Keyword Research in the AI Search Era: What You Think You Know vs. What You Might Uncover

The original version of this article ran on DirJournal back when keyword research mostly meant pulling search volume from Google Keyword Planner, sorting by competition, and picking the top of the list that you thought you could rank for. That approach is now actively harmful. Not outdated. Harmful. It produces content that ranks for nothing, gets ignored by AI Overviews, and wastes months of work before anyone notices.

The bones of good keyword research haven't changed in fifteen years. Relevance still matters. So does removing your own bias. The tools, the SERP, and the destination — all three have changed beyond recognition.

Here is what is still true, what is broken, and what most marketers are still getting wrong in 2026.

What Hasn't Changed

The principles are still the principles. If you target keywords your audience doesn't actually search, you fail. If you only chase the obvious head terms, you lose to bigger sites. If you let your assumptions about your business drive the research instead of letting the data drive it, you end up with content that validates what you already believe and helps nobody.

The "running shoes" vs "sneakers" problem from the original article is the same problem today, just with bigger stakes. A friend who sells industrial epoxy spent two years optimizing for "industrial floor coatings" before someone pointed out that roughly 80% of his actual customers were searching for "garage floor epoxy." Different volume. Different competition. Different intent. Same product.

That kind of blind spot is what keyword research is supposed to surface. That part of the job hasn't changed.

What's Broken About How Most People Still Do It

Sorting by search volume and picking the top five. This was suspect in 2010 and it is bad strategy now. The top-volume terms in any niche tend to be the most expensive to compete for, the worst converting, and increasingly the most likely to be intercepted by an AI Overview before the user clicks anything. According to a December 2025 Ahrefs analysis, the top-ranking page on queries with an AI Overview now gets 58% fewer clicks than equivalent queries without one. Even if you win the ranking, you are winning a smaller prize than you would have a year ago.

Treating "number of results" as a difficulty signal. "Returns more than one million results" used to be a rough proxy for competitive. It is now meaningless. A query that returns 4 million results might have a SERP dominated by Reddit threads ripe for displacement. A query with 80,000 results might be locked down by three .gov sites you will never outrank. The number on the SERP tells you nothing. The composition of the SERP tells you everything.

Tools that no longer exist or have been gutted. The original version of this article recommended Wordtracker, KeywordDiscovery, kwmap.net, seedkeywords.com, Google Insights, and Google Sets. Most are dead, abandoned, or shadows of what they were. Google Keyword Planner, the workhorse for over a decade, now hides volume in ranges for non-advertisers and rounds aggressively. If your research process has not been rebuilt in the last three years, it is running on broken inputs.

Intent Before Volume

If you change one thing about how you research keywords, change the order. Look at intent first. Use volume as a secondary filter.

Every query maps to one of four intents: informational (someone wants to learn), navigational (someone wants a specific site), commercial (someone is comparing options before buying), or transactional (someone is ready to act). Google's quality rater guidelines actually use an eight-part classification, but the four basics are enough for most decisions.

You don't need to guess at intent. Open the actual SERP and look at what ranks. If the top ten results are mostly product pages, the intent is transactional. If they are mostly comparison articles and "best X" listicles, the intent is commercial. If they are definitions, explainers, or how-tos, it's informational. The SERP is the answer key. Google has already done the classification for you. You are just reading it.

This is the single move that separates real research from theatre. Most people skip it.

The SERP is the Research, Not the Result

When you decide whether to target a keyword in 2026, spend more time on the SERP itself than in any tool's dashboard.

Look at who ranks. What content formats they use. What schema they implement. Whether an AI Overview is present and which sources it cites. Whether there is a featured snippet. What the People Also Ask box contains. Whether there is a video carousel, a local pack, a knowledge panel, an image block.

A SERP with no AI Overview and a sparse PAA box is an opportunity. A SERP with an AI Overview, a heavy PAA box, three video results, and a featured snippet is a battlefield where you will fight for a thin slice of remaining clicks.

The traffic context matters too. SparkToro's 2024 clickstream analysis with Datos found that of every 1,000 US Google searches, only 360 produced a click to the open web. The rest ended on Google itself, with no website visit at all. The remaining clickable traffic is now distributed across a more crowded SERP than ever. Pick your battles using what you can see, not what the volume column tells you.

Question Mining is Now Central, Not Optional

Voice search was supposed to make question-based keywords matter. It made them matter a little. Generative AI made them matter a lot.

When a user asks ChatGPT, Perplexity, Gemini, or Google's AI Overviews a question, the model is doing its own internal version of keyword research. It synthesizes what it knows about the topic and pulls from sources it considers authoritative. The questions being asked are largely the same conversational, long-form, problem-stated queries that the People Also Ask box has been surfacing for years. Mining those questions and answering them directly, in plain language, is now a primary path to AI visibility.

The data backs this up. A Yext analysis of 6.8 million AI citations across ChatGPT, Gemini, and Perplexity found that 86% of citations came from brand-controlled sources — websites and listings, not Reddit or forums. The brands showing up in AI answers are the ones answering the questions clearly, on their own pages, in their own words. Not the ones with the highest "search volume" on the head term.

Practical sources for question mining:

• The People Also Ask box on actual SERPs. Click through to expand the secondary questions, which are often more valuable than the surface ones.

• AlsoAsked.com and AnswerThePublic for bulk discovery.

• Reddit and Quora threads in your niche. Search [topic] site:reddit.com and read what real people are asking.

• Your own Google Search Console. The long-tail queries you are already getting impressions for are pure gold.

• Sales conversations and support tickets. The actual phrasing customers use is almost never captured in any tool.

That last one is underrated. The single most valuable list of keywords in any business is usually sitting in the support inbox.

Entities, Not Just Keywords

This is the conceptual shift that takes most practitioners longest to adjust to. It's also the one most working SEOs still under-invest in.

Search engines and LLMs think in entities — discrete people, places, things, concepts — and the relationships between them. The query "best CRM for small law firms" is not being processed as a string of words. It's being processed as [CRM software] + [small law firm] + [comparison intent] + [implicit attributes: affordable, easy setup, legal-specific features]. Your job is not just to use the keyword. Your job is to make sure your page is clearly about the right entities and connected to the right adjacent entities.

In practice:

• Use the exact terminology your industry uses. Not your branded variants.

• Mention adjacent entities — competitors, related concepts, key people, relevant standards — that someone writing a comprehensive piece would naturally include.

• Implement schema markup that explicitly identifies entities: Organization, Person, Product, Service, FAQ, Article.

• Build out Person schema with sameAs links to your verified profiles across the web.

Ahrefs research analyzing 75,000 brands found that brand web mentions correlate with AI citation rates at 0.664 — about three times stronger than backlinks at 0.218. Being talked about, in the right contexts, on the right surfaces, is now a stronger AI visibility signal than the link counts most SEO dashboards still emphasize.