Stop Paying for Social Media Analytics: Track TikTok, Instagram and YouTube Views With Claude or ChatGPT

TL;DR: the numbers you pay a tracking tool to show you are already in TikTok Studio, Meta Business Suite and YouTube Studio. Put them in plain files in your own repository, let your AI rebuild a database from them every week, and ask it questions in plain English. You keep the full history, the proof screenshots and the analysis, for free. The open-source skill statur-social-analytics sets it up with you.

The problem with paid social media analytics tools

Most analytics tools for creators and founders read the same numbers the platforms already show you, store them in their database, and rent the history back to you every month. Three things go wrong with that model.

They break when a platform changes a metric. Meta deprecated Instagram plays and impressions on April 21, 2025 and moved reporting to a new views metric (Emplifi’s note on the change). Every dashboard built on the old names had to be rebuilt, and comparisons across that date need care.

Your history sits in someone else’s database. What happens to it when you cancel depends on the vendor: one keeps it for a while and lets you export tables before you leave (an example policy), others keep less. Either way, the years of data that make a format review meaningful are outside your control.

The questions you can ask are the ones the vendor built. “Which of my formats keeps people past three seconds and also gets saved?” or “Which videos of this competitor beat their own average by three times this month?” need a query, and most dashboards give you charts.

What changed: your AI can read, write and query

Claude and ChatGPT can now do the three jobs a tracking tool does. They read your analytics through an API, an export file or the screen you already see in your browser. They write what they read into files. They run SQL on those files and explain the answer. The missing piece was an architecture: where the numbers go, how they are named, and which skill decides what with them.

The architecture: one folder, four kinds of files

Everything lives in one folder of a private git repository, for example content/analytics/:

1. accounts.csv, the only file you edit by hand. One row per account: your own accounts, client accounts and a short watchlist of accounts you study. For each one, how the numbers are collected, the period you worked on it, whether its views count in your totals, and whether it may be named in public.
2. raw/, one CSV per collection, named by date, platform, handle and kind (videos, daily stats, retention curves, audience). A raw file is never edited: the next reading is a new file. Git keeps every version, so you can follow a video from day 1 to day 90 long after the platform stops showing it.
3. proof/, two screenshots of the platform’s own analytics for each viral video, with a blurred copy you can show without naming the account.
4. analytics.db, a SQLite database rebuilt from the files at every collection, plus a generated TOTALS.md with your total views per account.

The files are the source of truth and the database is disposable. Delete it, change computer, change assistant: one command rebuilds it.

Five collection routes, best first

The skill picks, per account, the first route that works, and writes it down.

1. Official APIs. The YouTube Analytics API gives channel and video reports, audience retention, age, gender and country for channels you own or manage. Meta’s Graph API gives insights for professional Instagram and Facebook accounts. Stable, within the terms, one setup per account.
2. Native exports. TikTok Studio, YouTube Studio and Meta Business Suite each have a download button. You click, the assistant files the export under the naming rule.
3. The analytics screen in your own browser. For an owned account with no API, a browser agent opens the analytics pages you can already see and reads them at a human pace, read only, without your password. It is the same work as screenshotting each page by hand.
4. Public pages, for the watchlist. View counts shown on profile grids, read once a week or once a month for 5 to 20 accounts, never in bulk.
5. Screenshots. The fallback that always works, and the proof you want anyway.

What one collection gives you: a real account

We ran the method on the TikTok account @louis_korczowski_phd on October 25, 2026, reading TikTok Studio in the account owner’s browser. One pass returned:

• 238 published videos with views, likes, comments, shares and saves: 1,264,424 views in total.
• 364 days of account history, day by day: 469,385 views over the last twelve months, and followers going from 4,748 to 7,396.
• For the one video above 100,000 views at day 7: 122,242 views in its first seven days (140,334 at reading time), 48.81 seconds of average watch time, 14.9% of viewers to the end, 846 new followers, 86% new viewers, 57.5% of views from France, and a retention curve at 75% after three seconds and 48% after ten.

