Manual AI visibility snapshot
A complete, spreadsheet-friendly procedure for sampling buyer-style prompts across assistants, recording whether a brand is named, and reporting the uncertainty without implying a live automated service.
Read the sampling method · Size the sample
This is a manual measurement guide. You run the prompts yourself; the page does not submit prompts, call model APIs, or collect your brand name.
In short
This guide shows how to build a small, dated snapshot from a brand name, a category, buyer-realistic prompts, and repeated observations across the assistants you choose. The output records whether the brand was named, which other brands appeared, which sources were cited, the sample size, and an uncertainty range. It is a diagnostic starting point, not a tracked measurement programme. The sample is small by design, the range is wide, and the result describes one moment. The method is fully available on this page and requires no email address.
Who this is for
- Founders and heads of marketing who were told "I asked ChatGPT and you weren't there" and need to know whether that was a fluke
- In-house SEO leads who need something concrete to put in front of leadership before proposing a measurement programme
- Agency strategists sanity-checking a category before scoping a client engagement
- Anyone who has been shown an AI-visibility score by a vendor and wants an independent second reading
- Content and PR leads wanting a first look at which sources get cited in their category
When to use this
- You have never measured this and want a first baseline impression without buying anything
- You suspect a competitor is being recommended over you and want to know if it is real or anecdotal
- You want to see what an honest confidence range looks like before you trust anyone's dashboard
- You need a concrete artifact to open an internal conversation about whether this matters yet
How to run the snapshot
Start with two fields in your own worksheet: the brand name and the category a buyer would use to describe what it sells. Build a small set of buyer-shaped prompts, run each prompt repeatedly through the supported interfaces you are authorized to use, and record brand names, cited sources, and framing in the raw answers.
None of the prompts name your brand. Asking an assistant about you directly all but forces it to talk about you and produces a flattering number that means nothing. The prompts ask category questions, the way a buyer with a problem would.
Report the result as prose plus a small set of figures, each carrying the sample size and uncertainty interval behind it. There is no single composite score, because a composite hides which signal moved and therefore hides what you would do about it.
What your worksheet should include
Everything in the output is designed to be checkable. If a figure appears without the context needed to judge it, it should not be there.
- Whether the brand was named at all, and in what share of the sampled answers
- The other brands that appeared alongside it, with how often each was named
- Which sources were cited, where the assistant cited anything at all
- How the brand was framed when it appeared: recommended, listed, mentioned in passing, or cautioned against
- The prompts used, so you can judge whether they resemble your buyers
- The models and versions queried, the run date, the number of runs, and a confidence range on every figure
Every figure ships with its prompt set, model versions, run date, sample size and variance, or it does not ship.
The limits, stated plainly
This is the section that many free checkers leave out. These limits follow from using a small, manual sample and should be recorded with the result.
The most important consequence: if your result sits near the middle, the honest reading is usually "we cannot tell from this sample," not "you are doing moderately well." A wide interval is information. Treat it as such.
- A small exploratory sample produces a wide confidence range by construction. Choose and report the run count rather than hiding it.
- It is a single point in time. Assistants change under you, and a result from today may not describe next month.
- It is not personalized. Real users bring session history, region and account settings; a snapshot queries a defined interface under stated conditions.
- It cannot explain anything. It can show you a gap; it cannot tell you why the gap exists or what caused it to move.
- It measures the interfaces you query, not every surface a buyer might touch.
- It is not a substitute for a tracked programme with a frozen prompt set and scheduled runs.
How to read your result
Start with the interval, not the number. A figure with a wide range around it is a question, not an answer. Two brands whose ranges overlap heavily have not been shown to differ, no matter how far apart their point estimates look.
Then read the prompts. If they do not sound like your buyers, the result is measuring the wrong thing and the fix is a better prompt set rather than a bigger sample. This is the single most common reason a check surprises someone.
Then read the framing, not just the count. Being named in a list of eight is a different situation from being named as the recommended option, and both are different from being named as the wrong choice for the reader's stated constraint. Finally, read the cited sources. If the same handful of domains decide the answers in your category, that tells you more about where to spend effort than your own number does.
- Which metric answers which question
- When a mention is not a recommendation
- Why answer position is not a ranking
What to do next, depending on what you see
The right next step depends entirely on the shape of the result, and for a lot of people the honest answer is that no further spending is warranted yet.
