Almost Timely News: 🗞️ How To Do Feature Engineering with AI (2026-07-12)
Never ask questions you already have the data to answer
Almost Timely News: 🗞️ How To Do Feature Engineering with AI (2026-07-12) :: View in Browser
The Big Plug
👉 Please take my 6-question Reader Survey to tell me what you want!
Content Authenticity Statement
95% of this week’s newsletter was made by me, the human. The glossary at the end I made with Google’s Gemini 3.5 Flash based on the contents of the newsletter. Learn why this kind of disclosure is a good idea and might be required for anyone doing business in any capacity with the EU in the near future.
Watch This Newsletter On YouTube 📺
Click here for the video 📺 version of this newsletter on YouTube »
Click here for an MP3 audio 🎧 only version »
What’s On My Mind: How To Do Feature Engineering with AI
This week, if you haven’t taken it already, please take my Reader Survey. I’ll show you how I use the data and what you could be using with your own survey data using feature engineering.
Part 1: What is Feature Engineering?
If you’ve never heard the term, feature engineering is all about deriving new data from the data we have. I mentioned it in passing in the June 28th newsletter, but it’s worth reviewing again.
Broadly, in any given data set, there are two kinds of features, continuous and categorical, which is fancy for numbers and not-numbers. If you’ve used tools like Adobe Analytics or Google Analytics, these are metrics and dimensions.
We broadly define a number as anything you can do basic arithmetic on; website sessions is a number. You can add, subtract, multiply, and divide that number. A not-number is the opposite; a company’s industry is a piece of information you can’t do math on.
This is important because, as with everything, there’s nuance. There are things that look like numbers but don’t function like numbers. For example, in North America, there’s a system known as NAICS codes, the North American Industry Classification System. A company like Trust Insights has NAICS code 541611, management consulting. But this is a label - doing math on that number changes its value and meaning in unhelpful ways. It looks like a number, but it doesn’t behave like one.
Likewise, there are things that don’t look like numbers but function like them, like the day of the week. We say Tuesday, and if your week starts on a Sunday, a Tuesday is the third day of the week. Time in particular is rife with messy number-like things that don’t behave consistently. You can add and subtract days from a date like 2026-07-12 but you don’t really multiply or divide them often.
We need to understand this to understand that non-numbers in particular are packed full of information that we can extract and turn into more information. And once we have that new information, we can do more with the data we already have.
Now, I’ll borrow from the June 28th newsletter to give you a concrete example fo feature engineering, of deriving new data from existing data. Take this datestamp: 2026-06-28. We see this as a date, as a single piece of data. But we can decompose it into:
Year
Month
Day
Week of year
Day of year
Weekend or weekday
Day of week
Day of month
Week of month
Day of quarter
Week of quarter
Month of quarter
Quarter of year
Depending on what you’re doing, each of these new data points might be important and relevant, especially if, in the case of my reader survey, I’m looking for insights about who my readers are, how they behave, how I can serve them better, and if I’m blunt, how I can sell them more things. Knowing whether day of week or day of month or they of quarter changes their behavior might be important.
For example, if you work in B2B, month of year is really important. December is a month when a lot of folks don’t have a ton of sales activity going on in the Northern Hemisphere. People are winding down for the holidays; if I’m doing data analysis and I have datestamps, I can decompose my click rates with these additional data points to see if those points show any patterns I can’t observe from the datestamp alone.
In turn, this means that if I’m offering sponsorships to a sponsoring company, I might have to increase or decrease what I charge them based on the patterns I see in the data, but I won’t know that if I don’t engineer those features from the data.
Part 2: Why Feature Engineering Matters for Surveys
One of the most important reasons you need to know feature engineering for surveys is because every question you ask someone in a survey reduces the completion rate. The more questions you ask, the less participation you get. Thus, if we want high completion rates AND good quality data, we have to think about what we can feature engineer from the data we have. Never ask questions for information you can extract on your own.
The first step in the process is to inventory what data you do have. For the Almost Timely newsletter, what information do I already have that I don’t need to ask? I have email addresses, obviously. I have opens and clicks for each issue. I have dates and times for subscriptions and unsubscriptions. That’s already quite a lot of data I don’t need to ask for, which means the survey can be shorter.