Two details we measured that a tracking tool would not tell you. TikTok Studio stops drawing a video’s daily chart 365 days after posting, so the day-by-day history of an old video is gone unless you saved it. And the Business Suite content history on the Facebook page we tested went back about eighteen months, while the work we wanted to count was older: those videos had to be counted from their public view counts.

Use 1: track your own accounts

Collect weekly. TikTok returns its history day by day, so a weekly run gives you daily numbers without a daily job. Each run reads the last 14 days of account stats and the videos of the last 30 days, writes the raw files, rebuilds the database and commits.

Keep a viral line. By default, 100,000 views at day 7. When a video crosses it, the assistant takes two screenshots of the platform’s analytics (overview and audience) and stores its retention and audience in the base. Screenshots taken months later show different numbers, so the day-7 capture is the one worth keeping.

Count what is yours. If you work on several accounts, accounts.csv holds the period you worked on each one. A video counts in your total only if it was published inside that window, and a client who restricted public mentions stays anonymous in every proof.

Use 2: monitor accounts you study

Put 5 to 20 accounts in the watchlist: direct competitors, creators in your niche, founders whose formats you want to learn from. Read their public counts once a week or once a month. The base then answers the questions that matter for your next production day: which of their videos beat their own average by three times this month, which formats they repeat after a hit, how often they post. Public counts are rounded above a threshold (Instagram shows “318k”), so the watchlist is for comparisons and orders of magnitude, never for exact totals.

Use 3: replicate the winners, kill the losers

A format is a die and every video is one throw (the dice system). Judging a format needs several throws and mature data: at least seven days live and 1,000 views. With the base in place, two commands do in seconds what used to take an afternoon in a spreadsheet:

• dice prints, per format, the floor, the median, the typical range and the top on mature videos, with saves and shares per 1,000 views and three-second retention.
• winners ranks every mature video against its own format’s median, top and bottom of the list.

Then the decision moves to the skill that owns it. Each skill has one job and reads the same base:

The analytics skill never decides what to post. It measures, stores and answers. That separation is what lets each piece improve without breaking the others.

Questions you can ask your AI

Once the base exists, ask in plain English and let the assistant write the SQL. A few that founders ask in the first week:

• Which of my videos crossed 100k views in their first seven days, and what format were they?
• How many views did my account earn each month this year, old videos included?
• Which formats keep viewers past three seconds and also get saved?
• Do my videos over two minutes do better or worse than my 30-second ones, format by format?
• Where exactly do viewers leave my last viral video?
• Which videos of the accounts I watch beat their own average by three times this month?

The skill ships these as ready SQL in its references, tested on an invented sample.

Set it up in 30 minutes

1. Install the skill. In Claude Code or any agent that reads Agent Skills: npx skills add lkorczowski/content-strategist-AI-cookbook --skill statur-social-analytics. In claude.ai, upload the skill folder as a zip. In ChatGPT, paste its SKILL.md into the chat.
2. Ask your assistant to create the base: python3 scripts/analytics.py init content/analytics, in a private repository.
3. Fill accounts.csv together: your accounts, the route for each, and a watchlist.
4. Run the first full collection, then build and commit.
5. Ask your first question. Then schedule the weekly run, once you have checked it runs without you.

Rules that keep it clean

• Read only. The assistant never likes, follows, comments, posts or changes a setting.
• No passwords. You sign in in your own browser; the assistant never asks for a password or copies a session.
• Official APIs and exports first, human pace on any page, a short watchlist, no bulk collection.
• No invented numbers. Every figure in an answer comes from a raw file with its reading date, and rounded public counts stay labelled as rounded.
• Client numbers and screenshots stay private. Public proof uses blurred copies.

For your AI: the open-source skill statur-social-analytics walks you through the setup, the collection and the questions. How it fits with the other skills: architecture of the cookbook.