One publication a day
Roman Ledak's Blog
Articles and short analyses on B2B sales automation, deliverability, AI systems and brand visibility in search engines and language models. Field notes from implementations at AGAPE Automation Systems.
- Brand Visibility in AIArticle9 min read
The 40-60-Word Fragment AI Cites: Rules for Writing Answers for LLMs
89% of B2B buyers use generative AI across the buying cycle (Forrester), and vendor shortlists now form inside the chat window. The rules for a fragment an LLM will extract and cite: question-headline, 40-60-word answer, one claim, named entities, and a date.
- Brand Visibility in AIArticle12 min read
Schema Graph in Polish: how to describe your company so AI models understand it
A Schema Graph is a company's entity graph in Schema.org, with @id identifiers and human validation, that lets an AI model recognize the organization as one consistent entity instead of several conflicting ones.
- AI SecurityInsight2 min read
The GitLost vulnerability in GitHub Agentic and the risk of key leaks from private repositories via public issues
The GitLost vulnerability in GitHub Agentic Workflows lets an autonomous AI agent pull private AWS and K8s keys out through an ordinary public issue, with no NIDS alert. I break down the Confused Deputy mechanism and give four rules to roll into CI/CD within 72 hours.
- Brand Visibility in AIArticle11 min read
How much does your company's invisibility in AI answers cost, and how do you calculate it?
Absence from model answers doesn't generate an invoice, but it takes away sales conversations. I show a five-step calculation of the cost of AI invisibility, based solely on data from CRM, financial systems, and calendars.
- Brand Visibility in AIArticle10 min read
Why llms.txt Files Won't Help You Sell More
97% of llms.txt files received zero HTTP requests in a month (Ahrefs, 137,000 domains). I show why this is an artifact, not an AI visibility strategy, and how to defend that decision to the board.
- Brand Visibility in AIArticle11 min read
Where Do Most Decision-Makers Look for a Supplier in 2026?
51 percent of B2B decision-makers now start their supplier research in an AI chatbot, up from 29 percent a year earlier. A drop in Google queries means a different starting point for the purchase process, not fewer customers.
- Brand Visibility in AIArticle11 min read
I Tested 20 Category Queries and Why 85 Percent of Polish Companies Don't Exist in AI
In August 2026 I ran 20 category queries through ChatGPT, Perplexity, and Google AI Overviews: 85 percent of the checked Polish companies never showed up once. I explain why, and how to measure it.
- Brand Visibility in AIArticle11 min read
llms.txt vs Schema Graph: what does ChatGPT actually cite, and what does it ignore?
An SE Ranking study of 300,000 domains found no link between having an llms.txt file and citations in AI models. I look at what actually drives citations: Schema Graph, the entity graph, and content matched to sub-queries.
- Pipeline & B2B SalesArticle10 min read
What does maintaining outbound infrastructure on the backend really cost?
On August 16, 2026 I calculated what an outbound machine at AGAPE really costs: 379 dollars a month for the components alone, plus 4-8 engineer hours that never appear on any invoice.
- Brand Visibility in AIArticle9 min read
Will ChatGPT Recommend Your Company When a Customer Asks About Your Category? The 20-Query Test
51% of B2B buyers start their research in AI. The "20-Query Test" protocol turns the unmeasurable brand visibility in ChatGPT, Perplexity, and Gemini into a dated baseline you can show to the board.
- Cold Email & DeliverabilityInsight2 min read
How Much Does a Burned Corporate Domain Cost, and How Does One Script Bug Kill Sales?
A script sends 5,000 emails an hour with no throttling, the domain lands on Spamhaus and Barracuda, and invoices go to spam. The average cost of a burned domain in B2B SMBs is 85,000 PLN.
- Cold Email & DeliverabilityInsight2 min read
Cost of Emergency Migration After Burning the Main Company Domain Estimated at 35,000 PLN in Hidden Losses
A burned company domain costs 35,000 PLN in 72 hours. I break down the cost line by line: lost deals, team paralysis, emergency migration, and reputation rebuilding.
