So What? Marketing Strategy Review with Agentic AI

Katie Robbert – 00:00

Hey everyone, happy Thursday. Welcome to So What, the Marketing Analytics and Insights live show. I am joined by Chris and John. Howdy, fellas.

Christopher Penn – 00:45

Hello. We missed it. There you go.

Katie Robbert – 00:52

As a side note, Chris, I was thinking about the title of the show. We call it So What, and I think it is the Marketing Analytics and Insights Live part that we can absolutely change. I remember when we started the show, you wanted the acronym and it was MAIL—Marketing Analytics, Insights Live. I said that was dumb and we were never going to call it that.

As things evolve and as marketing steps more and more into the agentic AI world—which is what we’re talking about today—we could probably do something more with the name. We will keep the So What as we talk about marketing strategy review with agentic AI. Perhaps one of my takeaways offline will be to use what I learned in today’s episode to figure out what I can do with the branding of some of our outdated stuff without making it feel like a 180-degree turn to our users. With that preamble, Chris, where would you like to start today?

Christopher Penn – 01:54

This is part two of last week’s episode. Last week we talked about how we prompt agentic AI. You can see that on our YouTube channel if you missed it. We also talked about this on the Trust Insights podcast, In-Ear Insights, which you can find at TrustInsights.ai/podcast.

Agentic, for those who missed last week, is really about being good at delegation and project management. If you are good at those two things, you’re going to be great at using agentic AI. If you are not good at those things, this is a great time to get started.

Katie Robbert – 02:29

One of the things we highlighted on last week’s episode was that if you are not sure how to get started, we recommend using the Trust Insights 5P Framework. You can get the downloadable PDF at TrustInsights.ai/5pframework.

The 5Ps are Purpose, People, Process, Platform, and Performance. Purpose: What is the question you’re trying to answer or the problem you’re trying to solve? People: Who is involved internally and externally? Process: How are you doing this thing? Think about it like your standard operating procedure. Platform: What tools are you using? It’s not just where this will live, but what systems you have to interact with. Performance: Did you answer the question or solve the problem?

Christopher Penn – 03:13

Exactly. When we think about using agentic AI for marketing strategy, we have to think about what the project plan looks like and the things you would want to do. We said this week we would work on the Marketing Over Coffee podcast, which you can listen to at marketingovercoffee.com. John and I have been doing that for 19 years. It is older than one of my children and just under the age of my oldest child. It has been around for a long time.

We want the podcast to succeed. We want more listeners and more sponsors. We want sponsors to up-level what they are willing to invest in the show.

Christopher Penn – 04:07

To do that, we need to know what the current situation and landscape look like. How has the world of podcasting changed? We would want to know about the show itself and who our ideal customers are on both the listener side and the sponsor side. From those things, we could do something like a gap analysis. Generally, that is what we would want to do. Katie, is there anything you would add or take away from that general approach—besides the ICP, the competitive landscape, and the company profile?

Katie Robbert – 04:44

No, I think that’s a solid start. We always talk about the best way to kick off these projects is to get organized with your foundational information. It is the about section, the competitive landscape, and the customers. I think that’s a solid place to begin, with the caveat that it is just the beginning.

Christopher Penn – 05:09

Because this is agentic, we want to have the pieces in place first. You don’t want to just go to Claude Cowork and say, “Go build me a marketing strategy.” That is the 2026 equivalent of “Write me a blog post about B2B marketing.”

John Wall – 05:25

Right.

Christopher Penn – 05:25

You are just going to get a pile of slop. You have to think through the pieces we need. The good news is agentic software can extract some of them, but you have to check the extractions. Let’s walk through the pieces you would want to have in place to pull this off, and then we can actually do some of it. I say some of it because if we were to do them on the air, this would be a four-hour livestream and three and a half hours would be us sitting around waiting.

The first thing you will need is an agentic framework. The one we have been recommending for non-technical folks is Claude Cowork.

Christopher Penn – 06:05

If you are in the OpenAI ecosystem, you would use OpenAI Codex, the desktop application. If you were in the Gemini ecosystem, you would use Google Anti Gravity. If you are in the Microsoft Copilot system, there isn’t an equivalent yet, though I imagine there will be at some point.

