Keyword research without paid tools, and what you give up

The paid tools sell one thing that free methods cannot replicate: a monthly search volume next to every term. Everything else they sell, and most of what actually decides an editorial plan, is available at no cost from three sources almost nobody uses in order. The catch is worth stating first, because it is the whole trade: working free means never seeing a volume number, so priorities get decided by evidence of demand rather than by size of demand. That is a downgrade on paper and, on a small site, frequently an upgrade in practice.

How do you do keyword research for free?

By starting with your own data rather than a tool. A new site has none, so it harvests autocomplete and reads the results page; a mid-authority site is led by Search Console, which already shows which queries see it; a consolidated site is looking for gaps rather than building lists.

Quick answer

• Start with your own data. Search Console lists every query you already appear for.
• Autocomplete, done systematically, produces hundreds of real phrasings per seed.
• The results page itself shows what the query is understood to mean.
• What you lose is volume. You never learn whether a term gets 50 or 5,000 searches.

What kind of limitation this is

The assumption behind most tool subscriptions is that the hard part of this work is finding terms. It is not. Terms are abundant and cheap; on any subject you can generate hundreds in an afternoon.

The hard part is deciding which of them you can win and which are worth winning, and a volume number answers neither question. A term with 5,000 monthly searches that ten large sites already own is worth less to a small site than one with 80 that nobody has answered properly.

There is a second thing worth knowing about those numbers before paying for them. They are modelled estimates, not counts. Different tools report different volumes for the same term because they are each inferring from clickstream samples. Treating them as measurements is the most common way they get misused.

Source 1: your own data, which almost nobody reads

The first source and the highest-return one, and it is remarkable how many plans start somewhere else.

The performance report in Search Console lists every query the site has appeared for, with impressions, clicks and average position. Three patterns there give you what no external tool can.

Queries you already receive without having written for them. The cleanest signal available: demand exists and the site is already considered somewhat relevant. On one of my own sites, queries about SEO for Hispanic businesses were appearing with impressions and a position around 17 before a single word had been published on the subject.

Queries with high impressions and poor position. Demand exists and you are serving it badly. This is the best starting point for new content that exists, and it is why reading this report sits at item two of the basic SEO checklist, above writing anything.

Queries with good position and almost no clicks. Not a keyword problem at all. That is a title problem, and it is diagnosed differently, as covered in the Spanish guide on ranking first without clicks.

Source 2: autocomplete, done systematically

Typing a word into a search box and glancing at the suggestions is what everybody does. Doing it with a method is a different activity.

Combine the seed with letters and with modifiers. "seo a", "seo b", "seo c" through the alphabet produces separate branches. The same with how, what, cost, best, vs, free, near me.

Repeat by country. Suggestions differ between markets, and that difference is itself information: it shows which words each market uses for the same thing.

Save everything and clean afterwards. In this project's Spanish harvest, 14,715 unique suggestions produced 8,875 useful ones, so nearly four in ten were discarded as brands, duplicates or noise. Cleaning is part of the work rather than a setback.

That method is what revealed that the bare acronym "aeo" does not belong to this niche in Spanish: it is occupied by a clothing retailer and a customs designation. A volume tool would have shown a large and misleading number for that term.

And it is what tells you when a term you assumed is not searched at all. The seed "bilingual seo strategy" came back with zero suggestions belonging to this niche, while "translate website for seo" returned 45. Same reader, different words. No amount of volume data would have surfaced that, because the term with the volume was the wrong one.

Source 3: the results page itself

Search the query and read the page rather than the rankings. It answers the question a volume number cannot: what does the engine think this query means.

Look at what kind of result occupies the top. Product pages mean commercial intent. Guides mean informational. A map pack means local. If the results do not match the page you intended to write, you are not competing for what you thought.

Read the related searches and the People Also Ask box. Both are free lists of adjacent queries with confirmed demand, and they come from the same source the tools resell.

Notice whether an AI summary is present. That changes the value of the query before you write a word, and the reasoning for deciding what to do about it is in the Spanish guide on recovering traffic lost to AI.

When paid tools are worth it

When an opportunity has to be sized in advance for someone else's money. A client or a board asking for projections needs a number, however soft, and free methods do not produce one.

When you work across many sites and need speed. The tools compress hours of manual work, and at agency scale that time is the product.

