What you'll learn in this article
- Why the results page tells you more about a keyword's intent than any metric in the Planner
- How to spot intent dispersion, the single property that makes expensive keywords expensive
- What an unusually cheap CPC in a competitive market is actually telling you
- The grammatical tells that separate a buyer's phrasing from a researcher's
- The one-page test that kills more of my candidates than every other check combined
Most Google Ads keyword research tips you'll read describe a procedure: open the Planner, enter seeds, filter by volume, sort by competition, export. That procedure is fine. It's also not where the money is won or lost, because every competitor in your auction runs the same one and arrives at roughly the same list.
What separates a profitable keyword list from an expensive one isn't the sourcing, it's the judgement applied afterwards. And that judgement runs on signals that don't appear in any column: how a results page is composed, how a phrase is grammatically built, what an unusually low bid estimate implies about who else declined to bid.
These are five things I learned the slow way, by paying for the version where I ignored them. None of them are in the standard playbook. All of them are inference rather than data, which is precisely why they still work.
1. The results page is the cheapest intent research you will ever do
Before I add a term that will cost real money, I search it. Not to check rankings, but to read what Google decided to show, because that composition is the distilled outcome of an enormous quantity of behavioural evidence about what people actually wanted when they typed those words.
If the page is dominated by explanatory articles and definitions, the intent is informational no matter how commercial the phrase looks in a spreadsheet. If it opens with Shopping units and a dense block of ads, the intent is transactional and the auction already knows it. If it's a mix of comparisons, reviews and directories, you're looking at mid-funnel research traffic that will click, read, and leave.
The trap this catches is the term whose words sound like buying but whose audience isn't. "Cost", "price", "software", "system", "service" all read as commercial and frequently sit on pages full of explainers. I have bought that traffic more than once, at a good click-through rate, with a conversion rate near zero, and the results page had been telling me the answer for free the entire time.
There's a second read available. Look at who's advertising. If the same three well-funded competitors occupy every slot with tightly matched copy, that term has been validated by people spending more than me and my entry price will be set accordingly. If the ad block is thin or the copy is loose and generic, either the term is unproven or the market has tried it and quietly stopped.
2. Intent dispersion matters more than volume, competition or anything else
This is the single most useful concept I've developed for this work, and I've never seen it named in a guide. Ask, of any candidate: how many genuinely different things could a person want when they search this?
A term with one possible intention is a term I can serve with one ad and one page. A term with five is a term where at best a fifth of my clicks land somewhere relevant, and the other four fifths bounce while paying full price. Volume doesn't compensate for this, it multiplies it, which is why high-volume head terms are the most reliably expensive mistakes in a new account.
The practical test is quick. I try to write the ad. If one headline honestly addresses everyone searching that phrase, dispersion is low and the term is a candidate. If I find myself writing something deliberately vague to cover several audiences at once, dispersion is high and the vagueness is a confession, not a solution.
Dispersion is also why match type is a research decision rather than a setup decision. A dispersed term under broad match doesn't merely reach its own five intentions, it reaches the neighbours of all five. The same term pinned to exact reaches one narrow slice you can actually serve, and the reasoning behind that choice is something I break down in my guide to how keyword match types work.
When I do use a research tool for genuinely new territory, this is the lens I read the output through. Google's own guidance on this points in a compatible direction, advising that seeds describe the category the way customers think about it rather than mimicking full queries, and warning against terms that are too general to reveal intent. Their best practices for finding new keywords frame it as a specificity problem; I'd frame the same thing as a dispersion problem, and the practical instruction ends up in the same place.
3. A suspiciously cheap keyword is information, not an opportunity
Early on, a low bid estimate in an otherwise expensive market felt like a discovery. Everyone else is paying eight euros a click and here's a relevant term at ninety cents. I'd add it enthusiastically and learn the same lesson repeatedly.
Click prices in a mature market are the aggregate opinion of every advertiser who has already bought that traffic. When a term is priced far below its neighbours, the most economical explanation is that the market tested it and withdrew. Someone bid, measured, and decided it wasn't worth more. Cheapness in a competitive vertical isn't an inefficiency waiting to be exploited, it's usually a verdict.
