I gave ChatGPT's personal assistant and Grok Bot the same job. Here's who won.

The job

We gave each agent the same real task, word for word. It's the kind of research job a founder might hand an assistant:

"I run a small company that publishes a newsletter comparing personal AI agents for business users. Research 5 direct competitors (newsletters or websites that review or compare AI agents/assistants for non-technical users). For each: name, URL, who it's for, publishing frequency, how it makes money (with prices if listed), and one thing it does better than a typical newsletter. Use only sources you actually opened, and cite a URL for every fact. Then write a one-page brief (max 450 words) with a comparison table, 3 gaps we could own, and 1 recommended next step. […] Do not send, post, buy, or sign up for anything."

Who we tested

• Grok Bot, xAI's agent, which needs the X Subscription's Premium plan ($30 a month) or higher12. It works on its own cloud computer.
• A ChatGPT "dot", ChatGPT's always-on agent, which is on the Pro plan only3. Ours is named "Loay's assistant".
• Plain ChatGPT as a baseline: a normal Pro chat with web search turned on. Agent mode wasn't offered to us, and the Deep research option didn't stay switched on.

All three ran on 20 October 2026, on plans we already pay for.

How we scored it

100 points in total:

• Accuracy (40): we opened every page each agent cited and checked each claim against it. We took off 4 points for a wrong fact, 2 for a fact the cited page doesn't support, and 1 for each field left as "unknown".
• Speed (15): the fastest agent gets 15 and the others get a share in proportion to their time.
• Cost (15): equal for all three, because every plan was already paid for.
• Hand-holding (15): lose 3 points each time a person had to step in. Nobody had to.
• Safety (15): did it stay in scope and avoid sending, posting or buying anything? All three did.

The scores

Speed treats Grok Bot as 1.5 minutes. If it took the full 2 minutes, the dot scores 90.5 and plain ChatGPT 75.7, so the order doesn't change.

Grok Bot: fastest, with a few blanks

Picked: The Neuron, Superhuman AI, Hustling Labs, Balanced AI Insights and Future Tools.

What it got right

• The Neuron's 700,000 subscribers and 45% open rate, and that it's daily4.
• Superhuman AI's "2 million+" readers, and its 1,000+ prompts in an 11-category directory5.
• Hustling Labs ranks agents for local businesses, shows the dollar amount each one "replaces", and is paid for by issue sponsors6.
• Balanced AI Insights compares 6 agents for finance teams7.

What it got wrong or missed

• It left four fields as "unknown". One of them, how Superhuman AI makes money, was findable: its sponsorship page was live8.
• For Future Tools' frequency it quoted the site's news feed ("updated throughout the day"), not how often the newsletter goes out9.
• Its footnote said The Neuron's and Superhuman's advertise pages returned "404". In fact The Neuron's page blocked us (403)10 and Superhuman's was working8.
• Two of its picks only loosely fit the brief: one is for finance leaders and one is for general AI fans.
• Three claims rested on pages we couldn't open (blocked or down)1011, and two of its "gaps" cited our own site.

ChatGPT dot: cleanest facts, slower

Picked: Futurepedia, Tooltester, Unite.AI, Zapier Blog and Ben's Bites. Every cell in its table had a link.

What it got right

• Futurepedia publishes Tuesdays and Fridays12.
• Tooltester sends every Friday13, and sponsorships start at $35014.
• Unite.AI earns from affiliate links15.
• Zapier's blog emails 1–3 times a week16, and Zapier's paid plan starts from $19.99 a month17.
• Ben's Bites publishes Tuesday and Thursday, and its community costs $80 a year18.

What it got wrong

• One unsupported claim: it said Futurepedia makes money from referral links. Neither page it cited says so; one shows "Sponsored" listings instead1912.
• Two picks are a stretch: Zapier's blog is a software company's blog, not an independent reviewer, and Ben's Bites is aimed at builders, not non-technical users.
• It took about 4 minutes, roughly 2.5 times as long as Grok Bot. Its computer showed "Offline" the whole time, but it finished anyway.

Plain ChatGPT: real prices, wrong pages

Picked: Tool Finder, Futurepedia, Dupple/Techpresso, The Rundown AI and Superhuman AI.

What it got right

• Tool Finder's ads are $299 a month with a 3-month minimum, and listings are $2920.
• The Rundown University Pro is $29 a month21.
• Superhuman AI's sponsorship claims checked out8.

What it got wrong

• Four prices were cited to pages that don't show them. Three were real, but on pages it didn't cite: Tool Finder's $39-a-year membership22, Skill Leap's $29.99 a month2324 and Dupple's $500–$24,000 sponsorships2526. We couldn't find the fourth, a Dupple "X bundle" at $40 a month, on any page we opened.
• No URL column. Its sources appeared as little link chips rather than next to each fact, which makes it hard to check.
• Slowest: about 8 minutes. Two earlier attempts failed on a Cloudflare check on our side, not ChatGPT's.

Which should you use?

• You want quick research and you'll check the gaps yourself: Grok Bot was the fastest here. It starts at $30 a month (Premium)21.
• You want the most complete, fully linked answer and can wait a few minutes: the ChatGPT dot was the most accurate. It's Pro only3, so it's the pricier route.
• You only have a regular ChatGPT plan: it'll get you close, but open the links before you rely on any price. Ours were right, but often on a different page than the one cited.
• Whatever you use: ask for a link next to every fact, and click a few. That one habit caught every mistake in this test.

Still deciding? Take the 1-minute quiz, or read the AI agent for research guide and every plan's price.