SEO·October 2025·9 min read

The Evolution of Keyword Research in the Age of AI Search

Google stopped matching literal keywords to pages years before AI Overviews existed. RankBrain shipped in 2015, BERT in 2019, both built to understand concepts rather than exact words. Keyword research built around exact-match search volume was already behind; conversational AI search just made the gap visible.

The exact-match model was already broken

The pitch for "AI-era keyword research" usually frames semantic understanding as a brand-new shift search marketers have to catch up on. It isn't new. Google's own ranking systems guide documents RankBrain as a 2015 system built to help Google "understand how words are related to concepts," and BERT as a 2019 model that understands "how combinations of words express different meanings and intent."

Neural matching, also documented in that guide, does the same job at the page-and-query level: matching representations of concepts rather than strings of text. A workflow built entirely on matching a keyword string to a page has been fighting the underlying system for close to a decade.

What changed with AI Overviews is visibility, not mechanism. AI features use a query fan-out technique, issuing multiple related searches across subtopics before drafting an answer. That makes the semantic gap obvious in a way a ranked list of ten blue links never did.

Your query log is a better seed list than any keyword tool

Pull real queries from Search Console

Search Console's Performance report Queries tab shows the exact search terms already surfacing your pages, sorted by clicks, impressions, or CTR. Click through to the Pages tab from any query to see which URL is answering it today, and use the branded versus non-branded filter to separate returning-visitor traffic from genuine audience expansion.

A keyword tool's volume estimate is modeled and aggregated across a market. Your Search Console data is what real users typed to find you, which makes it a smaller but far more decision-relevant dataset than any third-party estimate.

Cluster by intent, not by word

Once you have the query export, group by what the searcher is trying to do, not by shared vocabulary:

Informational: "how does X work," "what is Y" — reward depth and clear structure over keyword density
Navigational: brand or product name searches — usually already won or lost, with little room left for content strategy to change the outcome
Commercial investigation: "X vs Y," "best X for Z" — reward comparison structure and specific criteria
Transactional: "buy," "pricing," "sign up" — reward clarity and friction removal over volume of content

Two queries with zero words in common can share the same intent, and a keyword-string clustering tool will never group them. Reading actual query text for intent takes longer per query and produces clusters that map to real content decisions.

A practical example: "how to reduce cart abandonment" and "why do customers leave without buying" share almost no vocabulary. Both are informational queries from the same buyer stage, and both belong on the same pillar page rather than two competing articles that split authority.

Finding content gaps without a magic tool

Content-gap analysis doesn't require a vendor's proprietary index. Pull your intent clusters, then check three things for each: does a page exist, does that page currently rank in Search Console for the cluster's core queries, and has search interest in the cluster moved recently in Google Trends.

A cluster with rising Trends interest and no existing page is a genuine gap worth prioritizing
A cluster with an existing page and flat or declining impressions signals a content refresh, not a new article
A cluster with strong impressions but weak CTR often means the title or snippet doesn't match the query's actual intent, which is a rewrite, not a content-gap problem

A keyword list tells you what strings people typed somewhere in a market. Your Search Console query log tells you what strings people typed to find you specifically. Only one of those is a decision input.

Topic structure is the real response to semantic search

Google's SEO Starter Guide recommends grouping "topically similar pages in directories," and notes that content organization affects how Google crawls and indexes a site. That's the practical version of the pillar-and-cluster model, minus the vendor branding: related pages, clearly linked, organized so both crawlers and readers can find the connections.

Internal links carry real weight here. The guide's own words: links help "connect your users and search engines to other parts of your site," and descriptive anchor text "tells users and Google something about the page you're linking to." A cluster of pages with vague "click here" links between them isn't a topic cluster, it's a folder.

Group content by the intent clusters from your query data, not by a generic category taxonomy
Link between cluster pages with anchor text that names the destination topic, not the page title
Route a portion of your internal link equity from high-traffic pillar pages toward newer cluster pages that need it

Why cluster freshness compounds

The ranking systems guide also documents a Passage Ranking System, which evaluates individual sections of a page rather than the page as a whole, and Freshness Systems, which surface newer content for queries where recency matters. A cluster of five well-linked pages, each updated on a rolling schedule, gives both systems more current material to work with than one page rewritten once a year.