To do this, we would need to have three key components: something that can build an ICP from external data sources, something that can build a company profile from external data sources, and something that can do gap analysis. Our first step is to build those skills. This is where agentic AI leans heavily on the skills you built from 2023 to 2025—good prompting and good frameworks.

Christopher Penn – 06:59

One of the frameworks we use all the time is the Trust Insights CASINO Framework, which stands for Context, Action, Scope, Intent, Narrator, and Outcome. It’s a six-part framework you can find at TrustInsights.ai/casino. If you build out those prompts, you can turn them into skills and plugins. Katie, you just did this yesterday.

Katie Robbert – 07:34

Full disclosure, for me it was more of an experiment. I spent quite a while working in Claude Desktop on a project involving our standard Trust Insights PowerPoint deck. I am the first one to say building PowerPoints is my least favorite activity. I thought the output was really good for a version 1.0.

After I finished, I basically did the same thing you do in OpenAI or Gemini.

Katie Robbert – 08:19

I told Claude to take the contents of the conversation and turn it into a skill I can reuse so I don’t have to keep prompting on how I want a presentation deck built. It did exactly that. If you have seen our previous episodes that deep dive on Claude Desktop—which you can find on our So What playlist on YouTube—it goes through its own steps. It comes up with its own to-do list, finds its own resources, and builds the skill.

Katie Robbert – 09:03

I asked if we could take it a step further and make it a plugin so I could share it with the rest of the team. It used a plugin builder skill, took the information, and built the necessary files. I could download the files, put them in Drive, and Chris could then download them for his own Claude Desktop instance. Now he can build the same decks I am building. It sounds convoluted, but it was actually very straightforward.

Christopher Penn – 09:39

I actually used it this morning for our upcoming GEO 101 for Marketers, which will be available tomorrow in the Trust Insights Academy for $99. It worked perfectly and came up with a terrific framework. I put the slides I wanted to keep in, and I was done in 15 minutes. I took a slide deck that was janky and rough around the edges and turned it into something polished.

Let’s look at what some of those ingredients might be. The first thing would be the prompts, like our Strategic Blind Spot Analysis. It’s just the CASINO Framework saying, “Here is the context, audience, scope, and intent.”

Christopher Penn – 10:25

I turned those into a plugin inside of Claude Cowork. That plugin comes with three different skills: the Gap Analysis skill, the ICP Builder skill, and the Strategic Profile skill. Katie, I know you’re going to ask—this is available in our internal Google Drive so you can install it.

Katie Robbert – 10:51

I will be installing it post-haste.

Christopher Penn – 10:53

What these came up with was a first round of outputs for those three documents. It spit out a profile, an ICP, and a gap analysis. John, I won’t show the rest of the gap analysis, but what did you see in it?

John Wall – 11:25

It was like I was in full panic mode. There were basic strategic gaps it found where I knew we weren’t doing a good enough job. But on the other side, there was stuff it pulled from the internet that I knew wasn’t true.

An easy example is it said, “You need to be doing XYZ because this person makes $1.3 million annually doing this.” I know for a fact that guy is looking for returnable cans on the weekend. There is no way he is making $1.5 million. I was worried we would get on the show and I would just be complaining about hallucinations. You guys walked me back off the ledge and explained that this stage is part of tuning and the process. We were able to grab the transcript and use that to train the model again to correct that bizarre information.

Katie Robbert – 12:46

Yeah.

John Wall – 12:46

One other hilarious story was that I tried to come up with a phony name to talk about people, but I realized people might still think it was a specific person. We have to come up with something very neutral. It was a process of figuring out what is reality versus what is stuff based on the internet—where everybody is lying.

Christopher Penn – 13:10

Exactly. One of the tools we used for that was our Fact Check skill. Katie, this is a bit of a surprise for you, but we’re going to give this away. You can get the Fact Check skill at TrustInsights.ai/factcheckskill. It will bring you to our Thinkific academy where you can download it for free. However, I am absolutely going to try to sell you our courses after you download it.

Katie Robbert – 13:47

Fair enough.

Christopher Penn – 13:51

What the Fact Check skill does is take the transcript and internally provided documents. It reconciles and checks every claim. You will see outputs saying a claim is completely ungrounded, partially grounded, or provides a URL that corroborates it. It’s an incredibly powerful tool. You can use it in Claude Code, Claude Cowork, and Google Anti Gravity.

