by Lisa Peyton, full-time faculty at the University of Oregon School of Journalism and Communication, AI marketing practitioner, and Forbes contributor

Yesterday morning I did something I tell my students to do all the time. I ran the experiment instead of reading about it. I wrote one research prompt, handed it to two AI agents along with my logged-in LinkedIn session, and asked each of them to audit the AI-for-content-marketing conversation on the platform. Same prompt, word for word. Same account. Same 30-day window. Then I sat back and watched ChatGPT Work and Claude Cowork take two completely different approaches to the same job.

If you’re running content for a brand or building your own thought leadership, you already know why this audit matters. LinkedIn is where our audience lives, and knowing which topics, formats, and voices are winning there is the difference between publishing into the void and publishing into a conversation. What you may not know yet is which AI agent to trust with that research. So I spent my Tuesday finding out, and I’m handing you the receipts.

Want to run this exact audit on your own niche? Download my new AI Audit Toolkit and follow along.

The Experiment: One Prompt, Two Agents, My Real LinkedIn Login

The assignment was a four-part audit covering the past 30 days: trending AI news and topics, top-performing posts and winning formats, tools and resources being shared, and influencers to follow. Both tools used their Chrome extensions to browse my logged-in LinkedIn session, which matters because LinkedIn is famously hostile to scrapers. No login, no data.

The prompt also carried strict sourcing rules, because accuracy is the whole ballgame in my shop:

• Every claim required a live LinkedIn URL
• No estimating or inventing engagement numbers (if a metric wasn’t visible, say “not visible”)
• Anything without a verifiable link had to be flagged, not smoothed over
• Each tool had to end with a methodology appendix documenting what it reviewed, what it couldn’t access, and how confident it was in each section

That last requirement turned out to be the most valuable part of the whole test. When you force an AI agent to show its work, the differences between platforms stop being vibes and start being data.

ChatGPT Work: Fast, Clean, and Careful

ChatGPT finished in 35 minutes. The report was polished, well-organized, and followed every sourcing rule I set. It reviewed 33 post cards across 12 documented searches, opened 12 post pages to verify them directly, and produced a report where every single link resolved to a real LinkedIn post. Zero fabrication. Zero flags needed. It even refused to guess at exact post dates, preserving LinkedIn’s relative labels (“2w,” “3w”) rather than inferring calendar dates it couldn’t confirm.

That is genuinely impressive behavior, and I want to give it full credit. Two years ago, “AI agent browses LinkedIn and returns 100% verifiable links” was science fiction. Today it’s a Tuesday.

But speed came with a ceiling. ChatGPT stayed inside LinkedIn’s search results and never ventured further. The result was a report built on roughly a dozen unique posts from about ten creators, recycled across all four sections. The influencer list was simply the authors of the top-posts list. The resources table was the same posts a third time.

Claude Cowork: Slower, Wider, and Willing to Show Its Work

Claude took almost two hours. Nearly four times longer. And the report explains exactly where that time went.

Claude reviewed roughly 120 to 130 posts across search results, my home feed, and individual creator activity pages. It visited eight creator profiles to pull follower counts and posting cadence directly. It fully captured 18 posts with clean permalinks, exact engagement figures, and capture timestamps. Where LinkedIn only showed relative dates, Claude decoded the millisecond timestamps embedded in each post’s ID to confirm every sourced post fell inside my 30-day window, then reported the exact dates.

It also did something I’ve never seen an AI research agent do unprompted. It disclosed its own environment problems. Mid-session, one of its browser tabs started auto-navigating through searches on its own, and a third-party extension overlay injected itself into the page. Claude flagged both in the methodology appendix, moved to a fresh tab, and kept going. That’s the kind of sh*t you only catch when the agent tells you, and it earned real trust points with me. (It also sent me straight to my extensions list for a cleanup, which I recommend before you hand ANY agent your logged-in browser.)