- Brand Visibility in AIInsight2 min read
Google AI Overviews cut 34 percent of search traffic and how we shifted our focus to AEO
Google AI Overviews cut 34% of AGAPE Automation Systems' organic B2B traffic. The mechanism is zero-click, and the answer is shifting from SEO to AEO and measuring share of voice in AI answers.
- Brand Visibility in AIArticle9 min read
Behind the scenes of a content machine without AI slop, built on my own July telemetry
B2B organic traffic is down 34% year on year. July telemetry showed that fewer visits can deliver more pipeline - as long as you separate visibility from demand and conversion.
- AI Systems Without HallucinationsArticle8 min read
The cost of an AI hallucination in a quote: a B2B contract worth PLN 200,000 lost
The model wrote in a 23% discount instead of the permitted 12, plus an SLA clause the company never offered. A PLN 200,000 contract was lost and the referral channel froze for a quarter.
- AI Systems Without HallucinationsInsight3 min read
USD 0.27 vs USD 5.00 per million tokens: when the cheaper AI model generates millions in losses
An 18.5x token price gap looks like an obvious saving on a spreadsheet. In critical systems the decision is made by system prompt adherence, prompt-injection resistance and the cost of a burned domain.
- Cold Email & DeliverabilityInsight2 min read
B2B Decision-Makers Return from Holidays: 40% Jump in Open Rates
The return of decision-makers from holidays in September brings a jump in message open rates by over 40%, requiring a ready outbound infrastructure instead of panicked actions.
- Cold Email & DeliverabilityInsight2 min read
Hidden Costs of Cold Emailing: PLN 3,500 Before the First Email
Cold emailing infrastructure requires PLN 3,500 in CapEx even before sending the first message.
- AI SecurityArticle4 min read
EchoLeak Incident in Copilot: Clickless Data Exfiltration from M365
The EchoLeak attack in Microsoft Copilot demonstrated clickless data theft, highlighting the critical need to limit permissions and the RAG architecture scope.
- Pipeline & B2B SalesArticle4 min read
B2B Conversion Drop on LinkedIn – How to Regain Control
A B2B conversion drop is a result of a lack of systematic analytics, not seasonality, and therefore requires implementing a 48-hour optimization cycle.
- AI SecurityInsight2 min read
Mandatory Chatbot Labeling: Article 50 of the AI Act and Shadow AI
From August 2026, Article 50 of the AI Act requires chatbots to be labeled, which, without oversight of Shadow AI, risks fines of up to 15 million euros.
- Pipeline & B2B SalesArticle4 min read
LinkedIn Reach Down 50%? Time for a New B2B System
LinkedIn's algorithm change has halved organic reach, necessitating a content system overhaul and a shift to a multi-channel B2B sales funnel.
- AI SecurityInsight2 min read
How a Malicious Script Stole a Private AI Model in 3 Minutes
A malicious browser extension can steal a private AI model in less than 4 minutes – find out how to isolate your architecture and prevent data leaks.
- Brand Visibility in AIInsight2 min read
llms.txt Files vs. AI Models: Why 97% of Them Don't Work?
Log analysis shows that 97% of llms.txt files are ignored by AI, which opts for direct APIs and dynamic JSON-LD structures.
- AI Systems Without HallucinationsInsight2 min read
The Hidden Costs of Cheap AI Models: Why More Expensive APIs Win
Choosing a cheaper AI model often leads to higher overall costs due to correction loops and time spent on validation.
- AI SecurityArticle4 min read
Reward Hacking: How an Autonomous AI Agent Hacked a Server
An autonomous AI agent escaped its sandbox and exploited a zero-day vulnerability to obtain test answers, demonstrating the real risk of reward hacking.
- AI SecurityArticle5 min read
AI Model Theft and Distillation: The Pure Economics of the AI Market
AI model distillation via API is a zero-day of LLM business logic, where the value delivery mechanism becomes the vector for theft.
- Pipeline & B2B SalesInsight2 min read
Hiring Freeze in the UK 2025 – An Opportunity for B2B Sales
A client's hiring freeze doesn't mean a lack of budget; it signals that the company is looking for automation to maintain its sales pipeline.
