# YC batch similarity Similarity between Y Combinator companies, measured across 6,142 companies in 48 batches from Summer 2005 onwards. Published with the report "YC batches got twice as likely to repeat an earlier YC company, from 18% to 47%" at https://fundingwatcher.com/research/yc-batch-similarity/ Data as of 2026-09-04. Licence: CC BY 4.0. Attribution: FundingWatcher. The measurement, the controls and the limitations are in METHOD.md. No vectors are published. ## Files ### companies.csv One row per company, 6,142 rows. domain bare domain after normalisation, the join key name company name as YC lists it batch YC batch, for example "Winter 2024" best_within_batch closest company in the same batch best_within_batch_domain best_within_batch_sim cosine, 4 decimal places best_prior_batch closest company in any earlier batch best_prior_batch_domain best_prior_batch_sim best_other_batch closest company in any other batch, earlier or later best_other_batch_domain best_other_batch_sim Companies in the first batch, and single-company batches, have empty prior or within-batch columns. ### batch_stats.csv One row per batch with 10 or more companies, 46 rows. Within-batch similarity and a size-matched random control. n_with_vec, n_total, coverage mean_pair_sim, median_pair_sim, std_pair_sim every pair in the batch mean_nn_sim, median_nn_sim each company's closest batchmate centroid_tightness cosine to the batch centroid p90_pair_sim, p99_pair_sim, max_pair_sim share_pairs_gt_065 .. share_pairs_gt_085 share of pairs over a threshold control_mean_pair_sim, control_std random draw of the same size batch_minus_control mean_pair_sim minus the control top_pair, top_pair_sim closest pair in the batch most_isolated, most_isolated_nn lowest best-in-batch match n_clusters_080, largest_cluster_080 connected components at 0.80 ### prev_batch_stats.csv One row per batch with 30 or more companies, 37 rows. Each batch against the batch immediately before it and against everything earlier. within_mean_pair, within_mean_nn prev_batch, n_prev, prev_mean_pair, prev_mean_nn, prev_median_nn prev_nn_random_control, prev_nn_control_std, prev_nn_minus_control n_prior_pool, prior_mean_nn, prior_median_nn prior_share_nn_gt_080, prior_share_nn_gt_085 against the FULL prior pool, not pool-size controlled prior_fixed500_mean_nn, prior_fixed500_std, prior_fixed_sample_n resembles_most_pairwise, resembles_most_nn closest earlier batch The `prior_share_nn_gt_*` columns rise with batch order because the prior pool grows. Use prior_twin_share_controlled.csv instead. ### prior_twin_share_controlled.csv The pool-size-controlled twin shares, one row per batch, 37 rows. Every column is the share of the batch whose best match into a fixed random sample of 500 companies clears the threshold, averaged over ten draws. batch, n, prior_n prior_n is the full earlier pool, for reference fx500_prior_80, fx500_prior_85 500 drawn from earlier batches fx500_any_80, fx500_any_85 500 drawn from any other batch fx500_later_80, fx500_later_85 500 drawn from later batches `nan` means fewer than 500 companies were available to draw from, so no controlled number exists. That covers batches before Summer 2013 in the prior columns and the last two batches in the later columns. ### pairs_top.csv The 200 closest pairs of companies inside the same batch, with YC one-liners. sim, batch, name_a, domain_a, one_liner_a, name_b, domain_b, one_liner_b ### cross_batch_pairs.csv The 1,000 closest pairs of companies in different batches, with YC one-liners. sim, old_vs_new, batch_a, name_a, domain_a, one_liner_a, batch_b, name_b, domain_b, one_liner_b `old_vs_new` is 1 when one company is from before 2020 and the other from 2023 or later.