Method
Every number in a Valzeo report comes from asking a live assistant a question a customer would actually type, and reading what came back. This page is the whole procedure, including the parts that limit what we can honestly claim.
The procedure
Questions are generated from the shape of the category, not written one by one. For study abroad that means destinations, cities, study levels, services and the constraints people attach — budget, visa odds, scholarships, a specific intake. A cycle is typically twenty questions, each weighted 1 to 3 by how much a customer asking it is worth.
The weights are our judgement and they are printed in the report so you can argue with them. Nobody publishes search volume for AI questions; anyone quoting you one made it up.
A firm that only places students in the UK is not scored on questions about Germany. We read your public site for the destinations, cities and services you actually cover, score you on the questions inside that scope, and report the rest as context with the note that they were excluded. The assumption list in the report states exactly what we read and from where — if we got your footprint wrong, that is the line that shows it.
Each question goes to the tiers people actually use, and the free and paying tiers are kept apart because they routinely give different answers. Every question is asked three times per tier, in separate passes, with no memory carried between them.
Three passes is the minimum that distinguishes a held position from a coin flip. A report built on one pass per question is a screenshot, not a measurement.
Every call is made with web search enabled, and we record how many searches the assistant actually ran. An answer produced with zero searches is the model reciting training data — it describes the web as it was months ago, not what your customer is told today.
Those answers are separated out, reported at the end with a caution, and never scored.
Assistants cite their sources, but through redirect links that name nothing. We resolve each one to the real page, fetch it, and check which brands it mentions. That produces the work list: the pages shaping answers in your category that never mention you, ranked by how many answers each one shaped, and split into a rival's own site and independent pages you can realistically get onto.
Identical questions asked minutes apart disagree. We measure that churn in the same cycle as everything else — in study abroad it has been running near 58% — and treat it as the floor. Any change smaller than the floor is reported as flat, including the ones that would flatter you.
Counting rules
Loose matching is the quiet way visibility figures get inflated, so the rules are strict and identical for you and for every rival:
| Rule | What it means |
|---|---|
| Whole names only | Every distinctive word in a brand name must appear. "Global Tree" is never counted as "Global Reach", and a shared word like "education", "overseas" or "global" can never carry a match on its own. |
| Aliases, stated | Trading names and common abbreviations count, and the exact list used is printed in the report's assumptions. |
| Position is recorded | Being named eighth is not the same as being named first. Every mention carries its position, and the visibility index weights it. |
| Competitors are discovered | The competitive set comes out of the answers themselves. Nobody supplies a list, so nobody can quietly shape the ranking. |
| Each pass counts separately | Nothing is averaged before you see it. The report shows every pass, and the spread across them sits next to the headline figure. |
Limits
See it applied
The sample walks through an actual report section by section, with real figures and the firms unnamed.