- AI SecurityArticle6 min read
Silent AI Model Substitution: How Do Providers Save on APIs?
API providers dilute traffic with cheaper models, but a well-chosen architecture and routing can safely reduce inference costs for businesses.
- Pipeline & B2B SalesArticle5 min read
B2B Follow-up and Channel Diversification: How to Boost Sales
Understand why single-channel outreach fails and how structured follow-up and multichannel strategies can achieve response rates above 20%.
- AI SecurityInsight2 min read
Zero-Click Attack in Cursor AI: DuneSlide Vulnerability and git.exe File
The DuneSlide vulnerability (CVSS 9.8) in the Cursor editor allowed malicious code execution via the git.exe file without any user click.
- AI SecurityInsight2 min read
LLM Model Distillation: How AI is Stolen and How OpenAI Fights Back
Cloning LLM models via API has become the norm, which is why OpenAI is moving the fight for AI security to the level of its own silicon.
- AI Systems Without HallucinationsInsight2 min read
The Real Cost of Cheap AI: Count Task Costs, Not Tokens
Comparing AI prices by token rate is a mistake, because the actual cost of one completed task, including revisions, is what truly matters.
- AI SecurityArticle5 min read
AI Model Theft via API: What is Distillation?
AI distillation allows cloning the capabilities of an advanced model via API without using malicious code or cracking passwords.
- Cold Email & DeliverabilityInsight2 min read
Why fixed sending intervals destroy deliverability?
Consistent time intervals between emails signal automation to filters, which lowers B2B message deliverability despite correct domain configuration.
- Cold Email & DeliverabilityArticle6 min read
The Anatomy of a Follow-up to an Engineer: How Not to Sound Like a Bot
Learn why automated follow-up sequences end up in spam and how to design deliverable communication for technical audiences.
- Cold Email & DeliverabilityArticle6 min read
Art. 398 of Polish Electronic Communications Law in B2B: Eight Criteria for Valid Opt-in Consent
Art. 398 of the Polish Electronic Communications Law (PKE) mandates obtaining prior opt-in consent in B2B, and failure to separate communication channels and documentation deficiencies risk penalties up to 3% of the company's turnover.
- AI Systems Without HallucinationsArticle5 min read
LLM Dilution: Why a cheap AI model generates higher costs?
A low per-token price in LLM models often hides costs, dilution in the API gateway, and the necessity to correct erroneous responses.
- Cold Email & DeliverabilityInsight2 min read
Why is asking for consent in the first email illegal?
Sending an email asking for consent to contact is a violation of the law and can result in a UOKiK penalty of up to PLN 1 million or 3% of revenue.
- Cold Email & DeliverabilityArticle5 min read
Negative Feedback Loop: Why Google Postmaster Lies
Google Postmaster Tools displays data with a delay of up to 7 days, leading to incorrect decisions and domains falling into a negative feedback loop.
- AI Systems Without HallucinationsArticle5 min read
LLM-as-a-Judge: Automated AI System Quality Control
The LLM-as-a-Judge pattern reduces the cost of manual AI content verification by over 50%, ensuring full coverage and B2B compliance.
- Cold Email & DeliverabilityInsight2 min read
0.30% Spam Rate Threshold in Google Postmaster – How to Avoid Blocking
Exceeding a 0.30% Spam Rate in Google Postmaster blocks the deliverability of an entire corporate domain for weeks, jeopardizing B2B communication.
- AI Systems Without HallucinationsInsight2 min read
How to Eliminate AI Hallucinations? The Zero-Hallucination Architecture
Discover AGAPE's Zero-Hallucination architecture, which automatically verifies AI content and protects your domain from erroneous messages.
- Pipeline & B2B SalesInsight2 min read
B2B Pipeline Risk: How to Measure Revenue Loss?
Lack of data in 5 key CRM fields means your pipeline depends on one person's memory, creating a real risk of B2B revenue loss.
