Ask ten working SEO practitioners how they choose which keywords to target and you’ll get ten slightly different workflows, Google Keyword Planner versus Semrush versus Ahrefs, spreadsheets versus AI-assisted sorting, competitor-first versus intent-first. Ask what they’re all actually optimizing for underneath those workflows, and the answer collapses down to almost the same three things every time: does a real, sizable audience search this, can this specific page realistically compete for it, and does ranking for it actually move the business. That gap, wildly different tooling stacked on top of nearly identical underlying judgment, is the real story of keyword research in 2026, and it’s a genuinely useful lens for a bigger argument: SEO as a discipline has changed its instruments far more than it’s changed its actual logic since Google’s algorithm went semantic more than a decade ago. Not every practitioner agrees with that framing, and the disagreement itself is worth taking seriously rather than smoothing over.

How keyword research actually used to work, and why it broke

In the earliest era of search optimization, through the late 1990s and early 2000s, keyword research was close to literal word-counting. Early search engines, including the first versions of Google, relied heavily on on-page text analysis to judge what a page was about, and SEO advice of the era recommended fixed keyword-density percentages, repeating a target phrase a specific number of times per hundred words. That approach was gameable in an obvious way, and marketers gamed it, saturating pages with unnaturally repeated phrases in a way that actively degraded the reading experience for the sake of a ranking signal.

Google’s own algorithm history is the clearest record of that era ending. PageRank, introduced with Google’s 1998 launch, already began shifting weight away from on-page text toward link-based authority signals. The 2003 Florida update specifically targeted keyword stuffing and thin, manipulative pages. And Hummingbird, Google’s 2013 algorithm overhaul, is widely treated as the real turning point: it moved ranking evaluation toward understanding the meaning and intent behind a full query rather than matching isolated keyword strings, laying the technical groundwork for everything that followed, including RankBrain’s machine-learning ranking signals in 2015 and BERT’s natural-language understanding upgrade in 2019. By the time HubSpot formally introduced the “topic cluster” content model in 2017, arguing that interconnected clusters of content built around a single broad topic consistently outperformed scattered, isolated posts, keyword density as a standalone strategy was already functionally dead. The tools built to serve the density era, and later the tools built to score “keyword difficulty” as a single number, had to keep up with a target that had already moved.

The part of keyword research every experienced practitioner distrusts

Ask around any working group of SEO practitioners today and one opinion shows up with striking consistency: keyword difficulty scores, the single number Semrush, Ahrefs, Moz and similar tools attach to a keyword, are treated as directionally useful at best and actively misleading at worst. That skepticism isn’t just internet grumbling. Moz’s own founder Rand Fishkin has publicly acknowledged real variation in how difficulty scores are calculated and interpreted across tools, and independent comparisons of difficulty-score methodologies across major platforms have found the same keyword can receive meaningfully different difficulty ratings depending on which tool generated the number, because each vendor is running its own proprietary, largely undisclosed formula on different underlying data.

That distrust has only sharpened as AI-assisted keyword tools have entered the workflow. A recurring, near-universal piece of advice among practitioners actively doing this work is blunt: don’t trust an AI tool to tell you keyword difficulty or realistic ranking chances, because a language model has no live search index, no real-time SERP data and no way to verify a ranking prediction against reality, it is producing a confident-sounding guess, not a measurement. The more durable use of AI in this workflow isn’t prediction, it’s organization: grouping large keyword lists into topics, spotting genuine content gaps, and cutting a list down by removing grammatically malformed, duplicate, or clearly mismatched-intent terms once a human has set the actual selection criteria. That’s a meaningful, real distinction that’s easy to blur in practice, using AI to think for you versus using AI to process volume faster once you’ve already decided what you’re looking for.

What the actual filtering process looks like, stripped of tool branding

Set aside which specific tool a given practitioner prefers, Google Keyword Planner, Semrush, Ahrefs, Ubersuggest, all of them are just different windows onto overlapping, imperfect data, and the underlying triage process that experienced practitioners converge on looks remarkably consistent:

  • Cut the long tail of near-zero volume immediately. A common, simple floor, filtering out anything under roughly 50 monthly searches, removes most of the noise from a raw keyword export before any real analysis starts.
  • Sort by the ratio of opportunity to effort, not by volume alone. Lower estimated difficulty paired with higher volume gets checked first, but the estimate is a starting filter, not a final verdict.
  • Manually check a real, bounded sample by hand. A recurring number in practitioner workflows is checking the top 30 to 50 keywords by hand rather than trusting any automated sort past that point, specifically asking whether the page actually being optimized genuinely matches what someone searching that term wants to find.
  • Group by topic before ranking individual terms. Thematic clustering, treating a page’s keyword targets as a related family of terms serving one search intent rather than a list of disconnected strings, is the direct practical descendant of HubSpot’s 2017 topic-cluster model, and it remains the dominant mental model for how to structure a page’s targeting today.
  • Check where a site is already ranking positions 5 through 20. This is repeatedly cited as the fastest, most reliable source of near-term wins, because a page already ranking on page one or the edge of it needs meaningfully less new authority to break through than a page starting from nothing.
  • Study competitor rankings, but treat it as a starting point, not an answer. Identifying which real competitors and adjacent sites rank for a target topic, then reverse-engineering which of their pages rank for which specific terms, is a consistently cited method for finding realistic keyword targets grounded in what’s actually working for comparable sites, rather than theoretical volume alone.

