AI Agents Can’t Find Your Business (Here’s the Fix) | Practical AI Ep 29

Practical AI: Episode 29

Claude Went to War. AI Can’t Find Your Business. Here’s What That Means.

Watch on YouTube

Published: February 20, 2026

TL;DR

Anthropic went from industry darling to crisis mode in one week after revelations that Claude was used in a Pentagon military operation that killed 83 people, while simultaneously losing OpenClaw’s founder to OpenAI and facing developer backlash over tightened usage limits. Meanwhile, a real-world experiment with AI agents revealed that 70% of local businesses are invisible to AI shopping assistants, signaling a massive shift in how consumers will discover and choose service providers.

Table of Contents

About This Show

Practical AI is a weekly live show (Fridays 11am CT) hosted by Olga Pechnenko and Chris Pearson that cuts through AI hype to deliver news, trends, and hands-on tips for builders and founders. Unlike technical AI podcasts, Practical AI focuses on business applications and ROI—what actually works, what’s hype, and what you can implement Monday morning.

What You’ll Gain

  • Understand Anthropic’s ethical crossroads as their $200M Palantir contract collides with their safety-first principles after Claude was used in a military operation. This case study reveals the impossible tradeoffs AI companies face when government contracts meet founding values.
  • Learn why 70% of local businesses are invisible to AI agents through a real-world experiment where an AI assistant tried to shop for car brake services. The results expose a massive blind spot that will determine which businesses thrive in the agent economy.
  • Discover the five specific optimizations that make your business website AI-agent ready, from plain-text contact information to transparent pricing—changes that take minutes but determine whether AI assistants can recommend you.
  • See how Lovable grew from zero to $300M ARR in 14 months by renaming a category rather than fixing a product, offering a masterclass in positioning that any founder can apply.
  • Gain insight into the OpenClaw crisis and what it means for developers who built workflows around Anthropic’s ecosystem, plus how OpenAI capitalized on the moment to pull ahead in the AI wars.

Biggest Takeaway to Implement: Run your business website through any AI chatbot and ask how GEO-optimized it is. If an AI agent can’t find your email, understand your pricing, or contact you without human intervention, you’re already losing customers to competitors who made themselves agent-ready.

PageMotor and Practical AI Updates

Free, informative, and FUN!

Frequently Asked Questions

Was Claude AI actually used in a military strike that killed people?

Yes. Anthropic’s Claude was used via their $200 million Palantir contract to coordinate drone strikes in a Venezuela military operation that resulted in 83 deaths. Anthropic reached out to Palantir after the fact to ask if their model was used, raising questions about their safety oversight. Read more below.

What happened with OpenClaw and why are developers angry?

OpenClaw founder Peter Steinberger joined OpenAI, moving the project to a foundation that developers believe will be heavily influenced by OpenAI’s interests. Simultaneously, Anthropic tightened usage limits without transparency, creating a double betrayal for developers who had built workflows around the Claude ecosystem. Read more below.

Why can’t AI agents find most local businesses?

A real experiment with 10 local auto shops revealed that 70% were unreachable by AI agents due to phone-only policies, JavaScript-hidden contact info, CAPTCHAs, and missing pricing. Businesses built on Wix were particularly problematic because contact details render via JavaScript that AI agents cannot read. Read more below.

What is GEO and why does it matter for my business?

GEO (Generative Engine Optimization) is the practice of making your website readable and useful to AI agents and AI-powered search. As consumers increasingly use AI assistants to shop and research, businesses that aren’t GEO-ready will become invisible to this new economy. Read more below.

How did Lovable grow so fast?

Lovable hit $1M ARR in 8 days and $300M ARR in 14 months by renaming their product from “GPT Engineer” (intimidating) to “Lovable” (approachable) and coining the term “vibe coding.” The product stayed the same—only the positioning changed. Read more below.

What is OpenAI’s Lockdown Mode?

A new enterprise security feature that prevents data exfiltration via prompt injection. Malicious actors were embedding invisible commands in PDFs that could instruct AI to steal company data. Lockdown mode disables external connections to prevent this attack vector. Read more below.

Practical AI: Claude Went to War, AI Can’t Find Your Business

Key Definitions

What is GEO (Generative Engine Optimization)?

The practice of optimizing websites and content for AI-powered search engines and agents rather than traditional SEO. GEO ensures AI assistants can read, understand, and act on your business information—including pricing, services, and contact details—without requiring human intervention.

What is Prompt Injection?

A security attack where malicious instructions are hidden in documents (often in white text invisible to humans) that AI systems read and execute. Attackers use this to exfiltrate sensitive data or manipulate AI behavior when companies upload documents to AI tools.

