How I Built an AI Research Analyst to Improve My Content Strategy

I stopped guessing what worked on LinkedIn and started knowing. Here is the high-level workflow — including illustrative data schemas and prompt logic — I used to turn over 4,300 anonymous engagement signals into a razor-sharp business strategy.

For years, I treated LinkedIn like a slot machine.

I would write a post, pull the lever (hit “Post”), and wait to see if the jackpot of “Engagement” would pay out. If a post got a high number of views, I felt like a genius. If it got almost nothing, I felt like I was shouting into a void.

But as a founder, “feeling” isn’t a metric.

I realized that while LinkedIn gave me vanity metrics (Impressions, Likes), it withheld the only data point that actually mattered to my business: Who is this audience?

Are the 10,000 views coming from potential Decision Makers, or just random noise?

Are the 100 likes on my conference photo coming from investors, or just my own colleagues being supportive?

I decided to stop guessing. I wanted to conduct a forensic audit of my entire 2025 content output. I wanted to know exactly which types of posts triggered which segments of the market.

To do this manually would have taken a team of analysts weeks. I did it myself, using AI as my Lead Research Analyst.

This was not a magic trick. It was a grind. It took me approximately 40 hours of data extraction, cleaning, and iterative analysis. But while a human team might have taken weeks to do the math, AI allowed me to guide the machine through the noise to find the signal.

Here is the exact workflow — down to the spreadsheet columns and prompt logic — that I used to turn over 4,300 anonymous engagement signals into a razor-sharp business strategy.

Step 1: Raw Material: The Data Schema

AI is only as good as the data you feed it. Most people fail here because they blindly and conveniently copy-and-paste text into their AI tool of choice. Others point AI at a set of files with relevant information, assuming it will make sense of everything. It can, to a point. But not enough to be impactfully useful.

To do a forensic audit, you need relational data.

I exported my year of data, but I didn’t dump it all into the AI at once.

Using the source data, I prepared two distinct CSV files to create a relational database structure:

File A: The_Content.csv (What I said)

• Post_ID: A unique identifier for the post.
• Date: To track seasonality.
• Post_Text: The full body copy (so the AI can analyze topics).
• Metrics: Impressions, Comments, Reposts etc.

File B: The_Audience.csv (Who reacted)

• Post_ID: To map the audience back to the specific post.
• Reactor: An anonymized string (e.g., User_6767) to protect privacy and ensure no Personally Identifiable Information (PII) entered the AI environment.
• Job_Title: The critical data point (e.g., “CISO,” “Sales Development Rep”).
• Headline: The bio text, which sometimes contains better context than the title.

The Engineering Insight: By preparing these files, I filtered out noise and allowed the AI to perform “joins” (like SQL) based on my precise guidance rather than just reading text.

Step 2: Lens 1 — Audience Segments: Role & Intent

First, I had to map the audience by Intent. A ‘Vice President’ could be a buyer, a partner, or a competitor. Titles alone weren’t enough. I created a ‘Data Dictionary’ to map job titles to six functional segments: Decision Maker, Decision Influencer, Investor, Peer, Partner, and Technology Community.

I ingested these datasets into a secure data analysis environment. But raw job titles are messy. A “Vice President” at a bank is a Decision Maker; a “Vice President” at a 2-person startup is a Peer.

I had to teach the AI my specific worldview. I created a “Data Dictionary” prompt — a set of IF/THEN rules to force the AI to categorize people correctly.

To be fair, AI was able to infer a lot of this logic, but it was not completely right.

Here is a sample of some logic I used (note that I have omitted some definitions as this is designed to be illustrative, not a complete guide):

“You are to act as a Market Researcher. I will provide you with a list of job titles. You must classify each reactor into one of the following segments based on these rules:

Segment: DECISION MAKER

• Keywords: CISO, Chief Information Security Officer, VP of Security, Head of Cyber, Director of Risk.
• Negative Constraint: Exclude anyone with “Sales,” “Account,” or “Consultant” in their title, even if they have “Director” (e.g., “Account Director” is NOT a Decision Maker).

Segment: INVESTOR

• Keywords: Partner, VC, Angel, Private Equity, Principal (at a Fund).

Segment: TECHNOLOGY PROFESSIONAL

• Keywords: Engineer, Developer, Analyst, Architect, Student.

Segment: NON-TARGET-AUDIENCE

• Keywords: Recruiter, Coach, SEO Specialist.”*

The Lesson: Without the “Negative Constraints,” the AI sometimes categorized Account Directors as Decision Makers simply because they had “Director” in their title. You have to tell the AI who you don’t want, not just who you do.