In terms of what I do need to ask for, let’s start with the first and second questions of the survey - your email address, and then your title and company domain name. That seems like a weird combination, right? Why did I do this? Well, email is important because if you win the drawing next week, I need some way of contacting you. But isn’t email enough to tell me a lot? Yes, but a lot of folks subscribe to this newsletter with their personal email accounts (as you should!) and so cspenn@gmail.com doesn’t lend a whole lot of insights.
So I ask for job title and company domain name. Why this? Look at how much you can unpack from these two fields. First, job title tells us about seniority in an organization - a marketing coordinator typically has less purchasing power and authority than a marketing director. It also tells us about the role a person plays in their industry. An HR administrator or a financial analyst do very different things than a marketing manager and have very different needs. If I have a collection of job titles, chances are I can infer things like ideal customer profile clusters.
The second thing I asked for is company domain name, because that’s very specific. There are a lot of things named Trust Insights, including a new feature in Apple’s iOS, which is a pain in my butt when it comes to media monitoring. But there’s only one TrustInsights.ai that’s relevant to the discussion, so if I have a company domain name, I have a more precise understanding of the company in question.
From that nugget, what can we know? With even just a few web searches, we can know the company size, the industry, what they sell, their approximate annual revenue for larger/publicly traded companies, approximate number of employees, their skill at marketing (proxies with things like domain rating for the domain from your favorite SEO software), their email marketing setup (based on their public DNS records), and if you’re really ambitious, their tech stack based on all the tracking software packages installed on their website.
Here’s the really cool part: you can obtain all that pretty much for free these days, at least for the publicly available data. In the past, you had to pay big data aggregators for that kind of information, but with generative AI, you can have it write the necessary code to go extract all that information.
You could, if you wanted to, even set up a search system where you took the person’s email, company, and title and did a quick search of LinkedIn to find their profile and attach the profile URL to your dataset.
Combine this with the data you already have, like who’s opened what, when, how often (all that comes from your existing email marketing system and can be feature engineered), as well as the email addresses that aren’t personal (for which you can extract the domain name), and you’ve got a really robust data set you can perform analysis on.
I remember a few years ago how dismayed a lot of data driven marketers were when the company Clearbit got acquired by HubSpot because Clearbit was a super useful utility for analyzing an audience of email addresses. Today, I can roll my own clear bit and I don’t have to worry about who’s acquiring what. I build my own software using AI.
Combine that with your content - which issues of your newsletter got opened and clicked on by whom, organized by topic - and you’ve got a demographic and psychographic profile of your audience.
When I started the survey for my newsletter, I did so with the express intent of updating my sponsorship kit, but in the process of doing all this feature engineering, I ended up with a dataset that’s ripe for deep analysis. To borrow from AI’s favorite phrasing (it’s not X, it’s Y), it’s not just data about what’s working, but data about what’s working with specific segments of my audience - because the audience isn’t one monolith. From all that data, I can extract and explore who exactly reads this newsletter and what resonates with them.
I can even pair that with purchase data from things like the Trust Insights CRM or our Trust Insights Academy or my Gumroad account for my books to see who’s buying what.
This all sounds magical, doesn’t it? Well, it’s not magic - it’s machinery.
Part 3: How To Build a Feature Engineering System
To build something like this, like what I just described, we have to clearly define the requirements. To do this, I use Katie Robbert’s 5P Framework by Trust Insights™, the master framework that our entire consultancy is built on: purpose, people, process, platform, and performance. In fact, you can use the 5P Framework by Trust Insights™ exactly as is for creating the system software you need.
I always recommend that people approach this in a three phase process - define the requirements in a requirements document (PRD, product requirements document in software engineering lingo) first. That’s the why. Next, you define the technical specifications of the software, the what. Finally, you define the workplan for implementing the software, the how. AI can build all that for you as long as you have rich, detailed explanations of what you want.
I’d start like this; yes, this whopping whole section is the prompt. Realistically, this entire newsletter issue is a prompt.