When competitor analysis is the point. Seeing which terms another domain ranks for is the one job free sources genuinely cannot do.

For a single small site with an owner who knows the business, none of those three usually applies.

The step no guide describes, measured on a real harvest

Every guide on this ends the same way: filter by volume and difficulty, prioritise, assign each keyword to a URL. The step nobody describes is the one that ruins most decisions, and it shows up only when you look at your own data.

The research behind this blog was built from autocomplete. In Spanish, 11,220 seeded queries produced 14,715 unique suggestions, of which 8,875 survived cleaning. Close to forty per cent is lost in the clean-up, and that share never appears in a guide because nobody publishes their harvest.

The important part came later. Some rows had been marked as inferred, which is what you write when a keyword seems reasonable but did not come out of the harvest. Months later, checking one by one, ten of twelve had never been seeded at all. The label did not mean weak, it meant nobody asked, and those are opposite things. Seeding the ten took eighteen seconds and moved the US corpus from 1,510 to 1,878 suggestions.

The rule that comes out of it: an absence in your list is only information if you searched for that word. Otherwise you are measuring your own seed list, not demand.

Where each source lives

Your own data, which almost nobody looks at and is the only exact one: how to set up Search Console.

Autocomplete and the results page, which cost nothing and tell you how people actually phrase the problem.

And the check that decides whether a query is yours at all, which is what type of page currently answers it. If the results are all national brands, the query is not yours yet, and the realistic ground is narrower: local SEO for small business.

What to do with the list once you have it is in basic SEO checklist and what is SEO. If the audience searches in Spanish as well as English, the list doubles in a way that is not a translation: SEO for Latino businesses in the USA and translating a website for SEO.

What "free" means on the first page of this query

Checked on 13 August 2026, every result that owns this query is a tool, and most of the free ones are the entry tier of a paid product. The numbers are in their own copy: five lookups a day here, three keywords a day there, a capped export somewhere else. That is a trial with a different name, and it runs out on the afternoon you actually start working.

The more telling part is what the page does not contain. Not one of those results names Search Console, which is free, has no daily cap, and is the only source on the list that holds your own queries rather than somebody's index. It gets left out because there is nothing to sell around it and because it only works for a site that already has data, which is exactly the site whose owner is searching this.

Mistakes that repeat

• Starting with a tool instead of Search Console. Your own data is free and specific.
• Treating volume as a measurement. It is a model, and tools disagree with each other.
• Chasing the biggest number. Winnable beats large on a site without authority.
• Skipping the results page. It is where intent is confirmed, and it costs a minute.
• Keeping every suggestion. Four in ten are noise; cleaning is part of the method.
• Assuming your phrasing is the market's. A seed that returns nothing is telling you something.

Data and transparency

That every result on the first page for this query is a tool, that most of the free ones cap usage at a handful of lookups a day, and that none of them names Search Console, is an own observation from 13 August 2026 and is recorded as a pattern in the results, not as a source.

The harvest figures, 11,220 seeded queries producing 14,715 unique suggestions with 8,875 useful after cleaning in Spanish, and the US corpus moving from 1,510 to 1,878 after seeding ten missing keywords, are this project's own measurements and the source files live in the site repository. No third-party volume figure is quoted, because two tools never agree.

The harvest figures come from this project's own keyword research: 14,715 unique autocomplete suggestions across Colombia, Spain and Mexico, of which 8,875 survived cleaning, and 1,510 unique suggestions in the US market. The files are in the repository. The queries appearing without published content on them come from Search Console for a site I own, read through a read-only credential. The observation that paid tools model rather than count search volume is drawn from the fact that they report different figures for the same term; no specific tool is named and no comparison table appears here because I have not run a controlled comparison. No claim appears about which method finds more terms: I have not measured that, and any number would be invented. Verified as of August 2026.

What this changes

The subscription question is usually framed as whether the tool is worth the money. On a small site it is worth asking what the tool changes about the decision.

If the plan would come out the same either way, the number is decoration. And it often would, because the terms a small site can realistically win are visible without it: the ones already appearing in Search Console at a bad position, and the ones autocomplete shows real people typing that nobody has answered properly. The month spent finding those is the same month either way. The difference is that one version ends with a spreadsheet of volumes nobody acts on.

Frequently asked questions