Run discovery broad and measurement narrow. Fan out from one seed through what Google already suggests, clean and group everything that comes back, and spend the expensive volume lookups only on the head terms that survive. That order is the whole method, and it is what keeps an agent-run version of keyword research from burning money.
Keyword research is finding and analyzing the exact words and phrases people type when they look for information, products, or services. It matters for three reasons: it finds real demand, it matches intent (buy, learn, or reach a site), and it points content at pages that can actually earn traffic. The stakes are not small. Ahrefs reports that about 90.63% of the pages in its Content Explorer index, a study of more than a billion pages, get no organic traffic from Google at all. Most of those pages answered a question nobody was asking.
The Standard Method
The guides from Ahrefs, Semrush, Conductor, and HubSpot teach the same four steps:
- brainstorm seed topics from what you sell;
- expand the seeds into specific queries with a planner or keyword explorer;
- analyze volume and difficulty;
- determine intent, so each page matches what its searcher wants.
That spine is sound. What changes when an agent runs it is the economics. A person clicking through a dashboard pays a monthly seat. An agent calling APIs pays per request, and a careless loop can make thousands of them before anything is written.
Discovery Broad: Fan Out from One Seed
Do not ask a model to invent keywords. A model asked to brainstorm produces plausible phrases, and plausible is not the same as searched. Real discovery reads what search already surfaces, then organizes it.
Start with one seed and expand it through three live surfaces:
- the autocomplete suggestions Google offers as someone types, which you can pull straight from Google's suggest endpoint;
- the People Also Ask questions around the topic;
- the related searches at the bottom of the page, which show lateral interest.
Each surface returns branches. Feed the branches back in as new seeds and the tree multiplies fast. Builders running these pipelines report depth three as the practical cutoff: new terms climb sharply through the third level, then duplicates take over, and deeper levels cost many more requests for few new phrases. Start there and adjust for your own seeds.
Clean, Then Cluster
A flat list of two thousand strings is not a plan. Normalize it first: lower-case everything, strip punctuation, and collapse near-duplicates that differ only by word order or a plural. In reported runs that step alone removes about a third of the raw list.
Then group by meaning. Each cluster gets one head term to target and a set of members to cover as sections and questions on the same page. Tag each cluster with its intent by looking at what the results page already rewards, guides, comparisons, or product pages. The rule is simple: one head term per page, with the cluster members covered inside it as sections and questions.
Validation Narrow: Spend Volume Only on Survivors
Suggestions are observed. Volume numbers are modeled.
A suggestion shows that a query exists and is being surfaced now. A volume figure is an averaged estimate, useful for deciding what to publish first and a poor basis for false precision. When a fresh long-tail term comes back with thin or missing volume, builders fall back on the page itself: its presence in autocomplete means real searches are behind it, a deep set of related questions marks sustained interest, and a crowd of advertisers signals commercial intent.
So run a stack, not a single tool. Discover broadly with fan-out that is cheap and live. Measure narrowly, with a volume provider, only on the shortlist that survived clustering.
One builder's local keyword tool shows why the order matters. Its guided report makes up to 19 calls per keyword, for volume, cost per click, difficulty, intent, trends, competing pages, and related terms. Those are exactly the metrics the guides teach. Run that report against every raw branch of a fan-out tree and the bill compounds before a single brief exists. Run it only against head terms and the cost stays proportional to what actually gets published.
Three Habits That Keep the Bill Proportional
- Cache every response per keyword and location, so a repeated lookup never bills twice.
- Dedupe before you fire a batch, so the tree never pays for the same query from two branches.
- Back off instead of retrying hard. On a rate limit or a timeout, wait, skip, and continue; a burst of retries is the fastest way to pay for nothing.
Builders who switch from model-invented phrases to rows returned by an API describe the change in the same terms: no invented numbers, just real search rows to build on. That is worth more than any single tool the loop replaces.
Research as a Loop the Team Owns
Wired this way, research stops being a dashboard export and becomes a loop: the agent gathers data, checks who ranks, drafts a brief, and hands it to a person for review.
Seed the loop from your own ground truth. Search Console shows every query your site has been seen for, including the ones that earn impressions but few clicks, which are often the cheapest wins on the list because Google already considers you relevant. Harvest those as seeds, expand them with a suggestions and results provider, then confirm the survivors with volume. Keep Search Console as the baseline: it tells you where you already stand, not where the rest of the demand is.
qequals covers the discovery half of that loop: live Google suggestions and live results for any seed, callable by an agent that has read its llms.txt file, with suggestions available on a key minted without an account. It has no volume model, so pair it with a volume provider for the final shortlist.
Fan out broad, measure narrow, cache everything.
