Direct Answer
A useful AI keyword research tool does more than expand one word into a large spreadsheet. It should use current search data to discover demand, remove irrelevant meanings, inspect each surviving SERP, classify intent and format, group equivalent queries, score opportunities transparently, and generate briefs and internal-link recommendations only after those checks pass.
The advantage is not automatic content. It is compression: hours of repetitive research become a structured decision queue that an experienced operator can inspect. Sam Oh's demonstration reportedly processed 11,800 houseplant keywords in about 20 minutes. That is a creator test on one setup, not a guaranteed runtime or quality benchmark.
Watch: This AI Tool Does Keyword Research for You
Credit: This guide is based on Sam Oh's Agent A build for Ahrefs, the supplied transcript, and current Ahrefs and Google documentation. The staged architecture, approval gates, claim controls, and validation checklist are editorial additions.
What the Tool Should Produce
| Output | Question it answers | Required evidence |
|---|---|---|
| Executive summary | Where is the clearest opportunity? | Scope, market, data date, counts, assumptions, and exclusions |
| Vetted keyword table | Which queries deserve review? | Volume, traffic potential, difficulty, intent, SERP, and relevance |
| Topic clusters | Which queries belong on one page? | SERP overlap, shared intent, entities, and parent topic |
| Competitor gaps | What do credible competitors cover that we do not? | Relevant domains, ranking URLs, positions, and business fit |
| Go / Maybe / Skip queue | What should a strategist inspect first? | Explainable factors, flags, and missing data |
| On-demand briefs | What would a genuinely useful page need? | Current top results, weaknesses, original angle, entities, and sources |
| Internal-link map | How should related pages connect? | Existing URLs, cluster role, destination relevance, and anchor rationale |
The Pipeline: Evidence First, AI Second
- Define the brief: niche, audience, country, language, business model, conversion, existing domain, exclusions, and risk level.
- Generate broad seeds: categories, products, problems, use cases, brands, jobs, and entities rather than long-tail article titles.
- Expand with live data: use Keywords Explorer and relevant competitor rankings from Site Explorer.
- Normalize and deduplicate: preserve metrics and provenance while merging exact duplicates.
- Filter relevance: classify ambiguous meanings and record why a candidate was removed.
- Inspect current SERPs: classify intent, page type, result ownership, publisher concentration, and special formats.
- Cluster by intent and SERP overlap: do not create separate pages merely because wording differs.
- Score and flag: calculate an auditable priority score, then apply branded, publisher, tool, safety, and evidence flags.
- Generate briefs on demand: spend model and data budget only after a reviewer approves a topic.
- Map links against the real site: recommend contextual links between approved existing or planned pages.
Ahrefs says Agent A has access to the full Ahrefs dataset and more endpoints than its public API or MCP. Its own published workflow describes the same core sequence: keyword expansion, candidate vetting, SERP analysis, topic clustering, optional briefs, and background jobs. If Agent A is unavailable, the architecture can also be built with the DataForSEO MCP workflow, but quotas, field names, costs, and capabilities will differ.
1. Fix Seed Selection Before Expansion
A seed is a discovery input, not a page title. The first version in the video proposed narrow phrases such as a product review or an exact brand comparison. Those may be target queries, but they are weak seeds because they produce a small and biased idea set.
Use broad but unambiguous concepts, then supplement them with competitor mining. For golf, categories such as equipment, apparel, courses, handicaps, and carts create useful branches. Competitor top pages can reveal less obvious demand that the standard category list misses.
Seed acceptance test
- Does it describe a category, entity, problem, job, or audience need?
- Does expansion produce enough relevant demand to justify the data call?
- Is the term so broad or ambiguous that it needs a qualifying context?
- Did it come from first-party knowledge, customer language, a competitor pattern, or model speculation?
2. Add Guardrails Before Calling a Keyword Winnable
The video's early prototype recommended navigational terms such as Topgolf and product-brand queries where official properties dominated the results. The corrected version added rules that a production tool should make visible:
| Guardrail | Signal | Default action |
|---|---|---|
| Branded SERP | Official site, profiles, support, stores, or product pages dominate | Skip for generic editorial acquisition; review only for legitimate comparison or support intent |
| Publisher dominance | Results are concentrated among large editorial brands | Maybe; require a clear evidence advantage and realistic authority assessment |
| Transactional dominance | Category, product, local pack, or marketplace results satisfy the query | Match the format or skip the article |
| Free-tool opportunity | Calculators, checkers, analyzers, templates, or generators rank | Route to a product or engineering brief |
| Video-heavy SERP | YouTube or video results occupy meaningful positions | Add a video deliverable rather than forcing text alone |
| High-risk topic | Health, safety, financial, or legal consequences | Require qualified review, authoritative sourcing, and stricter claim controls |
3. Use AI for Ambiguity, but Preserve an Audit Trail
Terms such as driver and iron can describe golf equipment, jobs, software, films, people, or household products. In the demonstration, the relevance stage reportedly removed 2,456 off-topic terms including delivery-driver and entertainment queries.
That is a strong classification use case, but discarded data should remain inspectable. Store the keyword, interpreted meaning, confidence, exclusion reason, source seed, and model version. Send low-confidence rows to review instead of silently deleting them. A bad false positive wastes review time; a bad false negative can erase an entire valuable cluster.
