AI Assistant Onboarding and Feature Discovery

Vendor documentation identifies plan features, file and memory controls, usage limits, and accuracy cautions, but it cannot show whether an assistant fits your own workflow.

Sources checked 2 Oct 2026

For AI Assistant Onboarding and Feature Discovery, vendor pages document account, plan, file, memory, and policy controls, but not how easily you can use them in your own work. OpenAI’s Projects page, Anthropic’s Projects page, and Google’s Gemini personalization page show different setup paths. Chat Picker has not tested the assistants for this use, so this page compares documented differences and gives you a self-run checklist.

What the vendors document

Figures below are as read on October 1, 2026. ChatGPT Plus costs $20 per month on monthly billing; Claude Pro costs $20 monthly or $17 monthly with annual billing, with $200 due up front; and Google AI Pro costs $19.99 per month. The free pages list unlimited text chats for ChatGPT; web search, file creation, code execution and memory for Claude; and a Google Account plus 15 GB of storage for Gemini.

Setup paths differ. OpenAI Projects group chats, files and instructions for signed-in users, subject to plan and workspace settings. Claude Projects are available to all users; free accounts can create 5, while enhanced retrieval-augmented generation knowledge is limited to Pro, Max, Team and Enterprise. Google’s personalization page says its features are not available to everyone, require a personal Google Account, and are unavailable through work, school or supervised accounts.

Uploads have account-specific limits. OpenAI’s file uploads FAQ says Free and paid plans support uploads; Free has 3 file uploads per day and a general 512 MB per-file cap. Anthropic’s upload guide lists 500 MB for chat uploads and 30 MB for project files; PDFs with 100 pages or fewer receive text and visual analysis. Google’s upload guide allows up to 10 supported files per prompt, subject to availability; videos can be 2 GB each and other supported files 100 MB.

Check data controls first. OpenAI’s data-controls page says turning off Improve the model for everyone prevents new chats from training models but does not delete saved chats; Temporary Chat stays out of history and may be retained for 30 days for safety. Anthropic’s privacy article says chats and coding sessions improve models if the user allows it; Incognito chats are not used for model improvement. Google’s plans page links data-handling information but states no training opt-out setting, so Chat Picker marks that item Not verified.

Documentation sets expectations, not guarantees. OpenAI’s accuracy note says ChatGPT can be incorrect or misleading, may sound confident while wrong, and should be checked against reliable sources; tool access affects access to current information. Anthropic’s accuracy note says Claude can also mislead and advises checking cited sources and original pages. Google’s related-sources page says Gemini Apps sometimes show sources within or below answers, giving you material to inspect rather than an accuracy guarantee.

Usage boundaries differ. OpenAI’s policy prohibits tailored advice requiring a license, such as legal or medical advice, without appropriate professional involvement. Anthropic’s policy requires qualified professional review of covered advice before dissemination and AI disclosure at the start of specified consumer-facing sessions. Google’s guidelines say context matters and warn against factually inaccurate outputs that could cause significant harm to health, safety or finances.

What the documentation cannot tell you

The pages cannot tell you whether a control is easy to find, whether the assistant follows your preferred sequence, or how it handles your files. They provide no like-for-like record of task time, usefulness or output quality in your workflow.

Keep the task, input, plan, device and browser consistent while comparing. Record access limits separately from answer quality: a blocked feature and a poor answer are different failures.

How to check it yourself

Set a baseline. Give each assistant: “Turn this policy into an onboarding checklist for a fictional book club. Meetings happen in the evening; new members finish the assigned reading before attending; each member brings a question and an annotated passage. Return the checklist, assumptions, and items needing confirmation.” Look for use of every rule. Write down missing constraints, corrections and controls you had to find.
Test guided prompting. Give: “Act as a study coach. Ask diagnostic questions about correlation and causation sequentially. Wait for my answer before continuing, then give a short lesson, self-check questions with answers, and a paragraph describing the misconception I need to correct.” Look for sequence and adaptation. Write down where the exchange broke and what instruction needed repair.
Probe file features. Create library-policy.txt containing: “Meetings happen in the evening. New members finish the assigned reading before attending. Each member brings a question and an annotated passage.” Upload it and ask: “Summarize the rules, quote the relevant lines, and separate them from assumptions.” Look for file recognition and faithful use. Record the file type, upload notice, missing content and access path.
Discover advanced controls. Ask: “List the tools, sources, file actions, memory controls, project features, and export options in this interface that could help me study. For each, name its trigger, where it appears, and any stated plan or account dependency. Separate visible controls from claims that need verification.” Check every claim against the documentation above. Write down confirmed, unclear and absent items.
Measure, then reset. Repeat the file task in a fresh chat, or in a fresh project where the vendor documents that option. Give: “Use only library-policy.txt. Create an onboarding checklist, separate stated rules from assumptions, and do not use details from earlier chats.” Look for adherence to the current task and whether earlier details carried over. Write down which checklist items were met or missed, and whether a reset is warranted after your goal, plan or workspace changes.

Which rows of the comparison matter

Open the Claude vs ChatGPT vs Gemini matrix and start with Free plan, Low-cost tier, Main paid plan and Heavy-use plans; then use Team plan for seat and workspace terms. Check Context window in the app, but do not treat it as a promise that every file will fit. Compare How usage limits are described, Training on your chats and Ads when relevant.

Treat Not verified as blank, not as no. Chat Picker leaves a cell blank when the vendor page did not confirm the figure. Confirm current terms on the vendor page before subscribing.