The step most keyword-research advice skips entirely

The most consequential disagreement isn’t about which tool to use, it’s about whether volume and difficulty are even the right first filter. A more business-first version of the process argued by experienced practitioners starts somewhere else entirely: understand exactly who the client sells to, who they deliberately don’t sell to, and what that buyer needs to learn and decide at each stage of their purchasing journey, before opening a keyword tool at all. Under that framework, keyword research becomes the process of finding what real people search for at each stage of a journey you’ve already mapped, then triaging that list not just by search volume and difficulty, but by how tightly each term matches your actual, intended customer versus a merely adjacent one.

This is repeatedly described as the real dividing line in the field: the difference between people who do SEO and people who deeply understand both SEO and the underlying business. Cutting keywords that are topically related but don’t serve your actual target customer is where that difference shows up in practice, and it’s a judgment call no keyword-difficulty score can make for you.

Where advertising data exists, that same group treats paid-search cost-per-click and conversion data as a genuinely useful cross-check on organic keyword priority, reasoning that a term already proven to drive paid conversions carries real business signal a pure search-volume number doesn’t capture on its own, and that this data can help build an actual return-on-investment case for the resourcing a keyword push requires, not just a ranked list of opportunities.

The classical-SEO argument, and where it actually holds up

Put the full picture together and a specific argument becomes hard to avoid: the actual discipline of keyword research, matching real, honestly assessed search intent to a page genuinely capable of serving it, grouping related intent into coherent topical structures rather than isolated terms, and prioritizing effort against realistic return, has not fundamentally changed since Google’s shift toward semantic, intent-based evaluation crystallized around Hummingbird in 2013. What has changed, repeatedly and dramatically, is the instrumentation: from manual density counting, to link-graph-weighted crawlers, to machine-learning ranking systems, to large-language-model-assisted keyword sorting today. Each new tool generation promises to finally automate the judgment call at the center of the process, and each generation, so far, still needs a human checking whether the resulting page actually answers the question a real searcher is asking.

The newest wrinkle is real and worth naming honestly rather than folding into the same old pattern by default: AI Overviews and other generative-search surfaces increasingly answer a query directly inside the search results page itself, and large language models increasingly “fan out” a single user question into multiple related sub-queries behind the scenes when constructing an answer, a genuinely different retrieval mechanism than a classic single-query search. That’s a legitimate new layer on top of the discipline, not evidence the discipline itself has been replaced. Content built for genuine topical depth and clearly matched intent, the exact output the classical process above is designed to produce, is also the content best positioned to be cited inside an AI-generated answer, because the underlying requirement, being the clearest, most directly useful source on a specific question, hasn’t actually moved.

The specific warning from someone who’s actually done this at scale

That argument isn’t just a pattern visible in hindsight, it’s being made directly, right now, by practitioners with genuine track records watching a new generation of AI-native “thought leadership” arrive without the underlying experience behind it. Eli Schwartz, author of the bestselling book Product-Led SEO and an SEO and AEO (answer engine optimization) consultant who has worked with LinkedIn, Anthropic, Tinder, Coinbase and more than 20 other companies on organic growth, posted a pointed warning on LinkedIn making close to the same argument from the inside of the industry.

Schwartz’s core point is about where real SEO judgment actually comes from: “There is a growing wave of SEO tools and thought leadership coming from people who have never actually done SEO,” he wrote, arguing that the real risk isn’t a lack of fresh perspective, it’s a lack of accumulated knowledge of how search actually works, specifically the kind of knowledge that only comes from having watched something go wrong. “They haven’t watched a site tank from a migration gone wrong or seen what happens when you buy links at scale,” he wrote, noting he’d seen both problems firsthand in the month before his post. He’s also direct about who’s most exposed to bad advice as a result: “executives and marketing leaders with minimal SEO experience themselves” who “don’t know enough to ask the right questions” when a confident pitch arrives.

