Clustering groups blockchain addresses that are likely controlled by the same entity, using heuristics such as common inputs or change address patterns; its results are probabilistic.
On Bitcoin and similar chains, the best known heuristic assumes that addresses spent together as inputs of one transaction belong to the same wallet. Others try to identify the change address that returns leftover funds to the sender. On account-based chains such as Ethereum, analysts look at patterns such as deposit addresses forwarding funds to the same exchange wallet, shared funding sources or repeated interactions.
Clustering makes analysis possible at scale: an exchange may use a very large number of addresses, and grouping them lets one label cover the whole entity. It is also central to reconstructing a fraudster's wallet structure. Each heuristic has known exceptions, however. CoinJoin transactions deliberately break the common-input assumption, and a wrong merge can attach a label to addresses that have nothing to do with that entity.
When a report relies on clustering, it should say which heuristics were applied and whether the cluster was checked manually. This matters especially when the opposing party contests a conclusion, because a single incorrect merge can change who appears to have received the funds.
See also: Address attribution, Blockchain analysis, Mixer and CoinJoin
General information, not legal or tax advice on your specific case.