Leading AI chat tools get broadly similar satisfaction scores despite huge differences in resources and benchmarks; most users run multiple services and switch freely, implying specialization sustains competition rather than winner-take-all consolidation.

Main Finding

Active, engaged AI chat users treat assistants as largely interchangeable utilities: multi‑platform use is pervasive, switching costs are low, and user satisfaction is driven as much by interface, policy, and product features as by raw benchmark performance. Despite large resource differences, the top three platforms (Claude, ChatGPT, DeepSeek) report statistically indistinguishable satisfaction, and competition is sustained by platform specialization rather than winner‑take‑all consolidation.

Key Points

• Sample and reach

  • Survey collected late 2025; total responses = 388; primary analytic (second‑wave) sample N = 237; complete demographic subsample n = 171; qualitative responses n = 329.
  • Respondents skew toward technology professionals and students; 79.5% report daily use; respondents from 37 countries (largest shares: North America 38.2%, South Asia 36.4%).

• Multi‑homing and switching

  • Mean number of platforms used = 2.83 (median = 3); 82.4% use two or more platforms; 52.9% use three or more.
  • ChatGPT is the most commonly used (66.7% of respondents) and acts as the default starting point for exploration; most switchers to other primary tools report coming from ChatGPT (e.g., 78.6% of those who switched to Claude came from ChatGPT).
  • Low switching‑in rate for ChatGPT (8.6%) vs higher switching‑in for newer entrants (e.g., Mistral 42.1%), consistent with first‑mover anchoring.

• Satisfaction and heterogeneity

  • The top three platforms (Claude, ChatGPT, DeepSeek) show statistically indistinguishable overall satisfaction despite large differences in funding, team size, and benchmark performance.
  • Platform choice clusters around different strengths: ChatGPT (interface/usability), Claude (answer quality), DeepSeek (word‑of‑mouth growth), Grok (content policy/permissiveness).
  • First‑mover anchoring: respondents who adopted ChatGPT as their first AI chat tool rated satisfaction 1.34 points higher than those who arrived from a competitor.

• Persistent frustrations

  • Thematic analysis of open responses identifies hallucinations (factual errors) and content moderation/filtering (policy limits) as the two most common user frustrations, representing a tradeoff between accuracy and permissiveness.

• Other findings

  • Adoption drivers and use‑case fit vary by platform (nine adoption drivers and six use‑case ratings were collected).
  • Multi‑platform use sometimes signals ongoing search for better fit (e.g., negative correlation between number of platforms used and satisfaction for Claude users).

Data & Methods

• Survey design

  • Instrument delivered via Qualtrics with four sections: demographics/usage; model selection (checklist of 7 platforms); per‑model evaluation blocks (within‑subject design); open‑ended questions.
  • Per‑model block items included: overall satisfaction (5‑point Likert), nine adoption‑driver importance ratings, six use‑case performance ratings, subscription plan, reaction to hypothetical 25% price increase, whether it was first AI tool, tenure, and switching history.

• Sampling and data collection

  • Convenience sampling through technology‑focused online communities (primarily Reddit) and professional networks.
  • Two waves: initial five‑model checklist (n≈151) with free‑text Other; mid‑survey addition of DeepSeek and Mistral for second wave (final analytic N = 237). The instrument change is reported and robustness checks were performed.

• Cleaning and quality checks

  • Excluded responses below 3 minutes, contradictory model selections, and straight‑lining.
  • Robustness checks: compared full vs partial completers and early vs late respondents; ChatGPT satisfaction stable across these splits.

• Analytical approach

  • Non‑parametric tests used for ordinal Likert data: Kruskal‑Wallis H tests for group differences; Mann‑Whitney U for pairwise comparisons with Bonferroni correction (adjusted α = 0.0024 for 21 pairwise tests).
  • Effect sizes: ε2 for Kruskal‑Wallis, Cohen’s d for pairwise, Cramér’s V for chi‑square.
  • Internal consistency: Cronbach’s α reported for use‑case (≈0.79–0.85) and adoption‑driver scales (≈0.75–0.80).
  • Open replies analyzed with inductive, keyword‑assisted thematic coding.

• Limitations noted by authors

  • Convenience, tech‑oriented sample (not representative of general population).
  • Mid‑survey instrument change (DeepSeek, Mistral added) — analyses involving those platforms should be interpreted cautiously.
  • Small per‑platform subsamples for some competitors reduce power for certain comparisons.

Implications for AI Economics

• Market structure and competition

  • Low switching costs and prevalent multi‑homing argue against rapid winner‑take‑all consolidation among engaged users; competition is likely to remain pluralistic, with suppliers competing on differentiated features and domain fit.
  • First‑mover anchoring (higher satisfaction among users who adopted ChatGPT first) suggests incumbency and default status provide durable advantages even when multi‑homing is common — incumbency effects coexist with low churn.

• Valuation and investment signals

  • Resource and benchmark advantages do not map directly to higher user satisfaction among power users. Investors and firms should account for product features (UI/UX, policy stance, integrations, niche performance) in addition to raw model capability.
  • Niche or specialized entrants can sustain viable positions by targeting specific adoption drivers (e.g., policy permissiveness, answer style, vertical integrations).

• Product strategy and prioritization

  • Product teams should prioritize features that drive everyday satisfaction (interface, reliability, content policy alignment, domain suitability) rather than optimizing only for automated benchmarks.
  • Addressing hallucinations and designing transparent, calibrated content‑filtering tradeoffs are high‑value engineering targets because they are primary user pain points.

• Policy and market monitoring

  • Regulators and antitrust analysts should incorporate multi‑homing rates, user satisfaction parity, and feature differentiation into assessments of market power — traffic or compute spend alone may overstate dominance.
  • Consumer welfare analyses should consider heterogeneity in user priorities (quality vs permissiveness vs cost) and the role of default/anchoring effects.

• Research and evaluation

  • Automated benchmarks remain necessary but insufficient; platform‑level, user‑centric evaluation protocols (covering interface, policy, multi‑turn reliability, and ecosystem features) are required to capture real‑world competitive dynamics.
  • Future empirical work should use more representative sampling and larger per‑platform samples to quantify heterogeneity across broader user populations and task domains.

Summary: Among active AI chat users, competition is sustained by specialization and product attributes rather than absolute benchmark superiority. Multi‑homing and low switching costs produce a pluralistic market, but incumbency effects and user experience factors (interface, policy, hallucination mitigation) strongly shape adoption and satisfaction.