Keyword Research with Google Search Console: 7 Methods

Can you do keyword research with Google Search Console?

Yes, for demand you already touch. Keyword research with Google Search Console means mining the Performance report for queries that already show your pages: questions, long-tail queries, queries that land on a loosely matching page, and queries that are new this period. You filter them with regex, split them by country and device, cluster them into topics and turn each cluster into a content brief. Search Console has no search volume and can’t show keywords you never appear for, so pair it with a keyword tool when you need to size a topic from scratch.

These are real searches on which Google already chose to show your site, so the topic fit is proven. The work is deciding whether each query deserves its own page, a section, or nothing.

We make GSC MCP Server, a read-only MCP server that brings Search Console and Bing Webmaster Tools data into Claude, ChatGPT, Cursor and other assistants, so the prompts below use its tool names. Every method also works in the Search Console interface, with more clicking and a 1,000-row ceiling.

Impressions as a demand signal: what to trust

Impressions are the nearest thing Search Console has to demand, but they count how often your page was shown, not how often a query was searched. Google’s John Mueller put it plainly in 2022: “It’s not the search volume.” Keep these traps in mind before you rank anything by impressions.

• Impressions depend on your position. Google counts an impression when your link appears on the current page of results. At position 3 your link is on the first page people see; at position 35, it registers only for the few who go that deep. The same query can gain impressions as you climb while demand stays flat, so a deep-ranking query’s impressions are a floor, not an estimate.
• Rare queries are anonymised. Google’s deep dive on performance data explains that queries searched by no more than a few dozen people over two to three months are hidden. They count in chart totals but drop out as soon as you apply a filter. The long tail is hit hardest.
• The interface stops at 1,000 rows. Tables and exports show at most 1,000 rows, and Search Console stores top rows rather than all of them. The API reaches up to 50,000 rows per day per search type; our Search Console API guide covers paging through them.
• Year-over-year impressions are distorted. Google’s data anomalies page says a logging error affected impressions, CTR and average position from 13 May 2025 to 27 April 2026, and that the fix may show up as a drop in impressions. Clicks were not affected. For year-over-year work, compare clicks, or simply whether a query appeared at all.
• Image-led searches have no queries. The Query dimension isn’t available for multimodal searches, such as Google Lens, so that demand never reaches a keyword list.

Method 1: find question keywords in Search Console

Question queries tell you exactly what people want answered, and they map neatly onto headings. In the interface:

1. Open Performance → Search results and set the date range to the last 3 months.
2. Click + Add filter, choose Query, then Custom (regex) and Matches regex.
3. Paste ^(who|what|when|where|why|how|which|can|does|do|is|are|should|will)\b and sort the Queries table by impressions.
4. Click a question, then the Pages tab, to see which URL Google shows for it.

The pattern is anchored, so it only catches queries that start with a question word. People also type “search console how to add user” or ask without question words at all (“crm cost small team”). The Search Console regex cheat sheet has the unanchored version and patterns for comparison, price and problem queries.

For each question, decide: if the page Google shows already covers the topic, add a heading and a two- or three-sentence answer; if no page does, carry it into method 3. Write answers for readers, not for markup: Google’s data anomalies page notes that FAQ rich results stopped appearing in Search on 7 May 2026.

The interface makes it awkward to exclude your brand at the same time. The API takes several filters, all of which must match, so an excludingRegex filter such as \b(acme|akme)\b strips branded questions out. Our guide to branded vs non-branded queries shows how to build that brand pattern.

Each 1,000-row page is one tool call. For more ready-made versions, see our Search Console prompts for Claude and ChatGPT.

Method 2: find long-tail keywords by word count

Search Console has no word-count filter, but regex can count words. ^(\S+\s+){3,}\S+$ matches four words or more, and ^(\S+\s+){6,}\S+$ matches seven or more, the conversational end of the tail. That end is growing in importance: Google’s metrics documentation says a follow-up question in AI Mode counts as a new query, so long, multi-clause queries now appear alongside classic keywords. To track AI surfaces themselves, see how to measure AI search visibility in Google and Bing.

