diff --git a/benchmark/benchmark_test.go b/benchmark/benchmark_test.go
index b495140..1337895 100644
--- a/benchmark/benchmark_test.go
+++ b/benchmark/benchmark_test.go
@@ -10,16 +10,38 @@ import (
 )
 
 func BenchmarkUnmarshalSimple(b *testing.B) {
-	d := struct {
-		A string
-	}{}
 	doc := []byte(`A = "hello"`)
-	for i := 0; i < b.N; i++ {
-		err := toml.Unmarshal(doc, &d)
-		if err != nil {
-			panic(err)
+
+	b.Run("struct", func(b *testing.B) {
+		b.SetBytes(int64(len(doc)))
+		b.ReportAllocs()
+		b.ResetTimer()
+
+		for i := 0; i < b.N; i++ {
+			d := struct {
+				A string
+			}{}
+
+			err := toml.Unmarshal(doc, &d)
+			if err != nil {
+				panic(err)
+			}
 		}
-	}
+	})
+
+	b.Run("map", func(b *testing.B) {
+		b.SetBytes(int64(len(doc)))
+		b.ReportAllocs()
+		b.ResetTimer()
+
+		for i := 0; i < b.N; i++ {
+			d := map[string]interface{}{}
+			err := toml.Unmarshal(doc, &d)
+			if err != nil {
+				panic(err)
+			}
+		}
+	})
 }
 
 type benchmarkDoc struct {
@@ -133,33 +155,32 @@ func BenchmarkReferenceFile(b *testing.B) {
 	if err != nil {
 		b.Fatal(err)
 	}
-	b.SetBytes(int64(len(bytes)))
-	b.ReportAllocs()
-	b.ResetTimer()
-	for i := 0; i < b.N; i++ {
-		d := benchmarkDoc{}
-		err := toml.Unmarshal(bytes, &d)
-		if err != nil {
-			panic(err)
-		}
-	}
-}
 
-func BenchmarkReferenceFileMap(b *testing.B) {
-	bytes, err := ioutil.ReadFile("benchmark.toml")
-	if err != nil {
-		b.Fatal(err)
-	}
-	b.SetBytes(int64(len(bytes)))
-	b.ReportAllocs()
-	b.ResetTimer()
-	for i := 0; i < b.N; i++ {
-		m := map[string]interface{}{}
-		err := toml.Unmarshal(bytes, &m)
-		if err != nil {
-			panic(err)
+	b.Run("struct", func(b *testing.B) {
+		b.SetBytes(int64(len(bytes)))
+		b.ReportAllocs()
+		b.ResetTimer()
+		for i := 0; i < b.N; i++ {
+			d := benchmarkDoc{}
+			err := toml.Unmarshal(bytes, &d)
+			if err != nil {
+				panic(err)
+			}
 		}
-	}
+	})
+
+	b.Run("map", func(b *testing.B) {
+		b.SetBytes(int64(len(bytes)))
+		b.ReportAllocs()
+		b.ResetTimer()
+		for i := 0; i < b.N; i++ {
+			d := map[string]interface{}{}
+			err := toml.Unmarshal(bytes, &d)
+			if err != nil {
+				panic(err)
+			}
+		}
+	})
 }
 
 func TestReferenceFile(t *testing.T) {
@@ -169,3 +190,38 @@ func TestReferenceFile(t *testing.T) {
 	err = toml.Unmarshal(bytes, &d)
 	require.NoError(t, err)
 }
+
+func BenchmarkHugoFrontMatter(b *testing.B) {
+	bytes := []byte(`
+categories = ["Development", "VIM"]
+date = "2012-04-06"
+description = "spf13-vim is a cross platform distribution of vim plugins and resources for Vim."
+slug = "spf13-vim-3-0-release-and-new-website"
+tags = [".vimrc", "plugins", "spf13-vim", "vim"]
+title = "spf13-vim 3.0 release and new website"
+include_toc = true
+show_comments = false
+
+[[cascade]]
+  background = "yosemite.jpg"
+  [cascade._target]
+    kind = "page"
+    lang = "en"
+    path = "/blog/**"
+
+[[cascade]]
+  background = "goldenbridge.jpg"
+  [cascade._target]
+    kind = "section"
+`)
+	b.SetBytes(int64(len(bytes)))
+	b.ReportAllocs()
+	b.ResetTimer()
+	for i := 0; i < b.N; i++ {
+		d := map[string]interface{}{}
+		err := toml.Unmarshal(bytes, &d)
+		if err != nil {
+			panic(err)
+		}
+	}
+}
diff --git a/internal/tracker/seen.go b/internal/tracker/seen.go
index e245aac..4b5d392 100644
--- a/internal/tracker/seen.go
+++ b/internal/tracker/seen.go
@@ -1,6 +1,7 @@
 package tracker
 
 import (
+	"bytes"
 	"fmt"
 
