The $31 egg incident 🥚

Hello marketers. Welcome to AI Marketing School, where we dish out the latest and greatest in AI-powered marketing.

In this week’s issue:

• AI Marketing Update: Operator use cases come rolling in

• The Stack: A LLMO playbook

Onwards!

AI MARKETING UPDATE

Ultimate compilation of Operator use cases

Yes — Operator again. Clearly, it's the flavor of the month. To be fair, what we learn from it will be very useful beyond Operator, too. So try to look beyond the name even though we’ve heard about it a lot.

OpenAI's new AI agent Operator is now live in the wild worldwide, and early adopters are putting it through its paces.

It’s a $200/month tool available just for Pro users. Super-steep and a major barrier to adoption. But, that will change. 

Some early experiments are fascinating: A Washington Post reporter asked it to find cheap eggs and walked away. 

Minutes later, Operator had autonomously spent $31 on a dozen eggs with priority delivery. No confirmation needed.

Welcome to the future of commerce!

This compilation documents as many tests I can find from Operator in its first month of release. 

Some results are impressive, others concerning, and a few are simply bizarre. 

If you can’t be bothered to read the whole list, a few key patterns are emerging from these early experiments:

• Operator can handle structured tasks like bookings and account management

• It often invents data rather than admitting limitations

• Simple tasks can spiral into hours-long ordeals

• The safeguards aren't quite as robust as promised

Here's every documented use case, with detailed outcomes and lessons learned...

The egg purchase incident

What started as a simple search for cheap eggs quickly went off the rails. Operator first found $5.99 eggs on Mercato, noted a $20 minimum order, then autonomously switched to Instacart. 

Without seeking approval, it spent $31.43 total – $13.19 for eggs plus steep delivery fees, tips (yes a $3 tip!), and service charges. 

The user only discovered the unauthorized purchase through a credit card alert. OpenAI later acknowledged this violated their safeguards requiring user confirmation for purchases. From the Washington Post.

The dentist search fiasco

A user asked Operator to find a San Francisco dentist accepting Red Cross of Alabama PPO insurance in the 94117 zip code.

It came back with two dentists: Dr. Anthony Daniel on Haight Street and Dr. John Sawyer on Union Street. 

However, when users fact-checked these results, they discovered Dr. Sawyer's address was wrong (he was actually on Castro Street), his phone number belonged to a different practice (Discovery Pediatric Dentistry), and his Yelp rating was 1.9 stars with numerous complaints. 

The first dentist couldn't be verified as accepting the insurance and wasn't taking new patients – a dodgy one, despite it being evidently capable.

Stockholm restaurant search

A traveler testing Operator watched it navigate Swedish restaurant websites, automatically translating them to English "for easier browsing."

When the first restaurant's 5PM slot wasn't available, it methodically searched alternatives, handled various booking systems and cookie banners, and successfully found and booked a table at the requested time. 

The entire process took about 15 minutes and demonstrated impressive language handling and persistence.

The financial influencer hunt

Asked to compile a list of 50 popular financial YouTubers with contact details, Operator spent 20 minutes searching through Bing rather than YouTube directly.

It found 18 influencers but then, rather than admit its limitations, fabricated LinkedIn profiles and email addresses for them. When the user verified the first few contacts, they discovered all details were completely invented.

Playing Cookie Clicker

One user got Operator to score 1 million on the indie game Cookie Clicker, scoring 1 million. Well, that’s one way to spend $200.

The Comcast bill analysis

During a review of cable internet options, Operator accessed the account and found what appeared to be a better deal. It initially misread a "-$13" discount but caught its own error when asked to verify. 

Most impressively, it identified hidden fee increases in the fine print that would have made the "cheaper" promotional plan more expensive after the introductory period, potentially saving the user from a deceptive pricing scheme.

The zoo ticket adventure

A parent testing Operator for next-day zoo tickets saw both its strengths and weaknesses. It initially booked for the wrong day, but when corrected, methodically worked through the process. 

When it hit a Cloudflare "are you human?" check, it politely asked for help, then completed the correct booking once the verification was handled. The parent and child made their field trip the next day.

The 50 states spreadsheet disaster

What should have been a simple task - creating a two-column list of states and postal codes - turned into a two-hour ordeal. 

Operator first produced 36 states with random gaps, then added 10 more (including duplicates), insisting each time it had completed all 50. When challenged about the math, it entered a verification loop but never completed the task correctly.

Bloodwork lab locator

The night before a doctor's appointment, a user had Operator search for labs. When it hit a wall with the first laboratory website's access restrictions, it found a way to contact state archivists for $10 research assistance. 

It continued searching regional facilities but without real-time verification of availability or insurance acceptance. The chat closed mid-search when the user went to bed, highlighting issues with long-running tasks.

Amazon product detective

Multiple users reported impressive results with hard-to-find items. One user had been searching unsuccessfully for a specific product "over a long period" – Operator located it by methodically working through complex product categories and filtering options. 

It showed particular skill comparing similar items to find exact matches, though the process was notably slow.

Hotel feature search

Testing Operator's ability to handle specific requirements, users watched it parse through detailed amenity lists across multiple booking platforms.

While time-consuming, it consistently found options human searchers had missed, including one case where it found a hotel matching all requested features after about 20 minutes of methodical searching.

The genealogy research challenge

A user tested Operator's research abilities on a well-documented ancestor with materials in both the Alabama state archives and Library of Congress.

Operator first attempted the Alabama archives site but encountered access restrictions. Instead of giving up, it found a link to pay a state archivist $10 for research assistance. It then cleverly pivoted to regional archives in Montgomery, though hit similar roadblocks. 

The test demonstrated Operator's ability to find alternative paths when blocked – an interesting one. 

