An AI employee for SEO does not work better across the board with a paid data tool. It works clearly better on some tasks and not at all better on others. In my case study from 17 October 2026, Sebastian, my AI employee for SEO and GEO, did seven tasks twice: first with free sources only, then additionally with SE Ranking. In 3 of 7 tasks (search volume, gaps compared with competitors, backlinks), the second run delivered verified, decision-relevant findings that the first could not deliver. In 2 tasks (own rankings, technical audit), the tool brought no meaningful gain. For AI search it added a narrow extra view, and the briefs ended without a verdict. To make the result worth anything, the run without the tool went first, the tool was technically locked out there, and checksums prove that the first run was not changed afterwards.
This article is about the method and about what came out of it. The figures on the tool itself are in my test report SE Ranking Review: API, MCP and Real Numbers. Who Sebastian is and how he works is described in Sebastian, my AI employee for SEO and GEO.
Upfront: I now have a partnership with SE Ranking, and I earn a commission through my links to SE Ranking*. SE Ranking had offered me a partnership; the test ran before we had agreed on anything, with the free trial I had set up myself. Only afterwards did we agree on the partnership: for this article and three others, SE Ranking provides me with an account free of charge. SE Ranking had no influence on the setup or the results.
* Affiliate link: if you subscribe or buy credits through this link, I earn a commission. Your price stays the same. I have a partnership with SE Ranking.
The question behind the test
For SE Ranking I had a trial account with 100,000 credits for 14 days, plus an API and an MCP server through which Claude can query the data directly. The obvious question was not "Is the tool good?" but: How much better does my AI employee work with it than without, on which tasks, and what does it cost? An AI employee makes a good test setup here because it can do the same task twice with the same work instructions, and because every number ends up in a file that you can check afterwards.
The setup: two runs, one order, one lock
Same tasks, defined in advance. Before the first run, the seven tasks were fixed, each with a yardstick and a hypothesis:
Run A first, run B afterwards. Both ran as separate, non-interactive Claude runs on my server, on the same evening: A from 19:55 to 20:15, B from 21:40 to 22:10. The order matters. Had B been allowed to run first, A would already have seen the tool's data, and the comparison would have been worthless.
A technical lock instead of a ban. In run A, the SE Ranking MCP server was removed at the start of the run with --disallowedTools: the start log of the run lacks all of its tools, while in the second run there were 208 (of the server's 236 tools; I had blocked the 28 deleting ones for both runs). So Claude could not call them, not even by accident. An instruction like "Do not use SE Ranking" alone would not have counted as a lock for me. The SE Ranking REST interface, by contrast, was ruled out only by the work instructions; according to the log, run A used neither it nor the API key.
Checksums. At the end of run A, Sebastian created checksums over all 71 files of the first run and wrote them into the log. At the beginning and at the end of run B, he checked them again: both times 71 of 71 unchanged. That proves B did not retroactively improve or worsen the results of A.
Only verified findings count. A finding from run B counted only if it was confirmed at a primary source: in Search Console or on the live page. Sebastian checked 15 of the 51 gaps found this way, retrieved 9 backlink sources live, and cross-checked every audit finding from B only against the live page.
Costs limited in advance and measured against the balance. Each block had a credit budget. Sebastian did not measure by the price list but by the difference in the balance before and after each block. That was a good call: the volume query cost 300 credits instead of the 10,000 expected according to the list, and prompt queries were not charged at all.
What went wrong and what the test cannot do
Four limits belong in every honest case study:
- Run B knew run A. This was intended: B was meant to show what the tool adds, and it was allowed to use the data from A as a basis. This is not a comparison of two completely independent runs.
- The blinding of the briefs was weak. You can recognise the briefs from B by the volume columns. In the end I skipped the blind rating and had Sebastian merge both briefs for each page. Task 7 therefore has no verdict.
- The trial account was limited. Domain queries returned 10 rows at most. According to SE Ranking, the limit does not apply to paid accounts; that has not been measured.
- A confounding factor in my own measurement setup. In run A, one of the measuring agents for AI search wrote: "The source kevinwelter.com belongs to the operator of this system". Subagents that start in the project folder load the project instructions and therefore know whom they are measuring for. The number of citations may be skewed in my favour as a result. The number "seen in the search results" is more robust, because the search queries themselves contained no brand name. In future, measuring agents will start in an empty folder without project instructions.