 	"github.com/pelletier/go-toml/v2/internal/ast"
@@ -29,67 +30,92 @@ func (k keyKind) String() string {
 	panic("missing keyKind string mapping")
 }
 
-// SeenTracker tracks which keys have been seen with which TOML type to flag duplicates
-// and mismatches according to the spec.
+// SeenTracker tracks which keys have been seen with which TOML type to flag
+// duplicates and mismatches according to the spec.
+//
+// Each node in the visited tree is represented by an entry. Each entry has an
+// identifier, which is provided by a counter. Entries are stored in the array
+// entries. As new nodes are discovered (referenced for the first time in the
+// TOML document), entries are created and appended to the array. An entry
+// points to its parent using its id.
+//
+// To find whether a given key (sequence of []byte) has already been visited,
+// the entries are linearly searched, looking for one with the right name and
+// parent id.
+//
+// Given that all keys appear in the document after their parent, it is
+// guaranteed that all descendants of a node are stored after the node, this
+// speeds up the search process.
+//
+// When encountering [[array tables]], the descendants of that node are removed
+// to allow that branch of the tree to be "rediscovered". To maintain the
+// invariant above, the deletion process needs to keep the order of entries.
+// This results in more copies in that case.
 type SeenTracker struct {
-	root    *info
-	current *info
+	entries    []entry
+	currentIdx int
+	nextID     int
 }
 
-type info struct {
-	parent   *info
+type entry struct {
+	id       int
+	parent   int
+	name     []byte
 	kind     keyKind
-	children map[string]*info
 	explicit bool
 }
 
-func (i *info) Clear() {
-	i.children = nil
+// Remove all descendent of node at position idx.
+func (s *SeenTracker) clear(idx int) {
+	p := s.entries[idx].id
+	rest := clear(p, s.entries[idx+1:])
+	s.entries = s.entries[:idx+1+len(rest)]
 }
 
-func (i *info) Has(k string) (*info, bool) {
-	c, ok := i.children[k]
-	return c, ok
-}
-
-func (i *info) SetKind(kind keyKind) {
-	i.kind = kind
-}
-
-func (i *info) CreateTable(k string, explicit bool) *info {
-	return i.createChild(k, tableKind, explicit)
-}
-
-func (i *info) CreateArrayTable(k string, explicit bool) *info {
-	return i.createChild(k, arrayTableKind, explicit)
-}
-
-func (i *info) createChild(k string, kind keyKind, explicit bool) *info {
-	if i.children == nil {
-		i.children = make(map[string]*info, 1)
+func clear(parentID int, entries []entry) []entry {
+	for i := 0; i < len(entries); {
+		if entries[i].parent == parentID {
+			id := entries[i].id
+			copy(entries[i:], entries[i+1:])
+			entries = entries[:len(entries)-1]
+			rest := clear(id, entries[i:])
+			entries = entries[:i+len(rest)]
+		} else {
+			i++
+		}
 	}
+	return entries
+}
 
-	x := &info{
-		parent:   i,
+func (s *SeenTracker) create(parentIdx int, name []byte, kind keyKind, explicit bool) int {
+	parentID := s.id(parentIdx)
+
+	idx := len(s.entries)
+	s.entries = append(s.entries, entry{
+		id:       s.nextID,
+		parent:   parentID,
+		name:     name,
 		kind:     kind,
 		explicit: explicit,
-	}
-	i.children[k] = x
-	return x
+	})
+	s.nextID++
+	return idx
 }
 
 // CheckExpression takes a top-level node and checks that it does not contain keys
 // that have been seen in previous calls, and validates that types are consistent.
 func (s *SeenTracker) CheckExpression(node ast.Node) error {
-	if s.root == nil {
-		s.root = &info{
-			kind: tableKind,
-		}
-		s.current = s.root
+	if s.entries == nil {
+		//s.entries = make([]entry, 0, 8)
+		// Skip ID = 0 to remove the confusion between nodes whose parent has
+		// id 0 and root nodes (parent id is 0 because it's the zero value).
+		s.nextID = 1
+		// Start unscoped, so idx is negative.
+		s.currentIdx = -1
 	}
 	switch node.Kind {
 	case ast.KeyValue:
-		return s.checkKeyValue(s.current, node)
+		return s.checkKeyValue(node)
 	case ast.Table:
 		return s.checkTable(node)
 	case ast.ArrayTable:
@@ -97,104 +123,135 @@ func (s *SeenTracker) CheckExpression(node ast.Node) error {
 	default:
 		panic(fmt.Errorf("this should not be a top level node type: %s", node.Kind))
 	}
-
 }
-func (s *SeenTracker) checkTable(node ast.Node) error {
-	s.current = s.root
 
+func (s *SeenTracker) checkTable(node ast.Node) error {
 	it := node.Key()
-	// handle the first parts of the key, excluding the last one
+
+	parentIdx := -1
+
+	// This code is duplicated in checkArrayTable. This is because factoring
+	// it in a function requires to copy the iterator, or allocate it to the
+	// heap, which is not cheap.
 	for it.Next() {
 		if !it.Node().Next().Valid() {
 			break
 		}
 
