Case Study Sefa Aydın 6 min read

Case Study: An AI Agent That Scores My Website Data Every Morning

Case Study: An AI Agent That Scores My Website Data Every Morning

Short answer

Here the server collects Google Analytics 4, Search Console and Clarity data every night with no personal data. Each morning an AI agent scores these numbers from 0 to 10 in 8 areas and on the 10 top landing pages, backs every score with numbers and gives at most 5 recommendations. It can only read and cannot change any setting.

On November 21, 2026 I built a new AI agent for my website. It has one job: read the website data every morning and score it. It does not come up with ideas and it does not suggest campaigns. It looks at the numbers, gives a score and explains the score. On its very first report this agent caught a tracking error that had gone unnoticed for weeks. In this article I explain how I use AI to score website data and what that finding meant.

What was the problem?

Website data lives in three separate places. Google Analytics 4 shows what visitors do on the site and what the ads cost. Google Search Console shows how the site appears in search. Microsoft Clarity gives heatmap signals that show where people click and where they get stuck.

Opening all three regularly and reading them together takes time. In most companies the dashboards get a look once a month, the most striking number gets discussed and the rest gets skipped. Almost nobody asks whether the tracking itself is right.

There is also an attention problem. The human eye looks at numbers that change. A number that stays the same every month but was counted wrong from the start never raises an alarm.

Yet every decision rests on that tracking. The ad budget, which page gets rebuilt, which service gets the spotlight. If the tracking is wrong, every decision built on it heads in the wrong direction.

What I wanted was a reader that looks at the data with the same rules every day and writes down what it sees, with reasons.

What system did I build?

The system has two parts, and the split is deliberate.

Collection: Every night the server pulls data from the three sources. It only takes totals by page, channel and search query. No personal data is collected. The query part of each address is removed, so any information carried inside a link never reaches the file either.

Scoring: Every morning an AI agent reads these collected numbers and gives each area a score from 0 to 10. These are the eight areas it scores:

AreaThe question the agent asks
Business leadsIs the signal for business inquiries from the site healthy?
User experienceDo visitors get stuck or give up on the page?
Message clarityDoes the page explain what it offers at first glance?
Organic searchHow are search visibility and clicks doing?
AI visibilityHow are visits from AI assistants doing?
Ad efficiencyAre the ads running on the right signal?
Traffic qualityAre the visitors actually relevant?
Technical healthAre the pages fast and free of errors?

The same scoring runs separately for the 10 pages with the most entrances. That shows whether a problem affects the whole site or a single page.

The report stays short: the score for each area, the numbers behind it, the agent's diagnosis and the recommendations. If a score looks unexpected, the number in the reasoning shows exactly where to look in the dashboard. Reading the report takes far less time than going through three dashboards one by one.

The agent follows three rules. It backs every score with numbers. It writes its own diagnosis. It gives at most five recommendations. The limit of five is deliberate. A list of fifteen items gets read and forgotten. Five items get done.

Keeping it away from ideas is deliberate too. The scoring agent's job is not to find new campaigns but to measure the current state honestly. An agent asked for ideas starts reading the numbers in a way that supports its own idea. I wanted to keep a clear line between the one that measures and the one that recommends.

How did I keep it secure and under control?

When I build an agent that touches data, my first question is this: what do I lose if this agent does something wrong? To make the answer "nothing", I set four limits.

  • Read access only: The Google account is accessed through a service account that can only read. It cannot change any setting, goal or ad.
  • Keys stay on the server: The access keys live only on the server. The scoring agent never sees them. It only reads the collected numbers.
  • No outside access: The agent cannot send email or connect to any outside service. It writes its report, and that is all.
  • Data is not instructions: Text inside the data files never counts as an instruction. Even if a page title or a search query looks like a command, the agent reads it only as data.

The last point matters more than it seems. Search queries and page names are text that comes from outside. Not trusting that text is a basic rule for any agent that reads data.

With these limits, a bad day for the agent never becomes a bad day for me. At worst it writes a wrong report and I notice while reading it. No setting, no ad and no data changes because of it.

What was the result?

The agent's first report flagged an inconsistency in the business leads area. Over the last 28 days, 24 events had been counted as leads. The agent looked at where those events came from and wrote this: 14 of the 24 events were actually survey forms filled in by people already taking part in a training.

So the lead count looked more than twice as high as it really was. More importantly, every decision about which page and which channel brings leads was being made from that inflated number. Some pages that seemed to bring leads were simply the form pages that participants filled in.

I fixed the tracking the same day:

  1. Forms filled in by participants moved to a separate event. They no longer count as leads.
  2. Intro forms filled in after a lead moved to a separate event called "qualified lead".

The aim is to see the right number when I look at leads, and to make page, channel and ad decisions based on real demand.

The lesson I take from this: a tracking error that people had missed for weeks was caught on the first day by an agent that reads the data by its rules every day. The agent is not smarter than me. But it looks with the same care every day, it does not get bored and it does not skip the quiet numbers.

How can this work in your company?

You can build this in any company. Sales, ads, an online store or customer support data, it does not matter. This is the order I follow:

  1. Separate collection from interpretation. Keep the keys on the collecting side. Let the interpreting agent see only numbers.
  2. Give read access only. The agent should not be able to change anything in the data source.
  3. Write down the scoring criteria. Decide up front how each area is scored and which number backs each score.
  4. Audit the tracking itself first. In the first week, have the agent ask one question: is what we count really what we mean to count?

The fourth step brings the fastest gain in most companies. Before any new analysis, you need to be sure the current count is right.

This setup fits best in companies whose data is spread across several dashboards and whose ad spend depends on tracking. In a company that runs ads, a conversion counted wrong means the budget quietly flows to the wrong place.

I wrote about where AI is strong in data work in AI for data analysis.

If you want a similar scoring system on your company's data, start with my AI consulting page.

Frequently Asked Questions

How does the AI agent score website data?

It reads the totals the server collects every night and scores them from 0 to 10 in 8 areas and on the 10 top landing pages. It backs every score with numbers, writes its own diagnosis and gives at most 5 recommendations.

Does the agent see personal data?

No. Only totals by page, channel and search query are collected. No personal data is taken, and the query part of each address is removed.

Can the agent change a setting in the Google account?

No. Access goes through a service account that can only read. The keys stay on the server and the scoring agent never sees them.

What error did the first report find?

It found that 14 of the 24 events counted as leads in the last 28 days were survey forms from people already in a training. The lead count looked more than twice as high as it was, and the tracking was fixed the same day.

Why does the agent not come up with ideas?

Its job is to measure the current state honestly. An agent asked for ideas may start reading the numbers to support its own idea. That is why measuring and recommending are kept as separate roles.

Want to learn to do this with AI?

In individual training sessions, live on your own project, and in corporate training for teams, I teach step by step how to do everything covered in this article with AI. Write in and I will get back to you the same day.

Message on WhatsApp See individual training Corporate AI training for your team
Sefa Aydın

Sefa Aydın · AI Trainer & Consultant

An AI trainer and consultant who has worked on the Turkey projects of world-famous luxury brands. He teaches, hands-on, how every kind of work is done with AI: design, video, branding and vibe coding.

About Sefa Aydın →   LinkedIn ↗

Related articles

Want to apply this to your own work?

Tell me what you are working on. I reply the same day with next steps.

Your details are used only to reply to you. See the privacy notice.

Message on WhatsApp