--- start of the plan ---
Purpose: The purpose of our feature engineering system is to create as much detail as possible from the data we have, using feature engineering, web search, web crawling with browser technologies like Camoufox or Playwright, topic modeling with YAKE or similar extraction with an LLM, metric extraction from log files, and FOSS Python libraries for feature extraction of things like dates, decomposing them into their components. The ultimate outcome will be a database that we can do regression-based analysis on to determine what’s working in our email marketing program as well as an update to our advertising sponsor kit. My definition of overall success for the Almost Timely newsletter is to grow it organically (unpaid) from 294,000 subscribers to 500,000 subscribers by January 1, 2027.
People: I have two sets of people data; first, I have people who responded to a survey. The format of this dataset is as follows:
Timestamp,Email Address,Job title and company domain/URL,Which topics do you most want to see MORE of in Almost Timely?,Which topics do you most want to see LESS of in Almost Timely?,Which of the following do you have budget authority/influence over at your company?,"Have you ever researched, tried, or purchased something after seeing it mentioned or sponsored in Almost Timely?","If you're open to it, would you leave a testimonial for Almost Timely?",Any other feedback you'd like to leave privately?
7/11/2026 13:45:03,cspenn@gmail.com,"Chief Data Scientist, TrustInsights.ai","Marketing strategy & industry trends, Marketing analytics & measurement, Generative AI tools & techniques","Career development, Politics","Marketing / advertising technology, Data & analytics / BI tools, AI / ML platforms","Yes, I researched/tried it but didn't buy",,The second is a large folder of logfiles from my email marketing software that needs to be ingested, tabulated, and prepared for feature extraction. The files are in the input/substack folder, and the record data format is as follows.
This is the record of delivers:
post_id,timestamp,email,post_type,post_audience,active_subscription
138352478,2023-10-27T22:48:08.386Z,cspenn@gmail.com,newsletter,everyone,falseThis is the record of opens:
post_id,timestamp,email,post_type,post_audience,active_subscription,country,city,region,device_type,client_os,client_type,user_agent
138590623,2023-11-05T11:11:08.411Z,cspenn@gmail.com,newsletter,everyone,false,US,,,,,,Mozilla/5.0Each newsletter issue in data/substack/posts.csv has a corresponding delivers.csv, opens.csv, and HTML text copy of that issue in the posts folder within the substack folder. The master subscriber list is in data/substack/email_list.almosttimely.csv
The advertising sponsorship kit is intended for sponsors who want to reach my audience. I don’t know who they are yet because we haven’t done the analysis.
Process: Here’s what we want to decompose from the different record sets. For the survey set, we want to decompose the email address into name (the part before the @) and domain. For the title and company domain, we want to determine the title, the seniority level of the title (one of: junior individual contributor (IC), senior IC, junior manager, senior manager, junior executive, senior executive, owner), the vertical of the title (example: finance, marketing, HR, etc.) within the company
From the company domain name, we want to use web search and web crawling/fetching to obtain the company name, size of the company in both annual revenue and number of employees, company industry (with NAICS code if possible), the company’s martech stack based on a crawl of their homepage and contact page to detect various marketing technologies, their DNS records to see what tech they’ve got set up, etc. There was a company years ago called BuiltWith that would do essentially this; we want to do the same thing.
From the qualitative fields in our survey, an understanding and tabulation of what people want more of, what they want less of, what they have purchasing power/authority for, whether they’ve ever bought from me, and whatever qualitative public feedback (testimonial) or private feedback they have.
From the combination of email and title/company domain, we’ll use an LLM to determine if the company domain is present; if it’s not, but it’s present in the email (and it’s not a consumer email provider like GMail/Hotmail/Yahoo etc., it’s a company domain like trustinsights.ai) then we’ll use the company domain provided in the email if it’s missing from the job title field.
For the substack data, we’ll do the same - if we have a company domain that isn’t a consumer email provider, we’ll reuse the same domain research logic.
From the combination of the data we then want to analyze, understand, and cluster our data to derive useful insights for what’s working for my email newsletter - topics, timing, industries, seniority, etc. - and what would be useful for my advertising sponsorship kit. In the input folder for this project are guides for both; analytics-catalog is a catalog of over 1,000 different analytics techniques in YAML format to help you choose algorithms that are the best fit for our combined data; input/analytics-catalog/index.yaml is the orchestration start that will help you find the individual techniques. input/newsletter-kit-best-practices is a catalog and checklist of best practices for newsletter sponsorship kits, along with your previous audit of the Almost Timely newsletter, gap analysis, and survey recommendations.
