Use Data Or Be Used By Data!

The April 7 issue of Seotistics is here for you!

The other time we talked about Claude Code... and now it's time for part 2!

We will cover business context and those juicy leaks from last week.

I still need some time to figure it out but hey, get this preview!

P.S. Won't link the actual leaks for clear legal reasons 👀

Please move this email to your Primary inbox or reply to it. This is to prevent Seotistics goes into spam by accident. Gmail users can read this tutorial to do it.

Read this in your browser

📣📣 Important Announcements 📣📣

You can now preorder my new course at 30% off until June 15, the official launch date.

This will be for Analysts or everyone wanting to explore Web Data in general, not only SEO.

The course is split in 4 amazing modules:

✅ GA4, GSC and other data sources: how to use them
✅ LLM Workflows, BigQuery, SQL & Python
✅ Dataform & Pipelines
✅ Real Use Cases at work

Oh, there are also the Seotistics sales on all the products until April 14!

Product updates will follow soon, very soon!

The Claude Code Leak

In case you didn't know, Claude Code got apparently leaked the other day.

They say it was "by accident" 👀.

While we don't actually have the actual model "data" that Claude uses (aka the brain), we have some more info on how Claude behaves.

For most normal people, this leak doesn't change anything, really.

But here are some findings worth mentioning:

✅ Insulting Claude or expressing disappointment is considered as feedback

✅ Claude can recognize if it can run tools in parallel (aka more at the same time).

Try asking "Examine these 2 texts at the same time."

✅ CLAUDE.md isn't advice, it's what it's actually used as a persistent memory layer

✅ Some phrases may trigger deeper reasoning (and no, not only ultrathink as we saw the other day).

If you tell Claude to do it in X steps or think thoroughly, it'll behave differently.

👀 Web search only gets the of a page... and

They also have a list of selected seed websites to fully search from, like GitHub or documentation stuff.

Hey schema enjoyers, where is your god now? 🤣

Wait Wait... Web Search!?

So... based on what I've noticed, Claude does NOT do any query fan-out, aka generating more questions to answer you.

This is quite different from what we have observed with ChatGPT, for example.

Claude rewards better prompting, unlike ChatGPT.

This is not necessarily a good thing.

And as I said, Claude knows that it can rely on websites like GitHub for documentation.

But ask Claude to use your website and look for quotes... it won't be the same.

Claude strips a lot from normal web pages and gives you some short quotes at best.

Oh, and there is no reference whatsoever to llms.txt as far as I know!

The Slash Commands

Most people open Claude Code and start typing like it's a chatbot...

which is fine but you're leaving a lot on the table.

If you type / in your instance you can invoke commands such as:

✅ /context → visualize how much context you have left (use this constantly)

✅ /compact → compress the conversation when you're getting close to the limit (avoid it in most cases, just run /clear)

✅ /clear → fresh start without closing the session

✅ /cost → check token usage and spending before you get a surprise bill

✅ /memory → edit your CLAUDE.md directly from inside the session

✅ /init → scaffold project documentation when you set up a new project

The ! prefix also lets you run bash commands directly (so !git status or !python my_script.py without leaving the session).

This one in particular is super useful because I didn't know before and had to open a new session lol

The (Business) Context

If you recall what I said before, LLMs need some context or the output will be utter crap.

It just doesn't work and you can test it yourself by using Claude Code on your data.

The output is often obvious or disappointing because LLMs are very good at making stuff up.

Once you read most of the LLM answers in depth, you will notice the output doesn't actually make sense.

It sounds so smart yet it's random advice.

Claude doesn't know what's a conversion for you or which metrics you care about.

And since LLMs are trained on the web... you can be sure they have been fed SEO propaganda for years to come.

To avoid all of this, you can provide .md files or use CLAUDE.md to make sure what you mean and what's needed.

For Analytics work, your CLAUDE.md should contain the things Claude can't infer from the files in your project:

✅ What counts as a conversion in your business (not only the event name, the actual meaning)

✅ Which data sources are canonical and which are approximations

✅ Your naming conventions for tables, fields, and models

✅ Any stylistic rules you want enforced (no em dashes, no AI syntax, etc.)

✅ The tools and versions in your stack

You can edit it with /memory from inside any session.

To check context drain, you can also run the /mcp command to see how many active MCPs you have.

Most of them will drain your context so I'd rather tell you to keep most of them disabled by default.

The CLAUDE.md Rabbit Hole

This magic file is Claude's persistent memory across sessions... which implies this stuff gets loaded into the context too.

You can also have multiple CLAUDE.md files at different folder levels.

The one in the root applies to everything.

The one inside a specific project subfolder applies only when you're working in that directory.

This is extremely useful if you have separate projects for different clients or data domains.

For example a structure like this one for an Ecommerce project:

What's the business and its business model
Flagship brand or product names
Projects, datasets, tables (like what is what) with some quick explanations
Preferred coding style or naming conventions
Common tasks/analyses
Personal preferences (e.g. skip the intro, give me the answer)
Mention where to get more details or info, e.g. check the docs folder inside the project

The mistake is overloading this file with a LOT of info.

You should rather provide general context here and then provide the details via separate files.

Company information is super important.

Specify your business model and be clear in who you sell to or how you make money.