The results per task
The costs: run A cost 0 euros and 0 credits, run B 19,155 credits. AI search cost the most at 9,800 credits, while volume, gaps and backlinks brought the most, together around 5,300 credits. In runtime, B was slightly longer than A at around 30 minutes against around 20. The queries themselves took seconds, and the time went into checking and evaluation.
What we made of it
An optimisation run with a control page. From four candidates checked in advance, Sebastian selected three existing SEO articles that had been unchanged since 22 September 2026 and were not part of any other test, plus a fourth page as a control that deliberately stays unchanged and gets no new links. The three pages received new sections on questions that searchers demonstrably have and that the pages had not answered so far. Title, description and main heading stayed the same, so that the effect of the content remains measurable. The run has been live since 17 October 2026 at 22:59.
A follow-up measurement that is still running. It is measured with the same Search Console query as before, for 28 days starting on the third day after going live. What counts is the change on the test pages minus the change on the control page; beforehand, Sebastian checks the trend of the whole website and the Google status dashboard for updates during the period. My hypothesis, not a promise: more search terms with impressions rather than a jump in position, and the most likely mover is a main term that was already at position 11. No effect is also a result.
I will add the result of the follow-up measurement here in November 2026.
Four new working rules. What the test showed about the way of working is now permanently part of Sebastian's rules: measure costs of credit APIs against the balance; run third-party audit crawlers on your own website only throttled and cross-check errors from fast runs live; use rankings from a rank tracker only after comparing them with Search Console; start measuring agents without project instructions. That feedback becomes a rule like this is the real difference between a tool and an employee that you manage.
A purchase decision that can wait. For my needs, volume, competitors and backlinks are wanted occasionally, before a topic plan or a revision, not every month. In my assessment a subscription is not worth it for that, but billing per query is. According to SE Ranking, it comes with no annual commitment.
How to set up such a test yourself
- Write down tasks and hypotheses before you connect the tool. Otherwise you measure what the tool is good at instead of what you need.
- Run the run without the tool first and freeze it. Checksums cost one line in the terminal and make the comparison verifiable.
- Technically lock the tool in the first run. In Claude Code, you can do this at startup with
--disallowedTools or permanently via deny rules in the settings.
- Check every new finding at the primary source. A number from a third-party database is an estimate until Search Console or the live page confirms it.
- Measure costs against the balance. For the volume query, the charge in my test deviated from the price list by more than a factor of 30, and prompt queries were not charged at all (according to SE Ranking, not the norm); other blocks matched exactly.
- Keep working with a control group. Only the optimisation run shows whether the better data also leads to more visibility.
Which free tools were used in the first run is described in SEO Audit Tools Compared. How an audit with an AI employee works step by step is shown in How to Run an SEO Audit With AI.
Frequently asked questions
How much better is an AI employee with an SEO tool?
In my test, clearly better on 3 of 7 tasks: search volume, gaps compared with competitors and backlinks. That is data that Search Console fundamentally does not have. For own rankings and the technical audit, the tool brought no meaningful gain, and for AI search in Germany a narrow extra view.
Why did the run without the tool go first?
So that it could not know the tool's data. Had the run with the tool gone first, its results could have influenced the second run. Checksums over all 71 files of the first run prove that the first run stayed unchanged afterwards.
Is it enough to forbid the AI employee a tool?
Not for a test. An instruction is a rule of behaviour that can fail to take effect at the wrong moment. The MCP tools were therefore technically removed at the start of the run. The same principle applies in everyday work: what really must not happen, I block through permissions, not through sentences in the work instructions.
What did the test cost?
The run without the tool nothing, the run with SE Ranking 19,155 credits from the trial account. On pay-as-you-go from 45 euros per 250,000 credits, that would have been about 3.50 euros of credit, just under 6 euros with the prompts charged as usual; the minimum deposit is 50 US dollars. The calculation is in the test report SE Ranking Review: API, MCP and Real Numbers.
When will the result of the optimisation run be available?
The follow-up measurement runs for 28 days from the third day after going live on 17 October 2026. I will add the result to this article, even if it shows no effect.
Where to go from here
If you are considering whether a data tool is worth it for your AI employee, first write down the questions that your free sources cannot answer. How Sebastian is built, with work instructions, review roles and approvals, and how you can hire such an employee for yourself, I show in my community Claude Practitioners. The basic structure of an AI employee is described in Creating an AI Employee: How to Start.