Platform: For this project, we’ll use Python 3.12 at the command line following the guides in docs/orientation.md and docs/rules-python.md plus your web search tools for other FOSS libraries that would be useful addons and to avoid reinventing the wheel. From Principle 12 in our rules, we never reinvent the wheel when there’s existing best in class FOSS options we can use instead.
head and tail are installed on this system; the Substack folder logs are each more than 20 MB, so instead of reading the entire file with grep, use head/tail for the first or last 10 lines to preserve your context window.
For this project, you should use the Haiku model for reading files, the Sonnet model for writing, synthesis, and code writing, the Opus model for planning and review, and the Fable 5 model for the advisor. Your fallback for Fable 5 is Opus 4.8.
For the LLM to use for inference, analysis of the data sets, and company research, you’ll use oMLX installed on this machine with the following credentials:
Endpoint: http://127.0.0.1:8998/v1
Model: Qwen3.6-35B-A3B-Mixed-4-8
API Key: 12345
This is the model you must use for all records in the datasets as it is private and local; we never use cloud models when handling PII/SPI/PHI (personally identifying information/sensitive protected information/protected health information). Feel free to use LiteLLM or similar libraries so you don’t have to reinvent the wheel here. You’ll process one row at a time; be sure to structure your prompts to maximize prompt caching and efficiency.
For the web search, I recommend the ddgs Python library, but feel free to use your web search tools for better FOSS options. DuckDuckGo is the preferred search engine for working with any kind of PII, such as searching for the URL of a person’s LinkedIn profile based on their email, job title, or company name.
Performance: There are two distinct goals for this project. The first goal is working Python software that performs all the functions described so far. You have succeeded when you have achieved 100% test coverage, 100% passing unit and E2E tests, 100% production smoke tests.
The second goal is useful outputs - the analysis of my newsletter and survey results for both the advertising kit and for what else I should be doing with my newsletter to make it more compelling, useful, and recommendation/referral-worthy so that readers are compelled to share it.
You have succeeded when my advertising kit meets the standards described in the checklist and the ideal customer profiles I provided would find it compelling and worth reaching out to me about sponsorship, and when the performance analysis gives me a complete strategic plan - the why, what, and how of helping my newsletter grow from 294,000 subscribers to 500,000 subscribers in the next 6 months.
Use your brainstorming superpowers to ask me up to 3 rounds of questions with up to 5 questions per round so that you have enough information to succeed at the task.
--- end of the plan ---
I could have it go from this, but even better would be building the PRD, spec, and workplan from this. Again, the more time AI has a chance to talk out loud, the better its results will be. Each planning document you create gives it another chance to think things through and revise.
Once the PRD, spec, and workplan are done, hand off the documentation to a smaller, faster AI system to do the build. For example, if you’re in the Claude ecosystem, you’d use Claude Sonnet 5; in the Gemini ecosystem, you’d use Gemini 3.5 Flash. In the OpenAI system, you’d use the brand new GPT-5.6 Terra model. As long as we’ve invested heavily in planning first, a smaller, faster, not as smart, not as expensive model will get the job done well.
Be sure you’re using a coding tool, like Claude Code, OpenAI Codex, Google Antigravity, Microsoft Github Copilot, etc. so that you get the best possible results; web-based chat tools are completely the wrong choice here.
What we should end up with, ideally, is all our planning documents. I used the Trust Insights planning plugin (not yet released, but soon!) and one of the five artifacts generated is this handy system diagram that I can use to sanity-check the AI’s plan:
What we see here is a robust, well-designed system - you can see the red section calls out where we build in protection for PII, keeping the data you’ve given to me safe. We can also see the AI’s choices of tools like Wappalyzer and dnspython, which show that it adhered to one of our key instructions, to not reinvent the wheel.
Part 4: Wrapping Up
To give you a sense of realistic expectations, the build for this kind of analysis system is likely to take 6-12 hours, depending on the AI system you use. It’s the kind of thing that, once you’ve had a chance to review all the plans and documentation, you then have the AI go build and tell you when it’s done. You don’t wait around to watch it, any more than you’d wait around to shoulder surf a development team doing the build for you.
