Skip to main content
Understanding the collection explains most of the surprises in the numbers. Five minutes here pays for itself the first time a chart looks wrong.

A tracking run

On a schedule, Surfaced takes your tracked prompts and asks each one of each tracked AI platform, from each market and in each language you have configured. One prompt, six languages, five markets and twelve platforms is 360 questions. Each answer is stored whole. From it, Surfaced records:
  1. Whether your brand was named, in the text or as a link.
  2. Which other brands were named, and how often.
  3. Where you were ranked, if the answer ranked anything.
  4. How you were described: positive, neutral or negative.
  5. Every source the answer cited, as a URL and a domain.
Every figure in the product is an aggregation of those five things. There is no modelling step and no estimation from search volume, which is why a number can always be traced back to the answers behind it.

What that means for your numbers

They are a sample. Thirty prompts is thirty questions out of everything a buyer might ask. The sample is yours to choose, and choosing it well is the highest-leverage thing you do in Surfaced. See Prompts. The same question asked twice gives different answers. These are generative systems; identical prompts produce different wording, different orderings and sometimes different recommendations. That is why prompts are asked repeatedly and why single-run movements are not signal. Platforms differ more than periods do. The spread between two assistants on the same day is routinely wider than the spread on one assistant across a month. Read the per-platform charts before concluding something changed in time. Markets are not interchangeable. Asking from Germany is a different question from asking from the US, even in the same language, because the engines route and ground differently. See Markets and languages.

What is not collected

Nothing about individual users. Surfaced asks the questions itself. It does not observe your customers’ conversations with assistants, and could not. No traffic data. Surfaced measures whether you are recommended and cited, not whether anyone clicked. Those are different questions and the second needs your analytics. Nothing behind a login. Both for the assistants, which answer as an anonymous user, and for the website audit, which reads what a crawler can reach.

How fresh the data is

Answers arrive per run. The date range you select is anchored on your most recent day of data rather than on today, and a run that is still in progress can be partially visible once it passes four-fifths complete. Both behaviours, and the plan history floor that bounds how far back you can look, are on Periods and filters.