Convergence and Differentiation in the Business AI Market: Evidence from a Longitudinal Product Panel

Gary Wang · Business AI Market Observatory · manuscript generated 2026-09-07 from dataset v0.1.2 (parsers price-1.7, feat-1.1, prov-1.3; detector diff-1.0)

Abstract

We build a provenance-first longitudinal panel of 125 commercial AI products from 121 companies across 18 market categories and observe them through 4,642 immutable page snapshots between 2022-01-01 and 2026-09-06, 4,392 of them archived captures from the Internet Archive and 250 live captures. Deterministic, versioned extractors turn each snapshot into 11,808 plan observations, 148,544 feature observations and 8,433 provider observations, each with a quoted evidence window, a confidence score and a review status; diffing consecutive snapshots yields 5,813 interval-dated pricing and feature events. We find (i) modest convergence in claimed feature sets (mean pairwise Jaccard 0.20 → 0.23), driven by the spread of enterprise-governance and agent features rather than by products becoming alike overall; (ii) a standardized headline subscription price (median entry price $20.0, 43% within $15–30) coexisting with a diversifying set of pricing models (entropy 2.79 → 2.68 bits); (iii) rapid imitation of the features that diffused, with median lags of a few months from the earliest observed adopter; (iv) concentrated, evidence-based provider dependence, with the top three providers accounting for 59% of named product→provider links among the 42 downstream products that name any provider; (v) no general drift toward vertical specialization in page language, but a broad move toward enterprise-governance positioning; and (vi) entry concentrated in coding, agents and infrastructure. All results are descriptive, and the panel's purposive, 2026-conditioned selection limits population claims; we discuss selection, survivorship, taxonomy subjectivity, pricing comparability, scraping reliability and causal overreach explicitly.

1. Introduction

The market for AI software products has changed faster than any standard data source can track. Pricing pages are rewritten quarterly, features are announced weekly, and products change the models they run on without press releases. Existing views of this market are either directories (a current snapshot with no history), analyst estimates (undocumented methods), or anecdotes. None supports quantitative claims about change.

This paper introduces the Business AI Market Observatory, a longitudinal research database and analysis platform, and uses its pilot panel to study six questions: whether products are converging in feature sets (RQ1), whether pricing is becoming more standardized (RQ2), how quickly competitors imitate newly introduced features (RQ3), how dependent downstream products are on a small set of model providers (RQ4), whether products move from horizontal toward vertical specialization (RQ5), and which categories experience the most entry (RQ6).

Three design commitments distinguish the dataset from a directory. First, raw evidence is separated from interpretation: the unit of storage is an immutable page snapshot; every derived value points at the snapshot and quotes the text it came from. Second, history is observed, not reconstructed: historical states come from archived captures with their own timestamps, never from today's page. Third, uncertainty is exposed: every observation carries a parser version and a confidence score, low-confidence rows enter a review queue, and the website's data-quality page reports staleness, missing fields, contradictions and backlog.

2. Market background

Between late 2022 and 2026 the business AI market moved through at least three phases visible in our data: an assistant phase (chat products priced at a flat ~$20/month, free tiers as acquisition funnels), a platform phase (per-seat team and enterprise tiers, governance features, API price lists denominated per million tokens), and an agent phase (agent modes, background and coding agents, computer/browser use, and hybrid subscription-plus-usage pricing). The panel spans foundation-model providers, horizontal application categories (coding, productivity, sales, marketing, support, design, image, video, data, analytics, research, agents, voice), vertical categories (legal, finance, HR) and infrastructure.

3. Data collection

3.1 Panel

The pilot panel contains 125 products from 121 companies (data/panel.yaml). A product is included when it sells an AI-native capability to businesses, developers or prosumers, has or had a public pricing or product page, and is prominent in 2023–2026 market coverage. For each product we record all known pricing-URL variants (pages move), a curated launch date with a precision flag (day/month/year: year 49, day 45, month 31) and its announcement URL, and, where applicable, an exit or acquisition date with its source. 7 products with documented exits were added deliberately.