The specific claims Schwartz calls out track closely with the classical-versus-instrumentation argument made above: promises that a site can rank for competitive keywords and inside LLM answers using nothing but AI-generated content, claims that visibility inside large language models comes from automated workflows spamming Reddit comments, and the broader dismissal that technical SEO knowledge doesn’t matter anymore because “Google search is obsolete.” His conclusion is blunt: “Implementing these plans will lead to disappointment and tears, as anyone who has been on the wrong side of an algorithm update knows,” and, most directly relevant to the argument in this piece, “SEO and AEO playbooks exist for a reason.” (Full post on LinkedIn; Schwartz’s profile is at linkedin.com/in/schwartze.)

That’s worth weighing seriously rather than dismissing as a veteran defending turf. Schwartz’s specific examples, a migration that tanks a site’s rankings, a link-buying campaign that triggers a manual penalty, aren’t hypothetical edge cases, they’re exactly the category of failure that keyword-density-era SEO and today’s AI-native shortcuts share: a technique that looks like it’s working right up until an algorithm update or a manual review reveals it wasn’t. The playbooks he’s defending, matching real intent, building genuine topical authority, avoiding manipulative shortcuts, are the same classical fundamentals this piece has traced from Hummingbird through topic clusters to today. The tooling keeps changing. The reason those specific failure modes keep recurring hasn’t.

The counterargument: maybe it isn’t one discipline anymore

Not every practitioner frames the AI-search shift as a new layer on top of an unchanged foundation, and the strongest pushback deserves real space here rather than a dismissive footnote. Muhammad Waqar, an SEO and GEO (generative engine optimization) strategist who works with local service businesses, argued in a separate LinkedIn post that the “SEO is dead” framing and the “nothing has really changed” framing are both wrong in the same way: both treat AI search as a variation on the old game, when his actual position is that it’s splitting into two separate disciplines with different underlying signals.

“Ranking on Google hasn’t gone anywhere,” Waqar wrote. “Getting quoted, cited, or recommended by AI tools is a distinct, second job, built on different signals.” His warning is specifically for businesses trying to run both as one strategy: “Businesses treating these as one strategy are likely to feel like they’re losing visibility even while doing everything ‘right’ by old standards. The ones who separate the two are the ones positioned to win over the next few years.” (Full post on LinkedIn; Waqar’s profile is at linkedin.com/in/muhammad-waqar-899380427.)

A comment on that same post, from Nikita Vlasyuk, CTO and co-founder at FeedHeat, added real numbers to the claim: AI-referred traffic to US retail sites has more than doubled year over year, and AI Overviews now appear on roughly half of non-personalized searches. Independent industry tracking backs the general shape of that trend even if exact figures vary by methodology and dataset, with reported AI Overview presence ranging from a conservative 15 to 25 percent of mixed-intent queries up to around 48 to 50 percent on informational-heavy queries as of early 2026, and multiple analyses finding that a meaningful share of what AI Overviews cite doesn’t overlap with a page’s top-10 organic ranking for the same query, reported anywhere from roughly 17 to 54 percent overlap depending on the study and time period measured. That volatility in the published numbers is itself informative: this is a genuinely new, still-shifting measurement problem, not a settled one.

Where does that leave the classical-SEO argument made earlier in this piece? Narrowed, in a useful way, rather than refuted. Waqar’s split isn’t a claim that intent-matching and topical depth stop mattering, his own framing still treats “ranking on Google” as a discipline with its own real, unbroken continuity. His actual claim is narrower and, on the evidence above, hard to dismiss: that a second, adjacent discipline, optimizing specifically for what gets a page cited inside an AI-generated answer rather than clicked from a results list, now runs on partially different signals and needs to be worked as its own effort rather than assumed to fall out automatically from doing classical SEO well.

A third view: the click decline predates AI Overviews by years

A third practitioner argument reframes the whole debate again, and it comes with real, checkable data behind it. Devang Panchal, a digital marketer working across SEO and PPC, argued in a widely discussed LinkedIn post that blaming AI Overviews for the collapse in organic clicks gets the timeline backwards: “The decline started two years before Overviews existed. People stopped clicking the moment Google started answering questions directly in the search bar. Featured snippets. People Also Ask. Knowledge panels. All of it trained users to expect the answer without the click. AI Overviews didn’t create this behavior. It just gave it a name marketers could point at. The real shift isn’t AI vs SEO. It’s answer-first search vs link-first search, and that shift started in 2019. If you’re only reacting to Overviews, you’re reacting three product cycles too late.” (Full post on LinkedIn; Panchal’s profile is at linkedin.com/in/devang-panchal-5960471b7.)

