HOW TO SELECT KEYWORDS FOR GOOGLE ADWORDS: THE FILTER I APPLY TO CUT 500 DOWN TO WHAT I LAUNCH

Every keyword project I take on ends up at the same uncomfortable moment: a spreadsheet with roughly 500 rows in it, and a decision I can't outsource. Knowing how to select keywords for google adwords isn't about generating more ideas, tools do that fine, it's about the ruthless cut from a bloated raw list to the lean set I'm actually willing to put budget behind. That cut is where campaigns are won or lost, and it's the part almost nobody documents honestly.

So this article is the filter itself, the exact one I apply on live accounts, told from experience rather than theory. I'm not going to hand you a keyword-generation tutorial. I'm going to show you the four passes I run over that 500-line list to get to the 30-60 terms I launch, and why each pass exists. The whole thing is built around one idea that changed everything for me: in 2026 the keywords I select are seeds for an inference engine, not fixed triggers, so the filter isn't about coverage, it's about signal.

Before I ever get to selection, the raw material comes from the usual places, mostly the Google Keyword Planner for search volume estimates, plus the search terms my existing accounts already convert on and a scan of the search results for the queries I care about. That generation step is easy. The hard, decisive step is the selection that follows, and it's where most advertisers quietly go wrong by keeping everything the tools handed them.

If you've ever stared at a giant keyword export and felt paralysed about what to keep, this is the method that unsticks it. It's how I pick keywords for Google Ads without second-guessing, and it produces a set I can defend line by line, because every term in it survived a filter I can explain out loud.

The 500-keyword problem nobody wants to name

Here's what actually happens. You run the Google Keyword Planner, pull competitor terms, add the phrases the client swears by, dump in a few variants from a scraping tool, and within an hour you're sitting on 400-600 keywords. It feels like progress. It isn't. A raw list of keywords that size is mostly noise: near-duplicates, terms with no real intent, vanity phrases with high search volume and zero buyer signal, and dozens of variants that Google's AI would have inferred on its own anyway.

The mistake I made early in my career was treating that list as an asset to preserve. I'd launch 300 keywords in a single Google Ads campaign because deleting them felt like throwing away work. What I learned the hard way is that a bloated keyword set doesn't just clutter reporting, it degrades matching. Because the system now reads the other keywords in an ad group and the landing pages behind them to interpret intent, a 300-term group hands the algorithm a confused, contradictory hypothesis. A tight group of relevant keywords gives it a sharp one. That's the real reason the cut matters, and it's why adding too many keywords hurts your Google Ads in ways that aren't obvious until performance quietly suffers.

There's a quality dimension to this too. Cramming loosely related terms into one ad group drags down relevance, and relevance is one of the pillars of Quality Score. A focused set of relevant keywords, matched to tight ad copy and a matching landing page, is exactly what earns higher Quality Scores, which in turn lower your costs. A sprawling list does the opposite: it spreads your relevance thin and a poor Quality Score follows. So the cut isn't just tidiness, it's a direct lever on what you pay per click.

So the question stopped being "which of these 500 can I justify keeping" and became "which of these does the AI genuinely need from me, versus which will it figure out itself." That reframing is what turned an overwhelming spreadsheet into a fast, repeatable decision, and it's the mindset the entire filter below is built on.

The filter: four passes I run every time

My entire method for how to pick keywords for google ads from a raw list is four sequential passes. I run them in order, top to bottom, and each one is a hard gate: a term that fails any pass comes out, no sentimentality. By the end, 500 is usually 30-60. Think of it as keyword selection as subtraction, I start from everything and remove until only the defensible terms remain.

Pass 1 Intent in one sentence

For every keyword, I ask: can I state, in a single sentence, what the person searching it wants? If I can't, it's out. "roofing" fails, I have no idea if that's a homeowner, a job seeker, or a student. "emergency roof repair near me" passes instantly, the intent is obvious. This pass alone typically kills a third of the list, because most raw keywords are ambiguous fragments, not real intents. Knowing the intent behind a term is the whole game, which is why understanding the different types of keywords and the buyer stage each one signals is the foundation the rest of the filter sits on. A good gut check here is your own product or service: if a keyword doesn't map to something you actually sell, its intent isn't yours to buy.

Pass 2 Redundancy against inference

Next I collapse duplicates, but not the way people usually mean it. I'm not just removing literal repeats, I'm removing terms whose meaning the AI already covers. If I keep "buy running shoes," I don't also need "purchase running shoes," "running shoes for sale," and "order running shoes online." Exact match now means same intent, so those four are one keyword to the system. Keeping all four just dilutes signal and splits data across a longer list of keywords than the account needs. I keep the cleanest phrasing of each distinct intent and drop the rest, because the inference engine expands that seed on its own. The old habit of exhaustively listing every variant is exactly the work the AI now does for free.

Pass 3 Evidence, not vanity

Now I check whether there's real evidence behind each surviving term: search volume from the planner, historical account performance, or an obvious commercial pattern. This is where high-volume vanity phrases die. A term with 40,000 monthly searches and no buyer intent is worse than a term with 90 searches and a credit card behind it. I'm not filtering by volume, I'm filtering by evidence of intent. Low search volume with strong intent stays; high volume with weak intent goes. I'll also cross-check the actual search query data from accounts I already run, and sometimes Search Console for the site's organic queries, because real recorded searches beat a planner estimate every time.

