Methodology: how Unvibe measures
Unvibe measures two different things: whether AI can read a website, which is checked directly and repeatably, and whether AI recommends the business, which is sampled and therefore noisy. This page says how each number is produced, what counts as a pass, and what the numbers cannot tell you.
What is measured directly
The checkup reads the live homepage, the pages customers ask about, robots.txt and the sitemap the way AI crawlers do: as plain HTML, without running JavaScript, and again with the real user-agent strings of GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Googlebot. Most checks are rules that give the same answer every time for the same page, for example whether robots.txt lets a crawler in, whether a sitemap exists, whether JSON-LD is present, or whether a link to prices is in the HTML. Each check and what passes is listed in The checkup and the score.
- A pass is a fact about the page. If a check fails, the report quotes what the crawler found, so you can see the evidence.
- The score is a weighted share of passing checks. Each check has a weight, the three areas are scored separately, and they combine at 40%, 35% and 25%. Checks that only apply to some kinds of business (product data for shops, address and opening hours for local businesses) are counted only when the checkup finds that kind of business.
- A checkup scores the homepage. The page audit reads up to 15 pages from the sitemap, but their results are not part of the score.
What is judged by an AI
One AI reading covers the homepage and the pages customers ask about (pricing, contact, about, help), exactly as a crawler receives them, plus the page addresses listed in the sitemap. The AI is asked, in a fixed format, first which market the business is in (with a quote that shows it and how sure it is), then what it offers, whether it is clear what it is and who it's for, which concrete reasons to choose it the pages state, whether any two statements contradict each other, and which questions buyers in that market ask that the pages answer. It must quote the pages for each answer.
- The business is judged from its pages only. What the AI remembers about the business is never used. Its general knowledge of the market is used for three things only, each labelled as such: what buyers compare, the alternatives they weigh, and the questions they ask.
- How sure the AI is about the market is checked too. "Sure" needs a quote from the pages that states the market; without one it counts as "fairly sure". When it can only guess, the category check fails and the fixes keep placeholders rather than a guess.
- What you told Unvibe is a comparison, not evidence. The niche you enter is compared with the AI's reading; it never replaces it. A default like "Business" is ignored.
- Quotes are checked. Every quote is compared with the text Unvibe fetched. A quote that isn't on the pages is dropped, and a question the AI calls "answered" without a real quote is downgraded to "partly answered".
- Four of the 17 checks come from this reading: what it is, who it's for, the category (which fails when the AI can only guess the market), and reasons to choose it. The audit's findings use the same reading, so the score and the findings never disagree.
- If the AI reading doesn't finish, those four checks are left out of the score and the findings that need judgement are marked as not judged. Nothing is guessed or counted as a failure, and the report says so.
- Everything else (crawler access, firewalls, the HTML, structured data, prices, links, contact details, actions an agent can take) is checked with fixed rules against what was fetched.
How findings are ranked
Each finding is critical (likely to make AI fail) or an opportunity (would make AI more confident), with a likely impact from 1 to 5 and an effort: about 10 minutes, an hour or two, or a day or more. Problems that stop AI reading the site at all come first, then other critical ones by impact, then opportunities. One cause is reported once: when the pages are empty without JavaScript, the checks that depend on their content wait until they can be read.
How findings are backed up
A finding links to other evidence in your workspace that points to the same problem: AI answers to your tracked questions that named someone else, claims a fact check found wrong, competitors that already get it right, and the prompts the fix would help you win. This evidence is shown next to the finding, never used to invent a new one.
How AI mentions are measured
AI mentions asks an AI your buyer questions and records what it recommends.
- Where the questions come from. Unvibe asks an AI for the 12 questions a buyer would type into an assistant when looking for a product like yours, using your category, audience, description and known competitors, in a mix of types (best of, comparison, pricing, how to, about you). You choose up to 8 to track and can add your own. The list is a suggestion, not a measurement of search volume; the separate search demand check uses search autocomplete signals to show which questions people actually type.
