How to Choose the Right SEO Keywords for Your Business Website

The standard way to pick keywords is to brainstorm a list, check the search volume, check how hard each one looks, and prefer the longer, more specific phrases because they are easier to rank for. That method was correct for about ten years. It is now the fastest route to work that ranks and returns nothing, because the phrases it steers you toward are precisely the ones search engines have started answering themselves.

Here is the number that reorders everything. Informational searches end without any click roughly 74 percent of the time. Transactional searches end without a click roughly 31 percent of the time. Same effort to rank, wildly different chance that ranking produces a visitor. Once you know that, checking volume before checking click potential stops making sense. At CIOR Digital, a digital growth agency in La Jolla, this is now the first question we ask about any keyword, before anything else.

This article covers the corrected method: what a keyword is now that engines resolve meaning rather than match text, which filter replaced volume, why long-tail stopped being a safe harbour, and what to do with the queries you choose. It also corrects itself, since the earlier version of this page gave the long-tail advice that the data has since overturned.

What a keyword is, and what changed

A keyword is a word or phrase someone types into a search engine. If you run a coffee shop in Chicago, phrases like best coffee shop in Chicago or coffee near me are keywords that could bring the right people to your site. That definition has not changed and it is still the right place to start.

What changed is what happens to the phrase after it is typed. Search engines used to match text: your page contained the words, the page was a candidate. They now resolve meaning, which is why a page can surface for a query it does not literally contain, and why two phrases that look different but mean the same thing increasingly return the same results. Semantically identical queries have effectively merged.

The practical consequence is that you are no longer choosing words. You are choosing questions you want to be the answer to. That sounds like a small reframe and it changes the whole selection process, because a question has properties a string of text does not: somebody is asking it for a reason, at a particular point in a decision, and either they will need to visit a website to resolve it or they will not.

The filter that replaced volume

The old order of operations was volume, difficulty, relevance. It made sense when every search with volume represented clicks that were available to win. Rank well enough and you got a share of them. The size of the prize was the size of the search demand.

That relationship broke. A large share of searches now end on the results page, because the answer appears there, and the share is not distributed evenly. It depends almost entirely on why the person was searching. We covered the mechanics of that shift in our piece on what changed in search this year, so the point here is narrower: it means search demand and click availability are now two different things, and the gap between them is the single most important thing to check before committing to a target.

Figures compiled from Searchlab and Semrush analysis. Independent measurement by Seer Interactive found AI Overviews appearing on around 36 percent of informational queries against around 5 percent of transactional ones, a different scale with the same ordering.

Local search deserves a note here, because the table above misleads slightly on it. A local query ending without a click to your website is not the same failure as an informational one ending without a click. The person may well have called you, tapped through for directions, or read your reviews, all inside the map results, and none of that registers as a visit. If you serve a defined area, judge local queries on calls, direction requests and profile actions rather than on sessions, or you will conclude that a channel producing customers is producing nothing.

So the four questions, in their corrected order. First, does this query still produce a click. Second, what is the person actually trying to do. Third, if they arrive, what happens next and is that worth anything to the business. Fourth, and only fourth, how much volume is there and how hard will it be. Volume has not stopped mattering. It has stopped being the thing you check first, and the businesses that choose targets that still return something are the ones that reordered those questions early.

Why long-tail is no longer the safe answer

The classic advice runs like this: short phrases are competitive, so target longer and more specific ones instead, where competition is lower and intent is clearer. It is sensible, it worked, and the earlier version of this page recommended it four separate times. It now points in the wrong direction, and the data on that is not ambiguous.

Longer, question-shaped, informational queries are the single most likely category to be answered on the results page. A query of seven words or more triggers an AI-generated answer roughly 74 percent of the time, against roughly 36 percent for a query of one or two words. Ahrefs found that 99.2 percent of the keywords triggering those answers are informational in intent. And low volume is not a shelter either: analysis of more than ten million tracked keywords found that nearly 60 percent of the phrases triggering an answer receive 100 or fewer monthly searches.

None of that makes long-tail worthless, and declaring it dead would be its own mistake. What it does is change what the work is for. Ranking for a long-tail informational query increasingly buys you a citation inside somebody else’s answer rather than a visit to your site, and a citation has measurable value: brands cited in AI answers earn around 35 percent more organic clicks and around 91 percent more paid clicks than brands that are not. That is a real return. It is simply not the return most people think they are buying, and choosing it by accident while aiming for traffic is how content programmes quietly fail.