John Wall – 14:33

There was one thing I wanted to ask you about regarding what is fact and what is not. For people who don’t know, 19 years ago we would actually go to Dunkin’ Donuts to record. We stopped doing that long before the pandemic. The dirty secret is we haven’t been in the donut store in over 12 years. Yet, I’ve never gone on record saying we stopped because I love that part of the legend. Our website still says “Recorded in a coffee shop.”

John Wall – 15:10

The AI comes up and says, “You need to bring in cameras next time you go to the coffee shop.” I have to say, “Well, no, we don’t go to the coffee shop.” How do we handle things that are “true” on the internet even though they don’t match the outside world?

Christopher Penn – 15:25

The Fact Check skill has a rule that user-provided data takes precedence over what web search tools find. We gave it the transcript where you literally said, “Dude, we haven’t been in the coffee shop since 2013.” The Fact Check skill identifies you as the subject matter expert on Marketing Over Coffee. If you say it, what you say takes first place, even if the internet contradicts you.

John Wall – 15:58

Right. So you have to have a subject matter expert fact-check it, otherwise you’re stuck with what the internet thinks is true.

Katie Robbert – 16:08

John, while I don’t love that we caused you a panic attack, you are highlighting what we always talk about. When we say “human in the loop,” this is what we mean. You have to have a human at critical points of the process, even with agentic AI that is supposed to run itself.

Katie Robbert – 16:36

You don’t want to cut your SMEs out of the conversation and assume the AI output is good. Our theme this week is: Don’t outsource your competitive advantage to AI. Your competitive advantage is your SMEs and your critical thinking.

John Wall – 17:12

Well, that’s just the normal underlying MOC panic. It’s always there.

Katie Robbert – 17:17

Then I can’t take credit for that.

Christopher Penn – 17:20

This is an example of what the Fact Check skill spits out. It found 27 errors across those three documents: two critical errors, seven high-level, and nine medium and low priority errors. We can then tell the tool to correct those documents and make the updates. You need to build the data foundation first.

Christopher Penn – 18:08

We haven’t even used Claude Cowork yet because these parts need to be in place. If you don’t do this work, you’re going to see some spectacular explosions. Now we have repaired research and our tooling. It is starting to sound like the 5P process.

We need two more things: requirements and a recipe. For an agentic task like this, we are going to use the 5P Framework as the project plan.

Christopher Penn – 19:02

Purpose: What does this company need to do to achieve its marketing goals? People: Here is who will be doing the work—some AI and some stakeholders. Process: Phase one is strategic background data, phase two is strategy development, and phase three is tactical development. Platform: Claude Cowork will be doing a lot of this. Performance: How do we know it’s done?

Because we took the time to get the data right, we can skip phase one. This is what we are calling the 5R Framework as of Monday.

Christopher Penn – 19:47

You have to have Research, Requirements, Rules about what the tool can and can’t do, a Recipe, and a Review of everything in the box.

Katie Robbert – 20:02

I don’t know if we’re going to keep that name, but we’ll figure it out.

Christopher Penn – 20:08

I’m going to save this as “Recipe MOC Review” and bring it into Claude Cowork. I’ll give it the three final fact-checked documents and the review instructions. I am giving it a folder to work in so it doesn’t just access the hard drive, which is a recipe for disaster. I’ll tell it phase one is complete and to begin at phase two. It’s now going to execute this flow to build the strategy, tactics, and work plan for Marketing Over Coffee.

While it does that, I see a comment from Brian. He wants to know if the next evolution of GEO is AO—Agent Optimization. As audiences use personal AI agents to search for us, we’ll need to optimize content so those agents deliver it.

Christopher Penn – 21:42

Brian, I’m going to slap you for yet another acronym.

Katie Robbert – 21:50

AO, AEO—honestly, we need to leave the acronyms alone. Let’s just stop.

Christopher Penn – 21:58

The answer is yes. However, it shouldn’t depart drastically from what you’re already doing because AI agents are already consuming your stuff. It’s just a question of the formats you are delivering it in. I was talking about this in the new webinar we’re releasing tomorrow.

You should be using tools like our free Trust Insights AI View. That tells you exactly what is on your website and how an AI agent sees it. You might want to consider delivering your site with an option for Markdown.