Step 3: Reality Check: The Loop

This is where a lot of the “40 hours” happened.

Classifying the “Unclassified” Audience

When I ran the first classification, 35% of my audience came back as “Unclassified.” The AI didn’t know what to do with titles like “Evangelist” or “Fractional Leader.”

I didn’t accept the error rate. I ran an iterative loop:

1. Export the rows tagged “Unclassified.”
2. Manually audited the top 50 recurring titles.
3. Update the Data Dictionary prompt with new rules.
4. Re-run the classification.

After over 20 rounds (yes, that is not a typo) of this “human-in-the-loop” debugging, I reduced the unclassified rate to under 5%. This ensured my data was statistically robust, not just a guess.

Classifying the Posts

I’d asked AI to make a few initial attempts at classifying my 52 posts into Topics. As reminder, these were:

1. Events/Networking
2. Personal Story/Reflection
3. Team Culture
4. Milestone
5. Industry Commentary

It made a fair attempt, but upon manual review, did not do a great job. It had trouble classifying many of the posts because it was not clear just from processing the text. AI lacked nuance and context. In the end, I had to manually insert a column in The_Content.csv file and classify each myself.

Refining the “Classified” Audience

AI didn’t get the classification of the audience right despite having done Step 2 above. I did spend a bit of time prompting AI to extract potential incorrect audience classifications, but found it to be too inconsistent. In the end, I had to:

1. Ask AI to use The_Audience.csv file as the basis, and insert its classification of each reactor in as a new column, and produce a new CSV.
2. I then painstakingly reviewed the resulting CSV file that AI had produced. This file had thousands of records, so it was not a small task.
3. Once done, I uploaded the new file back into the AI context and told it to treat the updated file as the new authoritative source of truth.

Step 4: Interrogation: Mapping Topic to Audience

Once the data was clean, I moved to the strategy phase. I didn’t ask “Which post was best?” because “best” is subjective. I asked specific questions to map Topic (from The_Content.csv) to Audience Segment (from The_Audience.csv):

• “Group all posts by Topic (e.g., Culture, Product, Industry News).”
• “For each Topic, calculate the ‘Intensity Score’: Total reactions from the ‘DECISION MAKER’ segment divided by the number of posts. Ensure that the score is normalized to account for the fact that some Topics are more frequently used that others — we do not want to skew the metrics because we are overly weighted on a specific Topic.”
• “Compare the Intensity Scores between the various Audience Segments and Topics.”
• “Calculate the ‘Specific Reaction Rate’ for each Seniority Level: Divide the number of reactions from a specific group (e.g., CXOs) by their estimated impressions. This measures efficiency rather than just volume.”

Human Interpretation: The Modes

It is important to note: The AI calculated the raw metrics, but it did not define the strategy. I had to look at the data clusters and manually define the “Influencer Archetypes” (Headhunter, Strategist, Raiser) based on the behavioral patterns the AI revealed.

The results were enlightening. My instinct (and ego) told me that my “Industry Commentary” — the deep, technical posts analyzing market trends — was my most valuable work. It felt thoughtful and authoritative.

The AI looked at the math and added nuance. It showed me that while those posts had lower volume (Intensity Score: 9.3), the Efficiency Rate (7.2%) was healthy. Decision Makers weren’t ignoring them; they were auditing them.

Meanwhile, Personal Stories — posts sharing vulnerable lessons — were generating nearly double the engagement from that same group (Intensity Score: ~15).

The AI forced me to confront a hard truth: Decision Makers engage with analysis to verify, but they engage with human resilience to trust.

Step 5: Lens 2 — Seniority Layer: Hierarchy & Behavior

The Segment lens showed me who they were, but it didn’t explain how they acted. I realized that a “Decision Maker” could be a C-Suite executive or a mid-level manager.

I instructed the AI to apply a second lens: Seniority Level. I forced it to parse every job title again, this time categorizing them strictly by hierarchy designed to mirror how LinkedIn classifies seniority levels: CXO, Senior, Manager, Director, and Entry Level.

This separation was the key. Lens 1 showed me that “Decision Makers” liked Milestones. Lens 2 revealed that within that group, “CXOs” were highly efficient Validators, while “Directors” were the Silent Evaluators.

Finding the seniority levels was not enough. I realized that a post with 100 likes from a specific group segment might just be a viral post that everyone liked. I needed to know what they liked disproportionately.

I added a calculation layer to the workflow:

• The Prompt: “Calculate the ‘Specific Reaction Rate’ for each segment. Divide the raw count of reactions from [Segment X] by their estimated share of impressions. Rank topics by this Efficiency Score.”