Exposure also varies enormously by industry, which matters if you are reading these averages and planning against them. Healthcare sits at the extreme, with AI answers appearing on around 88 percent of queries, followed by education at around 83 percent and enterprise technology at around 82 percent. Restaurants sit near 78 percent and insurance around 63 percent. A medical clinic and a local contractor are operating in genuinely different search environments, and the clinic should be far more sceptical of an informational content plan than the contractor needs to be. Check your own sector before assuming the general figures describe you.

The four intents, and what each one is now for

Search intent used to be a way of making sure your page matched what the searcher wanted. It is now the primary variable determining whether the searcher ever reaches a page at all, which promotes it from a quality check to a selection criterion.

The useful move is to stop asking which intent is best and start assigning each one a job. Informational queries build recognition and feed the systems that will later recommend you. Commercial investigation queries are where most businesses should concentrate, because the person is comparing options and usually needs to see the options. Transactional queries are the smallest pool and the most protected. Assign a purpose, and measure each category against that purpose rather than against a single traffic target.

• Informational. Someone wants to understand something. The job is citation, recognition and topical authority. Measure whether you are being referenced, not how many visits arrive.
• Commercial investigation. Someone is comparing options before deciding. The job is traffic and consideration, and this is where most effort belongs. If you are also weighing whether to buy those clicks instead, this is the intent where that question gets interesting.
• Transactional. Someone is ready to act. The job is conversion. Small volumes, high value, least exposure to being answered on the page.
• Navigational. Someone is looking for a specific brand or site. The job is defence, and if the brand is yours it behaves differently from everything else here.

One warning before you build a strategy on that table. The protected zone is real today and it is contracting. During 2025 the informational share of AI-generated answers fell from roughly 91 percent to roughly 57 percent of all triggers, while commercial rose to around 19 percent and transactional went from around 2 percent to around 14 percent. In finance, commercial queries triggering an answer grew 231 percent in six months. Anyone telling you commercial intent is a safe harbour is describing a boundary that is visibly moving, which is an argument for building on something more durable than an intent category.

Branded search is the exception

There is one category behaving completely differently from everything above, and it is the strategic payload of this entire article. Branded queries, meaning searches containing your business name, trigger an AI-generated answer only around 5 percent of the time. When one does appear on a branded query, click-through actually rises by around 19 percent, against a fall of around 20 percent on non-branded queries. Analysis across 700,000 keywords found branded search to be the only quadrant still gaining clicks.

The reason is straightforward. Someone searching your name has already decided which website they want, and no summary satisfies that. Google still treats brand intent as navigational, and navigational queries by definition want a destination rather than an answer.

The implication changes what keyword research is for. If the only reliably protected search is one containing your name, then a meaningful part of the work is no longer finding queries to compete on, it is creating demand for a query only you can satisfy. That is not a keyword research task in the traditional sense, which is exactly why most keyword research misses it. Businesses that build search visibility deliberately end up with a growing share of their traffic on terms no competitor can enter, and it is the most defensible position available in search right now.

Practically, that means the activities most likely to grow protected search are the ones keyword tools never surface. Being genuinely known in a defined niche. Being mentioned by name in places your buyers already read. Publishing something people cite you for. Doing work distinctive enough that clients describe you by name to other people. None of that appears in a research tool, all of it produces the one query type that still reliably ends in a visit to your site, and it is the reason informational content keeps its value even as its direct traffic falls.

How to find candidates

Start with what you already have, not with a tool. Search Console is the only source that tells you which queries you are already being shown for, and almost every site is being shown for dozens of things nobody noticed. A query where you sit in position 12 with meaningful impressions is a far better candidate than anything you will discover by brainstorming, because the hard part is already done.

Only after that does external research earn its place, and it should be aimed at gaps rather than at volume. What are competitors being found for that you are not, and of those, which ones survive the click potential filter. That second half is the step almost everyone skips, and skipping it is how businesses end up with a keyword list full of phrases that are easy to rank for because ranking for them is worth less than it used to be.

One habit is worth building above all the others, and it takes thirty seconds per keyword. Before committing to a term, search it and look at the page. Not the difficulty score, not the volume estimate, the actual results page. Is there an answer box above everything. How far down does the first ordinary result sit. Are the results dominated by large publishers, by forums, or by businesses that look like yours. That single glance tells you more about whether the keyword is winnable and worth winning than any metric a tool will sell you, and it is the step that separates people who choose keywords well from people who export lists.

