Guide Published 4 November 2026 15 min read

How to Choose the Best Keywords for SEO: Five Checks That Reject a Keyword

A phrase with KD 4 and a top ten held by Ahrefs, Semrush and HubSpot. Where the list comes from, what volume and KD measure, and the five checks that reject most keywords before a page gets written.

Roman Makuev
Last updated

Choosing keywords for SEO is the part of keyword research that comes after the list exists: you take a few hundred phrases from a keyword research tool and reject most of them. The best keywords for SEO are the ones left after five checks. There is demand for the phrase on the platform where your buyer searches; the SERP accepts the type of page you can build; the competition is one you can match, with sites of your weight and domain authority already ranking in the top ten; the phrase leads toward something you sell; and no page on your site owns it yet. Search volume and keyword difficulty feed into the first and third checks. Neither decides them. The aim is organic traffic that can convert, which is why relevance to the business outweighs both numbers.

The phrase this article targets is a fair test of that. Our parser shows “how to choose best keywords for seo” at 1,000 searches a month and KD 4, which reads as an easy win. On 31 October the top ten for it held Ahrefs, Semrush and HubSpot, three companies with keyword tools to promote, next to agency blogs that repeat the same nine steps. KD 4 measured something real. It wasn’t whether a new page gets in. Most of what follows is about reading those two numbers for what they are, and about the checks that sit before and after them.

Finding keywords for SEO beyond Google Suggest: what each source adds

Most keyword lists come from one place. A seed goes into a tool, the tool returns phrases from its database, and the database was built mostly from Google. The list then describes how people phrase things in a Google search box, which is most of the picture for a blog and a smaller part of it for a store, or for a product people research on YouTube before they ever type its category into Google.

Our keyword research parser queries the search box of each platform directly: Google, YouTube, Bing, DuckDuckGo, Amazon, eBay, Etsy, Google Shopping, Perplexity and Google Trends, plus two Google modes, A-Z Soup and Questions. Twelve sources produce a long list, and the length is the least useful part of it. The column that matters is Source, next to every row, because where a phrase came from says what kind of page it can become.

SourceWhat it adds to the listWhat it can’t tell you
Google, Bing, DuckDuckGo autosuggest (autocomplete)The phrasing people type into a search box; the base of any listVolume, which only Google publishes and only through Keyword Planner
A-Z SoupEvery letter and digit appended and prepended to the seed, 70-plus requests to Google Suggest for one seed; the largest set of long-tail rowsWhich of the near-duplicates are separate intents
QuestionsHow, what, why, can, should in front of the seed keyword; vs, cost, for, near after it. The same questions Google shows under People Also AskWhether a question deserves a page or a paragraph
Amazon, eBay, Etsy, Google ShoppingThe attributes buyers add right before they buy: size, material, use case, price bandAny search volume; none of these platforms publishes one
YouTubeSetup, how-to and comparison phrasing, the searches around the purchaseVolume
PerplexityPhrasing from an AI search box, longer and more conversational than what Google Suggest returns. When its API blocks the call, Google Suggest fills in and those rows are labelled GoogleVolume, and how many people asked
Google Trends, Rising and TopRelated queries from the last 12 months, including ones gaining demand before any tool publishes a volume for themCounts; Trends reports relative interest

Sorted with the multi-platform phrases first, the Source column reads as a first filter. A phrase that three platforms return independently is demand that exists everywhere, and it usually heads a cluster. A phrase only the marketplaces return is a product attribute: it belongs in a category filter or a product card, and a blog post written for it will sit on a SERP full of listings. Phrases that come only from YouTube and Questions are the informational layer, the material for guides, H2s and FAQ. Trends Rising rows are the ones to argue about, because they have demand and no number yet.

One source the parser doesn’t cover is competitor keywords. The keyword gap report in Ahrefs or Semrush lists the phrases your competitors rank for and you don’t, and in a narrow niche it is often the shortest route to a first list. Those rows go through the same five checks as the rest.