4. Make the Opportunity Score Explainable
Volume, intent, and difficulty are useful inputs, but a single opaque score hides too much. Separate the components so a strategist can disagree with the weighting.
| Factor | What to measure | Common mistake |
|---|---|---|
| Business value | Connection to product, lead, revenue, retention, or strategic authority | Ranking irrelevant traffic because volume is high |
| Demand | Volume, trend, traffic potential, geography, and cluster size | Treating one monthly estimate as exact demand |
| Intent fit | Whether the business can satisfy the dominant task and format | Publishing a blog post for a tool or local-service SERP |
| Competition | Page-level links, site strength, content quality, SERP stability, and specialist depth | Using keyword difficulty as the whole decision |
| Evidence advantage | Original data, experience, access, expertise, product, or process the page can add | Assuming a longer rewrite is differentiated |
| Production cost | Research, expert review, design, engineering, maintenance, and promotion | Counting AI generation as the total cost |
Use the weak-page opportunity method when page-level evidence suggests a strong domain is ranking accidentally. Use the BID framework to keep business potential, intent, and difficulty visible.
5. Generate Briefs Only After Approval
A useful brief can summarize the current SERP, dominant format, intent, weak coverage, questions, entities, credible sources, title options, URL structure, and internal-link candidates. It should also state what only the author can provide: testing, interviews, screenshots, data, expert review, or a product experience.
The live houseplant example proposed the phrase "vet reviewed" even though no veterinarian had reviewed the page. That is not a small copy error; it is a false credential and a publishing stop. The same rule applies to claims such as tested, certified, audited, medically reviewed, independently verified, or used by a named customer.
Mandatory brief gates
- Remove unsupported credentials, testing claims, statistics, and named endorsements.
- Require source URLs and distinguish source facts from model suggestions.
- Flag health, legal, financial, safety, and animal-care claims for qualified review.
- Do not use competitor word count as a writing target; write enough to complete the task.
- State the original contribution before drafting begins.
- Prevent publication when the only value is a reformatted summary of ranking pages.
6. Build an Internal-Link Map From Meaning, Not Geometry
A hub-and-spoke diagram is useful when it reflects real page relationships. The pillar should answer the broad task; spokes should satisfy distinct sub-intents; sibling links should exist only where they help a reader continue the journey.
Each recommendation should include the source page, destination page, proposed anchor, placement reason, page status, and whether the URL already exists. Do not generate reciprocal links by default. A visual line between two nodes is not a reason to add a link, and orphan prevention is not a license to manufacture irrelevant anchors.
A Safer Build Sequence
| Phase | Build | Acceptance test |
|---|---|---|
| 1. Thin slice | One niche, one country, up to five seeds, 200 candidates | At least 50 reviewed rows with provenance and reproducible metrics |
| 2. Relevance | Meaning classifier and exclusion log | Human-labeled sample reports precision, recall, and disputed rows |
| 3. SERP review | Intent, page type, brand, publisher, tool, video, and risk flags | Reviewer agreement on a stratified sample |
| 4. Clustering | SERP overlap plus semantic and intent checks | No obvious cannibalization or unrelated merged topics |
| 5. Scoring | Visible factors, weights, missing-data penalties, and reasons | Every score can be explained without asking the model |
| 6. Briefs | On-demand generation for approved topics | No unsupported claims; sources and original contribution are explicit |
| 7. Link map | Existing and planned URLs with contextual anchors | Every proposed link helps a reader and resolves to a valid page plan |
| 8. Pilot | Publish a small, human-reviewed cluster | Track indexing, rankings, conversions, citations, maintenance, and total cost |
Google's current guidance is compatible with AI-assisted research and structure, but warns against generating many low-value pages primarily to manipulate rankings or AI responses. The roadmap should therefore end with a small evidence-backed pilot, not a "publish all" button.
Watch the Related Videos
Complete AI SEO Course for Beginners
This AI Tool Generates Viral Thumbnails
Complete SEO Course for Beginners
Video Chapters
| Time | Section | Time | Section |
|---|---|---|---|
| 00:00 | One word to a keyword roadmap | 10:58 | Intent and the verdict engine |
| 01:22 | The first build prompt | 13:04 | Branded, publisher, and tool guardrails |
| 02:29 | Agent A and Ahrefs data | 15:20 | Filtering ambiguous keywords |
| 05:03 | The complete research process | 16:46 | Content briefs |
| 08:08 | Fixing seed selection | 20:10 | Hub-and-spoke map |
| 09:47 | Mining competitor top pages | 21:22 | Houseplant live result |
Sources and Further Reading
- Ahrefs: This AI Tool Does Keyword Research For You (primary video and supplied transcript)
- Ahrefs: 11 Ways to Automate SEO With Agent A
- Ahrefs: Agent A for Agencies
- Ahrefs: Keywords Explorer
- Ahrefs Academy: Introduction to Keywords Explorer
- Google Search Central: Creating Helpful, Reliable, People-First Content
- Google Search Central: Scaled Content Abuse
- Google Search Central: Optimizing for Generative AI Features
YouTube lists the primary video's publication date as 8 July 2026. This article was reviewed on 27 September 2026. Product access, included data, model choices, pricing, quotas, and interfaces can change. The runtimes, keyword counts, filtering results, and development spend described above are creator-reported examples, not guaranteed outcomes.