- How each question is asked. Every question is asked on its own, without naming your business and without your website's content, so the answer reflects what the AI already knows. The AI is asked for a short answer and the list of product names it recommends, most recommended first.
- What counts as a mention. You are mentioned if your name or the main part of your domain appears in the recommended list. If the AI recommends nothing at all, your name appearing in the answer text counts. Your position is your place in the recommended list.
- What is recorded. Whether you were named, your position, which products were named instead, and the full answer, so every result can be read.
- The "does AI know you" check asks what the AI knows about your product, from memory only, and records whether it says it knows the product.
How weekly tracking works
After your first run, Unvibe repeats the same questions every week on its own and saves each run. Movement is shown per question: up or down in position, newly named, or lost. Weekly runs do not use your monthly allowance.
How the AI visibility score is built
The score combines up to five parts, each out of 100:
| Part | Weight | How it is calculated |
|---|---|---|
| Recommended by AI | 45% | Share of answers that name you, weighted 2:1 toward the latest run over the one before it |
| Position when named | 15% | #1 scores 100, #2 scores 80, then 20 fewer per place, down to 0 |
| AI knows you | 10% | 100 if the AI says it knows your product, otherwise 0 |
| Facts AI gets right | 15% | Share of claims in the latest fact check that are correct |
| Visitors from AI | 15% | From AI traffic in the last 30 days: none is 0, fewer than 5 is 40, fewer than 20 is 70, 20 or more is 100 |
Parts you have not set up are left out and the remaining weights are scaled up to fill the score, so the score is comparable but not the same as a fully set up one. The report says which parts are missing.
How fact check works
Fact check asks an AI what it knows about your product from memory, then compares each claim in the answer with your live site. Each claim is marked correct, wrong or not on your site, and facts on your site that the AI left out are listed. The comparison is done by an AI told to be strict and literal, so borderline claims can be judged either way; the report shows the claim and the evidence side by side.
How fixes are verified
A fix is only counted as done when a later checkup of the live site confirms it: a check passes, or the audit no longer raises the finding. Marking a fix as done, or a coding agent marking it, records your intention; the next checkup, from the app or from a coding agent, tests the live site and reports what it confirmed, what still shows the problem, and anything new since the last checkup. Confirmed fixes are kept as work completed.
What the numbers cannot tell you
- AI answers vary. The same question can produce a different answer on another day or another run. Unvibe asks each question once per run, so a single run can over- or understate how often you're recommended. Trends across weeks are more reliable than any one run, and the visibility score averages the last two runs for this reason.
- A sample is not the whole market. Unvibe asks the questions you track, not every question buyers ask, and asks them without personal context, location or history, which real users have.
- Assistants differ. Results come from one leading AI model asked without web search, and from a web-searching model for the cited-sources view. They do not describe what every assistant would say.
- Readable is not the same as recommended. A perfect checkup score means AI can read and use the site. It does not guarantee any assistant will recommend the business.
- Some firewalls can't be tested from outside. If a site turns away every visit that claims to be a crawler, even Googlebot, it is checking where crawlers come from. The real crawlers usually get in, so Unvibe doesn't report this as a block or count it in the score; the report says it couldn't be tested.
- Unvibe does not use a browser agent yet. "Usable by agents" checks whether the things an agent needs, like a link to prices or a labelled main button, are in the page's HTML. It does not yet run an agent through a real task.
- Competitor names are matched by text. A business with a very common name can be matched wrongly.
Questions
Can I check how a number was produced?
Yes. The report shows what the crawler found for each check, the answer text and recommended list for each AI mention, and the claim, verdict and evidence for each fact check.
Does Unvibe change my website?
No. Unvibe only reads your public pages. You or your coding agent make the changes.
Why did my visibility score change when I changed nothing?
AI answers vary between runs. Look at the weekly trend, and at which questions moved, rather than one run.