In fairness, there’s also a lot of prerequisites I glossed over intentionally, because they’re foundation materials that we’ve covered together many times before, like using deep research tools to build best practices guides and checklists that help keep AI on the rails, or installing command line utilities and applications so that AI doesn’t try to custom code everything, wasting your token budget and resources, or setting up local models to protect your data.
What we’ve walked through today is essentially custom software development using the 5P Framework by Trust Insights™ as our north star. What’s powerful about this is that it’s made for me, by me, to handle my specific problems. The lessons in it for you, the key takeaways, are to take what’s useful from my process for yourself.
Use the 5P Framework by Trust Insights™ to come up with what it is that you want.
Use the practice of developing a requirements document, a technical spec, and work plans to have AI think things through.
Use the idea of having AI build you a system diagram that you can see so that you can sanity check it.
And then use agentic coding tools to bring your ideas to life.
You do not have to type a single line of code at all to make this work for you. You just have to go through and think through and plan ahead and build that really robust project plan that will deliver great results because you’ve thought of everything and you’ve asked AI to push back and say, “Hey, I got some questions for you”.
What you didn’t see in this process was, when I kicked it off in Claude Code, it had like 9 different questions for me. What about this? What about this? Did you think about this? Why am I doing this? And it did a great job helping me think through how I wanted this software to work.
So take this issue, take everything that I’ve shown so far, and try it on a problem that you’ve been needing to solve, but you haven’t been able to solve yet.
How Was This Issue?
Rate this week’s newsletter issue with a single click/tap. Your feedback over time helps me figure out what content to create for you.
Here’s The Unsubscribe
It took me a while to find a convenient way to link it up, but here’s how to get to the unsubscribe.
If you don’t see anything, here’s the text link to copy and paste:
https://almosttimely.substack.com/action/disable_email
Share With a Friend or Colleague
Please share this newsletter with two other people.
Send this URL to your friends/colleagues:
https://www.christopherspenn.com/newsletter
For enrolled subscribers on Substack, there are referral rewards if you refer 100, 200, or 300 other readers. Visit the Leaderboard here.
ICYMI: In Case You Missed It
Here’s content from the last week in case things fell through the cracks:
Optimization vs. Innovation: Why the Next Big AI Leap Isn’t in the Models Themselves
Stop Measuring the Wrong Metrics: Why ROI is Failing Your AI Strategy
Why You Must Use Lateral Thinking to Prevent AI From Stifling Your Creativity
Stop Boosting Your Competitors: Why Listicles Are Killing Your AI Search Visibility
Stop Wasting Your AI Budget: How to Plan with Smart Models and Execute with Cheap Ones
Almost Timely News: 🗞️ The Biggest Problem with AI Today (2026-07-05)
On The Tubes
Here’s what debuted on my YouTube channel this week:
My Merch Shop
I’ve been adding so much stuff that I’ve decided to bundle it all in what I call a Merch Shop, because otherwise there’s literally too much to keep track of and I run out of space in my own newsletter.
So welcome to the Merch Shop!
Skills for Claude and Agentic AI:
👉 LLMs.txt Skill - generate an llms.txt file for your site based on your real data.
Job to AI Skill and Plugin - unpack any job description into what AI can and can’t do
Books:
Courses:
Subscriptions:
Recent Talks
These are just a few of the classes I have available over at the Trust Insights website that you can take.
👉 New! From Text to Video in Seconds, a session on AI video generation!
Never Think Alone: How AI Has Changed Marketing Forever (AMA 2025)
Powering Up Your LinkedIn Profile (For Job Hunters) 2023 Edition
Building the Data-Driven, AI-Powered Customer Journey for Retail and Ecommerce, 2024 Edition
The Marketing Singularity: How Generative AI Means the End of Marketing As We Knew It
Advertisement: New GEO 201 Course
In GEO 101, the first course I built on the basics of GEO, I taught you about presence, appearance, and relevance, the three phases of GEO, and what you need to do in each phase to align with how AI search operates.