3.2 Sources and snapshots

Three collection methods write to one immutable snapshot table. Archived captures are listed through the Internet Archive's CDX index (status 200, at most one per month per URL) and fetched as original bytes; the archive timestamp is the observation time. Live captures respect robots.txt, use a descriptive User-Agent and a per-host delay, and treat anti-bot responses as terminal: 21 live attempts met anti-bot controls and were not retried, 55 returned errors, 1 were disallowed by robots.txt. A manual ingestion path stores analyst-recorded quotes as snapshots with the same provenance. In total 4,642 snapshots cover 122 products' pricing pages; 403 sources were registered including 96 announcement pages cited for entry/exit dates.

3.3 Extraction

Extraction is deterministic and versioned (Methods §4). The pricing parser (price-1.7) produced 11,808 plan observations across 108 products; a heuristic-assisted review pass (Methods §4a) verified 7623 and rejected 297 (worked examples, ROI claims, and one bad capture found in a Fintool snapshot that had captured an unrelated Microsoft support page), and 7811 rows remain queued, including 1380 rows flagged for an unresolved dual-price ambiguity. The feature lexicon (feat-1.1) covers 32 features (27 tracked for diffusion); the provider lexicon (prov-1.3) covers 19 providers. Mean parser confidence is 0.655 for plans and 0.708 for providers.

4. Taxonomy

Taxonomy v1.0 has 18 categories with one-sentence definitions and an orientation (horizontal, vertical, infrastructure). Assignment is curated from each product's stated job-to-be-done, multi-label with one primary label; the revision log lives with the data. The only departure from the study design is the addition of voice_audio for speech products. We treat taxonomy subjectivity as a robustness dimension: category-level results are reported with the multi-label and primary-only variants, and the market map offers a taxonomy-free view.

5. Measurement

Pricing. monthly_seat_usd is filled only for USD per-seat or flat subscriptions billed monthly (or an explicit annual-billing monthly equivalent, flagged) — never for token, usage, credit or non-USD prices. A product's entry price is its lowest positive normalized price. Pricing model is a coarse label (seat, flat, usage, hybrid, mixed, custom-only, free-only) from the units present on the page.

Features. Presence of a lexicon pattern on the pricing page. We measure claims, not capabilities.

Convergence. Mean pairwise Jaccard similarity of tracked-feature sets across products observed in a quarter, with a product-level bootstrap 95% interval.

Diffusion. Adoption share by quarter; interval-dated added events; Kaplan–Meier time from first snapshot to adoption with left-censored adopters excluded and counted.

Provider dependence. Evidence-backed product→provider edges; link-count HHI; multi-homing.

Positioning. Enterprise-governance terms per 1,000 words and a 0–1 vertical-specificity index.

6. Market evolution (RQ6, RQ5)

Multi-label totals across the panel: agents 48, infrastructure 28, productivity 25, coding 15, sales 14, foundation_model 13, data 12, research 12, video 11, marketing 11, image 10, customer_support 9, design 8, legal 6, analytics 6, hr 5, finance 5, voice_audio 4. Coding, agents and infrastructure receive the most entry, and 2023–2024 is the modal entry period for AI-native products in this panel. The vertical share of entries by year is 2021 0%; 2022 14%; 2023 9%; 2024 7%; 2025 0%.

Within products, positioning moved toward enterprise governance rather than toward verticals: of 118 products observed for three or more quarters, 52% increased enterprise-governance language (mean +1.99 terms per 1,000 words), while only 36% became measurably more vertical and 38% less (mean change in specificity 0.006).

Documented exits and acquisitions: MosaicML Platform (acquisition, 2023-06-26); CoCounsel (Casetext) (acquisition, 2023-08-17); Pi (Inflection) (exit, 2024-03-19); Rockset (acquisition, 2024-06-21); Adept (exit, 2024-06-28); Weights & Biases (acquisition, 2025-05-05); Windsurf (formerly Codeium) (acquisition, 2025-07-14). Three of these (Adept, Inflection, Casetext) are acqui-hire or absorption patterns in which the product page disappears while the technology persists inside an acquirer — a form of exit that a directory would record as a broken link.