Long-tail queries are rarely worth a page each. Their value is in the modifiers they share:

• Audiences: “for freelancers”, “for small charities”, “for property managers”.
• Jobs and use cases: “to track invoices”, “for onboarding”, “without coding”.
• Integrations and alternatives: “with Shopify”, “instead of Excel”.
• Constraints: “free”, “offline”, “under £50”, a place name.

A modifier that recurs across dozens of queries is a subtopic. If one page already gets those impressions, it probably needs a section; if the queries scatter across unrelated pages, you may need a new one. Remember the anonymisation trap: the rows you see are a sample of the tail, so compare modifiers with each other rather than calculating a share of all traffic.

With an assistant, run query_search_analytics with the query and page dimensions and the four-word pattern as an includingRegex filter, then ask it to list each recurring modifier with its number of queries, total impressions and the pages that receive them.

Method 3: queries with impressions but no dedicated page

Google sometimes shows your site for a query using a page that only loosely matches it: the home page, a category page, or a broad guide that mentions the topic once. The impressions prove relevance; the weak match usually caps the ranking. The signs:

• The query’s main words aren’t in the ranking URL or title, or the page serves a different intent (a feature page for a “how to” question).
• Average position is mostly beyond 20. Queries at 8–20 on a well-matched page are quick wins for the existing page instead; the striking distance keywords playbook handles those.
• The ranking URL changes between periods, or several of your URLs share the impressions. Run find_keyword_cannibalization before creating anything, and see the keyword cannibalisation guide if two pages are competing.

In the interface you check this one query at a time (click the query, then the Pages tab). Through the API, a single request with the query and page dimensions returns up to 25,000 query–page pairings per request, minus anonymised queries. Our server pages through them 1,000 rows at a time, and each 1,000-row page is one tool call. Here is the kind of output to expect; the numbers are illustrative.

The rule: create a new page when the intent is distinct (a different audience, format or job), expand the existing page when the query is a sub-question of its topic, and ignore queries you rank for by accident. An assistant only sees URLs through our server, not your page content, so paste in your sitemap or a list of titles, or in Claude Code let it read the content files in your repository.

Method 4: spot emerging queries early

New queries are often the earliest sign of new demand, such as a product launch or a regulation. There are three ways to find them.

In the Performance report

Open the date filter, choose the Compare tab and pick “Compare last 3 months year over year” or “Compare last 3 months to previous period”. Turn on the Impressions metric and sort the Queries table by the impressions difference. Queries with nothing in the earlier period are new.

In Search Console Insights

The Insights report has a “Queries leading to your site” card with a Trending up view, and Google’s launch post suggests trending-up queries as a source of content ideas. Two limits: trending items are ordered by the absolute change in clicks, not the percentage, so big existing queries crowd out small new ones; and a query with impressions but no clicks yet won’t trend at all.

With an AI assistant

compare_search_periods with dimension: "query" compares the top 1,000 queries by clicks in each period, using previous_period or previous_year (the same weekdays 52 weeks earlier), and returns the biggest gainers and losers by click change. Gainers with no previous data are new queries. Like Insights, it ranks by clicks, so for brand-new queries that only have impressions, pull both windows with query_search_analytics and compare the lists.

A year modifier is a useful extra filter: \b202[67]\b finds people planning ahead (“best crm 2027”). Roll it forward each January. Before you call any query emerging, check its position: if impressions rose because you climbed from 40 to 12, that is ranking, not new demand. Rising impressions at a stable position are the real signal. The opposite pattern, a page steadily losing clicks, calls for a content decay audit.

Both 90-day windows in that prompt start after 27 April 2026, so their impressions are comparable. For a year-over-year version, judge on clicks and on whether a query appeared at all.

Method 5: split queries by country and device

The same topic is often searched differently in different markets. Add the Country dimension (in the interface, the Countries tab, or a Country filter followed by the Queries tab) and look for:

• Queries whose impressions come mostly from one country that isn’t your main market: a candidate for a localised page or section.
• Spelling variants (optimise and optimize, colour and color) that tell you which spelling each market uses in titles and headings.
• Currencies, units, local laws, tax terms and place names that a generic page doesn’t address.
• Queries in other languages, which signal unserved demand.

The Device dimension (desktop, mobile, tablet) is about format rather than topic. If a cluster’s impressions are overwhelmingly mobile, put the answer first and keep tables narrow; if desktop dominates, a downloadable template or a detailed comparison may fit better.