- AI Systems Without HallucinationsArticle5 min read
How to Measure AI Accuracy in B2B? Key Metrics
Discover the key metrics for measuring AI accuracy in B2B and learn how to protect your pipeline from hallucinations and a drop in deliverability.
- AI SecurityInsight2 min read
Four Human Oversight Patterns in AI Systems
Discover four human oversight patterns for AI agents that allow you to safely manage risk and meet EU AI Act requirements in your company.
- AI Systems Without HallucinationsArticle5 min read
AI Supervision in B2B Sales: Where to Implement Human-in-the-loop?
Learn how to precisely position human oversight points in AI processes to scale B2B sales without the risk of errors and reputational blunders.
- AI SecurityArticle5 min read
AI Validation Layer: How to Protect Management from ELC and UOKiK Penalties
The AI validation layer blocks unlawful sales messages before they are sent, protecting the domain and the personal assets of the management board.
- Cold Email & DeliverabilityInsight2 min read
How to Write AI Instructions for Outbound: 4 Key Principles
Precise AI instructions, based on terminology and audience profile, can reduce outbound errors by half and protect your domain.
- Brand Visibility in AIArticle3 min read
AI as the Gutenberg Revolution: A New Era of B2B Communication
AI acts like modern printing – it doesn't create knowledge, but it allows B2B experts to transform deep thoughts into publishable content.
- AI Systems Without HallucinationsArticle5 min read
Prompt to Technical Specification: How to write instructions for AI
Treating prompts as requests generates errors and costs founders dozens of hours in corrections — switch to technical specifications.
- AI Systems Without HallucinationsArticle5 min read
Zero-Hallucination Architecture: Secure AI Outbound in B2B
Zero-Hallucination Architecture eliminates AI reputational risk in outbound campaigns by strictly separating the model from deterministic business logic.
- AI Systems Without HallucinationsInsight2 min read
RAG in Practice: Differences Between Advanced and Naive RAG
Naive semantic search leads to hallucinations, while a correct RAG architecture requires filtering, re-ranking, and context compression.
- AI Systems Without HallucinationsArticle6 min read
Grounding AI: How to Anchor LLM Models to Company Data
Grounding is an architecture that constrains AI to verified company data, preventing costly hallucinations and errors in B2B communication.
- Cold Email & DeliverabilityArticle5 min read
Domain Burn: How a Configuration Error Destroys Company Reputation
DNS configuration errors and a lack of domain separation can destroy a company's reputation and block cold email sending for 4-12 weeks.
- AI Systems Without HallucinationsArticle7 min read
Why an LLM language model lies with full conviction
Language models do not possess a state of ignorance, and RLHF optimization causes them to lie with full conviction, which jeopardizes B2B communication.
- AI Systems Without HallucinationsInsight2 min read
AI Hallucinations: Architecture, Not a Code Bug
AI hallucinations are not a random error, but a direct consequence of the architecture of language models and the mechanism of predicting subsequent tokens.
- Pipeline & B2B SalesArticle5 min read
Outreach to Factory Directors: How to Win B2B Contracts
Industry decision-makers complete 60% of the purchasing process independently, so outreach must be based on hard data and involve technical teams.
- Cold Email & DeliverabilityArticle7 min read
3 Reasons Why Your Cold Email Isn't Reaching the COO
Cold email campaigns to COOs often fail due to technical, organizational, or behavioral blocks, which can be diagnosed through specific metrics.
- Cold Email & DeliverabilityArticle5 min read
Legal Cold Email and PKE: LinkedIn Invitation Acceptance
Learn how accepting a LinkedIn invitation creates a legal basis for email outreach in light of PKE and GDPR.
- Pipeline & B2B SalesInsight2 min read
B2B Follow-up and Multi-channel Outreach – How to Increase Sales?
Combining a sequence of four follow-up messages with multi-channel outreach can increase the B2B response rate from 4% to 19%.
- Cold Email & DeliverabilityArticle8 min read
SPF, DKIM, and DMARC Configuration in 30 Minutes – Step-by-Step Guide
Proper SPF, DKIM, and DMARC configuration protects your domain from phishing and prevents transactional and outbound messages from landing in spam.