7. Pricing (RQ2)

The headline subscription price is remarkably standardized: the median entry price is $20.0 in 2026Q3, and 43% of normalizable entry prices fall between $15 and $30. Around that anchor, the form of pricing diversified: pricing-model entropy fell from 2.79 to 2.68 bits as usage-metered, credit-based and hybrid subscription-plus-usage models spread. Free tiers (37%) and contact-sales tiers (49%) are both widespread, i.e. the modal product bundles a freemium funnel with an enterprise quote.

Printed subscription prices changed less often than the volume of plan restructuring suggests: the detector recorded 119 confident price changes (52% increases; median absolute change 33.3%) against 915 plan additions and 828 removals. Token-metered API price lists, which fell steeply over the period, are excluded from the change count because the parser cannot reliably pair a price with a model across page layouts; they are preserved verbatim in the api_usage plan rows.

8. Feature diffusion (RQ1, RQ3)

Similarity rose over the observation window (linear trend +0.0004/quarter, t = 0.69), and the mean number of tracked features per product rose from 3.6 to 7.2. The balanced sub-panel of 91 products observed in every quarter from 2024Q4 to 2026Q3 moves from 0.21 to 0.25, so the trend is not purely compositional. Decomposing by feature:

Convergence is concentrated in enterprise governance (SSO/SAML, compliance certifications, admin consoles, no-training promises) and in agent language; modality features (voice, video, image generation) remain differentiating.

For the features that diffused, imitation was fast: median lags from the earliest observed adopter are typically under a year. Two cautions apply. The earliest adopter in our panel is the earliest observed, and many adopters already showed the feature at their first snapshot (left-censoring), so lags are upper-bounded by the panel's start rather than by true introduction dates. And a feature "appearing" on a pricing page is a marketing decision that can lag the capability itself.

9. Product similarity

TF-IDF embeddings of 488 product-year documents (122 products) yield 10 clusters (silhouette 0.08), whose top terms are: cluster 0: papers, consensus, literature, researchers, candidate; cluster 1: dimensions, weaviate, vector, gcp, database; cluster 2: agent, agents, service, insights, workflows; cluster 3: agents, agent, mcp, code, agentic; cluster 4: perplexity, legal, law, firm, firms; cluster 5: ideogram, stability developer, jetbrains ides, stability, marketing software; cluster 6: ach, grammarly, counter, checker, writing tone; cluster 7: video, image, videos, banana, nano banana; cluster 8: tokens, model, inference, models, input; cluster 9: chatgpt, browser, gpt-, canva, intelligence. Mean pairwise text similarity by year is 2022 0.150, 2023 0.124, 2024 0.139, 2025 0.148, 2026 0.186. The t-SNE layout on the website is a visual aid only; we do not interpret distances as economic substitutability.

10. Provider dependence (RQ4)

42 of 106 downstream products name at least one provider or model on their own pricing or product page. Among named links the top three providers hold 59% (HHI 1537.0 on link counts). Multi-homing is the norm among disclosers (0: 64, 1: 18, 2: 12, 3: 8, 4: 3, 9: 1 providers named per product), consistent with products marketing "model choice" as a feature. The share naming any provider went from 10% (2022Q1) to 40% (2026Q3). These are disclosure-based measures: they cannot see undisclosed dependence, and they count products, not revenue or usage.