-		k := string(it.Node().Data)
-		child, found := s.current.Has(k)
-		if !found {
-			child = s.current.CreateTable(k, false)
+		k := it.Node().Data
+
+		idx := s.find(parentIdx, k)
+
+		if idx < 0 {
+			idx = s.create(parentIdx, k, tableKind, false)
 		}
-		s.current = child
+		parentIdx = idx
 	}
 
-	// handle the last part of the key
-	k := string(it.Node().Data)
+	k := it.Node().Data
+	idx := s.find(parentIdx, k)
 
-	i, found := s.current.Has(k)
-	if found {
-		if i.kind != tableKind {
-			return fmt.Errorf("toml: key %s should be a table, not a %s", k, i.kind)
+	if idx >= 0 {
+		kind := s.entries[idx].kind
+		if kind != tableKind {
+			return fmt.Errorf("toml: key %s should be a table, not a %s", string(k), kind)
 		}
-		if i.explicit {
-			return fmt.Errorf("toml: table %s already exists", k)
+		if s.entries[idx].explicit {
+			return fmt.Errorf("toml: table %s already exists", string(k))
 		}
-		i.explicit = true
-		s.current = i
+		s.entries[idx].explicit = true
 	} else {
-		s.current = s.current.CreateTable(k, true)
+		idx = s.create(parentIdx, k, tableKind, true)
 	}
 
+	s.currentIdx = idx
+
 	return nil
 }
 
 func (s *SeenTracker) checkArrayTable(node ast.Node) error {
-	s.current = s.root
-
 	it := node.Key()
 
-	// handle the first parts of the key, excluding the last one
+	parentIdx := -1
+
 	for it.Next() {
 		if !it.Node().Next().Valid() {
 			break
 		}
 
-		k := string(it.Node().Data)
-		child, found := s.current.Has(k)
-		if !found {
-			child = s.current.CreateTable(k, false)
+		k := it.Node().Data
+
+		idx := s.find(parentIdx, k)
+
+		if idx < 0 {
+			idx = s.create(parentIdx, k, tableKind, false)
 		}
-		s.current = child
+		parentIdx = idx
 	}
 
-	// handle the last part of the key
-	k := string(it.Node().Data)
+	k := it.Node().Data
+	idx := s.find(parentIdx, k)
 
-	info, found := s.current.Has(k)
-	if found {
-		if info.kind != arrayTableKind {
-			return fmt.Errorf("toml: key %s already exists as a %s,  but should be an array table", info.kind, k)
+	if idx >= 0 {
+		kind := s.entries[idx].kind
+		if kind != arrayTableKind {
+			return fmt.Errorf("toml: key %s already exists as a %s,  but should be an array table", kind, string(k))
 		}
-		info.Clear()
+		s.clear(idx)
 	} else {
-		info = s.current.CreateArrayTable(k, true)
+		idx = s.create(parentIdx, k, arrayTableKind, true)
 	}
 
-	s.current = info
+	s.currentIdx = idx
+
 	return nil
 }
 
-func (s *SeenTracker) checkKeyValue(context *info, node ast.Node) error {
+func (s *SeenTracker) checkKeyValue(node ast.Node) error {
 	it := node.Key()
 
-	// handle the first parts of the key, excluding the last one
+	parentIdx := s.currentIdx
+
 	for it.Next() {
-		k := string(it.Node().Data)
-		child, found := context.Has(k)
-		if found {
-			if child.kind != tableKind {
-				return fmt.Errorf("toml: expected %s to be a table, not a %s", k, child.kind)
+		k := it.Node().Data
+
+		idx := s.find(parentIdx, k)
+
+		if idx >= 0 {
+			if s.entries[idx].kind != tableKind {
+				return fmt.Errorf("toml: expected %s to be a table, not a %s", string(k), s.entries[idx].kind)
 			}
 		} else {
-			child = context.CreateTable(k, false)
+			idx = s.create(parentIdx, k, tableKind, false)
 		}
-		context = child
+		parentIdx = idx
 	}
 
+	kind := valueKind
+
 	if node.Value().Kind == ast.InlineTable {
-		context.SetKind(tableKind)
-	} else {
-		context.SetKind(valueKind)
+		kind = tableKind
 	}
+	s.entries[parentIdx].kind = kind
 
 	return nil
 }
+
+func (s *SeenTracker) id(idx int) int {
+	if idx >= 0 {
+		return s.entries[idx].id
+	}
+	return 0
+}
+
+func (s *SeenTracker) find(parentIdx int, k []byte) int {
+	parentID := s.id(parentIdx)
+
+	for i := parentIdx + 1; i < len(s.entries); i++ {
+		if s.entries[i].parent == parentID && bytes.Equal(s.entries[i].name, k) {
+			return i
+		}
+	}
+
+	return -1
+}