Here is how to do keyword research using AI without producing a list of terms nobody searches: use the model for the parts it is good at — generating seeds, clustering by intent, and reading SERP patterns — and use a real search index for every number. Models cannot know search volume. They can only guess, and a confident guess formatted as a table is the single most common way AI keyword research goes wrong.
The six-step process below takes about two hours for a new site and produces a prioritized build list rather than a spreadsheet.
Step 1: Use AI to generate seeds, not keywords
Seeds are the three to six core concepts your business sells against. Ask the model to produce them from your positioning: paste your homepage copy, your two best product pages, and a description of who buys, then ask for the concepts a buyer would search at each stage — problem-aware, solution-aware, and vendor-aware.
This is a job models do well, because it is a reasoning task over text you supply. Expect output like “warehouse inventory software”, “cycle counting process”, “barcode scanning ROI” rather than long-tail phrases. Long tail comes from the index in the next step, not from the model.
Step 2: Expand seeds against a real index
Feed those seeds into a keyword tool that queries an actual search index and returns volume, difficulty, CPC, global volume, and the SERP features present. Multi-seed exploration matters here — running four seeds together surfaces the overlap between them, which is usually where the best terms live.
Set the country index before you run anything. This is the step people skip, and it silently ruins the whole exercise: a Singapore or UK site queried against the US index returns near-empty data and makes a healthy market look nonexistent. SEO Rocket returns up to 150 ideas per search with country-specific indexes, and filtering, sorting, and CSV export are free after the first query.
Treat every number as an estimate. Volume figures are typically modeled twelve-month averages, so seasonal terms read low out of season. Difficulty is a proprietary formula, useful for ranking terms against each other and not much else.
Step 3: Let AI cluster by intent
Now hand the raw export back to the model. This is where AI earns its place — sorting 800 rows into intent groups is tedious for a human and fast for a model. Ask for four buckets:
- Informational — learning the topic. Volume-heavy, conversion-light, good for authority and internal linking.
- Commercial investigation — “best”, “vs”, “alternatives”, “review”. Lower volume, much higher value, frequently under-served.
- Transactional — “buy”, “pricing”, “near me”. Product and service pages, not blog posts.
- Navigational — brand terms. Ignore unless the brand is yours.
Then verify a sample by hand. Open the live results for six terms across the buckets. If a term you labeled informational returns nine product pages, the model was wrong and no article will rank. Ten seconds of SERP checking per term catches the errors that would otherwise cost you a month of writing.
Step 4: Add the competitor layer
Seed-based research finds what you thought of. A content gap report finds what you did not. Run your domain against three to five close competitors and pull the terms they rank for and you do not, with per-rival position columns so you can see whether one site owns a topic or all of them cover it.
Read the position columns, not just the presence of a row. If all five competitors rank between #3 and #9 for a term, the topic is established and page one is crowded but reachable. If one competitor sits at #2 and the rest are past #40, you are usually looking at a term where one strong page owns the topic — which is good news, because a single page is easier to outbuild than five.
Rows where three or more competitors rank and you do not are the strongest signals in the entire dataset — proven demand, proven that content can rank for it in your niche. Prioritize those above anything the model generated on its own.
Step 5: Judge winnability against the weakest page-one result
Difficulty scores are a starting filter, not a verdict. The real test is the site sitting at position ten. Open the SERP and look at that page specifically: how long is it, how many referring domains does the URL have, does it actually answer the query? If the weakest page-one result is a thin 600-word post on a domain comparable to yours, that term is winnable regardless of what the difficulty score says.
Benchmarking against the median or the market leader produces paralysis. You are not trying to beat the strongest competitor. You are trying to be better than whoever is currently tenth.
The mistakes AI makes that you have to catch
Four errors recur often enough to check for every time.
- Invented volumes. If you ask a model for search volume, it will produce a number. That number is fabricated. Any figure not traceable to an index export should be deleted, not adjusted.
- Merged near-duplicates. Models happily collapse “project management software” and “project management tools” into one row. Those are frequently different SERPs with different intent. Check before consolidating.
- Stale assumptions. A model’s sense of what ranks comes from training data that predates today’s results. It cannot see that an AI Overview now answers your target query, or that a competitor published a definitive guide last month.
- Over-clustering. Ask for tidy groups and you get tidy groups, including ones that only exist because you asked. Verify a cluster by checking whether the top results genuinely overlap across its terms; if the URLs are entirely different per keyword, it is not one cluster.
None of these are reasons to skip AI. They are reasons to keep the index as the source of numbers and the live SERP as the source of truth about intent, with the model doing the sorting in between.
Step 6: Turn the list into a build order
Score the survivors and publish top-down. A simple rubric that works:
- Proven demand — appears in the competitor gap, or has meaningful volume in your country index
- Winnable — the page-one floor is beatable with the resources you have
- Commercially relevant — a reader of this page could plausibly become a customer
- Cluster fit — strengthens a topic group you are already building rather than starting an island
Group the winners into clusters of five to eight pages around a pillar topic. Clusters compound: internal links pass context between them, and complete coverage of a topic is itself a signal. Scattered high-volume posts across unrelated topics rank more slowly, which is why volume-sorted spreadsheets so often underperform.
What to do with the finished list
Save it somewhere that feeds the rest of your workflow rather than a dead spreadsheet. In SEO Rocket the shortlist becomes a project keyword pool that both the AI writer and the rank tracker read from, so the terms you researched are the terms you write about and the terms you monitor — no re-entry, no drift.
Then set expectations honestly. Even a well-chosen keyword takes eight to sixteen weeks to show meaningful movement on a competitive term, and daily fluctuation of two or three positions is normal noise throughout. Research quality determines your ceiling; patience determines whether you ever reach it.
