How can I find keyword opportunities?

Most keyword research produces a spreadsheet of terms with volume attached and no indication of which ones you could actually win. That is a list, not an opportunity set.

What counts as an opportunity

An opportunity needs three things at once:

• Demand — people search it, or ask an assistant about it.
• Winnability — you could realistically reach the top five within a quarter or two.
• Value — ranking would produce something you care about.

Drop any one and the term is a distraction. High volume you cannot win is a vanity target. High winnability with no commercial relevance is traffic that never converts. The seven sources below tend to produce terms with all three, which is why they beat starting from a volume-sorted export.

1. Striking distance queries

The highest-return source, and it uses data you already have.

In Search Console, export the last three months of queries and filter to average position between 8 and 20. These are terms where Google has already confirmed you are relevant, but you sit far enough down that you receive almost nothing. The page exists; the gap is depth rather than existence.

This is the single most reliable source of quick wins in SEO. Improving an existing page on a proven term outperforms publishing a new page on an unproven one, almost every time, at a fraction of the cost.

2. High impressions, low click-through

Same export, different sort: impressions descending, CTR ascending.

The top of that list is pages being shown to thousands of people who consistently choose someone else. The ranking work is already done. What is failing is the title and description — the twenty seconds of writing that converts a ranking into a visit.

This is the cheapest traffic available in SEO, and it is routinely ignored because it does not feel like keyword research. It is: you are finding demand you have already earned and are failing to collect.

3. Competitor gaps

Search Console has a structural blind spot — it only knows terms you already rank for. Everything you are missing entirely is invisible to it, which is precisely the set you most need.

Closing that requires competitor data. Take three to five genuine search competitors — the sites that rank for your topics, not necessarily your commercial rivals — and find terms where two or more rank and you do not. Overlap matters: a term several competitors cover is validated demand, whereas one competitor covering it is often a one-off.

Our keyword gap analysis guide covers the method in detail, including how to cluster raw gaps into the handful of pages that would actually close them.

4. Question mining

Questions are where specific, winnable, high-intent demand lives, and they are cheap to find.

• Autocomplete. Type your core term plus each letter of the alphabet. Tedious, free, effective.
• People Also Ask. Expand a box and it generates more. Fifteen minutes produces a substantial map of the question space.
• Reddit and industry forums. Real phrasing from real people, including the objections and edge cases marketing copy never covers.
• Your own support inbox. The most underused source in the discipline. Every recurring customer question is a page nobody in your category has written properly.

Question-shaped content has become disproportionately valuable, because a page answering one question completely is the format both featured snippets and AI engines prefer to extract.

5. The prompt gap

The newest source, and invisible to every keyword tool.

When a buyer asks ChatGPT “what’s the best X for Y” and it names three companies, no search happened on a results page. No volume was recorded anywhere. But a shortlist was formed, and if you are not on it you never entered the funnel.

Finding these requires running your category’s real questions against the assistants and recording who gets named and what gets cited. The questions where competitors appear and you do not are prompt gaps, and each one is a content brief. Our guide to increasing AI citations covers what to do with them.

6. Seasonality and timing

Annual demand is predictable and most teams still publish late. Anything with a recurring peak — tax deadlines, back to school, regulatory changes, industry events — should be live and indexed three months before the peak, not during it.

Check the twelve-month volume curve rather than the average. A term averaging 500 searches a month might be 200 for eleven months and 4,000 in one, and that distribution changes the whole plan.

7. Internal site search

People who are already on your site, telling you in their own words what they cannot find.

Every internal search with no useful result is a page you should have. This source has zero acquisition cost, perfect intent signal, and is checked by almost nobody. Turn it on if it is not already recording.

Scoring what you find

You will end up with more candidates than capacity. Score each on four axes and take the top of the ranked list rather than the top of the volume list.

The last axis is the one most often skipped and it matters most. Twenty related terms covered thoroughly build topical authority that makes the next twenty cheaper. Fifty scattered terms build nothing and compete with each other. Depth beats breadth, and it beats it by more every year.

Turning a list into a plan

A scored list is still not a plan, because keywords and pages are not the same unit. Twenty terms rarely need twenty pages — they usually need five.

Cluster by intent, not by string similarity. “How much does X cost”, “X pricing” and “is X expensive” are one page. “X pricing” and “X alternatives” are two, despite sharing a word. The test is whether one page could satisfy someone searching either term without compromising for both. If it would have to hedge, split it.

Verify the cluster against the results page. Search two terms you have grouped together. If the top five results are largely the same URLs, Google agrees they are one topic and one page will serve both. If the results are entirely different, they are separate pages regardless of how similar the words look.

Assign each cluster to an existing page or a new one. Default to existing. A cluster that maps onto a page you already have is an improvement task with a much shorter payback than a publishing task.

Sequence by compounding. Start with the cluster closest to what you are already known for, then work outward. Authority radiates from strength; a site known for one thing can extend into adjacent topics far more cheaply than into unrelated ones.

What a finished plan looks like

For each cluster: the primary question, the supporting queries it covers, whether it is a new page or an improvement, the specific gap it closes against the current top results, and one thing it will contain that nobody else has. That last field is not optional decoration — it is the difference between a page that gets published and a page that gets cited.

Original data is the most reliable way to fill it. The Princeton and IIT Delhi GEO study found that adding statistics, quotations and cited sources were the highest-performing methods for being quoted in AI answers, improving visibility by roughly 30 to 40 percent. The same specificity that earns an AI citation earns a featured snippet and a link, so it pays three times from one piece of work.

One practical warning about volume figures: treat them as order-of-magnitude signals rather than forecasts. Every tool models them from clickstream samples, they vary substantially between providers for the same term, and conversational search has dispersed a great deal of demand into phrasings too long and too varied to be reported at all. A term showing zero volume is not necessarily a term nobody searches — it may simply be one nobody has measured.

Finally, revisit the whole exercise quarterly. Rankings drift, competitors publish, demand disperses into new phrasings, and a gap list older than about ninety days should be re-validated before you commit a quarter of writing capacity to it. The quarterly cadence matters more than the thoroughness of any single pass: a rough list refreshed four times a year beats an immaculate one built once and quietly followed for eighteen months.

Sources

• Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande — “GEO: Generative Engine Optimization”, ACM SIGKDD 2024 (arXiv:2311.09735); Statistics Addition, Quotation Addition and Cite Sources as top-performing methods, ~30–40% relative visibility improvement.
• Google Search Console — the only first-party source of query, impression, click and position data.
• SERP feature tracking, 2026 — AI Overview presence averaging ~32.5% of monitored queries within a 31.02%–34.40% band.

Frequently asked questions