Not always, and the exceptions are worth having. New vocabulary that hasn't been priced yet, terminology specific to a niche the big spenders haven't segmented, and phrasings that are genuinely commercial but read as informational to automated tools all produce real bargains. The distinction I use: is this term cheap because nobody has found it, or cheap because everybody has? New vocabulary is usually the former. An old term with obvious commercial phrasing at a tenth of market price is almost always the latter.
The same reasoning applies in the other direction. When a term is far more expensive than its volume seems to justify, the market is telling me it converts. That's not a reason to avoid it, it's a reason to check whether my conversion rate and margin can survive the price, which is a different and much more answerable question. I go deeper into reading these price signals in my breakdown of what keyword cost per click actually reflects.
4. The grammar of a query predicts its value better than its vocabulary
When people ask me how to do Google Ads keyword research that doesn't just replicate what a tool suggests, this is the habit I point to. Two phrases can contain identical words and describe completely different people, and the difference sits in the structure rather than the terms.
Questions are the clearest example. A query framed as a question is almost always a person acquiring understanding, not a person acquiring a product. Someone asking how a thing works is at the start of a process; someone naming the thing plus a qualifier is somewhere near the end of it. The vocabulary can be identical while the distance to a purchase differs by weeks.
Then there's the qualifier position. A modifier before the noun usually describes a category being explored. A modifier after it usually describes a specification being matched, and specifications imply someone who already knows what they want. That asymmetry has been more reliable in my accounts than almost any tool-supplied intent score.
Plurals carry a signal too. Plural forms lean towards browsing and comparison, singular forms towards a decision already narrowed. It's soft, it's not universal, and it's still worth noticing when a keyword is otherwise borderline.
The most valuable structural tell is the presence of a constraint. Location, budget, compatibility, timing, quantity: any constraint means someone has moved from wondering to specifying. The words around the constraint barely matter. A constrained query with a fraction of the volume routinely outperforms the unconstrained head term it descends from, which is one of the reasons my keyword lists tend to be smaller and stranger than the ones tools generate.
5. The page test, which kills more candidates than everything else combined
The last check is the simplest and the one I resisted longest, because it disqualifies keywords I want to keep. Before adding a term, I ask which existing page I would send that click to, and whether that page answers the specific thing the query asks.
Not a page that's roughly on topic. The page. If the honest answer is a category page, a homepage, or a page that mentions the subject among six others, the keyword isn't ready. The traffic will arrive, find something adjacent to what it asked for, and leave, and the reporting will describe this as a keyword problem when it is a destination problem.
I apply this before launch because afterwards it's contaminated. Once a term is running, sunk cost turns "there's no page for this" into "let's see how it does". It never does well, and the diagnosis is delayed by however long you spend testing something you already knew the answer to.
What this changes is the order of operations. Keyword research stops producing a list of keywords and starts producing two lists: terms I can serve today, and terms I could serve if I built something first. The second list is a content brief, and treating it as one rather than as keywords to add anyway has saved me more budget than any bidding adjustment I've made. It also means research and account structure stop being separate exercises, a link I develop further in my approach to building a keyword strategy that survives contact with a live account.
The uncomfortable version of this test: if you can't name the page, you're not researching keywords, you're collecting them.
Putting the five together, and why the order matters
These aren't a checklist to score a keyword against. They're a sequence, ordered by how cheaply each one eliminates a candidate. The results page read takes seconds and removes the terms whose intent is wrong at the root. Dispersion removes the ones I can't write a single honest ad for. Price anomalies flag the terms whose economics need explaining before anything else. Grammar sorts what survives by distance to a decision. The page test decides what actually launches this week versus what goes on the build list.
Run in that order, most of the work is elimination, which is the part guides consistently under-serve because eliminating candidates doesn't look like productivity. A session that ends with two hundred keywords feels successful and mostly produces future obligations, since every term still needs an ad, a destination and enough budget to reach a verdict.
None of this replaces the account's own evidence. Once a campaign is live, its search terms report outranks every inference here, because those queries have already been bought and already produced outcomes. These signals are for the moment before you have that data, or the moment you're moving into territory the account has never seen, and that's exactly when the standard guides leave you with nothing but volume and competition columns to reason from.
So, condensed: read the results page before trusting the phrase, count the intentions before counting the searches, treat a cheap click in an expensive market as a verdict rather than a bargain, read the structure of the query rather than its vocabulary, and refuse to add anything you can't name a destination for. Five habits, all of them inference, all of them learned by paying for the alternative.