The top piece of feedback we got at Trust Insights about it was, “okay, great, but how do I tell my boss that we’re ‘winning’ at GEO?“ After I quelled my murderous rage at your boss on your behalf, Katie and I sat down and worked out a straightforward, aligned methodology for doing this.
GEO 201 is based on the three phases, what you can control and what you can genuinely see - and critically, what you can’t. Because there is absolutely no way to say your brand “ranks higher” in AI search, period, end of story.
But you can say and show with confidence what you’ve done and how you show up for presence, appearance, and relevance with tools you’re probably already paying for, and based on how AI search systems really work.
👉 GEO 201 is available now for USD 149.
Get Back To Work!
Folks who post jobs in the free Analytics for Marketers Slack community may have those jobs shared here, too. If you’re looking for work, check out these recent open positions, and check out the Slack group for the comprehensive list.
Director Of Product Marketing And Commercialization at Enterprise Health
Director, Product Marketing - Customer & Market Insights at Attentive
Marketing Director - Solutions Architect at Creative Planning
Disclosure: I source these links from LinkedIn every week on the following criteria: New in the past seven days, Easy Apply on, remote roles, USA geography.
How to Stay in Touch
Let’s make sure we’re connected in the places it suits you best. Here’s where you can find different content:
My blog - daily videos, blog posts, and podcast episodes
My YouTube channel - daily videos, conference talks, and all things video
My company, Trust Insights - AI help
My podcast, Marketing over Coffee - weekly episodes of what’s worth noting in marketing
My second podcast, In-Ear Insights - the Trust Insights weekly podcast focused on data and analytics
On Bluesky - random personal stuff and chaos
On LinkedIn - daily videos and news
On Instagram - personal photos and travels
My free Slack discussion forum, Analytics for Marketers - open conversations about marketing and analytics
Listen to my theme song as a new single:
Social Good: Ukraine 🇺🇦 Humanitarian Fund
The war to free Ukraine continues. If you’d like to support humanitarian efforts in Ukraine, the Ukrainian government has set up a special portal, United24, to help make contributing easy. The effort to free Ukraine from Russia’s illegal invasion needs your ongoing support.
👉 Donate today to the Ukraine Humanitarian Relief Fund »
Events I’ll Be At
Here are the public events where I’m speaking and attending. Say hi if you’re at an event also:
MAICON, Cleveland, October 2026
SMPS AI Conference, Austin, November 2026
MarketingProfs B2B Forum, Boston, November 2026
There are also private events that aren’t open to the public.
If you’re an event organizer, let me help your event shine. Visit my speaking page for more details.
Can’t be at an event? Stop by my private Slack group instead, Analytics for Marketers.
Required Disclosures
Events with links have purchased sponsorships in this newsletter and as a result, I receive direct financial compensation for promoting them.
Advertisements in this newsletter have paid to be promoted, and as a result, I receive direct financial compensation for promoting them.
My company, Trust Insights, maintains business partnerships with companies including, but not limited to, Amazon, Talkwalker, MarketingProfs, Agorapulse, The Marketing AI Institute, Spin Sucks, and others. While links shared from partners are not explicit endorsements, nor do they directly financially benefit Trust Insights, a commercial relationship exists for which Trust Insights may receive indirect financial benefit, and thus I may receive indirect financial benefit from them as well.
Thank You
Thanks for subscribing and reading this far. I appreciate it. As always, thank you for your support, your attention, and your kindness.
Please share this newsletter with two other people.
See you next week,
Christopher S. Penn
Glossary
I generated this glossary using Gemini 3.5 Flash from this issue.
5P Framework by Trust Insights™
Quick Translation: A planning recipe that aligns marketing projects before you buy software or hire staff.
Definition: A Trust Insights planning framework built on five pillars: Purpose, People, Process, Platform, and Performance.
Why it Matters: It stops you from buying tools (Platform) before defining your goals (Purpose), your team (People), your workflow (Process), and your metrics (Performance).
BuiltWith / Wappalyzer
Quick Translation: Tools that scan websites to list their installed software.
Definition: Software databases that inspect public website code to identify active analytics, CMS, and marketing tools.
Why it Matters: They let you audit competitor tech stacks and qualify leads by showing you their active software (e.g., Salesforce, Shopify).
