Every keyword tool on earth is selling you a reconstruction. The originals are free. Google's autocomplete is live demand, straight from the query stream. Search Console is what your site already ranks for. Trends is the seasonality map. And Reddit is where your buyers say what they actually call the thing - before any SEO got near the wording.
We got this wrong first. Our first keyword doc lived in a paid tool export - 4,000 rows, volume columns, difficulty scores, and almost no sentence anyone would ever type. It took one humiliating realization to fix it: the tool knew what people search. It had no idea what our readers search. That difference is the whole post.
What follows is the five-step stack we use now. Total cost: zero. Total time: about an hour for a content cluster, and by the end you'll have a keyword list with buying language attached - which most volume exports still can't give you.
Step 1 - Mine Search Console for the keywords you already own
Open Performance, filter for queries where your average position sits between 5 and 20. These are your "almost" keywords: Google already trusts the page enough to show it on page one or two, and the gap between position 8 and position 3 is usually a content upgrade, not a backlink campaign.
Use the regex filter (the Queries tab supports regular expressions) to pull question-form queries - ^(how|what|why|when|can|does|is) - and you have next month's FAQ content planned in ninety seconds. Google states some rare or anonymized queries never appear in this report (source below), so treat it as your strongest sample, not a census. The receipts live here: this is the only keyword source in the stack that comes from Google's own serving logs about your pages.
Step 2 - Autocomplete and People Also Ask: Google's live demand feed
Type your seed term into Google. Before you press enter, the dropdown shows you the exact strings real searchers use, ordered roughly by frequency, refreshed continuously. This is a volume database that updates daily, costs nothing, and can't be stale. The alphabet method - seed + a, seed + b, seed + c - takes twenty minutes and surfaces the long-tail.
Here's why the long-tail matters, with a number: in Ahrefs' landmark study of search demand, 94.74% of all keywords get ten or fewer searches per month (source below). That study is from 2018 and we cite it deliberately as a landmark - no newer census has replaced it, and the structural fact it maps (search demand is a long tail, not a head) has only deepened as query diversity grew. The head terms in a paid tool are the 5% everyone fights over. Autocomplete hands you the 95% nobody's watching.
People Also Ask boxes do the same job for questions - and each expanded question seeds two more. Ten minutes of clicking gives you the question tree for the topic, which is your H2 structure before you've written a word.
Step 3 - Google Trends: the shape of demand
Two keywords can have identical volume and completely different futures. Trends shows you which one is a rising or falling query - seasonality, regional concentration, breakout terms. The classic expensive mistake it prevents: building a cluster around a term that peaked two years ago, then wondering why the traffic forecast never arrives. Trends is also the only tool in this stack that shows demand for things people search before they have product words - problems, symptoms, event names.
Step 4 - Reddit and forums: the language mine
Search your topic on Reddit with site:reddit.com plus your seed term. Ignore the rankings advice for a moment and read the words. Buyers write "keeps crashing when I export," not "export reliability issues." The thread titles are your meta descriptions written by the market itself - and since Google began surfacing Reddit content heavily in results (the licensing deal that made forum content a first-class result type, widely reported in 2024), the language gap between forums and marketing copy is now an SEO variable. It stopped being merely a copywriting one.
The Receipt: take one thread title, one autocomplete string, and one GSC query for the same topic. You now hold three independent confirmations of demand language. A keyword tool can show you the number. Only this stack shows you the sentence.
Step 5 - Validate before you write
Before committing a cluster to the calendar, check what the results page actually rewards. Run the keyword through a SERP preview against the current top results: if page one is all video, or all forum threads, or all product pages, your content format decision makes itself. Then draft title and description variants in the same preview before publishing - the click is won or lost in those two lines, not in the keyword list.
The Devil's Advocate
The honest case against this stack: it has no difficulty scores, no volumes, no competitor gap analysis - and it costs an hour of your time per cluster, every time, while a paid tool amortizes its cost across hundreds of lookups. Agencies running keyword research at scale would drown in this workflow. That's true. This stack is built for site owners and small teams who publish a handful of clusters a month - people who need ten good keywords with real language attached, not ten thousand rows.
And yes - autocomplete strings come without numbers attached. You trade the illusion of precision for recency and truth. My take: a volume number that's a monthly-average estimate from a click model is worse than ten confirmed query strings from Google itself. An estimate comforts; a string converts.
The 2am Version
Search Console: what you almost rank for. Autocomplete and PAA: what people type right now. Trends: which demand is rising. Reddit: the exact words buyers use. SERP preview: what format wins before you write. Zero dollars, one hour, and the output list is better than a tool export because every row carries a human sentence with it.
Analyses - the RankForge frame
Data lens. Two anchors carry this post: Google's own documentation on query sampling in Performance reports, and Ahrefs' 2018 long-tail census - outside our preferred recency window, cited as a landmark because its structural finding has strengthened, not weakened, as query diversity grew. The Reddit visibility claim rests on a commercial deal and observed results-page changes, both widely reported; we've flagged it as reported rather than measured. Nothing in the stack hides behind a dashboard - every step is reproducible in your browser tonight.
Risk lens. The failure mode here is sampling complacency: GSC omits rare and anonymized queries, autocomplete personalizes by location and history, Trends normalizes rather than absolutes. Build a cluster on one source alone and you inherit its blind spots - the stack works because five flawed instruments triangulate. The second risk is time-blindness: an hour per cluster is cheap monthly, ruinous daily. This is a research workflow, not a research job.
Ethics lens. Keyword research content usually ends in a tool affiliate link - the recommendation and the commission share a wallet. This post recommends the tools that pay us nothing, and the only links out are to our own free checker and Google's docs. Education, not advice: the stack shows you demand; choosing what to build remains yours.
RF stance. Originals over reconstructions. Because autocomplete, GSC and Trends are Google's own demand signals - free, current, and impossible for a third-party crawler to fully replicate - starting keyword research anywhere else means buying a copy of data you could read at the source.
Change-my-mind. A reproducible test showing a paid volume database predicting content performance materially better than a GSC-plus-autocomplete cluster built for the same site and period - same content quality, same publishing cadence, performance measured over six months. Publish that comparison and the workflow earns a paid-tool appendix.
Sources
- Google Search Central - Search Performance report documentation: regex filtering support and the omission of anonymized or rare queries from results. support.google.com
- Ahrefs keyword study (2018) - 94.74% of keywords receive ten or fewer monthly searches; landmark long-tail census, retained per our recency-tier rule because no newer study has replaced it as the reference measurement. ahrefs.com
- Google Trends - relative search interest over time and across regions; methodology notes on normalization. trends.google.com
- Reddit-Google content licensing and expanded forum visibility in results (2024) - widely reported commercial agreement; cited as reported context for the language-mining step, not as a measured ranking factor.



