Keyword Research in the Age of AI Search

Keyword research used to run on a simple formula: find phrases with high volume and low competition, write the page, collect the traffic. That formula is now broken — not because keywords stopped mattering, but because the relationship between "people search for this" and "a visit arrives at your site" has been severed for a large and growing slice of queries. AI Overviews answer questions directly on the results page. Chat assistants answer them without a results page at all. The searcher gets served; your analytics get nothing.

The wrong conclusion is that keyword research is dead. The right conclusion is that it needs a new filter: not just "do people search this?" and "can I rank?" but "if I rank, does a human actually arrive?" That third question changes almost everything about which keywords are worth your time.

What actually changed

Three shifts, all pointing the same direction:

Answers moved onto the results page. Definitional, factual, and simple how-to queries increasingly get an AI-generated answer above the organic results. Zero-click searches were already the majority of queries before AI features; the trend has only steepened for informational queries. Ranking #1 for "what is DNS propagation" now means being a citation under an answer most people won't scroll past.
Some questions never reach a search engine. A meaningful chunk of "how do I..." and "explain..." queries now go to ChatGPT, Claude, Perplexity, and their kin. This traffic doesn't show up as lost clicks; it shows up as declining impressions on queries that used to grow.
The queries that survive got longer and weirder. People trained by conversational AI now type full, specific, multi-constraint questions into search too. The long tail didn't just grow — it fragmented into queries your keyword tool has never seen and reports as "volume: 0."

Which keywords lost value — and which gained it

Think of every candidate keyword as sitting somewhere on an "AI-answerable" spectrum. The research skill of this decade is placing queries on it correctly.

Devalued: the commodity informational query

Definitions, conversions, general explanations, consensus how-tos. "What is a CDN," "how many ounces in a liter,""how to compress an image." These have one correct, stable, source-independent answer, which is precisely what AI features are best at delivering inline. Content targeting these still has a role (supporting topical authority, earning citations), but as a traffic strategy it's a melting iceberg. If your content plan is a hundred glossary posts, you're building for the past.

Holding or gaining: everything that resists a synthesized answer

Experience queries. "Is X actually worth it," "X review after one year," comparisons between specific products in a specific context. AI can summarize reviews; it can't have used the thing. First-hand accounts with real details — costs, screenshots, failure stories — are what both humans and AI systems go looking for.
Local and personal-context queries. Anything where the answer depends on being there or knowing the situation.
Transactional and navigational queries. People buying something or going somewhere still have to click. Commercial-intent keywords have quietly become more valuable relative to informational ones.
Fresh and volatile topics. Where the answer changes monthly, synthesized answers lag and hedge; a maintained page wins.
Deep, situational how-tos. Not "how to back up WordPress" but "restoring a WordPress site when the backup plugin's restore fails midway." Specificity is a moat: the narrower the situation, the more the searcher needs a practitioner's page, not a paragraph of consensus.
Your brand and category queries. As generic traffic gets absorbed, being searched for by name is the most defensible demand there is.

Volume numbers now lie in both directions

Keyword tools were always estimates; now their headline metric misleads structurally.

Overstated: a "5,000 searches/month" informational keyword with an AI Overview on the SERP might deliver a small fraction of the clicks it did three years ago. Volume measures questions asked, not visits available. Before targeting anything informational, search it yourself and look at the page: is there an AI answer? A featured snippet? People Also Ask boxes? Count what's between position one and the fold. The realistic click pool for that query might be a tenth of the volume figure.

Understated: the specific, long, conversational queries that still convert beautifully often report as zero volume because each individual phrasing is rare. But a hundred zero-volume phrasings of the same underlying need add up to real demand your tool can't see. Google Search Console is your corrective lens here: the queries you already get impressions for include long-tail phrasings no tool tracks. Mine them monthly.

The practical replacement metric is roughly: expected clicks = volume × (share of clicks the organic results actually get on this SERP) × (your realistic CTR at an achievable position). You'll estimate every term loosely — that's fine. Even loose arithmetic kills the fantasy keywords a raw volume sort produces.

Research the need, not the string

When queries fragment into thousands of phrasings, targeting exact strings stops scaling. The unit of research becomes the underlying need — the situation a person is in — and the unit of content becomes the page that resolves that situation thoroughly.

A workable workflow:

Start from problems, not tools. List the 20–40 concrete situations your customers find themselves in (from support tickets, sales calls, community forums, your own memory of being the customer). These are your topics; keywords are just the phrasings of them.
Expand each into query phrasings. Use your keyword tool, Search Console, autocomplete, People Also Ask, Reddit threads in your niche. You're mapping the vocabulary of the need, including the zero-volume variants.
Score each need, not each keyword, on three axes: click survival (would a searcher still need to visit a page?), business proximity (does this reader ever become a customer?), and your right to win (can you say something a synthesizer can't — experience, data, tools, opinions?).
Write one authoritative page per need, structured so every major phrasing is addressed — which mostly means clear H2/H3 questions and direct answers, not keyword-swapped near-duplicate pages. The near-duplicate play is doubly dead: it fails helpfulness standards and fragments your own authority.
Recheck the SERP before writing. The page you're about to spend two days on: search its target queries and see what the results page looks like today, not what the tool's screenshot implied.

Being the citation is the new consolation prize (and sometimes the prize)

When an AI answer does sit on top of your query, the remaining click flow goes disproportionately to cited sources. Getting cited isn't mystical: answers cite pages that state facts clearly, early, and quotably; that carry evident expertise signals; and that are technically easy to ingest (server-rendered, crawlable, structured). This changes on-page writing more than research: lead sections with the direct answer, use precise claims over vague ones, attach names and dates to assertions. A cited paragraph in an AI answer sends fewer visitors than a #1 ranking used to — but those visitors arrive pre-warmed, and the citation itself builds the brand searches that pay off later.

What this means for a small site's content plan

Concretely, if you run a small business site or a niche blog:

Cut the glossary-content roadmap unless it serves internal linking or sales enablement. Its traffic story is over.
Shift the ratio toward commercial, comparative, and experience content. A "we tested five of these, here's the spreadsheet" post is worth ten "what is" posts now.
Double down on what only you can publish: your data, your prices, your failures, your local knowledge. This is simultaneously AI-resistant and citation-attractive — the rare strategy that wins both ways.
Watch impressions and branded search, not just clicks. If clicks on informational queries sag while branded queries and direct traffic grow, your content is still working; the value is arriving through a different door.
Keep a Search Console mining habit. Monthly, export queries with impressions and no dedicated page. That list is keyword research no competitor can buy, because it's specific to the demand your site already attracts.

The honest summary

Keyword research isn't dead; the lazy version of it is. Sorting a tool export by volume and writing whatever floats to the top was always a mediocre strategy — AI search just turned mediocre into futile. What replaces it is closer to actual market research: understanding real situations, judging honestly whether a click survives the results page, and concentrating your writing where experience and specificity still beat synthesis. The sites hurting most right now are the ones that industrialized commodity answers. The ones holding steady wrote things a language model can summarize but never replace. Be the second kind.

Photo: Nothing Ahead on Pexels