Prioritize refresh cycles by cluster performance, not by publish date. A pillar page carrying most of a cluster's traffic earns a shorter refresh interval than a supporting article three links deep.

Structured data: use it for what it still does, not for what it's rumored to do

FAQ schema is the clearest example of a tactic that outlived its original purpose. As of May 7, 2026, FAQ rich results no longer appear in Google Search, per Google's own deprecation notice. The FAQPage type itself is still valid schema.org markup, but the specific rich-result payoff that made it a popular SEO tactic is gone.

Google's generative AI optimization guide is direct about the rest: structured data "isn't required for generative AI search." It still matters for standard rich results, and Google's general structured data guidelines set a clear bar regardless of which type you use — mark up only content visible on the page, keep it a true representation of that content, and specify all required properties for the rich result type you're targeting.

Picking the right type instead of defaulting to FAQ

Schema type	Fits when	Doesn't fit when
FAQPage	Page genuinely answers several distinct questions in Q&A format	You're adding it purely to try to win a rich result that no longer exists
QAPage	Users can submit and see multiple community-contributed answers to one question	There's only one authoritative answer with no user submissions — Google explicitly excludes this case
Article / BlogPosting	Standard editorial content with a clear author, publish date, and headline	The page is primarily a product listing or transactional page
Product	Page sells a specific item with price, availability, and reviews	The page is informational content with no purchase path

Measuring visibility that a rank tracker can't see

Search Console's Generative AI performance report shows impressions and clicks your pages receive specifically from AI Overviews and AI Mode, broken out by page, device, and country. A traditional rank tracker, built to check position for a keyword string, has no visibility into whether your content was cited inside an AI-generated answer.

Google's preferred sources feature adds another visibility layer: users can mark a domain as preferred inside AI Mode and AI Overviews, and preferred content gets a highlighted badge. It's domain-level only, so a well-optimized subdirectory can't qualify on its own — the whole domain has to earn that trust.

Google's quality rater guidelines, the document behind E-E-A-T, are explicitly "not a guide to ranking first in Google." Raters use them to evaluate how well the ranking systems are already performing, which means E-E-A-T is a description of what good content looks like, not a checklist you implement on a Tuesday.

A quarterly workflow that replaces the keyword-list habit

Export the last quarter's query data from Search Console, filtered to non-branded queries
Cluster by intent using the four categories above, not by shared words
Identify clusters where impressions are climbing but clicks aren't — a signal AI Overviews or a SERP feature may be absorbing the click
Check the Generative AI performance report for the same clusters, if available on your property
Map each cluster against your existing content and internal link structure, and flag gaps for the next content cycle
Re-run the whole export next quarter and diff against this one, rather than starting from a fresh keyword tool export each time

The workflow trades a bigger keyword list for a smaller, higher-confidence one built from what already works. It also produces a paper trail: every content decision traces back to a query your own site already receives, not a volume estimate from a tool with no visibility into your actual audience.

FAQ

Is keyword research dead in the age of AI search?

No. The exact-match version of it stopped reflecting how Google understands queries around 2015. Intent-based research using your own query data replaces it, it doesn't disappear.

Should I still add FAQ schema to pages?

Only if the page genuinely has Q&A content you want marked up for clarity. As of May 7, 2026, it no longer produces a rich result in Google Search, so it shouldn't be your primary SEO tactic.

Does structured data help my content get cited in AI Overviews?

Google states structured data isn't required for generative AI search. It's still worth using for standard rich results and to help Google parse page content accurately.

What's the difference between FAQPage and QAPage markup?

FAQPage covers a set of question-and-answer content authored by the page owner. QAPage is for pages where users submit and see multiple community-contributed answers to a single question — Google says not to use QAPage for standard FAQ content.

How do I measure visibility in AI Overviews if I can't rank-track it?

Use Search Console's Generative AI performance report, which shows impressions and clicks attributed specifically to AI Overviews and AI Mode, rather than a third-party rank tracker built for traditional position tracking.

Is E-E-A-T something I can implement directly?

Not as a checklist. It's the framework Google's quality raters use to judge how well ranking systems are performing, which makes it a description of content quality signals, not a discrete ranking factor you toggle on.

References

From the Destm engineering archive. For current work on this topic, start at Solutions or the blog.