Public nucleotide archives contain annotated marker loci within long sequence records, but inventories based only on short records can overlook them. We quantified this difference for Orchidaceae using a frozen European Nucleotide Archive catalogue and a predeclared, annotation-based extraction protocol. The catalogue contained 578,690 records; 555,683 candidate accession versions were retrieved. Parsing yielded 98,534 marker regions from 71,358 parent records, of which 96,963 (98.406%) passed syntactic quality control. The primary counting unit was the NCBI taxon ID–marker combination, not an accepted species or specimen. Short-record extraction represented 39,298 combinations and long-record features represented 4,208, yielding a union of 41,875. Of these, 2,577 were present only as embedded features: 6.154% of the union and a 6.558% increase over the short-record inventory. For the core plant markers matK and rbcL, 613 of 11,082 combinations (5.531%) were embedded-only. The hidden fraction ranged from 6.088% to 9.601% across 18 predeclared quality-control scenarios. Post-review parent-length reclassification with a fixed region pool yielded 5.953–6.357%, whereas jointly changing the boundary and maximum region length yielded 4.414% at 2,500 bp and 6.681% at 10,000 bp. An internal audit using independently implemented code verified all 98,534 extracted regions against the preserved EMBL records without discrepancies. The complete candidate-record snapshot and region-level verification evidence have been preserved. These findings identify a methodological source of undercounting in archive inventories. They do not establish species-level barcode coverage, biological correctness of deposited annotations, or improved identification performance.