DIY SEO: the path we took ourselves.

Step 1 – The math: agency vs. doing it yourself

The reason to do SEO yourself is rarely curiosity – it is the bill. Ongoing agency support is usually billed monthly, plus setup and, if ads run in parallel, the ad budget on top. Our total data costs for everything this guide describes, on the other hand, are logged precisely:

On top of that come running fixed costs you probably already have or need anyway: an AI subscription for the coding agent and hosting for the website – together, an order of magnitude below the price of a single agency hour. Current prices are on the providers’ websites; here we only state figures we paid and logged ourselves.

What the math doesn’t show: time. The path in this guide is not a push of a button but focused work in sessions of a few hours. The difference from an agency is that every one of those hours builds knowledge that stays with you.

Step 2 – The toolkit

Four tools, all linked in the sources below – exactly the combination this guide was made with:

Set them up in this order: install Claude Code, add the plugin from its GitHub repository, enter your DataForSEO credentials in the plugin’s connection, set a daily limit (ours: 10 US dollars – we never reached it), verify your website in Search Console.

Step 3 – Keyword research with live data

The most expensive mistake in DIY SEO is building pages on gut feeling. Before we build, every page gets four measured numbers: search volume, cost per click, competition and difficulty – plus the search intent and what Google shows on the results page (an AI Overview? Videos? Question boxes?).

Two directions supply the candidates: your own topics (what can we do, what do our clients ask?) and your competitors’ rankings – for a few cents, the API shows you the 100 search terms that bring each competitor the most traffic. For us, four competitors produced 361 terms; 229 of them were missing from our list, and after the truth filter from step 4, 20 were left. Four real finds from these measurements – German search terms, because our market is Germany:

The first row is the core of the method: find a term that 4,400 people search for every month, with a measured difficulty of almost zero – and that the big players still only serve in passing. You don’t spot gaps like that by thinking. You spot them in data.

Step 4 – Prioritize instead of guessing

Hundreds of measured terms become a ranked order – for us, through an open formula. Each term gets five sub-scores from 0 to 100, weighted into an overall score:

More important than the formula is the filter in front of it: the truth list. A table that records, for every product feature, what demonstrably exists, which wording is allowed for it – and what must not be claimed. Every search term has to get past it. That is why terms with six-figure volume were cut for us: everything containing “kostenlos” (free – our product is paid, so visitors would leave disappointed), everything about features that don’t exist, and the entire “make money online” topic cluster that competitors use to buy reach with no connection to their product.

Step 5 – The content plan

For us, two files carry the whole plan, both versioned in the project: a master keyword list (every term with all its metrics, target page and priority) and a content map (every topic cluster with exactly one target URL, primary and secondary terms, title and notes). The iron rule: one cluster, one URL. Build three pages for “Rechnung schreiben” (writing an invoice) and you compete with yourself.

The least conspicuous trick is a check script that runs with every build: it verifies that every target page really exists and that no cluster points to two URLs – and a deliberately strict test breaks as soon as the page count no longer matches the documented state. Sounds pedantic, but it prevents the silent drift out of date that otherwise kills plans like this. You can simply hand this exact discipline to a coding agent – see the prompt in the next step, point 7.

Step 6 – Building pages with the coding agent

This is where the agent pays off: for us, it takes hours, not weeks, to get from a measured term to a published page. The morning this guide was written is the example – in the morning the measurement showed the calculator gap, in the afternoon the tool was live:

So that the result doesn’t read like interchangeable AI text, the agent needs rules – ours come from real rounds of corrections: overly colorful areas became white with fine lines, two accent colors became one, every exaggeration was cut, and every change was checked with a screenshot before it counted. Distilled, that gives you the prompt to start with – copy it, adapt it, get going:

You are my coding agent for the website of my self-employed business.
Before you build anything, these rules apply:

1. Measure first, then build. Every new page first needs a
   measured keyword with search volume, competition and difficulty.
   No pages "on a hunch".
2. Only claim what the product can do. Create a truth list:
   feature, evidence in the product, allowed wording. Everything
   else goes – even if it sounds better.
3. Banned words: "guaranteed", "fully automatic", "legally compliant"
   and any promise a court could examine. Build a script that
   checks the copy for them automatically.
4. Design: white surfaces and fine lines instead of tinted boxes,
   exactly one accent color, text first. It has to look good on
   mobile first, not last.
5. One page per topic, one topic per page. The page title says
   what I do and for whom – not "Welcome".
6. After every change: take a screenshot, check it yourself, only
   then show it to me. If something can't be backed up, tell me
   instead of making it up.
7. Keep a list of all pages and keywords as a file in the project,
   and a test that breaks when the counts go out of date.

Now ask me, one question at a time: What do I offer, and for whom?
What can my offer demonstrably do? And which three topics do my
clients search for most often?

If you don’t want to use a coding agent, a website builder gets you there too – that path is in the chapter Building a website with AI. The rules from the prompt apply there just the same.

Step 7 – Measure and report honestly

SEO works in months, not days – anyone who promises you faster results is selling something. That is why doing it yourself includes an honest baseline. Ours, from Search Console, for the first 24 hours after setup:

Zero clicks, position 78 – this is what day 1 really looks like, and that is exactly why it is here. One observation from our measurements belongs here too, in fairness: the classic SEO search terms themselves are currently losing volume by double digits – part of search is moving to AI answers. So we build pages that serve both: classic rankings and being quotable in AI Overviews (clear statements, backed-up numbers, named sources). This guide is the example of that, too.