DataForSEO + Claude Code: Your First Analysis in Ten Minutes

Full transcript

1,237 words · script draft, under review.

Intro

You have credentials for DataForSEO from episode three. You have a clean workspace from episode six. Now let's make them do something useful, fast.

In this episode, I describe a small analysis in plain English, and Claude Code writes the script, shows me the plan, asks before it spends a cent, runs it, and summarizes the result. About ten minutes, and about twelve cents.

But the most important part comes after the summary. Because the first time I trusted an AI-written analysis without checking it, it told me something that wasn't true. I'll show you exactly how that happened, and how to catch it.

The task, in plain language

Our demo shop is example-shop dot test, a made-up store that sells coffee equipment. I have ten keywords I care about. Burr grinder. Espresso machine for home. Gooseneck kettle. That kind of thing.

The task I give Claude Code is one paragraph. For these ten keywords, get the monthly search volume and the live Google results page in the United States. Store everything in the database. Then tell me who ranks, and which features appear on each page.

Notice what the prompt does not say. It doesn't say which endpoints to use, how to parse the response, or what the tables are called. The instruction file in the workspace already holds that, and the rules: cache first, keys never touched, and no paid call without my go.

That's the point of the setup from episode six. The prompt can be short, because the rules are already written down.

The plan comes first

Claude Code doesn't start by calling anything. It answers with a plan, and a script.

The plan has four steps. Check the cache. One search volume task with all ten keywords in it. Ten live results pages, one per keyword, at depth ten. Then parse everything into the tables we already have, and write a short summary.

And I read the script. Not every line in detail, but enough to answer three questions. Which endpoints does it call? How many times? And where does the raw response go before anything is parsed? If I can't answer those three from the code, I don't approve it.

The cost gate

Then the gate. Before any paid call, it has to show me what it will call and what it will cost.

One search volume task covers up to a thousand keywords, at about nine cents. Ten live results pages at depth ten cost about three tenths of a cent each, so three cents. Total, about twelve cents. Then it stops and waits.

Twelve cents is nothing. The gate isn't about this run. It's about the habit. The same prompt with ten thousand keywords, or a retry loop that never ends, uses exactly the same code path. The gate is what makes you look before it costs real money.

My own setup goes one step further. The AI tool I use simply can't spend money on its own. Paid calls need my explicit go, every time. I recommend the same for you.

Cache, then run

After my go, it runs. Every response is saved to the cache folder before it's parsed, with the keyword, location, depth and date in the file name.

That matters right away. If the parser crashes, or I want a different summary, or I ask a follow-up question, the script reads from the cache. I never pay twice for the same page. Run it again a minute later, and the gate says: ten cache hits, zero paid calls, zero cost.

Reading the summary

And here's the summary. These are made-up numbers for illustration, not real data.

For each keyword, a volume, the domains in the top results, and the features on the page. A short paragraph on top: competitor-a dot test appears on seven of ten pages. AI Overviews show up on the informational keywords. Shopping results dominate the product keywords. And example-shop dot test, our own site, is in the top ten for only four.

Ten minutes ago, I had a list of keywords. Now I have a database with volumes, ten saved results pages, and a first picture of the market.

This is where most people stop. Copy the summary into a report and move on. Don't.

AI writes, you verify

Here's the rule for the whole course. AI writes, you verify.

The model is very good at writing a script that runs. It's less good at noticing when the script runs fine but measures the wrong thing. A summary written in confident sentences looks exactly the same whether the numbers behind it are right or wrong.

So before you believe a conclusion, you check two things. The code: does it count what you think it counts? And the numbers: does one raw response, opened by hand, agree with the summary? Here's why I'm strict about this.

The depth-ten trap

This happened to me, on one of my own sites, very recently. I asked for the same kind of analysis you just saw. Nine keywords. The AI-written report came back with a clear conclusion: my site isn't in the top ten for five of the nine.

That's a serious finding. It sounds like a ranking problem, and it would have sent me off rewriting pages.

So I opened the raw responses. And the numbers didn't add up. Some of those results pages had only six or seven organic results in them. Not ten.

Here's why. The request used depth ten. And depth counts every element on the page, not just the blue links. An AI Overview counts. Each "People also ask" box counts. So on a page full of features, depth ten gave me somewhere between six and nine organic results. My site wasn't missing from the top ten. It was simply outside what we had asked for.

The conclusion was an artefact of the request, not a fact about the search results.

And the real finding was different, and more interesting. On several of those keywords, my site was cited inside the AI Overview itself. Visible at the very top of the page, in a place the script never counted as a ranking.

How to verify, in five minutes

So here's my verification routine. It takes about five minutes.

First, read the request parameters in the code. Depth, location, language, device. Most wrong conclusions start there.

Second, ask the model to count the organic results per keyword. If any page has fewer than ten, a "not in the top ten" claim isn't proven.

Third, open one raw response from the cache yourself, and find one row of the summary in it.

Fourth, ask: what else could explain this result? A missing ranking might be a short page, a different location, or a feature that holds the site.

The fix in my case was cheap. Re-run only the short pages at depth twenty, about six tenths of a cent each, and through the same cost gate.

Where this goes next

In episode forty-eight we automate the whole analysis: scheduled pulls, a report every week, and the same cost gate.

But in between, the rest of this course teaches every method by hand first. Not because AI can't do it, but because you can only verify what you understand. When the model says something surprising, you'll know where to look.

Wrap-up

Let the AI write the script, gate every paid call, and check the numbers before you believe them. The prompt and the cost-gate checklist are linked in the description. Next episode: how to read Search Console properly, and what clicks, impressions, CTR and position really mean. See you there.