The Result: This was the mathematical breakthrough. It revealed the “CXO Validator” effect. It proved that while “Industry Commentary” had low volume, it had a high efficiency among executives. Without this step, I would have killed my most strategic content theme.

Step 6: Visualizing the Strategy

I asked the AI to analyze and visualize the trade-off between “Reach” (Vanity) and “Depth” (Value).

“Create a scatter plot. X-Axis is Total Impressions. Y-Axis is Total Reactions. Color the dots by Topic. Size the dots by the number of Decision Makers who reacted.”

In seconds, it generated a chart that tempered my “viral” ambitions. It showed clearly that the posts with the highest reach (Culture) often had the smallest dots (few Decision Makers). It proved the ‘Reach Trap,’ but also identified the ‘Universal Alpha’ — Milestone posts that achieve both viral reach and elite executive efficiency.

I also instructed it to compare ‘Viral Reach’ vs. ‘Momentum Trust.’ This revealed that while viral posts filled a stadium, the ‘VIP Box’ (CXO engagement) was virtually identical in size to my smaller posts. I was trading signal density for noise.

“Scatter plots were good for ‘Reach vs. Depth,’ but I needed to see the ‘hotspots’ of my network.

I instructed the AI to generate a Heatmap:

• X-Axis: Audience Segment (ordered by Seniority).
• Y-Axis: Content Topic.
• Color Intensity: Average Reactions per Post.

This visual immediately exposed the ‘Silent Evaluator’ pattern (a zone where Directors viewed content but rarely clicked) and the ‘Milestone Alpha’ (a glowing red hot spot where every segment engaged).”

Step 7: Reviewer: Refining the Narrative

The audit gave me the raw numbers, but numbers don’t persuade humans — narratives and storytelling do.

Once I had the insights, I didn’t stop using the AI. I pivoted its role from Data Analyst to Ruthless Reviewer. I fed the raw findings back into the model with a specific prompt:

“Act as a cynical storytelling expert. Review these findings. Identify the counter-intuitive truths that contradict standard LinkedIn advice. Create distinct ‘Laws’ or ‘Paradoxes’ to package these insights for an executive audience.”

This is how we arrived at terms like “The Reach Trap” and “The CXO Validator” found in the final report. The AI didn’t just find the data; it helped frame the argument so it couldn’t be ignored.

Step 8: Stress Test: Simulating Stakeholders

Writing the strategy is only half the battle. Ensuring it is bulletproof is the other.

Before publishing, I ran the findings through a “360-Degree AI Review.” I didn’t ask “Is this good?” (which invites flattery). I simulated specific, critical personas to stress-test the content.

The Prompts I Used:

• The Initial Audit: “Review all three articles as an: 1) Auditor for accuracy, consistency, and fidelity. 2) Strict editor as per a tier-one news outlet for quality, readability, and engagement. 3) Public Relations Expert to ensure we do not make any communications-related errors and cause ourselves to have PR problems to address. 4) Legal, Risk & Compliance Expert to ensure adherence to regulations and contractual obligations. Provide recommendations for potential changes that need to be made.”
• The Gap Check: “I asked you to act as: 1) Auditor for accuracy, consistency, and fidelity. 2) Strict editor as per a tier-one news outlet for quality, readability, and engagement. 3) Public Relations Expert to ensure we do not make any communications-related errors and cause ourselves to have PR problems to address. 4) Legal, Risk & Compliance Expert to ensure adherence to regulations and contractual obligations. Are there other roles you should be covering to ensure we perform a thorough review?”
• The Final Audit: “Yes, go with your recommendation. Run checks for 1) Data Science Validity. 2) Brand Alignment. 3) DEI/Inclusion. for a complete 360-degree review.”

The Result: This process caught various items that needed to be addressed, which I did.

The AI helped me build the argument, but the AI Critics saved me from potential risks.

Outcome: A Momentum Model

Because of this audit, I am not guessing in 2026. I have a strategy derived from hard data, not intuition. The data dictated three distinct “Modes” of operation:

• To Hire (The Headhunter): Post Team Culture (High Reach, Talent focus).
• To Position (The Strategist): Post Events (High Status, Decision Maker focus).
• To Fundraise (The Raiser): Post Milestones (High Velocity, Investor focus).

I fired my gut instinct. I hired the data. And for the first time, I have data-backed clarity on who is listening to what.

Ian Yip is the founder and CEO of Avertro, a venture-backed cybersecurity software company.