What is an Agentic AI?

AI systems that can take autonomous action on behalf of users—browsing websites, sending emails, making purchases, and completing multi-step tasks without constant human direction. Unlike chatbots that only respond to prompts, agentic AI proactively executes workflows.

What is Token Routing?

The practice of directing AI requests to different models based on complexity and cost. Companies route simple queries to cheaper models and complex reasoning tasks to expensive ones, optimizing their AI infrastructure costs while maintaining quality.

Quotable Moments

“Last Friday we were so high on Anthropic, they could do no wrong. It was the honeymoon stage. And then… record scratch.” — Chris Pearson on Anthropic’s dramatic fall from grace

“Out of 10 websites, only 3 were able to work with the agent. The other 7 are not getting our business. Period.” — Olga Pechnenko on real-world AI agent shopping results

“All these cute little tricks we used to do—hiding pricing, call for quote—all out the window. AIs need information. If you don’t give it to them, they move on.” — Chris Pearson on the death of traditional marketing tactics

“If this Russian girl who is super non-technical is already using her AI agent to do this stuff, imagine what’s happening with people way ahead of me.” — Olga Pechnenko on the coming wave of AI-assisted consumers

0:52 Anthropic’s Rapid Fall: From Darling to Drama

The Palantir Contract

Anthropic holds a $200 million contract with Palantir for government work. After a military operation in Venezuela resulted in 83 deaths, Anthropic asked Palantir if Claude was used—raising questions about why they didn’t know beforehand.

The company founded on “safer AI” principles now faces an impossible choice. Defense Secretary Pete Hegseth reportedly stated that if AI companies won’t provide military access, the government doesn’t want to work with them. OpenAI and xAI are already positioning themselves as alternatives willing to take contracts Anthropic might abandon.

The irony is stark: the company that talks about safety the most became the first major AI provider publicly linked to military casualties. Whether Claude was directly responsible for targeting decisions or used for coordination, the “safety-first” brand now carries an asterisk.

8:54 GPT-5.2 Breakthrough: AI Solves “Impossible” Physics

Key Stat: 12 Hours vs. 15 Years

GPT-5.2 independently proved a physics formula in 12 hours that humans had deemed unsolvable for 15 years. The academic consensus held that a particular value was zero; the AI triangulated from available data and determined it was non-zero.

This marks a shift from AI as a “guessing mechanism” to AI as genuine reasoning system. The model wasn’t just pattern-matching—it synthesized information across domains to reach a novel conclusion that humans couldn’t. For researchers and builders, this signals that AI is moving from assistant to collaborator in ways that will reshape scientific discovery.

10:50 Anthropic Sonnet 4.6: 1 Million Token Context

Context Window Expansion

Sonnet 4.6 offers 1 million tokens of context—equivalent to roughly 10 copies of The Great Gatsby. Developers who have been “chunking” large tasks to fit within smaller context windows can now process entire codebases in a single reasoning window.

12:16 OpenClaw Rug Pull: Developer Backlash Intensifies

The Double Whammy

Peter Steinberger joined OpenAI, moving OpenClaw to a “foundation” that developers believe will be controlled by OpenAI interests. Simultaneously, Anthropic tightened usage limits without transparency—a black box that subscription users cannot monitor.

The developer community that championed Anthropic over OpenAI feels betrayed. They built workflows around OpenClaw and Claude, and are now watching it move to their least-favorite provider while facing unexplained usage restrictions. Some developers are already migrating to Codex as an alternative.

15:52 OpenAI Eyes $850B Valuation

The Funding Numbers

OpenAI is closing a $100 billion raise at an $850 billion valuation, backed by Amazon, SoftBank, Nvidia, and Microsoft. This comes as Meta launches $65 million in super PACs to back AI-friendly lawmakers.

The valuation raises eyebrows given OpenAI’s revenue, but in a week where their competitor stumbled badly, the timing couldn’t be better. Meta’s lobbying push aims to preempt California’s AI restrictions by getting federal law that supersedes state-level regulation.

18:35 OpenAI Lockdown Mode: Blocking Data Exfiltration

The Prompt Injection Threat

OpenAI’s new Lockdown Mode addresses a vulnerability where attackers embed invisible commands in PDFs using white text. When companies upload these documents to AI, the hidden prompts can instruct the AI to exfiltrate sensitive data to external addresses.

An HR manager uploading 100 resumes for screening could unknowingly trigger hidden instructions that steal employee Social Security numbers. Lockdown Mode disables external connections, essentially creating a “dumb AI” that can’t be manipulated to leak information. The tradeoff: enterprise users sacrifice AI capabilities for security, potentially giving advantage to startups who don’t face the same compliance requirements.