Camoufox / Playwright
Quick Translation: Playwright automates web browsing; Camoufox hides the automation to avoid bot blocks.
Definition: Playwright programmatically controls web browsers to perform tasks. Camoufox modifies the browser profile to mimic a real human user.
Why it Matters: They allow automated programs to scrape search data and check sites without getting blocked by security screens.
Categorical Features (Dimensions)
Quick Translation: Text-based labels or categories you cannot process with math.
Definition: Qualitative data points (like industry or country) that group or segment information rather than measure quantities.
Why it Matters: Google Analytics calls these “Dimensions” (e.g., Traffic Source, Device Type). They let you filter and segment reports.
Continuous Features (Metrics)
Quick Translation: Pure numbers that allow math calculations.
Definition: Quantitative data points (like revenue or pageviews) that measure numeric values.
Why it Matters: Google Analytics calls these “Metrics” (e.g., Sessions, Conversions). You use them to compute totals, averages, and rates.
CRM (Customer Relationship Management)
Quick Translation: The main database that stores customer contact details, communication history, and purchases.
Definition: A software database (like HubSpot or Salesforce) that manages all customer interactions and relationships.
Why it Matters: It connects marketing campaigns to real sales revenue, showing which ads generate paying customers.
Demographic vs. Psychographic Profile
Quick Translation: Demographics track who buys; psychographics explain why they buy.
Definition: Demographics measure objective facts (age, job title, company size). Psychographics analyze subjective traits (values, interests, lifestyle).
Why it Matters: Demographics help you target your ads; psychographics help you write copy that converts.
DNS Records (Domain Name System)
Quick Translation: Public settings that connect domain names to servers and secure outgoing emails.
Definition: Public database records that point web domains (e.g.,
google.com) to IP addresses and authorize mail servers (via SPF/DKIM).Why it Matters: Scanning a company’s DNS records reveals their email platforms and tests their deliverability setup.
Domain Rating (DR)
Quick Translation: A 0-to-100 score estimating how much search engines trust a website.
Definition: An SEO authority score (from tools like Ahrefs) that evaluates the strength of a website’s backlink profile.
Why it Matters: It acts as a proxy for a company’s SEO strength and overall digital marketing maturity.
E2E (End-to-End) Tests
Quick Translation: Testing a complete user path—like purchasing a product—to ensure the system functions.
Definition: A testing method that runs a software program from start to finish to confirm all parts work together.
Why it Matters: It mirrors a test-buy (submitting a form, verifying the receipt, checking the CRM record) to catch issues.
Feature Engineering
Quick Translation: Calculating new, more insightful data fields from existing ones.
Definition: The process of extracting new data columns from raw inputs using domain logic.
Why it Matters: It unlocks hidden patterns. Turning a date like
2026-06-28into “Weekend” or “Q2” helps you find timing-based behavior trends.
FOSS (Free and Open Source Software)
Quick Translation: Software that is free to use and modify without licensing costs.
Definition: Software distributed under licenses that allow anyone to access, edit, and share the source code.
Why it Matters: It reduces expenses by letting you use existing libraries (like Python or R) instead of building custom tools.
ICP (Ideal Customer Profile) Clusters
Quick Translation: Data-driven groups of target companies sharing similar traits.
Definition: Groupings of target accounts sorted by traits like industry, revenue, and team size using clustering algorithms.
Why it Matters: It helps you target accounts scientifically rather than relying on guesswork.
LiteLLM
Quick Translation: A wrapper library that lets code talk to different AI systems using a single language.
Definition: A software tool that translates API calls to support multiple AI models (OpenAI, Claude, local models) with identical code.
Why it Matters: It lets you swap AI models instantly to save money or boost speed without rewriting your software.
LLM (Large Language Model)
Quick Translation: The core AI engine behind writing assistants, chatbots, and text categorization.
Definition: An AI system trained on vast text data to process, generate, and summarize human language.
Why it Matters: LLMs automate qualitative data analysis, categorize customer feedback, and write copy at scale.
Log Files
Quick Translation: The raw journal tracking every single action on your server or platform.
Definition: Automatically generated text files that record events, errors, and logins chronologically.
Why it Matters: They provide the unvarnished facts behind marketing metrics, listing exact emails sent, delivery bounces, and device details.