Two details change what a run returns. The market setting (US, UK, Australia, Indonesia) applies to most sources, while Etsy and Perplexity are global. And synonym discovery is skipped for acronyms such as SEO, PPC, CRM and SaaS, because the expansions it finds for them are wrong more often than right. A seed like ours, with SEO in it, gets its breadth from A-Z and Questions. The parser’s own hint on the seed field says to try both word orders, and it’s there for a reason: Suggest completes “keywords for seo” and “seo keywords” differently.

Search volume: a range from Google, and a blank everywhere else

Google is the only platform that publishes search volume, through Google Keyword Planner, and an account without active ad spend sees ranges such as 1K to 10K instead of numbers. Every volume in a third-party tool is a model built on top of that: Planner data, clickstream panels, samples of Search Console. Ahrefs compared its own estimates with Search Console impressions on 72,635 keywords and found them roughly accurate in 60% of cases, against 45% for Keyword Planner. Semrush ran a similar study in 2022, also against Search Console, and found its own data the most accurate. Each vendor wins its own study. What the studies share is the reference: impressions in Search Console. Planner also shows cost per click (CPC), the one number in it tied to money: advertisers bid high where a query leads to a conversion, so a high CPC on a low-volume phrase is a sign of business value. Search Console is where I check the shortlist, on the client’s property or on any site I can see that already ranks near the topic, before a phrase gets a page.

Amazon, YouTube, Etsy and eBay publish no volume at all. A number next to an Amazon keyword in any tool is either Google’s volume for the same string or an estimate of an estimate. The parser leaves those cells empty and fetches Google’s volume and KD only for the rows you select, at 1 credit per 100 keywords. Empty cells are the normal state of a keyword list. Over 90% of the keywords in Ahrefs’ US database get fewer than 10 searches a month. Backlinko’s study of 306 million keywords put 91.8% of queries in the long tail, carrying 3.3% of total volume. Google says 15% of the queries it sees every day are new, which is demand no database can have a number for yet.

A blank volume is unknown demand. Zero demand is a phrase that Suggest, Questions and Trends don’t return either.

A list sorted by volume hands the decision to whichever rows happen to have a number. The rows without one are where most B2B demand sits: a phrase a buyer types twice a month can still be the one that books the call, and the way to measure it is to publish and read the impressions in Search Console four to eight weeks later.

Volume also counts searches, and a search can end without a click. Ahrefs’ example is “does a frog have a backbone”: about 700 searches a month and almost no clicks, because the results page answers it. AI Overviews extend that to more queries. Pew Research tracked 900 US adults in March 2025 and found they clicked a traditional result on 8% of visits when an AI summary appeared, against 15% when it didn’t. I don’t discount volume by a fixed factor for this. I note in a column whether the SERP carries an overview and compare shortlisted phrases with that column visible, because a question-shaped query with an overview and no commercial angle is the one most likely to bring impressions and nothing else. The queries an overview closes least well are the ones that need a comparison between options, a calculation, a file, or experience someone had first-hand, and on an informational list those are the ones I keep.

Keyword difficulty is one vendor’s model; the SERP is the check

Keyword difficulty is a tool’s estimate of the competition for a keyword. Every tool computes it its own way, and the vendors say so. Ahrefs bases KD on how many backlinks the pages in the current top ten have. Semrush offers Personal Keyword Difficulty, which changes with the domain asking: Exploding Topics showed “SEO tools” at 85% in general, 13% for semrush.com and 91% for a sporting-goods store. One phrase, three numbers, all correct inside their own model. KD describes the model and, in Semrush’s case, the domain, as much as it describes the keyword.

The thresholds in most guides, KD under 30, or under 40 for a new site, work as a filter on a list of two thousand rows. Nobody can open two thousand SERPs, and a filter that throws away the obvious fights is fine there. For the twenty phrases left at the end, the number has done its job and the results page takes over. Three things to read in it first: which domains hold the top ten and whether a site of your weight ranks among them; whether those domains sell the subject of the query, the way keyword-tool vendors sell keyword research; and what type of page they are, which is the next section.