One more detail worth noting: my own AI Unleashed article surfaced organically in Claude’s searches, and instead of padding the results with it, Claude included it with a disclosure that it belonged to the requester. Chef’s kiss.

The Scorecard

Dimension	ChatGPT Work	Claude Cowork
Runtime	35 minutes	~1 hour 55 minutes
Posts reviewed	33	~120–130
Posts fully verified with URLs	12	18 (plus 8 creator profiles)
Unique creators cited	~10	25+
Biggest post surfaced	94 reactions	2,246 reactions
Link accuracy (my manual check)	100%	100%
Date handling	Kept relative labels	Decoded exact dates from post IDs
Items flagged as unverified	0	~9, clearly marked
Link usability	Copy-and-paste required	Clickable inline

Both tools passed the accuracy bar. Neither invented a single post. That alone is a milestone worth celebrating, and it means the comparison gets decided on quality of research rather than trustworthiness.

The 94-Reaction Problem

Now for the finding that decided this matchup for me.

ChatGPT’s #1 “top-performing post” had 94 reactions. Claude’s top ten included posts with 2,246, 1,302, 1,049, and 929 reactions. Claude found Becca Chambers’ viral “AI writing has a shape” chart, Ethan Mollick’s exponential-growth argument, Ryan Levesque’s newsletter post citing the research where readers identified AI writing from structure alone 93% of the time, and OpenAI’s own ChatGPT Work launch livestream. Those are the posts that dominated this niche over the past month. ChatGPT’s report missed every one of them.

Neither tool was dishonest about this. Both appendices admitted that LinkedIn offers no platform-wide engagement sort, so “top-performing” can only mean “the best of what I surfaced.” The difference is that ChatGPT accepted the first net it cast, while Claude kept casting: feed, profiles, activity pages, follow-up searches on the stories it found. One answered the prompt. The other investigated the question.

The influencer sections told the same story. Claude’s list of voices to follow, with Ethan Mollick at the top, mapped closely to the people I follow and cite in my own work. ChatGPT’s list was, again, the authors of the twelve posts it happened to verify.

Two Ways to Tell the Truth

Here’s the nuance I’ll be chewing on for a while. ChatGPT produced zero unverified items partly because it only reported what it could easily verify. Claude carried nine flagged items because it cast a wider net and told me exactly which fish got away. Call it compliance by omission versus compliance by disclosure.

Both are honest. But as an editor-in-chief, I know which one I want on my team. A report that says “I saw this compelling post but couldn’t capture the link, verify before using” gives me a decision to make. A report that quietly drops everything inconvenient gives me a blind spot.

What This Means for Your AI Stack

If you’ve read my work on model routing (the owl and the rottweiler will be familiar to my regulars), you already know where I land: the skill isn’t picking one tool, it’s knowing which job goes to which tool. This experiment gave me a clean routing rule for research:

• ChatGPT Work is my pulse check. Thirty-five minutes, clean links, a reliable read on what a quick search surfaces. Perfect for a fast weekly scan.
• Claude Cowork is my monthly audit. When I need to know what’s genuinely winning with my audience, the extra ninety minutes buys me 4x the coverage, the posts that matter, and a methodology appendix I can stand behind.

You already have this muscle. You route work between team members based on their strengths every single day. Routing between AI agents is the same call, and now you have the data to make it.

Ready to put your own niche under the microscope? Grab my new AI Audit Toolkit and run this experiment yourself.

And if head-to-head experiments like this are your jam, this is exactly what we do every month in my Advanced AI for Content Marketing Alumni Meetup, and my weekly AI Marketing Brief delivers the next one straight to your inbox. I can’t wait to see what we discover together.

---

Made with my team of AI superheroes and my own skills. Every opinion, edit, and fact-check is mine. I am an AI practitioner, professor, and pioneer helping marketers put AI to work with purpose. Find more resources at lisapeyton.com/ai-marketing-resources or connect with her at linktr.ee/lisapeyton.