Pass 4 Inference load

The final pass is the one the inference angle forced me to add. For each remaining keyword I ask: how hard is the AI going to have to guess to serve this well? A term that gives the system a clean, unambiguous starting point stays. A term that forces it to guess wildly, vague, off-theme, or disconnected from my landing pages, comes out even if it survived the first three passes. This is where I read each keyword as a hypothesis I'm handing the algorithm, and I only keep the ones that give it a hypothesis worth extending. In practice this is where choosing the right keywords for Google Ads becomes a discipline in its own right: the right term is the one that tells the system precisely who I want, so the inference lands on Google search where my buyers actually are.

Running the filter on a live account

In practice the four passes take me about twenty minutes on a 500-line list, and the discipline is in not skipping ahead. I resist the urge to judge search volume first, because Pass 3 only makes sense after intent and redundancy have already thinned the field. Filtering for evidence on a list still full of ambiguous fragments just wastes time scoring terms I was going to cut anyway. Order is the whole reason the filter is fast.

Once the set is down to its final 30-60, I don't dump it into one ad group. I sort the survivors into tight, single-intent groups, because that structure is what lets the AI infer cleanly from each seed. This is the point where keyword selection hands off to organisation, and the two are really one continuous decision. The way I think about grouping and sequencing the final terms is part of my broader keyword strategy in Google Ads, where the selected set becomes the input to how I stage match types and open up to broad.

Structure also protects Quality Score. When each ad group holds a handful of relevant keywords that share one intent, I can write ad copy that speaks directly to that intent and point it at a landing page that matches. That alignment between keyword, ad copy and landing pages is what the system rewards with higher Quality Scores and cheaper clicks. A single bloated group can't do that, the ad copy has to hedge across too many intents, and the quality score sags for every keyword in it.

What I deliberately don't do anymore is pad the launch set "just in case." The old instinct was that more keywords meant more coverage. In an AI-matched world it means more noise. Every term I add past the necessary minimum is a slightly muddier signal for the system to read, so the filter's job is as much about what I refuse to keep as what I select.

Selection isn't done until the negatives are drawn

A point I learned to treat as non-negotiable: choosing what you bid on is only half of keyword selection. The other half is choosing what you refuse to serve on, and that means negative keywords go in before launch, not after the budget's already leaked. When I widen a term to broad and let the AI infer around it, I'm granting it freedom, and negative keywords are the fence that keeps that freedom pointed at buyers instead of browsers.

My rule is simple: every intent I keep gets a matching set of exclusions built at the same time. If I'm bidding on "running shoes," I'm pre-adding negatives for "free," "repair," "how to clean," and the job-seeker and research queries that share the words but not the intent. This is where the search query view from existing accounts earns its keep again, it shows me the real off-target searches to block before they cost me anything. Negative keywords aren't a cleanup task, they're part of the selection itself.

The reason this matters more in 2026 is the same reason the whole filter changed: because matching is now inference, a keyword without its negatives is an open-ended instruction to explore. With the negatives drawn tightly around each intent, the AI explores inside guardrails I chose. That combination, a lean set of selected terms plus disciplined exclusions, is what keeps broad match from turning into uncontrolled spend, and it's why I never consider a keyword truly "selected" until its negatives exist too.

The filter doesn't end at launch

Here's the part that makes the whole method work: selecting the final set isn't the finish line, it's the opening hypothesis. Once traffic arrives, the search terms report tells me what the AI actually matched my selected keywords to, and that's the real test of whether my cut was right. Terms I kept that pull garbage get corrected with negatives; winning queries the system found that I never listed get promoted into their own keywords.

This is why I don't agonise over keeping a borderline term in the initial 500-to-60 cut. If I'm genuinely unsure, I lean toward cutting, because the report will surface anything valuable I missed. The selection filter and the post-launch loop are two halves of the same process: I select tight to give the AI a clean start, then let the data expand and correct the set. Getting the underlying keyword choices for Google Ads right upfront just means the loop starts from signal instead of noise.

If there's one takeaway on how to select keywords for google adwords in 2026, it's this: don't try to keep everything that might work. Run the filter, cut hard to the terms whose intent you can name and whose evidence you can point to, hand the AI a clean set of seeds, and let the search terms report finish the job. The skill isn't generating 500 keywords, it's having the nerve to launch 50. For the mechanics behind match types and matching that this selection feeds into, Google's official documentation on keywords in Search campaigns is the reference I point clients to.

One last thing worth saying plainly, because it's easy to forget under all the AI talk: your selected keywords determine who ever sees your ads in the first place. No amount of clever bidding rescues a bad list, and no algorithm compensates for a keyword set built on noise. A single well-chosen google ads keyword, tied to matching ad copy and a landing page that delivers on its promise, still outperforms a hundred loose ones. So I treat every term I keep as a deliberate bet: it earned its spot by naming a clear intent, showing real evidence, and giving the system a clean hypothesis to extend. Select tight, fence it with negatives, and let the data widen the map from there, that's the whole method, and it's held up on every account I run.