The keyword of this article fails the second read and passes on a different ground. KD 4 reflects a top ten with few links to its pages. The pages themselves belong to Ahrefs, Semrush and HubSpot, brands with every reason to rank for the query and years of authority on the subject. A page that repeats their nine steps has nothing to win against them. The case for writing it at all is the material those pages can’t carry: data from a parser that queries twelve platforms, and rejections from a real keyword list, which is section five.

A fourth read decides whether the page is worth writing at all: what the top ten lacks. Open the ten pages side by side and note four things for each.

What to note for each pageWhat a gap looks like
When each page was last updatedMost of the top ten untouched for a year or more, in a topic where the tools, the documentation or the SERP itself has changed since
Who publishes itVendors explaining the workflow of their own tool, with no independent page among them
What the evidence isScreenshots of a tool and generic examples; no data of the publisher’s own and no worked case with numbers
How alike the pages areThe same steps in the same order on every page

A top ten that agrees with itself this closely is where a page with something new has a chance, and that gap matters more for the choice than the KD. It is also the case for this article: the pages above it run the same chain from seed to volume to KD to intent, and none shows a real list being cut. The read cuts the other way too. If the top ten already carries first-hand data, a recent study and a tool, a page without any of those has nothing to add however low the difficulty.

One more reason to open the live page. Ahrefs reported on 22 October 2026 that Google has started serving different results to suspected SERP data collectors, affecting between 5% and 20% of results depending on the provider. Tools filter and replace what they detect. For the shortlist, the results page you check by hand in the right country is the one to trust.

Which page type the SERP accepts

The search intent label a tool puts on a phrase is a classifier reading the words. The results page shows the search intent Google settled on, and the type of page in the top ten decides what can rank there, whatever the quality of a page of another type.

What the top ten showsSearch intentThe page that can rank
Guides, how-tos, definitionsInformationalAn article; a service page here gets nowhere
“Best X” lists, comparisons, reviewsCommercial investigationA comparison or a review that names alternatives, yours among them
Category pages, product pages, marketplaces, service pagesTransactionalA category, product or service page; an article sits below them at best
One brand’s pagesNavigationalNone, unless the brand is yours
A mix of the aboveSplitThe type you can build best, with a lower ceiling; the top three usually show which reading Google leans to

The rejection I make most often at this step is a service page for a guide SERP. The phrase sounds commercial, it has a service word in it, and the top ten is ten articles explaining the thing. Writing the article and linking it to the service page is the move that can rank there.

Does the keyword lead to what you sell, and is it already taken on your site

Business fit, the relevance of a keyword to what the site sells and to its target audience, is the check every guide mentions and few apply to their own list. Ahrefs scores it from 0 to 3. One question does most of the work: can the page link, naturally and in one click, to something you sell? A phrase that can’t pass it brings traffic from an audience that was never going to buy, however clean its numbers.

The second half of the check is the one guides skip: whether a page on your site already answers the phrase. In Google Search Console, Performance report, filter the query to the phrase and open the Pages tab. If a URL already gets impressions for it, that URL is the page to improve, and a new one will split the impressions between the two. On large sites, Query groups in Search Console Insights, rolled out from October 2025, clusters similar queries automatically, and the group a candidate phrase falls into shows which existing page Google already associates with it. The same report works as a source in the other direction: queries where the site already appears at positions 11 to 30 with one click or none are demand Google already connects to you, and they usually belong to a page that exists.

Here is the check run on a real list: the keyword plan for this blog, about fifty-five phrases with KD mostly under 15. Sixteen were taken, this article among them. Some of the rejections:

KeywordVolume (US) / KDWhy it was rejected
what are local citations880 / 4Local SEO; nothing on the site for the reader to go to next
seo link building service1,000 / 11The searcher wants to buy link building, and we don’t sell it
customized seo590 / 4Service intent; an article would compete with our own Strategic SEO page
canonical in seo590 / 5Already answered by our article on duplicate content
how to manage your backlinks590 / 7Owned by the backlink audit article
backlink exchange1,300 / 5A link scheme; not something to rank for under our name
ahrefs vs moz4,400 / 25Tool-comparison SERP, and the only rejection where KD was part of the reason

Six of the seven had a KD of 11 or less and would have passed any difficulty filter. On a list that has already been cut by KD, which is most lists by the time anyone reads them, the rejections come from fit and from ownership. Two of these phrases would have produced a second page on a question the site had already answered, and the existing page would have lost part of what it had.

