DeepSeek Just Changed Analytics


Use Data Or Be Used By Data!

The August 3 issue of Seotistics is here for you!

DeepSeek is the buzz of the moment and it's not even random hype.

Chinese models are revolutionizing the LLM world because they are sooo cheap and good.

This has some consequences on Web Analytics too, mind you!

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​

πŸ“£πŸ“£ Important Announcements πŸ“£πŸ“£

The Seotistics Summer Sales will start this Wednesday.

I will send you an email on Wednesday with the valid promo code to get 30% off on all products.

So prepare your wallets in advance πŸ‘€

Including my new course below, so have a look now:

I am also working on new updates and content as we speak.

As for conferences, I will be present at:

  • ​Belgrade SEO (you can guess the city lol)- 18th September 2026 (use code MARCO for a good % off)
  • ​SEO On Stage (Athens) - 24th October 2026

And some more to be announced later I guess πŸ‘€

OK... What Happened?

On July 31, DeepSeek released the public beta of V4 Flash 0731.

No new architecture.

It's still the same 284B-parameter mixture-of-experts model with a 1-million-token context window.

What changed is the post-training: significantly better agentic behavior, coding, and tool-calling.

They also added native support for the OpenAI Responses API format, which means it's a drop-in replacement for any workflow already calling OpenAI.

The pricing: $0.14 per million input tokens, $0.28 per million output tokens. Cache hits cost $0.003 per million input tokens.

For context: GPT-4o is roughly 5-10x more expensive depending on the task, Claude Sonnet is in a similar range.

Look, the model quality debate is interesting but not the point here.

The point is that the price-to-capability ratio just moved in a direction that makes some analytics use cases genuinely viable that weren't before.

Model Input /1M Output /1M 100K queries Tier
GPT-4.1 Nano $0.05 $0.20 $0.10 Budget
Gemini 2.5 Flash-Lite $0.10 $0.40 $0.20 Budget
⭐ DeepSeek V4 Flash 0731 $0.14 $0.28 $0.28 Budget
GPT-4o mini $0.15 $0.60 $0.30 Budget
Gemini 2.5 Flash ‑ $0.15 $1.25 $0.30 Mid
GPT-4.1 Mini $0.40 $1.60 $0.80 Mid
Claude Haiku 4.5 $1.00 $5.00 $2.00 Mid
Gemini 2.5 Pro $1.25 $10.00 $2.50 Premium
Claude Sonnet 5 † $2.00 $10.00 $4.00 Premium
GPT-4o $2.50 $10.00 $5.00 Premium
Claude Opus 4.8 $5.00 $25.00 $10.00 Top-end

* "100K queries" = cost to classify 100K GSC queries (~2M input tokens, ~20 tokens/query). † Claude Sonnet 5 introductory price, rises to $3/$15 after Aug 31 2026. ‑ Gemini 2.5 Flash scheduled for deprecation Oct 16 2026. GPT-5 excluded β€” pricing unconfirmed across sources. Pricing as of August 2026.

The table above doesn't include caching and it's just an estimate, please consider it as an indication!

P.S. Let me know if some of the numbers are wrong too, I am testing some new method to extract that data πŸ‘€

What The F Is This Jargon

You'll see these terms like MoE and parameters around everywhere, so let me explain them easily.

Parameters are roughly the "knowledge" stored in a model.

More parameters generally means more capability, but also more compute required to run each query, which means higher cost.

284 billion is a very large number.

Here's the twist: V4 Flash uses a Mixture of Experts (MoE) architecture.

Instead of activating all 284B parameters for every single token it processes, the model is split into many specialized sub-networks ("experts") and only a small fraction of them activate for any given input.

Think of it like a company where you have 284 specialists on payroll, but each customer call only routes to 2-3 of them depending on the topic.

The full team is there but most of them are dozing off (for now).

Post-training means the model is improved after it's trained at specific tasks.

Some Use Cases For You (Again)

Since it's dirty cheap and actually better than previous models, I urge you to reconsider some use cases.

It's important to remember that most of these tasks are handled manually today or with worse processes πŸ’€

This is your chance to actually upskill and make the difference.