Christopher Penn – 22:50

If you use Cloudflare on a paid plan, it gives you a “Markdown for agents” switch that automatically converts HTML. That is okay, but you’re better off building this into your own site so you have total control.

Christopher Penn – 23:45

I rebuilt this into my WordPress theme over the holidays. I have an AI tools page that includes a Markdown version. I have a gigantic block at the end of the text that influences AI agents when they consume it. Katie had an amusing run-in with one of my prompts recently that recommended pineapple on pizza because it was embedded in my email.

Katie Robbert – 24:21

Google Gemini in our Gmail summarizes email chains. I was looking at a series of emails and the summarization listed an action item for me: “Chris recommends that you put pineapple on your pizza.” I wondered what was happening in my inbox. Chris puts “Easter eggs” for AI bots that pull data.

Katie Robbert – 25:14

It’s like using a fake name for a subscription to see who is selling your data. He puts it in white text at the bottom of an email. Humans can’t see it, but machines read the text regardless of color. I was confused why he was recommending pineapple on pizza in the middle of a technical discussion with clients.

Christopher Penn – 25:42

The version of Gemini that does summarization in Google Docs is the Flash version. It is a “non-thinking” model, so it is relatively fast and inexpensive, but it can be hijacked through prompt injection. I am injecting my love of pineapple on pizza to everyone who gets my emails.

John Wall – 26:21

From what I’ve seen, your activities fall under AO. What would be specific to agents that isn’t already part of optimizing for LLMs?

Christopher Penn – 26:41

It’s not significantly different. You might have more prompt injections to direct agents. Internal linking is also huge. If you don’t have good internal linking, models might grab a single episode but miss related content. You should revisit your internal linking strategy so that if an agent pulls an episode with Seth Godin, the other nine episodes with him are linked and provide additional context.

Christopher Penn – 27:30

The biggest thing of all is making your site lightweight. Google announced that they will only crawl the first two megabytes of data from your site. That includes stylesheets and images. If your site is top-heavy with extra junk, Google will stop crawling before it ever gets to the content.

Christopher Penn – 28:08

Check your source code to see how many scripts are running. If a bot comes to your site, how long does it have to wait for actual content? You may need to slim down your WordPress theme or CMS.

Katie Robbert – 28:47

If you want an audit for yourself, contact John Wall at TrustInsights.ai/contact.

Christopher Penn – 28:55

Claude has gone through the instructions and come up with the Marketing Over Coffee revised strategy. It says the commercial opportunity is larger than ever and the US podcast advertising market is big money. There is an opportunity for AI positioning.

One of the big three things it says repeatedly is that we have to start putting show episodes on YouTube.

Christopher Penn – 29:49

YouTube is the largest podcast engine on the planet for discovery. It also suggests launching content atomization immediately. It recommends using tools like Opus Clips or Descript to put 30-second snippets on LinkedIn. It also suggests doing more with the website transcripts and positioning documents.

Katie Robbert – 30:44

Marketing Over Coffee is not a video-based podcast. You have 19 years of podcasts without associated video. Suddenly moving people to YouTube feels a little jarring. If I were responsible for this, I’d be worried about editing both audio and video. Is it enough to just post 30-second shorts for discovery, or do we need a whole channel with full episodes?

Katie Robbert – 31:35

I feel like those are two different things.

Christopher Penn – 31:38

The general best practice is to have both full episodes and promotional shorts because they reach different audiences.

Katie Robbert – 31:52

I’m talking about practicality and bandwidth. If forced to choose, would shorts be enough? You can pull a snippet from a recording on StreamYard without having to edit the entire video.

Christopher Penn – 32:31

Anything is better than nothing at this point because it’s a missed opportunity. But the best practice remains the two-channel approach.

Katie Robbert – 32:51

I understand the best practice, but realistically, what do I have time to execute against?

Christopher Penn – 33:18

One option for older podcasts is to programmatically place the MP3 against a video loop. You don’t have to record new video for the past 800 episodes.

Christopher Penn – 33:40

You can create audiograms with looped video. I was playing with a piece of Python code that takes a ten-second loop of an espresso machine and extends it underneath the audio track.

John Wall – 34:03

The tools are so close. I have an audiogram running, but I haven’t cracked the workflow to get it to automatically post. It’s frustrating that this isn’t already baked into my hosting provider.

