The same question gets different answers
Ask an AI system the same question twice and you can get two different sets of brands. So "are we in the answer" has no yes or no. It has a rate: across many runs, you appear in some share of them.
That is why Shruwd asks every question repeatedly, at different moments, rather than trusting a single run. And it is why a result reads 32% (24–41%): the share, and the range the evidence behind it can support. Two results with the same middle can deserve very different trust. 32% (29–35%) is a number to act on. 32% (10–60%) is barely more than "somewhere in the middle".
Why a number is sometimes withheld
When there are not enough answers behind a number yet, Shruwd shows insufficient data instead. It is not an error, not a loading state and not zero.
A percentage from a handful of answers invites a decision the data cannot carry, and "we have not measured this yet" must never look the same as "we measured this and it is low". You will meet the state most on single questions, which take a few weeks to fill in. Your question set as a whole is ready much sooner.
Why there is no visibility score
A single blended number is easy to chart and impossible to act on. When it drops you cannot tell whether you were named less often, named later in the answer, or cited less, and when it rises you cannot tell whether you did something or a competitor slipped.
So the parts stay apart: how often you are named, how much of the naming goes to you, how early you appear, and how often your pages are the cited source. Each can be checked against the answers it came from.
When a number has moved
Two measurements almost always differ, and most of the difference is noise. Shruwd says a number moved only when the change clearly stands out from that noise and is large enough to matter.
Either test alone gives wrong answers. With enough data, a trivial change looks real. With too little, noise looks like progress. And because many questions are compared every week, a few would pass by luck alone, so Shruwd accounts for that across the whole set before anything is labelled.
The visible effect is that a change you can see in a chart may not be called movement, and a fix may come back as did not move. That is a result, not a failure: most single-page changes do not shift a whole question set much, and a tool that says so is the only kind whose "it moved" means anything. When a provider changes the model behind its answers in between, the comparison is marked model changed rather than credited to you.
Where the answers come from, and what they are a proxy for
Shruwd never signs in to ChatGPT or any other chat app, and never reads anyone's logged-in session. That is a choice, not a gap: automating a chat app breaks its terms, gets accounts banned, and turns a measurement product into an arms race with the thing it measures. ChatGPT's answers come through OpenAI's API. Google's AI Overviews come through a data provider that reads Google's public results page, the way rank trackers do.
The cost of that choice is stated plainly. An answer collected this way is a proxy for what a person sees. It has no signed-in history, no memory of earlier chats and no personalisation, so it will not match any one person's screen word for word. What it gives in exchange is the thing a measurement needs: the same question, asked the same way, every week, so that a change in the numbers reflects a change in the answers and not a change in who was asking.
What is exact
Not everything here is an estimate. Your server logs are a record, not a sample: a crawler either fetched a page or it did not, and your server either answered or refused. So crawler numbers carry no range.
They carry a different safeguard. A user agent is one line of config to fake, so a hit counts as an AI crawler only once it is verified to come from the vendor it claims. Hits that fail are shown and never counted. The AI crawler reference lists each crawler.
Names are matched strictly
A brand counts as named only when one of its names actually appears in the answer. There is no guessing at near-matches. A near-match that quietly counts the wrong company corrupts every number downstream and is close to impossible to spot afterwards. Names Shruwd sees but does not recognise are suggested to you, never added on their own.
Findings say how sure they are
Every finding says how it was reached: seen directly in your own data, reasoned from evidence you can check, or a lead worth testing. And new findings arrive a few at a time, because a list of forty gets nothing done.
Frequently asked questions
Does Shruwd scrape ChatGPT?
No. It never signs in to ChatGPT or any chat app. ChatGPT’s answers come through OpenAI’s API, and Google’s through a data provider that reads its public results page, as rank trackers do. That makes them a proxy for what a signed-in person sees, and Shruwd says so: the value is that the same question is asked the same way every week.
Why does Shruwd show "insufficient data" instead of a number?
Because there are not enough answers behind it yet to support one. The state is not an error and not zero, and it clears as more runs come in.
Why is there no AI visibility score?
A blended score cannot be acted on, because a change in it does not say what changed. Shruwd reports the parts separately: how often you are named, your share of the naming, how early you appear, and how often your pages are cited.
How does Shruwd decide that a fix worked?
After you mark a fix applied, it gives the AI engines time to catch up, then measures the affected questions again, more thoroughly than a normal week. The number counts as moved only if the change clearly stands out from the noise and is large enough to matter.
Are the crawler numbers estimates too?
No. They are counted from your own server logs, so they carry no range. A hit only counts once it is verified to come from the vendor it claims, because a user agent can be faked.