NAICS Codes
Quick Translation: Government-standard number codes that label a company’s industry.
Definition: A six-digit classification system that categorizes businesses based on their primary economic activity.
Why it Matters: It standardizes muddy data. You get a clean industry label (like
541611for Management Consulting) instead of inconsistent manual entries.
oMLX
Quick Translation: Software that runs open-source AI models privately on your own computer instead of the cloud.
Definition: A local server engine that runs open-source Large Language Models on private hardware.
Why it Matters: It protects customer privacy. By keeping data local, you avoid violating privacy laws (GDPR, CCPA) when using AI.
PII / SPI / PHI
Quick Translation: Sensitive customer data regulated by strict privacy laws.
Definition: Information classes (Personally Identifiable, Sensitive Protected, Protected Health) that identify individual identities, finances, or medical statuses.
Why it Matters: Mismanaging this data leads to major fines under GDPR, CCPA, or HIPAA. You must secure it.
PRD (Product Requirements Document)
Quick Translation: The brief or blueprint detailing what a software project must do before developers start coding.
Definition: A planning document that outlines a product’s purpose, features, and success metrics.
Why it Matters: It aligns marketing goals with developer execution, avoiding costly miscommunications and delays.
Prompt Caching
Quick Translation: An AI memory feature that avoids re-reading instructions for every single task, saving time and money.
Definition: An optimization that saves the static part of a prompt in memory so the AI doesn’t process it repeatedly.
Why it Matters: It cuts API costs and speeds up execution when processing bulk items (like categorizing thousands of surveys).
Qwen3.6-35B-A3B-Mixed-4-8
Quick Translation: The specific model name of an open-source AI engine.
Definition: An open-source model variant from the Qwen family. “35B” denotes 35 billion parameters (its intelligence capacity), while the remaining tags indicate speed tuning.
Why it Matters: It identifies the exact model version running on your system, helping you balance speed and quality.
Regression-Based Analysis
Quick Translation: A statistical method showing how strongly input factors (like ads) drive outcomes (like sales).
Definition: A statistical technique that measures the relationship between independent inputs and a dependent outcome.
Why it Matters: It proves marketing attribution by identifying which channels predict purchases.
Smoke Tests (Production Smoke Tests)
Quick Translation: A fast test to confirm the core basics of a website or app work and don’t immediately crash.
Definition: A preliminary test suite that verifies critical system paths function after a deployment.
Why it Matters: It serves as a sanity check. For example, testing if a landing page button redirects correctly before starting a campaign.
Tech Stack (Technology Stack)
Quick Translation: The suite of software tools your business uses to operate.
Definition: The combination of software applications, platforms, and databases used to run business functions.
Why it Matters: Your “MarTech Stack” represents tools like HubSpot, WordPress, Google Analytics, and Stripe working together.
Topic Modeling (e.g., YAKE)
Quick Translation: An AI method that groups thousands of written comments into main themes automatically.
Definition: A natural language processing method that extracts keywords and groups documents by theme. YAKE is a lightweight tool for this.
Why it Matters: It analyzes customer feedback instantly, identifying top requests or complaints from survey comments.
Unit Tests
Quick Translation: Automated tests checking if a single, tiny code function works correctly.
Definition: A testing practice that evaluates individual, isolated pieces of code in isolation.
Why it Matters: It ensures minor updates won’t break specific parts of your site, like verifying form validation.
Web Crawling / Web Scraping
Quick Translation: Using code to copy information from websites into a spreadsheet automatically.
Definition: The automated process of fetching web pages, extracting content, and saving it to a database.
Why it Matters: It replaces manual data entry. A crawler can capture contact details or tech setups from hundreds of sites in seconds.
What stood out to me is that the AI wasn't really the story here.
The planning was.
The quality of the final system depended on the quality of the thinking that happened before a single line of code was generated. It feels like another reminder that AI often amplifies human clarity more than it replaces it.
The part AI quietly changes here is the division of labor: it’s genuinely good at generating candidate features, and genuinely bad at noticing when one of them leaks the target. So the human job moves from inventing features to auditing them: is this real signal, or is it the answer smuggled in through a side door? Generation got cheap. Judgment about leakage did not.