11. Robustness

• Composition. Quarterly samples grow with archive coverage; the balanced sub-panel (§8) and product-fixed-effects regressions address this. Tracked-feature count on quarter index: 0.2633 (SE 0.0254, p = 0.0) with product FE; 0.228 (SE 0.0239) with category FE. Log entry price with product FE: 0.0195 per quarter (SE 0.0165, p = 0.2372).
• Extraction noise. 795 feature events reversed at the next snapshot were flagged as flicker and excluded from diffusion statistics. Contradictory URL variants (24 product-months) are surfaced rather than resolved silently.
• Parser versions. All results are keyed to parser versions; re-running with a new parser appends new observation rows and leaves old ones auditable.
• Manual review. An analyst read roughly 500 plan cards verbatim, derived content-based rules for two recurring error classes (worked-example text captured as a plan; ambiguous dual-price strings), and applied them across all 11,808 rows (scripts/apply_pilot_review.py, Methods §4a): 7623 verified, 297 rejected, 1,380 flagged as ambiguous. The pass surfaced two further, distinct engineering bugs, both fixed and regression-tested before this release: a later, unrelated price in the same text window could leak its billing period onto an earlier price token (pricing parser v1.7); and the provider lexicon's bare 'Nova'/'Titan' patterns matched a demo person's name, a different vendor's own same-named model, and an unrelated deal codename (provider lexicon v1.3, found via a Hebbia product page). A third bug, found independently while re-running the pipeline rather than during content review, re-inserted a bookkeeping row on every rerun for snapshots with no parsed plans; it was fixed, 2,332 stale duplicate rows were removed, and idempotency is now covered by a regression test. All three are reported here rather than quietly fixed, because a pipeline that finds and discloses its own errors is more trustworthy than one that reports none.

12. Limitations

Selection is purposive and conditioned on 2026 prominence; archive coverage is uneven across products and years; extraction is lexical (claims on pricing pages, not verified capabilities); pricing comparability is limited to USD subscriptions; provider evidence is disclosure-based; launch dates are curated with declared precision. Every statistic is descriptive.

13. Response to Reviewer 2

"Your sample is selected on prominence in 2026 — selection bias." Correct, and we do not claim population coverage. The panel file states the inclusion rule; every count is labelled as describing the panel. Mitigations: 18 categories with a minimum of several products each; big-tech, incumbents and startups all present (startup 101, big_tech 4, incumbent 16); and results that depend on composition (convergence, pricing dispersion) are re-estimated on a balanced sub-panel and with product fixed effects.

"Survivorship bias — dead products are missing." Partly mitigated, not solved. We added 7 products from public exit/acquisition announcements, and the schema records exits as first-class events with sources. The paper's entry counts are therefore lower bounds on entry and the exit counts are severe under-counts; we say so wherever they appear.

"Taxonomy is subjective." It is curated and we say so. We publish definitions, orientation and a revision log; assignments carry a method and confidence so alternative assignments can coexist; the market map provides a taxonomy-free similarity structure as a check.

"Historical coverage is incomplete." Yes: quarterly sample sizes are reported next to every series, early quarters rest on few products, and analyses require at least eight observed products per quarter. Interval-censored event dates make the coarseness of monthly captures explicit rather than hiding it behind point dates.

"Pricing isn't comparable." We agree, which is why the normalized field is NULL for 43% of plan observations. Token, usage, credit and non-USD prices are preserved verbatim and summarised as pricing models, never converted.

"Scraping is unreliable." Every fetch attempt is logged with its outcome; anti-bot walls are respected; extraction confidence is stored per row; low-confidence rows are queued; flicker is flagged; and a manual ingestion path exists for pages automation cannot reach. The data-quality page publishes staleness, missing fields and backlog.

"You are implying causation." We are not. Adoption lags describe timing, not imitation as a mechanism; regressions are labelled descriptive; the market map is a visualisation; no market-share claims are made anywhere.

14. Conclusion

A provenance-first longitudinal panel makes the evolution of the business AI market measurable without pretending marketing pages are clean data. The pilot shows a market converging on a standardized headline price and a shared enterprise-governance vocabulary while differentiating on modality and diversifying in pricing form, with rapid imitation of the features that diffuse and concentrated, disclosed dependence on a few model providers. The schema and pipeline are built to scale collection to thousands of products; the pilot's value is that every number in it can be traced to a captured page.

Data and code availability

Dataset release v0.1.2 (SHA-256 4255c131f876f02c…); CSV bundle under data/releases/, schema in docs/SCHEMA.md, data card in docs/DATA_CARD.md, code and tests in the repository.