• Search Console firstFilter for queries with impressions and an average position between 8 and 25. These are the closest wins you own.
• Apply the click potential filterFor each candidate, check what the results page actually looks like. Not the difficulty score. The page.
• Map competitor gapsWhere are they visible and you are not, and does the query pass the same filter.
• Check ownershipDoes an existing page of yours already target this. If so, improve it rather than creating a competitor to yourself.
• Then sort by volume and difficultyLast, not first, and only among the candidates that survived the previous steps.

Choosing a research tool

The question people ask is which tool is best. The more useful question is what you need to be able to answer, because the honest situation is that free tools now cover more of the basics than they used to, and paid tools earn their price on a narrow set of things that happen to matter more than they used to.

What the free stack genuinely cannot tell you is competitive: which queries a competitor is visible for, how a results page has changed over time, and whether an AI-generated answer appears on a term and who it cites. Those three are the difference, and note that the third one did not exist as a category two years ago. If you are choosing today, judge tools on whether they report AI answer presence and citation, because that is now a core part of the selection filter and the older tools are catching up unevenly.

Where the chosen keywords go

A keyword list is not a plan, and the gap between the two is where most of the value leaks out. The rule that matters more than any other is one page, one intent. Two pages targeting the same query do not double your chances, they split the signals that either page needed, and the damage is not neutral but negative.

Before creating anything new, check whether an existing page already covers the query. Improving a page that already has age and links usually beats publishing a competitor to it, and the decision of which route to take should be written down rather than made in passing. We covered how that register works, and the boundary rules that keep adjacent pages from colliding, in our guide to how to plan the content around them.

A question that comes up immediately: how many keywords should you be working on. Fewer than most lists suggest. A site publishing one piece a month can meaningfully pursue perhaps a dozen targets a year, and a list of two hundred exported from a tool is not a strategy but a way of feeling productive. Pick the smallest set you can actually build proper pages for, get them right, and add more once the first set is performing. The constraint is production capacity, not research capacity, and confusing the two is why so many keyword lists are never used.

• One primary intent per page. Secondary phrases are fine when they mean the same thing. A different intent needs a different page.
• Map to existing pages first. Update beats create in most cases, and it is faster.
• Write the boundary rule. If two pages sit close, define in a sentence what each one owns.
• Keep a register. Every published URL and the query it owns, in one place someone else could read.
• Match the format to the intent. A comparison query wants a comparison, not an essay.

Knowing whether the choice was right

You will know long before traffic tells you. Impressions in Search Console move first, usually within weeks, and they answer the only question that matters early: is this page being shown for the query it was built for. If impressions are climbing on the target term, the selection was sound even if nothing else has moved yet. If they are flat after two months, the problem is the choice rather than the execution, and that is a cheap thing to discover early.

For anything informational, add a second signal that no analytics tool will give you: whether you are being cited. Ask the question your page targets in an AI assistant, periodically, and record whether your business appears. It is manual and slightly unsatisfying and it is currently the only way to see the outcome that informational content is now actually producing. Building that into your reporting is worth the effort, because otherwise a flat traffic line looks like failure when it may be nothing of the kind.

• Weeks 2 to 6: impressions on the target queryBeing shown at all is the first confirmation that the choice made sense.
• Months 2 to 4: average positionMovement matters more than the absolute number. Position 30 to 14 is progress even though neither converts.
• Months 4 onward: clicks, but only where clicks were the goalJudge informational pages on citation and branded search lift instead.
• Continuously: branded query volumeThe one line that should be rising regardless of which individual terms won.

Mistakes that cost more than they used to

The first is the one this page used to recommend: choosing long, specific, informational phrases because they look easy, without checking whether the resulting ranking produces anything. It was reasonable advice for a decade. It is now the most common way to spend six months building pages that rank exactly as intended and deliver almost nothing measurable.

The second is targeting the same query from two pages, which happens most often by accident when nobody keeps a register. The third is judging every page by traffic when the categories now have genuinely different jobs, which leads businesses to cut the informational work that was quietly feeding their branded search, and then to wonder why the branded line stopped growing.

There is a fourth that is less a mistake than an omission: treating the keyword list as settled. Everything described in this article moved measurably in eighteen months, and the terms that passed the click potential filter last year will not all pass it next year. A list reviewed once a year is enough for most businesses, and a list never reviewed at all is how a site ends up optimised for an environment that stopped existing.