John Wall – 34:43

The huge win would be clicking a button in StreamYard to create a 30-second clip and have it automatically post. The tools are moving so fast that I can see this being totally different within a year.

Christopher Penn – 35:09

In Claude Cowork, you can use Claude Code to build that. YouTube has an API. Since your site has a direct link to the MP3, a piece of code could check for the latest post, grab the MP3, make the audiogram, and post it via the API to YouTube. You wouldn’t have to edit one more thing.

John Wall – 36:17

We’re inches away.

Christopher Penn – 36:26

Claude Cowork has finished all three pieces. You might ask how this is different from how we always do it. Our mistake in this process was not starting with the SME interview. If you start with a 30-minute conversation with the subject matter expert, that transcript becomes the first data piece.

Christopher Penn – 37:14

Then Claude Cowork can follow the recipe to do research on pieces it doesn’t have and execute the rest of the strategy autonomously. We started with the AI research first instead of the human expert.

Katie Robbert – 37:48

Growth is learning from mistakes. For us, it was a great learning point on how to make these automated systems even more effective. Next time, we’ll start with that SME interview and give the transcript to the plugin.

Christopher Penn – 38:24

Exactly. That is what makes it agentic. Once you have the infrastructure and the 5Rs, the agents can do the work for us over and over again.

Katie Robbert – 38:54

What do you think, John? Was that as painful as you thought?

John Wall – 38:57

I still have to edit today’s show.

Katie Robbert – 39:02

There is the human in the loop.

John Wall – 39:04

Exactly. This has been a great exercise. We are crossing these barriers at a regular pace now.

Katie Robbert – 39:38

I used the plugin I built for Claude Desktop this morning to redo my speaker kit. It hadn’t been updated since 2018. Back then, I had never been on stage, so it was a pretty lame kit. Fast forward eight years and all the talks, workshops, and virtual sessions I’ve done are now relevant.

Katie Robbert – 40:32

I have a lot of information on the website and in knowledge blocks. I’ve found Claude Desktop helpful because you can give it a massive amount of information from different places. The web version sometimes struggles with that volume. I told it I hate making PowerPoints and it offered to do that for me. It turned a sad, five-slide deck into an almost 20-slide deck that I’m really proud of.

Katie Robbert – 41:56

It went from a very basic deck to something truly useful. It suggested things I hadn’t thought to include, like bio versions for event programs and current speaking topics.

Katie Robbert – 42:53

It used the 5P Framework and session abstracts. It identified my authority as a speaker through my developed courses and select speaking engagements. It even included session ratings and feedback.

Katie Robbert – 44:07

It included audience metrics and placeholders for speaking reels. I set this up in Claude Desktop this morning and walked away. This is what the process looks like. Using the Trust Insights 5P Framework sets you up to get really cool outputs.

Christopher Penn – 45:27

Brian was asking which version of Claude you used.

Katie Robbert – 45:30

It was Claude Cowork Opus 4.6.

Christopher Penn – 45:33

In today’s smartest model, GLM 5, Opus 4.6, and Codex 5.3 are well-tuned for agentic use. Opus in particular is starting to exhibit judgment—it will tell you if you forgot something or offer suggestions.

John Wall – 46:22

It’s interesting for it to find outliers and say something about them.

Christopher Penn – 46:31

It is starting to express judgment. To be clear, these are not sentient or self-aware. They have no actual autonomy and are not alive. But the capabilities to express judgment are there. Don’t be surprised when they do, and be ready to push back.

The 5P Framework is the perfect agentic prompting framework. We also have the 5Rs: Research, Requirements, Rules, Recipe, and Review. These are the five ingredients you need for an agentic project.

Christopher Penn – 47:40

If you don’t have those, you run the risk of sloppy outputs or going off the rails. Any final words?

John Wall – 47:58

Start playing with these tools now. Do not wait. The bus is moving.

Christopher Penn – 48:05

Our friend Tom Webster has been using Claude Code to rewrite his data processing infrastructure. We are all on board this train and going as fast as we can. Thanks for tuning in. Be sure to subscribe and check out the Trust Insights podcast and our weekly email newsletter. Join our free Analytics for Marketers Slack group at TrustInsights.ai/analyticsformarketers. See you next time.