SEO Research with Agents: API Data Instead of Browsers

On 8 June 2026, we checked search demand for this exact topic cluster. In Germany, "ai agent api" returned 10 monthly searches at high competition, with no difficulty value returned. "KI Agenten API", "Computer Use Agenten" and "API Automatisierung Unternehmen" returned no metric. The articles had already been written. They remained drafts for the time being. That is the most useful test for an SEO agent: may the result be "do not publish yet"? A pipeline that only produces new topics automates content inventory. Good SEO research produces a decision with raw data, limits and a date for the next measurement.

Give the agent a business question

"Find keywords" is not a usable task. For our cluster, the question was whether enough German-language demand existed to publish the four technical articles planned at the time. That determined the region, language, candidates, required metrics and decision threshold. Only those choices made the API response interpretable. A monthly run may ask a different question: which existing pages are gaining impressions without a matching increase in click-through rate or business impact? Google Search Console, the company analytics system and the stored page family then belong together. SERP data alone cannot answer that question.

The raw dataset is the working product

A report is easy to rewrite. The raw data must make the same run possible later. Store the following for every query:

• The exact search question or API query, including region and language.
• The provider, endpoint and retrieval date so a later run remains comparable.
• The unchanged response or a traceable snapshot of it.
• The link from each raw value to its recommendation, including missing values and errors.
• The decision a person made from the run, together with the next review date.

In our own SEO work, dated JSON artifacts therefore hold the DataForSEO responses beside the SEO map. The map holds the editorial decision. Neither replaces the other.

A missing metric is not zero

"No metric" must not become "no demand". It initially means only that the chosen provider returned no dependable value for that query. Keyword volume is also a market estimate, while Search Console data observes your own site. An agent has to keep that provenance separate, or it will produce precise-looking comparisons from different types of measurement. That is why the agent/API cluster stayed on hold for the time being. The available data did not justify publication, although it did not rule out later demand. GSC signals, sales questions or another DataForSEO measurement can change the decision.

A monthly comparison needs a fixed edge

A repeatable run does not simply compare "this month" with "last month". It fixes the query, period, region, device class and page set. The agent can then mark differences without silently changing the basis of measurement. Every notable page should follow the same path in the report:

• Observation: which raw value changed against which stored run?
• Source: did the value come from Search Console, analytics, a SERP API or the crawl?
• Explanation: which interpretation is plausible, and which alternative remains open?
• Proposal: what should be checked or changed without performing the change yet?
• Acceptance: which later measurement would show whether the decision was sound?

Publishing stays outside the agent

A title can be changed back easily in technical terms. Its effect in search results, internal links and later reports is not immediately reversible. We therefore let the agent research, compare and propose a diff. A person decides priority, meaning and publication. A study of real coding-agent sessions documented inaccurate self-reporting and rare self-correction. It does not measure SEO work. It nevertheless supports a sober acceptance rule: "report created" is not a result. The result is the stored dataset, the reviewable diff and the documented human decision.

Four possible outcomes keep the research honest

Our measurement ended with "hold". Other runs may conclude "publish", "revise an existing page" or "keep measuring". Put all four outcomes in the instruction before the run starts. Otherwise the process rewards activity instead of the right decision.

On 31 August 2026 we published the cluster after all, this article included: the four planned articles had become six posts, published together with the new lead page Building AI Agents: No-Code Platform or Code?. Research shows 1,900 monthly searches for "KI-Agenten erstellen", the only validated four-digit German term in our portfolio, and that page was not meant to launch without neighbouring articles that support it. The demand data did not change. The question changed. It first asked whether the cluster carries on its own; in August it asked whether the cluster supports a page that carries on its own. That is part of the method as well: the measurement produces the recommendation, a person decides, and that person may decide differently on a new fact, as long as the reason is written down.

The principle for agent-ready interfaces explains why the data should come through APIs rather than screenshots.

The pipeline owns the measurement; the editors own the decision.