diff --git a/rust/lance-index/benches/inverted.rs b/rust/lance-index/benches/inverted.rs
index 415c1bc3fc4..f08d711ce5b 100644
--- a/rust/lance-index/benches/inverted.rs
+++ b/rust/lance-index/benches/inverted.rs
@@ -14,7 +14,6 @@ use futures::stream;
 use itertools::Itertools;
 use lance_core::cache::LanceCache;
 use lance_core::ROW_ID;
-use lance_datagen::{array, RowCount};
 use lance_index::prefilter::NoFilter;
 use lance_index::scalar::inverted::lance_tokenizer::DocType;
 use lance_index::scalar::inverted::query::{FtsSearchParams, Operator, Tokens};
@@ -27,6 +26,8 @@ use lance_io::object_store::ObjectStore;
 use object_store::path::Path;
 #[cfg(target_os = "linux")]
 use pprof::criterion::{Output, PProfProfiler};
+use rand::{rngs::StdRng, Rng, SeedableRng};
+use rand_distr::Zipf;
 
 fn bench_inverted(c: &mut Criterion) {
     const TOTAL: usize = 1_000_000;
@@ -43,16 +44,32 @@ fn bench_inverted(c: &mut Criterion) {
         ))
     });
 
-    // generate random words using lance-datagen
     let row_id_col = Arc::new(UInt64Array::from(
         (0..TOTAL).map(|i| i as u64).collect_vec(),
     ));
 
-    // Generate random words with 1-100 words per document
-    let mut words_gen = array::random_sentence(1, 100, true);
-    let doc_col = words_gen
-        .generate_default(RowCount::from(TOTAL as u64))
-        .unwrap();
+    // Generate Zipf-distributed words to better reflect real-world term frequency.
+    const VOCAB_SIZE: usize = 100_000;
+    const MIN_WORDS: usize = 1;
+    const MAX_WORDS: usize = 100;
+    const ZIPF_EXPONENT: f64 = 1.1;
+    let vocab: Vec<String> = (0..VOCAB_SIZE).map(|i| format!("term{i:05}")).collect();
+    let word_zipf = Zipf::new(VOCAB_SIZE as f64, ZIPF_EXPONENT).unwrap();
+    let mut rng = StdRng::seed_from_u64(42);
+    let mut docs = Vec::with_capacity(TOTAL);
+    for _ in 0..TOTAL {
+        let num_words = rng.random_range(MIN_WORDS..=MAX_WORDS);
+        let mut doc = String::with_capacity(num_words * 8);
+        for i in 0..num_words {
+            let idx = (rng.sample(word_zipf) as usize).clamp(1, VOCAB_SIZE) - 1;
+            if i > 0 {
+                doc.push(' ');
+            }
+            doc.push_str(&vocab[idx]);
+        }
+        docs.push(doc);
+    }
+    let doc_col = Arc::new(LargeStringArray::from(docs));
     let batch = RecordBatch::try_new(
         arrow_schema::Schema::new(vec![
             arrow_schema::Field::new("doc", arrow_schema::DataType::LargeUtf8, false),
@@ -86,32 +103,48 @@ fn bench_inverted(c: &mut Criterion) {
     let no_filter = Arc::new(NoFilter);
 
     // Get some sample words from the generated documents for search
-    let large_string_array = doc_col.as_any().downcast_ref::<LargeStringArray>().unwrap();
-    let sample_doc = large_string_array.value(0);
+    let sample_doc = doc_col.value(0);
     let sample_words: Vec<String> = sample_doc
         .split_whitespace()
         .map(|s| s.to_owned())
         .collect();
+    let sample_words_len = sample_words.len();
+    const TOKENS_PER_QUERY: usize = 15;
+    const QUERY_SET_SIZE: usize = 1024;
+    let mut query_rng = StdRng::seed_from_u64(7);
+    let mut queries = Vec::with_capacity(QUERY_SET_SIZE);
+    for _ in 0..QUERY_SET_SIZE {
+        let mut query_tokens = Vec::with_capacity(TOKENS_PER_QUERY);
+        for _ in 0..TOKENS_PER_QUERY {
+            let word_idx = query_rng.random_range(0..sample_words_len);
+            query_tokens.push(sample_words[word_idx].clone());
+        }
+        queries.push(Arc::new(Tokens::new(query_tokens, DocType::Text)));
+    }
+    let mut query_idx = 0usize;
 
     c.bench_function(format!("invert_search({TOTAL})").as_str(), |b| {
-        b.to_async(&rt).iter(|| async {
-            // Pick a random word from our sample
-            let word_idx = rand::random_range(0..sample_words.len());
-            black_box(
-                invert_index
-                    .bm25_search(
-                        Arc::new(Tokens::new(
-                            vec![sample_words[word_idx].clone()],
-                            DocType::Text,
-                        )),
-                        params.clone().into(),
-                        Operator::Or,
-                        no_filter.clone(),
-                        Arc::new(NoOpMetricsCollector),
-                    )
-                    .await
-                    .unwrap(),
-            );
+        b.to_async(&rt).iter(|| {
+            // Cycle through pre-generated queries to avoid skewing benchmark results.
+            let query = queries[query_idx % queries.len()].clone();
+            query_idx = query_idx.wrapping_add(1);
+            let invert_index = invert_index.clone();
+            let params = params.clone();
+            let no_filter = no_filter.clone();
+            async move {
+                black_box(
+                    invert_index
+                        .bm25_search(
+                            query,
+                            params.clone().into(),
+                            Operator::Or,
+                            no_filter.clone(),
+                            Arc::new(NoOpMetricsCollector),
+                        )
+                        .await
+                        .unwrap(),
+                );
+            }
         })
     });
 }