Research findings

Six research questions, answered descriptively from a panel of 125 products, 4,642 immutable snapshots (2022-01-01 → 2026-09-06) and 11,808 plan observations. Every claim below links to a page with the underlying series, sample sizes and evidence. Read the methods and the Limitations section of the paper before quoting any number.

RQ1 — Are AI products converging in feature sets?

Partly, and mostly through governance and agent features. Mean pairwise Jaccard similarity of tracked feature sets across observed products rose from 0.20 in 2022Q1 (n=45) to 0.23 in 2026Q3 (n=115); the quarter-level linear trend is +0.0004 per quarter (t = 0.69). The mean number of tracked features claimed per product rose from 3.6 to 7.2. In the balanced sub-panel (91 products observed in every quarter 2024Q4–2026Q3), similarity went from 0.21 to 0.25. Features whose prevalence grew most between 2022Q1 and 2026Q3: AI agents (+49%), Workflow automation (+32%), MCP support (+25%), Audit logs (+22%), Image generation (+20%), Coding agent (+19%). Falling: Knowledge retrieval (RAG) (+0%), Canvas / artifacts (+2%), Custom / fine-tuned models (+3%). Convergence here means products claim increasingly overlapping capability sets on their pricing pages; it does not mean the capabilities perform alike. → Feature diffusion

RQ2 — Is pricing becoming more or less standardized?

The headline subscription price is standardized; the pricing model is not. Among products with a normalizable entry price, the median was $12.0 in 2022Q1 and $20.0 in 2026Q3 (n=47); 43% of entry prices sat in the $15–30 band and 13% were exactly $20. The coefficient of variation rose (1.78 → 2.36), and pricing-model entropy fell (2.79 → 2.68 bits) as usage, credit and hybrid subscription-plus-usage models spread alongside seat pricing. Free tiers were offered by 37% and contact-sales tiers by 49% of products in 2026Q3. The detector found 119 confident changes to a printed subscription price, of which 52% were increases (median absolute change 33.3%), plus 915 plan additions and 828 removals. → Pricing

RQ3 — How quickly do competitors imitate newly introduced features?

Fast for the features that diffused at all, but the panel's start date left-censors the earliest movers. 27 tracked features have at least one adopter; for those with ≥3 in-panel adoptions, the median lag from the earliest observed adopter is: AI agents 30.0 months (62 adoptions, 8 already present at first observation); Workflow automation 32.4 months (55 adoptions, 34 already present at first observation); Third-party integrations 33.0 months (46 adoptions, 61 already present at first observation); Knowledge retrieval (RAG) 33.0 months (45 adoptions, 43 already present at first observation); SSO / SAML 32.4 months (45 adoptions, 28 already present at first observation); Admin console 38.6 months (39 adoptions, 15 already present at first observation); Audit logs 43.2 months (38 adoptions, 5 already present at first observation); Credit-based usage 35.8 months (38 adoptions, 48 already present at first observation). Kaplan–Meier median time from a product's first snapshot to adoption, where estimable: Third-party integrations 17.9 months (events 46/61); Workflow automation 26.8 months (events 55/88); Knowledge retrieval (RAG) 32.5 months (events 45/79); Credit-based usage 37.5 months (events 38/74); SSO / SAML 41.0 months (events 45/94); AI agents 42.8 months (events 62/114). → Feature diffusion

RQ4 — How dependent are downstream products on a small set of model providers?

Where products name a provider, dependence is concentrated. 42 of 106 downstream products (40%) name at least one foundation-model provider or model on their own pages. The top three providers account for 59% of named links (HHI on link counts 1537.0): openai 24, anthropic 17, aws_bedrock 10, google 10, azure 8. Multi-homing is common among those that name anyone: 0 provider(s): 64 products, 1 provider(s): 18 products, 2 provider(s): 12 products, 3 provider(s): 8 products, 4 provider(s): 3 products, 9 provider(s): 1 products. The share of downstream products naming any provider rose (10% in 2022Q1 → 40% in 2026Q3). Absence of a mention is not independence: many products never disclose their model stack. → Providers

RQ5 — Do products move from horizontal toward vertical specialization?

Not in this panel's page language; the movement is toward enterprise governance, not verticals. Among 118 products with three or more quarters of observation, mean within-product change in vertical specificity was 0.006; 36% became more vertical and 38% less. Enterprise-governance language (SSO, SOC 2, audit logs, data residency…) increased for 52% of products (mean change 1.99 terms per 1,000 words). The vertical share of curated entries by launch year: 2021: 0% (8 entries); 2022: 14% (21 entries); 2023: 9% (54 entries); 2024: 7% (15 entries); 2025: 0% (1 entries). → Research charts

RQ6 — Which market categories experience the most entry?

Coding, agents and infrastructure, with a 2023–2024 peak. Multi-label entry counts by category across the panel: agents 48, infrastructure 28, productivity 25, coding 15, sales 14, foundation_model 13, data 12, research 12. 7 documented exits or acquisitions were added deliberately (MosaicML Platform, CoCounsel (Casetext), Pi (Inflection), Rockset, Adept, Weights & Biases, Windsurf (formerly Codeium)). Because the panel was selected in 2026, these counts describe the panel, not the population; see the survivorship discussion in the paper. → Timeline

Product similarity (market map)

TF-IDF embeddings of 488 product-year documents form 10 clusters (silhouette 0.08). Mean pairwise text similarity by year: 2022 0.150 (n=62), 2023 0.124 (n=84), 2024 0.139 (n=107), 2025 0.148 (n=116), 2026 0.186 (n=119). → Market map

Robustness snapshot

Product-fixed-effects regression of tracked-feature count on quarter index: 0.2633 per quarter (SE 0.0254, p = 0.0, N = 1636); with category fixed effects instead: 0.228 (SE 0.0239). Log entry price with product fixed effects: 0.0195 per quarter (SE 0.0165, p = 0.2372, 79 products). 795 feature events were flagged as flicker and excluded; 7811 rows are in the review queue; a heuristic-assisted review pass (Methods §4a) verified 7623 plan observations and rejected 297 (worked examples, ROI claims, and one bad capture).