Updated 5 October 2026.
A brand name appearing in an AI answer is not enough to prove that the answer refers to your business. This guide sets out practical classification rules for names, aliases and ambiguous mentions. It complements our general tool selection guide, rather than providing another vendor ranking.
The examples below are fictional. They are not customer results or findings from Semly reports.
Create a dictionary of names
Document the main brand name, official domain, accepted abbreviations, previous names and relevant product brands. Keep competitor entries separate.
An alias should have a reason to be included. Do not add every generic word that happens to occur in your company name. If the dictionary changes, record the date and decide how historical observations will be handled.
Resolve ambiguous mentions
Consider a fictional business called “North”. An answer containing “north of the city” is not a reference to that company. “North, the accounting service at north.example” would be stronger evidence of the intended brand.
For ambiguous cases, inspect surrounding context and any linked domain. If the evidence remains insufficient, classify the observation as unresolved. Do not force it into a positive result to improve a headline percentage.
Domain agreement is supporting evidence; a spelling match alone can still refer to another business.
Distinguish what the answer does
| Classification | Practical rule |
|---|---|
| Mention | The answer identifies the intended brand |
| Recommendation | The answer presents that brand as an option for the user's need |
| Citation | The answer links to a page associated with the brand |
| Product reference | The answer identifies a product or variant under a defined rule |
| Unresolved | The available context does not establish identity |
An answer may satisfy more than one rule. Store the classifications separately. A linked source need not be a recommendation, and an unlinked recommendation need not be a citation.
Review a sample of classifications
Before interpreting a trend, manually review a sample containing positives, negatives and unresolved observations. Ask reviewers to apply the written rules independently and discuss disagreements.
Record how errors are corrected and whether correction affects only one observation or also the recognition rule. Do not present a manual sample as proof that all collected answers are accurate.
Keep denominators consistent
Specify whether a metric counts answers, questions, mentions or links. Define the treatment of failed collections and repeated observations.
For example, two mentions in a single answer need not count as two answers containing the brand. Mixing those units makes percentages difficult to interpret. For differences between indexed answers and custom question monitoring, read our measurement-method guide.
Separate branded and unbranded questions
A question that includes the company name tests the response about a known brand. A question without the name can test whether the system introduces it as an option.
Keep these groups separate when interpreting discovery. Correct information in response to a branded question does not establish spontaneous recommendation.
Document changes before comparing periods
Keep the name dictionary, classification rules, questions, markets, languages and systems with the measurement record. If one changes, explain how that affects comparability.
Use a visibility observation to decide what to inspect on the public website. Do not publish private answers or report findings as a case study without separate authorisation.