- Not named anywhere, and the category prompts look right: start with entity clarity and whether you are legible to a retriever at all.
- Named sometimes, wide interval: you have a sampling problem before you have a visibility problem. Work out how many runs the question needs.
- Named, but framed poorly or for the wrong use case: this is a positioning and category-language problem, not a measurement one.
- Competitors dominating the cited sources: the leverage is in earned coverage on the sources that actually get cited.
- You want to keep watching this: set a defensible baseline before you change anything.
- You want to do this properly without buying anything: the spreadsheet method works further than most people expect.
"Not yet" is a legitimate conclusion. If AI answers are not where your buyers are, a single check is enough for now.
When this snapshot is the wrong instrument
A free snapshot is the wrong tool more often than it is the right one, and pretending otherwise would make the rest of this site dishonest.
If any of the situations below describe you, the check will give you a number you should not act on. The measurement hub and the sample-size calculator are the better starting points.
- You need to detect a small change over time. A capped one-off sample cannot resolve small differences; size the run first.
- You have to report to leadership on a cadence. That needs a frozen prompt set and scheduled runs, not a repeated snapshot.
- You need to compare several brands rigorously. Fair comparison requires identical prompts, runs and windows.
- Your category prompts are ambiguous or contested. Fix the prompt set before you measure anything with it.
- You are trying to prove that a change you made worked. That is a causal question with confounders a snapshot cannot rule out.
- An assistant is stating something false about your brand. That is an incident, not a measurement problem.
How your data is handled, and why the result is free to see
The procedure is fully available without an email address. Record the observations in your own worksheet so you can inspect the prompts, raw responses, sample size, and uncertainty together.
RankRift does not collect or store the brand, category, prompts, responses, or results from this manual procedure. You control the worksheet and its retention.
This page collects no snapshot inputs or results. Keep your worksheet private unless you deliberately publish it, avoid confidential or personal data in prompts, and review each provider's current data-handling terms before sending a prompt.
Related pages
- How to measure AI visibility properly
- Sample-size calculator
- The measurement methodology
- Prompt-set library by category
- Free AI-visibility checkers compared
- Glossary of AI visibility terms
Frequently asked questions
Why don't you ask for my email before showing the result?
Because the complete procedure is published on the page. You run it in your own worksheet and can inspect the prompts, models, sample size, and confidence range without giving RankRift personal data.
My result was different from what I see when I ask ChatGPT myself. Who is right?
Both observations may be valid. Assistant output varies by phrasing, session, region, personalization, interface, and date. Compare the recorded conditions and increase the sample before treating either result as representative.
Can I run the check on a competitor's brand?
You can, and your worksheet will show how often assistants named that brand under the prompts you sampled. Read it as a statement about model output rather than about the competitor's quality, and be careful about repeating it externally without the prompt set and confidence range attached. A figure quoted bare, stripped of its sample size and interval, becomes a claim you cannot defend.
How small is the sample, exactly?
Choose a small exploratory batch only for a first look, record the exact count beside every figure, and use the sample-size calculator when the decision requires more precision. If you need to know what sample size your own question requires, the calculator answers that directly.
Does the check use the same assistants my customers use?
You choose which authorized interfaces or official APIs to sample and record each one. Do not treat an API result as identical to a consumer product: report only the surfaces and conditions you actually queried.
Will the check tell me how to appear more often?
No. The worksheet shows what appeared in the sampled answers and, where citations exist, which sources those answers used. Turning that into action is a separate job, and the honest levers are slow ones: entity clarity, presence in the sources models actually cite, retrievability of your pages, and real third-party coverage. Nothing on this site can promise a change in what any model says about you.
Is a result from this good enough to put in a board deck?
As a conversation-opener, yes, provided you carry the confidence range and sample size into the slide with it. As a metric you commit to reporting quarterly, no. A number that moves on a small sample will embarrass you the first time it moves for no reason. If it is going in a recurring report, freeze a prompt set and size the runs first.
Get a number, then decide if it deserves a programme
The manual snapshot gives you a first reading with its uncertainty attached. If the reading matters to you, the next step is sizing a proper measurement, not buying anything.
Use the manual method · Size your runs with the calculator · Learn the measurement basics