The practical consequence is simple: treat it as a long shortlist, not a tested buying guide. That distinction matters most for Windows users and Microsoft 365 administrators, because a generic chatbot subscription and a Copilot deployment are different purchases with different data boundaries, licensing conditions, and failure modes.
OpenAI’s reported 900 million weekly ChatGPT users are real. OpenAI announced that milestone on February 27, 2026, and TechCrunch independently reported it the same day. But that verifies ChatGPT’s scale, not the guide’s broader conclusion that a set of rivals has taken a measured, comparable share of the market — or that one assistant can be crowned the best replacement for every kind of user.
The “tested” rankings cannot be reproduced
Techshali says it tested more than 25 assistants over four weeks using the same prompts and grading notes, including a 1,200-word draft, a 40-page PDF analysis, a Python debugging task, citation checks, a math problem, and voice chat. That is the right general shape for a comparison. The missing material is what turns a claim of testing into evidence: the exact prompts, source files, dates of each test, model and plan selected, usage limits encountered, raw outputs, scoring rubric, and weights behind the final rankings.
Without those, statements such as Claude needing “the fewest rewrites,” Gemini handling a noisy voice note correctly, or Cursor finding a three-file bug are anecdotes, not comparative results. They may be true for the author’s account and moment in time, but no reader can determine whether the outcome came from the model, a premium plan, a temporary model rollout, a different system prompt, or the evaluator’s subjective preferences.
The article also lists six scoring criteria — answer quality, context handling, price, privacy, ecosystem, and freshness — but does not disclose how any were weighted. That omission can reverse an outcome. A Windows developer working entirely in Visual Studio and GitHub may reasonably value repository integration, audit controls, and identity management above literary writing quality. A regulated organization may weight regional processing and contractual data protections so heavily that a technically strong hosted chatbot is eliminated before benchmarks matter.
There is another warning sign in the copy itself. Several sections repeat data fragments and sentences, including the traffic-share claims and free-tier descriptions. That may be a publishing or formatting error, but it undermines confidence in the article’s claim that every entry was subjected to a consistent editorial and technical review.
A useful comparison would publish a spreadsheet or repository containing the prompts, outputs, test dates, plan tiers, human graders, and a clear declaration of conflicts or affiliate relationships. Techshali publishes none of those materials. Readers are being asked to accept the rankings on trust.
Model names are real, but the product claims are too broad
The guide is not inventing the major model releases it mentions. Google’s Gemini 3.1 Pro was announced on February 19, 2026, and Anthropic released Claude Opus 4.8 on May 28, 2026. Microsoft also rolled GPT-5.5 Instant into Microsoft 365 Copilot and Copilot Chat experiences during 2026. Those are verifiable product events.
The trouble starts when model availability is presented as if it describes a uniform consumer experience. It does not.
Take Claude’s context window. Techshali describes Claude Opus 4.8 as having a one-million-token context window and says users can effectively paste an entire book into a conversation. Anthropic’s own current materials distinguish among surfaces and plans: its Opus 4.8 product page promotes a one-million-token context window, while its Claude consumer help documentation says paid Claude plans use a 200,000-token window, with a 500,000-token option documented for certain Enterprise Sonnet use. Anthropic’s developer and cloud offerings also have their own terms, limits, and pricing.
That does not make Claude a poor choice for document work. It means “one million tokens” is not an answer until the buyer specifies whether they mean the Claude app, API, Amazon Bedrock, Google Cloud, Microsoft Foundry, and which account tier. A ranking that treats the largest published context figure as the default experience risks sending users toward a plan that does not deliver it.
Microsoft Copilot needs the same treatment. Techshali calls it the best choice for Microsoft 365 work, which is broadly defensible: Copilot’s value comes from its placement in Word, Excel, Outlook, PowerPoint, Teams, OneDrive, and Microsoft 365 identity and compliance controls. But the guide collapses free Copilot, Copilot Pro, consumer Microsoft 365 subscriptions, and commercial Microsoft 365 Copilot licensing into a single entry with a rough “around $20” price.
Microsoft’s current U.S. consumer pricing lists Microsoft 365 Premium at $19.99 per month, while Microsoft 365 Personal and Family have lower prices and different AI allowances. Microsoft’s support documentation also makes clear that Copilot features in desktop apps, advanced features, agents, and higher usage limits depend on the subscription and, in Family and Premium plans, are generally limited to the subscription owner. Commercial Microsoft 365 Copilot licensing is a separate decision, commonly listed at $30 per user per month for enterprise licensing, with business-specific bundles and terms.
