How an SEO agent turns noisy keyword data into a defensible content decision
See how an SEO agent used Scrollport to research keyword demand, test a content idea and produce an evidence-backed article under editorial supervision.
By Scrollport
In brief
An SEO agent used four tools from DataForSEO, Serper and Exa through one Scrollport account to test keyword demand, inspect current results and find source material. The bounded task was supervised and refined through three rounds of editorial feedback. Paid tool execution cost $0.0800.
We wanted to show how we used Scrollport ourselves to research and improve Scrollport’s SEO, and to check whether this workflow was worth publishing. We set the business question, audience and existing page owners. The agent then used four tools from DataForSEO, Serper and Exa to test the idea, inspect current search results and find source material.
Scrollport let the agent discover the tools it needed, see each price and run the complete task through one account and balance. The agent gathered and organised the evidence while we supervised the work, challenged weak conclusions and decided what was worth publishing.
We used Scrollport to test our own article idea
We did not begin with a keyword export. We began with a practical question: could a real SEO workflow show why shared access to specialist tools makes an agent more effective? We also needed to know whether this article would add something useful for readers or merely compete with our existing SEO agents guide.
That meant researching three things before writing: how people search for the topic, what the current results already answer and which claims or recommendations needed current support. Keyword data was evidence for that decision, not the decision itself. One export can contain unrelated meanings, the wrong audience and several phrases that express the same need.
The agent worked from a bounded brief
We supplied Scrollport’s positioning, the intended audience, the existing page owners and the decision that had to be made. The agent turned that context into a research plan and then ran it. This was the working brief:
Help us decide whether Scrollport should publish an article about using
Scrollport itself to improve our SEO for a US English audience.
The article should answer three questions:
1. Does this workflow show a useful advantage of Scrollport?
2. Is a new article distinct from our existing SEO-agent guide and tool pages?
3. What current evidence should support the article?
Start with: seo agents, ai seo agents, automated seo agent,
ai agent seo tools and keyword research agent.
Generate related ideas, remove phrases that do not describe Scrollport's
product or a job its tools can support, then validate the exact shortlist.
Inspect current result pages wherever page type or intent could change the
decision. Check the existing canonical owner before recommending a new page.
Find current, credible sources that can support the article.
Return observed evidence, interpretation and a page-ownership recommendation
separately. Show the current price before every paid run. Do not publish or
change a live page without editorial review.The field test ran on 24 August 2026 in Codex using Scrollport CLI 0.2.1. The agent useddiscover to find relevant tools, inspect to read each live contract and price, and run to gather paid evidence.
| Provider | Scrollport tool | Job | Cost |
|---|---|---|---|
| DataForSEO | Keyword ideas and keyword overview | Expand and validate search demand | $0.0548 |
| Serper | Google search | Inspect four current result pages | $0.0056 |
| Exa | Web search | Find and check current source material | $0.0196 |
The agent could get to work without waiting for three separate provider setups. Recreating the same chain directly would normally mean three provider accounts, three API credentials and separate subscription or prepaid-credit arrangements. Scrollport replaced that setup and maintenance work with one pay-as-you-go balance. Product positioning and publication remained supervised decisions.
The first 100 ideas were mostly noise
The first paid run expanded five seeds into 100 ideas. The highest-volume rows included AI writing products, people’s names, jobs and meanings of “agent” unrelated to automated SEO. A simple relevance filter left only two clearly useful phrases: seo automation and seo automation tool.
Observed evidence: the keyword-ideas run cost $0.0368 and returned 100 rows from 10,935 available ideas.
Interpretation: expansion improved recall but had poor precision for the product question. The export was useful as raw material, not as a content plan.
Recommendation: apply product-fit and intent filters before spending money validating every row. Keep the rejected terms so a reviewer can see what the agent removed and why.
Exact validation changed the shortlist
The agent checked ten exact phrases in one consistent US English request. Six returned usable database metrics. Four had no result in that market and language combination, which is missing evidence rather than proof of zero demand.
| Keyword | Monthly volume | Difficulty | CPC | Intent |
|---|---|---|---|---|
| seo agents | 27,100 | 19 | $45.83 | Commercial |
| seo automation | 2,400 | 14 | $23.61 | Commercial |
| seo automation tool | 480 | 19 | $14.26 | Commercial |
| ai keyword research tool | 170 | 68 | $21.44 | Commercial |
| ai keyword research | 140 | 54 | $66.34 | Commercial |
| automated keyword research | 90 | 58 | Not returned | Commercial |
Observed evidence: the exact overview cost $0.0180.
