method // how we measure
supereo asks AI engines the questions your buyers ask, judges every answer, and grades every change on the questions it aims at against the rest of your panel. This page is how, with the numbers the product runs on.
Every tracked brand gets a fixed panel of buyer questions: the 20 its free audit wrote, 4 more written from what the audit learned about the brand, and 6 that name the brand (what it is, what it costs, how it compares, alternatives to it, reviews, and whether it fits): 30 to start, 40 at most. You confirm the panel and its competitors before it locks.
Which question is asked moves a number more than chance does, so the panel is versioned. Each question is a series of its own: adding, editing or removing one makes a new version, and only that question's series starts again. How often AI recommends you counts the questions that don't name you; the ones that do are reported beside it.
Every question runs on six engines, 3 times a week each, and never twice on one engine within 30 minutes. Each is asked the way a buyer reaches it, with nothing forced:
The free audit asks its 20 questions once each, on five of these engines: every one but Claude.
Every answer is read by supereo's judge, which makes three calls about your brand, and the same three about each competitor:
Recommended is the number supereo leads with. If you think the judge got an answer wrong, say so on it: the answer is judged again with your note, and the counts follow when the second reading agrees with you.
A rate is the share of the last 4 weeks of answers that recommend you, shown as answers out of every 100. Its range is where it likely sits, its 95% range: 1.96 standard errors either side. Only run-to-run luck within a question moves a fixed panel's rate by chance, so the range comes from each question's own spread, with half a hit added on each side so a question answered the same way every time still carries some.
Every change that aims to win answers gets a test: supereo's own, and a change you ship from supereo's recommendations. Its test is registered when supereo recommends the change, so its before is ready by the time it ships, or as it ships when it wasn't. A fix isn't tested: it puts something right and stays. A test holds nothing: another change may aim at the same page or question, and each change is credited with what it adds. Its rules are saved when it is registered, before its change is live, and never edited after.
A change that was live before supereo started measuring it, because supereo never read its page without it or holds no answer to the questions it aims at from before it, has no before to compare with: it shows its trend, those questions against the rest of your panel week by week, and never gets a verdict. A test whose change isn't live 28 days after it was registered ends.
From the first answers after its clock starts (below), the test shows where it is heading: the difference so far, averaged over its engines, with its range. Before day 14 it never uses a verdict word, and nothing is decided from it.
From day 14 after its clock starts, supereo grades the test each time its answers come in, and writes the verdict as soon as the range is clear. If it isn't clear by day 60, supereo writes the verdict then from the answers it has. The range is made to be read that often: however many times a test is graded, a Worked comes up by luck at most 1 in 20. The size reads as answers, as in “10 more of every 100 answers recommend you”.
| Verdict | When |
|---|---|
| Worked | The whole range is above zero, and the estimate is at least 10 more of every 100 answers. |
| No change | The whole range is below 10 more of every 100, and it includes zero. The change stays, unless you've set supereo to roll such changes back. |
| Made it worse | The whole range is below zero. The change is recorded to be undone: supereo reverts its own, and tells you about yours. |
| Too close to call | Anything else by day 60. The verdict says so instead of rounding up. |
A change that hurts shouldn't stay live for weeks. While a test runs, supereo looks at most every 84 hours, twice a week, once each question it aims at has 5 answers on each engine since its clock started. If the whole range is below zero at 99%, a stricter bar than the verdict's, the test ends at once as Made it worse, and the change is recorded to be undone. The harm stop can only ever say Made it worse: nothing is called a win early.
Every change that worked is measured again every 28 days after its verdict: the questions it aims at over 7 days, against the test's own before, with the rest of your panel pooled again then.
A test's change goes live when supereo's fetch of the page sees it, when you mark it shipped and the next fetch shows it, or when supereo's own pull request is served. Its window starts when AI first reads the changed page, as your Cloudflare or Vercel edge sees it: an AI company's own agent opening it for someone's question, never for one of supereo's own, so a read that comes while supereo asks that company a question whose answer pulled the page (within 2 minutes of the ask, or until its answer is back when it takes longer) doesn't start it. Answers between the change and that read count on neither side. With no edge connected that sees the page, or no such read within 7 days, the window starts when the change went live. The page is fetched daily while its test runs. An edit to it by anything but the change is a note on the test's timeline: the test runs on, and its verdict says it includes the edit. If the change itself comes off the page, the test ends with what it saw.
Pages puts each page of your sites beside what AI did with it over the last four weeks. A read counts when the request came from the AI company's own addresses and your Cloudflare or Vercel edge answered it with the page; a read for a person is one made while someone asked that AI a question, never one supereo's own questions caused. Cited counts the tracked answers that list the page among their sources, each answer once however it writes the address.
Recommended you counts the answers that cited a page and recommended you. The citation and the recommendation come in the same answer, which links them; it doesn't prove the page made the AI pick you.
Off-site lists the pages off your sites that your tracked answers cite. Every week supereo reads the 50 they cite most, every cited page on your profiles and your network, the pages your first analysis aims at and any page you add, and reads a page again only when its text changed. A page on one of your profiles or network sites is yours by its address; any other names you or names only your rivals by what supereo read on it. A wrong fact is a statement on the page that gives you a value your own facts don't, and a listing's rating comes from the page's own structured data.
The network figure counts the tracked answers that cited a page of the sites that point to you and how many recommended you, against the answers that cited none of them. The two go together in the same answers, which links them; it doesn't prove the sites made the AI pick you.
A verdict comes only from supereo's own question runs, because only they can be held against the rest of your panel. Visitors confirmed from AI and real users' reads of your pages are shown beside your numbers, and never counted in a verdict.