One keyword or a cluster?

A page ranks for a set of queries, and the phrase you choose is the head of that set. The way to find which phrases share a page is the SERP: if two phrases return mostly the same URLs in the top ten, Google treats them as one intent, and they go on one page as primary and secondary keywords. If the URLs differ, they’re two pages. Our clustering tool groups a list by stems, by meaning or by that SERP overlap, and the overlap grouping is the one to trust for page decisions.

This article is an example. Its primary keyword is “how to choose keywords for SEO”, and “best keywords for SEO” and “finding keywords for SEO” return the same kind of pages, so they’re secondary keywords here. “Researching keywords”, 1,300 searches and KD 10 on the same list, returns library guides on academic search next to glossary entries. Its volume would fit; its SERP doesn’t, and it stayed off this page.

Keywords for AI answers: what can be measured

Google documents that AI Overviews and AI Mode may use a “query fan-out” technique, issuing several related searches across subtopics to build one answer. Those searches are ordinary queries. ChatGPT behaves similarly on the occasions it searches: Ahrefs analysed 66.6 million logged-out ChatGPT responses from August and September 2026 and found fan-out queries in 14.4% of them, about 1.6 queries each, with “2026”, “official” and “price” among the most frequent words. Those modifiers are the ones the Questions and A-Z sources already add to a seed.

What can’t be measured is the AI side of demand itself. Search Console counts AI Mode clicks and impressions inside the Performance report’s Web totals, without a label. Its generative AI report, available worldwide since August 2026, shows impressions by page, country, device and date, with no query dimension. Ahrefs, explaining how it estimates AI search demand, states that apart from Microsoft Copilot no major AI platform shares query data, and its own “AI adjusted volume” is Google volume multiplied by a ratio per platform. Even the vendor with the largest keyword database anchors AI demand to Google.

What I don’t have is a number for how much demand on a B2B topic now starts in a chat window and never reaches a search box, and no tool I know of has one either. Until a platform publishes query data, keywords for AI answers are Google keywords with the modifiers fan-out adds, and the same five checks apply to them.

Checklist: reject a keyword if…

CheckWhere to lookReject if
1. DemandSource column, Trends, sales and support questions; the live SERP for clicksNo source returns it and no customer has asked it, or an AI Overview answers it in full and it has no commercial angle
2. Who ranksTop ten domains, checked by hand in the right country; what the top ten lacksNo site of your weight is there and the page has nothing the top ten lacks
3. Page typeTypes of page in the top tenThe top ten is a type you can’t build
4. Business fitYour service, product or category pagesThe page can’t link to something you sell in one click
5. OwnershipSearch Console, Pages tab for the query; Query groupsAnother URL of yours already gets impressions for it
Then: clusterSERP overlap with phrases already chosenNot a rejection: if its top ten matches a chosen phrase, it becomes a secondary keyword on that page

Choosing keywords FAQ

Can SEO keywords be phrases?

Most are. Short-tail keywords, head terms of one or two words, are the top of the demand curve and the hardest to rank for; long-tail keywords are easier and more specific, and nearly everything a site wins is a phrase. Word order matters less to Google than to the tools: it usually reads “seo keywords how to choose” and “how to choose seo keywords” as one intent, while autosuggest completes each differently, which is why seeding both orders returns a wider list.

Is a keyword with no search volume worth targeting?

When it comes from customers, yes. A phrase heard on sales calls or in support tickets, returned by two or more sources and missing from every database, is often the most specific demand on the list. Publish, then read the impressions in Search Console after four to eight weeks; that figure is the only volume the phrase will ever have that is not an estimate.

Choosing is the first half of the work. Grouping the chosen phrases into pages and writing a brief for each is the second, and in our SEO service both happen before anyone opens a document to write.