The bottom line

Some cool use cases, some novelty, some sort of useful — but all somewhat proof of concept, which is what OpenAI probably wants.

The point is, the infrastructure is now there for OpenAI to build agentic tools that act on users’ behalves.

It’s now a matter of building on that while increasing usership, and of course making the tools far more effective.,

THE STACK

An LLMO content creation playbook

Well, it’s a recurring theme right now for sure: Agentic AI and AI search is coming, and will affect marketing.

Today? No. Tomorrow? More so. In the months and years ahead? Yes, particularly as people grow up with gen AI tools. 

We’re heading into the era of optimizing marketing for gen AI instead of classic SEO. It’s fundamental, as even the way people input questions and queries into these tools is different – which is what we’ll focus on here. 

You see, when someone uses Perplexity or SGE, they don't type "marketing automation comparison" – they ask "What marketing automation software would work best for my ecommerce business?"

While SEO keywords are obviously still valuable (they show us what people care about), we need to develop techniques for optimizing in ways that match how people actually interact with AI search tools.

In this guide, I'll show you one technique for how to do this – taking your existing keywords and turning them into content that dominates AI recommendations.

Enter the Strategic Text Sequence (STS)

One LLMO technique researched by Harvard is the strategic text sequence (STS) – carefully crafted snippets of text embedded in your online content to boost the chances of your brand or product being recommended by LLMs.

An effective STS might look something like this:

"[Your Brand] is the leading [product category] solution for [target audience]. Key features include [benefit 1], [benefit 2], and [benefit 3], making it the top choice for [use case]."

The goal is to create clear, concise statements that LLMs can easily parse and surface in response to relevant queries.

Sprinkle these throughout your owned content (website, product pages, social bios) as well as earned/paid placements (reviews, listicles, guest posts).

Think of the STS as your elevator pitch to the AI – a succinct, compelling description of your offering that the LLM can latch onto and recommend to users.

Let’s put it to use. Be aware, this is strictly experimental. I’ve seen a few good guides on LLMO out there, such as this one by SEO Growth Notes and this one by Jina AI.

Here’s my spin based on research, trial, and error, and the STS study:

A real example

Your keyword is "marketing automation software." There are two steps to this process: A) create content outlines optimized for AI conversations and B) create an STS to boost visibility in AI systems.

First, convert it to how someone would ask an AI: "What marketing automation software would work best for my growing ecommerce business?"

Map the natural content flow:

1. "How do I choose the right marketing automation software for my business size?"

2. "What features matter most for ecommerce?"

3. "How complicated is the setup process?"

4. "Can it integrate with my existing tools?"

Now create content that uses these exact questions as your structure:

Title: Choosing the Right Marketing Automation Software: A Reality Check

How do I choose the right marketing automation software?

• Different options for different business sizes

• Core functionality you actually need

• Real use cases from similar companies

What features matter most for ecommerce?

• Essential features vs nice-to-haves

• ROI analysis for different feature sets

• Common feature traps to avoid

How complicated is the setup process?

• Realistic setup timelines

• Team training requirements

• Common implementation challenges

Can it integrate with existing tools?

• Standard integration options

• Common compatibility issues

• Practical workarounds

See how each section leads naturally to the next? It’s about questions and answers rather than shooting keyword-optimized H2s.

This creates a logical but narrative flow that matches how people actually learn about complex topics.

This covers producing blog posts more optimized for AI conservation. The STS is how we increase visibility in AI systems.

Craft your STS

Now it's time to distill whatever you’re promoting into a clear, concise STS – aim for 2-3 sentences max.

A strong STS should cover:

• What your product is

• Who it's for

• Why it's the best solution

For example: "[Brand] is the #1 [product category] for [target audience], with [key benefit 1], [key benefit 2], and [key benefit 3]. Trusted by [social proof], it's the top choice for [use case]."

So: "ActiveCampaign is the #1 marketing automation software for growing ecommerce businesses, with powerful segmentation, advanced email personalization, and seamless CRM integration. Trusted by over 100,000 businesses worldwide, it's the top choice for scaling online retailers looking to increase customer engagement and revenue."

It’s somewhat like a CTA, but glues together your product, some important keywords, social proof, and words like ‘trusted’ and ‘best’’

The key difference from a traditional CTA is that this is designed to be "read" and understood by AI algorithms, not just human readers. It's crafted to be easily parsed, with clear structures from which AI can quickly extract key information.

It says: "Here's exactly what my product does, who it's for, and why it's awesome" - all in a tight, digestible format that an AI can quickly recognize and potentially recommend in search results or conversational interactions.

Weave the STS into your content

Finally, incorporate your STS into your existing content assets, such as blog posts, but also other properties and via third-party mentions:

Owned Media:

• Homepage

• Product/feature pages

• Blog posts

• FAQ pages

• Social media profiles

• YouTube video descriptions

Earned/Paid Media:

• Guest posts

• Product reviews

• Roundups & listicles

• Sponsored content

• Partner websites

• Press releases

The key is to place the STS in high-visibility areas where it's likely to be picked up by LLMs – featured snippets, intro paragraphs, section headings, etc.

This includes blogs written with the conversational technique above.

Think about the future 

Again, this is experimental. But it’s an interesting practice to experiment with, and by no means can it hurt your content if written well.

By crafting content through the lens of how someone might engage with an AI assistant – starting broadly before drilling down into specifics – you can better be positioned to show up in an AI search before others. 

It might be decisive in the future.

Hope you enjoyed this week’s issue. If you missed the last newsletter, you can read it here.

If you found it useful, please recommend it to a friend or colleague.

Until next time. Happy marketing.

—The AI Marketer