The specific year he names checks out against independently tracked data, not just anecdote. Research from SparkToro and Jumpshot found that zero-click searches, where a user gets what they need directly from the results page and never clicks through to a website, passed the 50 percent mark for the first time in 2019, driven by exactly the SERP features Panchal names: featured snippets, People Also Ask boxes and knowledge panels answering queries directly on the results page. By 2020, some tracking put the zero-click share of all Google searches at close to 65 percent, well before AI Overviews existed in any form. Separately tracked data from that same 2016 to 2019 window shows organic click-through rates declining steadily even as paid-ad click-through rates rose, a pattern search-industry analysts have summarized as stable rankings, rising impressions, declining clicks, the same shape Panchal is describing, with a specific, checkable start date attached to it.

That timeline matters for how the earlier arguments in this piece should actually be read. It doesn’t contradict the classical-SEO argument that intent-matching and topical depth remain the durable core of the discipline, and it doesn’t contradict Waqar’s argument that a second, AI-specific discipline is now splitting off with its own signals. What it does is push the actual inflection point years earlier than the debate usually places it. If answer-first search behavior was already dominant by 2019 or 2020, then AI Overviews are less a cause of the click decline than the most recent, most visible symptom of a shift that had already been running for half a decade by the time most marketers started talking about it, which is a genuinely different, more uncomfortable claim than “AI killed SEO,” and harder to dismiss because the underlying zero-click data was published years before generative search existed to blame.

Common myths about keyword research

Myth: a lower keyword-difficulty score means a term is genuinely easier to rank for. Difficulty scores vary meaningfully between tools because each vendor calculates them differently on different underlying data, and independent comparisons have found the same keyword scored differently across platforms. Treat the number as a rough triage filter, not a verified prediction.

Myth: AI tools can reliably tell you your actual chance of ranking for a term. A language model has no live index of current search rankings and no way to verify a prediction against reality. Practitioners consistently recommend using AI to organize and cluster large keyword lists, not to forecast competitive outcomes.

Myth: more keywords targeted per page means more ranking opportunity. The dominant modern model, descended directly from HubSpot’s 2017 topic-cluster framework, treats a page’s keyword targets as one coherent, related family serving a single clear intent, not a checklist of disconnected terms stuffed into one page.

Myth: ranking well in classical Google search automatically means good visibility inside AI-generated answers too. Multiple industry analyses have found meaningful gaps between a page’s top-10 organic ranking and whether it gets cited inside an AI Overview for the same query, suggesting the two are related but not identical signals worth tracking and optimizing separately.

Myth: AI Overviews caused the decline in organic click-through rates. Zero-click search behavior, where a user never clicks through from the results page, passed 50 percent of all Google searches in 2019, years before AI Overviews existed, driven by featured snippets, People Also Ask boxes and knowledge panels answering queries directly on the results page.

Frequently asked questions

What’s the minimum search volume worth targeting for keyword research?
There’s no universal number, but a commonly cited practical floor among practitioners is cutting anything under roughly 50 monthly searches immediately, simply to remove noise from a raw export before deeper analysis, then adjusting that floor up or down based on the site’s actual scale and niche.

Why do experienced SEOs distrust keyword-difficulty scores from tools like Semrush and Ahrefs?
Because each tool calculates difficulty using its own largely undisclosed formula and underlying data, independent comparisons have found meaningfully different difficulty scores for the same keyword across different platforms, and the scores don’t reflect live, verified ranking outcomes.

What is a topic cluster in SEO?
A content structure, formally introduced by HubSpot in 2017, where a single pillar page covers a broad topic and multiple related pages on specific subtopics link back to it and to each other, signaling to search engines that a site covers a subject thoroughly rather than through scattered, disconnected posts.

Has AI search changed keyword research fundamentally?
It’s added a real new layer, AI Overviews answering queries directly on the results page, and language models fanning a single question into multiple related sub-queries, but the underlying requirement, genuinely matching content to real search intent with topical depth, is the same quality that positions content to be cited by an AI-generated answer. Some practitioners argue this new layer is significant enough to treat as its own discipline, distinct from classical ranking, rather than an extension of it.

What is GEO, and how is it different from traditional SEO?
GEO, generative engine optimization, refers to optimizing specifically for visibility inside AI-generated answers, being quoted, cited or recommended by tools like AI Overviews or chatbots, rather than for a ranked position in a traditional search-results list. Practitioners are actively debating whether GEO is a genuinely separate discipline with different signals or an extension of classical SEO fundamentals.

When did zero-click search behavior actually start becoming dominant?
Research from SparkToro and Jumpshot found zero-click searches passed 50 percent of all Google searches in 2019, several years before AI Overviews existed, driven by featured snippets, People Also Ask boxes and knowledge panels answering queries directly on the results page.

For more on how SEO fundamentals hold up in the AI search era, see our companion piece Why SEO Is Stronger Than Ever in the AI Era, and our companion analysis ChatGPT Is Now Legally a Search Engine in the EU on how AI visibility pipelines are reshaping search strategy.