Data cleansing
Data cleansing is removing broken, duplicate and misread records from a dataset before you count anything from it.
Data cleansing is removing the broken, duplicate and misread records from a dataset before you count anything from it. It's the dull step between collecting data and reporting on it, and it decides whether you can trust the number at the end.
For AI visibility, the data is answers. Some of them shouldn't count, and if they do, your numbers drift in ways nobody can explain.
What needs cleaning in AI answers
A few problems come up again and again:
- Failed answers, where the platform returned an error or nothing at all
- Answers about a different business with the same name as you
- Ordinary words that happen to be a brand's name
- The same brand counted under two spellings, as if it were two brands
- Duplicate prompts that ask the same question twice and double its weight
For example, Harrow & Pine tracks a competitor called Harrow Coffee. Answers about coffee shops in north-west London keep mentioning Harrow, the area. Before cleaning, Harrow Coffee's visibility looks like 25%. Once answers about the place are left out, it's 9%.
Leave gaps as gaps
Cleaning also means not filling in what's missing. If a platform didn't answer one day, the honest record is a missing day. Guessing a number for it makes the chart smoother and the data less true.
In AnswerOrbit
If we fail to read an answer, it doesn't count, and we never fill in a missing day with a guess. Answers clearly about a different business with your name don't count either, and the ones we can't be sure about are marked "possibly another business" and left out. See How we count.
In Competitors, the Unmatched tab lets you map a name to the right brand, and Excluded lets you leave out a name that isn't a competitor. Each time, we recount your answers. The Coverage tab in Topics flags duplicate prompts.