22:28 Meta Ray-Ban “NameTag”: Real-Time Facial Recognition

Meta is planning a “NameTag” feature for Ray-Ban smart glasses that identifies strangers in real-time—pulling up names, LinkedIn profiles, and other information as you look at people. The privacy implications are staggering: if this ships, anonymity in public spaces effectively ends.

24:39 Sam Altman Calls for Global AI Regulation

At the India AI Summit, Sam Altman called for urgent IAEA-style international AI regulation—while simultaneously closing a $100 billion funding round. The classic tech titan playbook: establish dominance, then advocate for regulations that hamstring competitors.

26:29 Real Experiment: AI Agent Shops for Car Brakes

The Knox Experiment

An AI agent named Knox was tasked with finding brake shops and Tesla repair services. Of 10 local businesses contacted, only 3 (30%) were reachable by the AI agent. The other 7 required phone calls, had broken forms, or hid contact information behind JavaScript.

Olga’s AI assistant Knox has his own email and follows approval workflows. When tasked with finding brake services for Chris’s car, Knox found 8 shops but could only email 2 of them—the others required phone calls or had contact information invisible to AI agents. For Tesla repairs, Knox found independent shops at half the dealership price and caught a double-billing error on an existing quote.

The Wix Problem

One Tesla shop had an email visible to humans but invisible to AI agents because Wix delivers contact information via JavaScript after the HTML loads. AI agents reading raw HTML see nothing. This single technical choice made the business unreachable.

The winning business, Brakes To Go, had pricing on their website and required zero human interaction. They got rated first by the AI agent and are likely getting the business. The other 7 shops don’t even get to compete.

35:44 5 Ways to Make Your Business AI-Agent Ready

The New Hierarchy of Local Search

First place: Pricing on site, zero human interaction required. Second place: Email works, AI can negotiate. Invisible: Phone-only or broken code.

1. Plain-text contact information. Put your email in raw HTML, not hidden in JavaScript widgets or CAPTCHA-protected pages. AI agents read HTML; they can’t execute JavaScript.

2. Transparent pricing. “Call for quote” worked when humans had patience. AI agents don’t. A price range with an asterisk beats no price at all.

3. Complete Google Business Profile. AI agents trust this structured data. Fill out every field.

4. Real FAQs. Answer specific questions: “Do you fix Audis?” AI agents use this to match your business to user needs.

5. Fast email response. AI agents track latency. Respond in 30 minutes, not 24 hours.

The actionable step: paste your website URL into any AI chatbot and ask how GEO-optimized it is.

51:36 Lovable: Zero to $300M ARR in 14 Months

The Growth Trajectory

Lovable hit $1M ARR in 8 days, $30M by month 4, $100M by month 8, and $300M ARR by month 14. Current valuation: $6.6 billion. Marketing budget: zero. Sales team until $100M: none.

Founder Anton Osika originally named the product “GPT Engineer"—technically accurate but intimidating to non-developers. He did two things: renamed the product “Lovable” and coined the term “vibe coding.” The product itself didn’t change. GPT Engineer was “the right tech in the wrong costume.”

The Margin Challenge

Lovable operates at 20% gross margins—far below the 70%+ SaaS standard. As a wrapper around OpenAI, Anthropic, and Google models, their profitability is hostage to inference costs they don’t control. Their path to 70% margins requires cheaper models, proprietary routing, and enterprise contracts.

The company must “refine product market fit every 3 months” to stay ahead of Cursor, Bolt, and Google. Speed is their only moat. But for founders, the playbook is clear: name the category, don’t fix the product.

1:05:07 AI Funding This Week: $2.4B Across 54 Companies

Weekly Funding Breakdown

Total AI funding: $2.4 billion across 54 companies. Down from last week’s $30B Anthropic-driven spike, but the 12-week average holds at ~$9B/week. AI represented 38% of all global funding this week.

Biggest rounds: World Labs ($1B) for spatial AI/3D robotics, Nesa ($600M) for AI-native enterprise applications, and Braintrust ($150M) for AI observability. The “agentic AI” label is becoming a funding magnet. Geographic distribution: US led with 38%, Asia-Pacific second with $640M.

1:15:02 PageMotor Demo: AI Builds Live Websites in Minutes

Chris demonstrated the PageMotor workflow: describe a website to Claude, run a conversion skill, and import directly to a production CMS. What took 6 hours in January now takes minutes. Four live websites have been built by beta users in under two weeks—not prototypes, but production sites with SEO optimization and structured HTML that AI agents can read. When tested against GEO criteria, PageMotor sites scored 9/10 without additional optimization. Beta is open at pagemotor.com/beta.

Keep Learning