Interpretation: “seo agents” had much stronger measured demand than the noisy export made obvious. “Automated keyword research” was smaller but described this specific workflow. Difficulty did not settle the choice because page purpose and existing ownership still had to be checked.
Current results settled page ownership
Four live Google result pages supplied the missing context. The broad “seo agents” results were about AI agents, guides, products and real-world use. The “seo automation” results mixed tools, tutorials and workflow lists. Results for “AI keyword research tool” were dominated by standalone tools and commercial comparisons. “Automated keyword research” had a narrower mix of workflows, automations and tools.
Observed evidence: four result-page runs cost $0.0014 each, or $0.0056 in total.
Interpretation: broad SEO-agent education, exact tool selection and this field-tested workflow are different jobs. They should not be forced into one page, but they should not create duplicate owners either.
A third provider helped the agent support the article
Once the page decision was clear, the agent used two Exa searches to find current source material for the draft. A broad search exposed the language practitioners use for supervised AI workflows. A narrower search found Google Search Central’s guidance on using generative AI content and creating helpful, reliable, people-first content.
Observed evidence: the two source searches cost $0.0196 in total.
Interpretation: finding a source is not the same as approving it. The agent widened the search and brought back candidates; editorial review checked which sources were credible, relevant and accurately represented before they were used.
The content decision
| Search intent | Canonical owner | Decision |
|---|---|---|
| How SEO agents work | SEO agents guide | Strengthen the existing owner. Do not publish a second broad guide. |
| Automated keyword-research workflow | This field test | Publish the method, evidence, collaboration and limitations. |
| Keyword-research tool or API | Keyword discovery capability | Keep exact tool intent on the live capability page. |
| Available SEO-agent toolkit | SEO and search category | Let the category own the current set of capabilities. |
The total paid evidence cost was $0.0800: $0.0548 through DataForSEO, $0.0056 through Serper and $0.0196 through Exa. Free discovery and contract inspection calls are not included. The recommendation was to strengthen one existing owner, add this narrower proof page and reject duplicate content.
Three rounds of editorial feedback improved the article
The first review checked the research conclusion and page-ownership plan before drafting. It confirmed that the field test was distinct from the broader SEO-agent guide and that the workflow, not a single keyword, was the useful story.
The second review asked for plainer English, clearer supervision boundaries, a third provider in the workflow, no public run IDs, a cleaner image and desktop tables without horizontal scrolling. The third reduced repeated emphasis on the reviewer and kept the agent’s efficient automated execution at the centre of the story. Each round changed the draft.
A reusable supervised SEO-agent workflow
- Define the decision. Give the agent the business question, audience, market, conversion and existing pages.
- Let the agent assemble the toolchain. It can discover suitable tools, inspect live contracts and prices, and start gathering evidence.
- Expand once. Generate enough language to reveal the space, then stop treating the export as a backlog.
- Reject semantic noise. Remove wrong meanings, wrong audiences and jobs the product cannot satisfy.
- Validate and inspect. Check exact demand, expected page type, competing quality and credible supporting sources.
- Review the recommendation. Challenge the interpretation, decide page ownership and approve or reject the draft.
- Revise before publication. Record substantive feedback and make the final editorial check explicit.
Use Scrollport to research [TOPIC] for [PRODUCT] among [AUDIENCE] in
[MARKET AND LANGUAGE]. Help me decide [DECISION]. The desired conversion is
[ACTION], and the existing possible page owners are [URLS].
Generate related search language, then remove wrong meanings, wrong audiences
and needs the product cannot satisfy. Validate exact demand only for the
shortlist. Inspect current result pages where page type or intent could change
the recommendation. Find current, credible sources for any claims the content
will need to support.
Return observed evidence, interpretation and recommendation separately. Assign
one canonical owner to each shared intent, flag cannibalisation, list rejected
options and explain what new evidence would change the decision. Inspect the
current tool contract and show the price before every paid run. Stop for editorial
review before drafting, and again before publication.An agent can use Scrollport’s keyword demand, Google result-page and web search capabilities to gather this evidence through one account. You still own positioning, source judgment, page purpose and publication.
What this field test does not prove
- It is one US English snapshot from 24 August 2026, not a permanent result.
- Search volume, CPC and difficulty are third-party estimates, not traffic or revenue forecasts.
- Result pages vary by time, location, device and personalisation.
- The ten-query shortlist was deliberately small and did not test every possible phrase.
- The workflow does not prove that the chosen page will rank, earn citations or convert.
- The agent did not publish autonomously; the decision and article were reviewed.
Results and prices can change, so an agent should inspect the live catalog immediately before repeating the workflow. For the broader operating model, read the SEO agents guide. To reproduce the research, start with the live SEO and search toolkit.