Content Audit/Research

Auditing is self-explanatory, instead of auditing content manually, you could:

  • pull the data from somewhere like BigQuery
  • define what to calculate
  • get some tables as output
  • LLM to interpret the results based on your guidelines

As usual, don't use MCPs or pull data straight from Google APIs like you read on LinkedIn (for serious projects).

There is no data retention and the process will be a nightmare over time.

I am not a huge fan of writing content with AI BUT in the right hands, this can also work.

The issue is that the right hands aren't who you think πŸ’€

Automated Reporting

This is the easiest yet more valuable use case.

Apart from dashboards, which I have discussed in my latest newsletter, normal reporting is needed.

You can create a process in a software like n8n and schedule it to send an email to the right person.

The actual part where AI is present would be in summarization.

Chatting With Data?

You'll hear people say this unlocks "just ask your data a question in plain English."

And technically, yes, you can wire up a chat interface over your BigQuery data.

But look, the problem with conversational analytics isn't the model.

It's that most analytics data is badly structured, inconsistently named and missing the business context that makes a question meaningful.

And natural languages like the one you are reading are terribad at being clear.

Tell a machine you want a glass of water.
​
What's a glass? What's water? How is the water inside the glass?
​
Use code and that's a whole different story...

There are some instances where this can be good but ultimately not your life goal.

Cheaper But Not More Popular

Yes, in terms of output and cost, this is one if not the best model.

BUT... LLM usage isn't as high as you imagine.

At least, based on my purely personal evidence, most companies still ignore these topics.

The bar is quite low if you have some willingness to work on projects with LLMs.

I gave you even too many examples today... and these are already enough!

We continue next week, I got so much to do now!

πŸ‘₯ Join Our Community

Our Discord community offers a small place where we can talk business and web data.

If you hate all the noise of social media, then this place is for you.

I will start posting more there as we have a forum channel now.

This is the best way to stay updated in real time on Seotistics:

πŸ”Ž Analytics For SEO Ebook - Course / Ebook

You will:

βœ… Use GSC and GA4 Data to their fullest potential

βœ… Learn Python/SQL for your needs

βœ… Get a complete blueprint for auditing websites

βœ… Learn how to 10x your productivity

βœ… Learn BigQuery to work on large websites

I teach you what's needed to go from 0 to a professional Data Analyst.

Even if you leave SEO, the foundations are the same for other jobs!

Also in ebook:

Think Like A Web Analyst

This course teaches you to:

βœ… frame Analytics problems

βœ… understand which metrics matter

βœ… managing Web Analytics projects successfully

πŸ“š Recommended Reads - Peak Content πŸ—»

Read these peaks (and documentation):

As usual, my most recent LinkedIn content is here.

❗️ Feedback and Recommendations

If you have ideas/recommendations for the next issues of Seotistics, you can simply reply to this email.

Marco Giordano
​
Data/Web Analyst

Follow me on πŸ”½πŸ”½πŸ”½:

Bernerstrasse SΓΌd 167, Zurich, Switzerland
​Unsubscribe Β· Preferences​

Seotistics - Web Analytics + Business + Strategy

The Seotistics newsletter is written by Marco Giordano, a Data/Web Analyst with the goal of combining business and web data. Tired of the usual boring Analytics content without any business impact? Seotistics teaches you how to use Analytics, web data and even content in your workflow while helping you with Strategy.

Read more from Seotistics - Web Analytics + Business + Strategy

Use Data Or Be Used By Data! The July 27 issue of Seotistics is here for you! Dashboards are everywhere and tbqh often overengineered for simple use cases. I've audited too many organizations having this issue here. So what can we do instead? The answers here are simple and feasible, you just need to know them! 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...

Use Data Or Be Used By Data! The July 20 issue of Seotistics is here for you! LLMs are everywhere and they have some interesting applications for our daily tasks in Analytics. Yeah, they won't make you lose your job as the news claim... but for sure you can't sit watching and sacrifice opportunities. 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 πŸ“£πŸ“£...

Use Data Or Be Used By Data! The July 13 issue of Seotistics is here for you! Scraping is one of the most important skills you must have as an analyst AND marketer. But what changed after LLMs? Well, a lot! This issue is about YOU scraping websites to get data, not how LLM providers scrape your website. 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 πŸ“£πŸ“£...