was passing started failing, with the evidence we read.

Something that was in the way is fixed

The same check, passing again. Each one is sent once, when it happens.

There is a budget on the alerts, and it is at most four a month. You can set it to none, two, four or eight, and when it is spent we still write the alert down and list it in your next month report; we simply do not mail it. We never send one because a number wobbled: a count that moved inside its usual range is what sampling looks like, and mail about it would train you to ignore us.

A paused account gets none of this, and neither does an address that has unsubscribed. Nothing we send is a newsletter, nothing is a promotion, and there is no mail you have to receive to keep your account working.

Where the money goes

Every call we make records what it cost. One free check is about thirteen cents of vendor spend, which is why there is a daily cap on the free checker and why a full day turns into a next morning queue instead of a worse measurement. Free checks are limited to 3 a day from one place, so the measurement stays honest.

We do that because sampling honestly is not free, and a product that quietly cuts the number of checks to save money would be reporting a worse measurement while charging the same. If we ever change how often we sample, this page says so before your dashboard does.