With an assistant, ask query_search_analytics for the query and country dimensions, then for each top query’s share of impressions by country and any British and American spelling pairs. Countries use ISO 3166-1 alpha-3 codes (such as gbr or usa) and devices are DESKTOP, MOBILE or TABLET.

Method 6: Discover and News for topic interest

Discover and Google News are not keyword sources in the usual sense, but they show which topics your audience engages with when nobody typed a query.

• Discover has no query data. The Discover performance report groups by page, country, date and Discover appearance, and only appears once a property reaches a minimum number of Discover impressions. Group your Discover pages by topic to see which subjects and angles earn feed clicks.
• The News tab in Google Search is a search type in the main Performance report, and it does have queries.
• Google News (news.google.com and the apps) has its own report with no query data; you can group it by page, country, device, date and News appearance, according to Google’s help page.

The useful comparison is between surfaces. A topic that does well in Discover but earns few web search impressions may be interesting but rarely searched: good for timely articles, weak for evergreen guides. The reverse suggests steady search demand without feed appeal. Our Google Discover guide covers the report in more depth.

Method 7: Bing queries as a second sample

Bing Webmaster Tools reports its own queries from a different set of searchers, so it works as a second, independent sample of demand. A query that is thin or missing in Google’s data can show up clearly in Bing’s, and a topic that appears in both is a safer bet. Bing’s Search Performance report now holds 16 months of data, matching Search Console.

Bing has no RE2 regex filter, so to apply the question or long-tail patterns, pull the queries and let the assistant match them. get_search_queries with provider: "bing" returns top Bing queries with clicks, impressions and position; compare_search_periods works for Bing too, though Bing may not serve data from a year earlier. Bing also has a free Keyword Research tool for terms you don’t appear for yet; our server doesn’t read it, so use it in the Bing interface.

Bing’s index matters beyond Bing too: see Bing Webmaster Tools and AI search visibility.

Cluster queries into topics and turn them into content briefs

Methods 1 to 7 produce lists of queries. Clustering turns them into decisions. Google now does some of this for you: Query groups, announced in October 2025, use AI to group similar queries in the Insights report’s “Queries leading to your site” card, with Top, Trending up and Trending down views. Google says they are only available to properties with a large volume of queries; they show clicks only, and they aren’t in the API.

For everyone else, an assistant can cluster the rows you pulled above. Ask it to group by intent (one page could satisfy every query in the group), not by shared words, and to give each cluster its total impressions, clicks, best position and the URLs that currently receive them. Each cluster then gets one of three outcomes: a new page, a new section on an existing page, or no action.

What goes in the brief

• Primary query and working title, chosen from the cluster’s highest-impression query at a typical position.
• Secondary queries and questions, written as the page’s H2s and H3s.
• Intent and format: guide, comparison, template, tool or landing page.
• Boundaries: the existing page it must not compete with, and what to leave to that page.
• Internal links from pages that already rank for neighbouring queries; see internal linking with Search Console.
• Success measure: the cluster’s impressions after 8 weeks, then clicks.

When you still need a keyword research tool

First-party data covers demand you already touch. You need something else when:

• The site is new and appears for few queries. Google notes that a new site can take up to a week just to start generating data.
• You want a topic you don’t appear for at all, where Search Console has nothing to show.
• You need search volume to size a topic, or difficulty and competitor rankings to judge whether it is winnable.

Free options come first. Google Keyword Planner gives average monthly searches with a Google Ads account, Google Trends shows relative interest over time (an index, not a volume), and Bing’s Keyword Research tool covers Bing demand. Paid suites such as Ahrefs and Semrush, or a pay-per-request API such as DataForSEO, add difficulty, competitor gaps and SERP data. Our comparison of Search Console MCP vs Ahrefs and Semrush MCP shows how to combine first-party and third-party data in one assistant.

The order matters: start from Search Console, where demand is proven for your site, then size the best clusters with a tool. By our count, one pass through all seven methods with GSC MCP Server takes roughly 15–20 tool calls, more if you page through large query lists. The Free plan includes 30 calls a month, and paid plans start at $9.

Frequently asked questions