For a Windows or Microsoft 365 shop, that is not a detail. It is the buying decision. “Copilot” does not automatically mean a chatbot can read a user’s mailbox, summarize a Teams meeting, analyze a local workbook, or access tenant data. Those capabilities depend on the account type, license, admin configuration, data location, permissions, and the app surface where the prompt is issued.
Traffic statistics do not establish chatbot market share
The guide’s market-share section is written with more certainty than the data supports. It claims that ChatGPT’s web-traffic share dropped from about 87% in January 2025 to roughly 54% by May 2026, while Gemini reached 2.9 billion visits and Claude reached 952.6 million. It acknowledges that Similarweb, Statcounter, Apptopia, and Sensor Tower measure different things, then treats their directional results as mutually reinforcing.
They are not directly comparable measurements.
Web visits exclude native-app use, API traffic, enterprise deployments, embedded assistants, and assistants accessed inside products such as Windows, Microsoft 365, WhatsApp, Instagram, Search, or X. Mobile app market share measures a different population again. A user can visit ChatGPT on the web, use Copilot inside Edge, use Gemini through Google Search, and never appear as a clean “market share” unit in any one dataset.
This is particularly relevant to Microsoft. Techshali describes Copilot’s web-visit share as small while correctly noting its broad distribution through Windows and Office. Those facts do not sit comfortably in a single consumer-chatbot ranking. Copilot’s strategic position is not measured well by standalone web visits because its most consequential use happens inside applications, managed identities, and enterprise workflows.
The defensible conclusion is narrower: ChatGPT remains enormous, and competitors are growing across multiple surfaces. The submitted numbers do not prove a single, comparable chatbot-market-share league table.
Privacy advice needs a deployment boundary, not a country label
The guide is right to flag DeepSeek’s hosted service as a data-residency concern. DeepSeek’s privacy policy says personal information is stored on servers in the People’s Republic of China. That is sufficient reason for organizations to keep confidential customer, employee, financial, and source-code data out of the hosted service unless their legal, security, and procurement teams explicitly approve it.
But the guide goes too far when it implies that running an open-weight model locally through Ollama makes privacy concerns disappear. Local inference can sharply reduce exposure, especially when the model, documents, embeddings, and user interface remain on an offline Windows device. It does not by itself create a compliant system.
Administrators still need to account for downloaded model provenance, Windows endpoint security, disk encryption, user permissions, backup destinations, browser extensions, telemetry, update services, remote APIs, and any connected tools. If Open WebUI, a coding extension, or an MCP server can reach the internet, the system is no longer meaningfully described as fully offline. Local AI is a deployment architecture, not a privacy guarantee.
The article is also too coarse when it says U.S. tools store data on U.S. infrastructure and Mistral stores data in the EU. Enterprise customers can often select regions, invoke cloud-provider controls, sign data-processing agreements, or use specific service tiers with different retention and training terms. Consumer products are usually less negotiable. The relevant question is not the vendor’s home country; it is the exact service, account type, processing region, retention setting, contractual terms, and connectors enabled.
What Windows users should do instead of following a single ranking
The most reliable recommendation in Techshali’s article is its least flashy one: use the tool that fits the job. The problem is that the article then assigns winners without enough auditable evidence to justify the confidence.
For Windows users, the initial shortlist should be based on where work already lives. Microsoft 365 users should test Copilot inside the Word, Excel, Outlook, Teams, and OneDrive workflows they actually use — not merely compare its standalone chat window with Claude or Gemini. Developers should test GitHub Copilot, Cursor, Claude Code, or other coding agents against a non-production branch of a real repository and measure review burden, test failures, rework, and elapsed time.
The METR randomized trial cited by Techshali offers a useful corrective. Its early-2025 study found that experienced open-source developers working in familiar mature repositories took 19% longer with the AI tools available at the time, despite believing afterward that they were faster. METR’s February 2026 update also said newer tools may improve results but that its later experiment had selection effects serious enough to make the estimate unreliable. The lesson is not that coding assistants fail; it is that vendor claims and felt productivity are insufficient measurements.
A sound trial has a modest scope:
- Test two or three candidates against the same real but low-risk task, using the exact licensing and account configuration you would deploy.
- Record the model selected, prompts, outputs, human edits, verification time, and errors found after the first answer.
- Keep sensitive data out of consumer chat services unless the organization has approved the service, region, retention terms, and connector access.
- Re-test after major model or plan changes, because a chatbot ranking can become stale faster than a Windows feature update.
ChatGPT does not need to be “replaced” for users to benefit from specialized tools. Claude, Gemini, Perplexity, Copilot, local open-weight models, and coding agents can each be useful alongside it. But a guide that claims to rank them on testing and data should show its work. Until Techshali publishes that work, its list is best read as a catalog of products worth evaluating — not evidence that its winners have won.