The Myth of Long-Tail Keywords, Revisited: Voice Search Faded and AI Took the Long Tail

What this post’s own Search Console data shows

Before rewriting this post, we pulled its Google Search Console data for the last six months. It had 2,970 impressions and zero clicks. That’s not unusual for an old post. The queries were.

Here are a few of them, copied exactly:

• google search central voice search conversational language long-tail keywords
• google search central voice search long-tail keywords official
• google search central voice search seo conversational long-tail keywords
• google search central voice search conversational long tail keywords seo official

There were about forty variations like that, and almost nothing else. Nobody types “google search central … official” into a search box. I can’t prove where these came from (Search Console doesn’t say), but they look exactly like what an AI system generates when it breaks one question into several searches and goes looking for an authoritative source. Our page showed up for those searches at an average position of 5 to 9, got looked at by a machine, and never got a click.

I’ll admit that’s a slightly embarrassing way to open a post about long-tail keywords. It’s also the best evidence I have for what the rest of this post argues.

What we said in 2024, and what we got wrong

The 2024 version of this post argued that voice search and long-tail keywords weren’t the holy grail SEO blogs made them out to be. The core of it was my own habit: I don’t use voice search in public or at work, and when I search for business I type short, precise queries. So I was skeptical of the advice that every business should rewrite its content around long conversational phrases.

I still think that was mostly right. But it was right partly for the wrong reason. The long, conversational query didn’t fail to arrive. It arrived through a different door, typed into an AI chat box rather than spoken to a smart speaker.

We’ve also removed a table from the 2024 version, and there’s a note about it at the end of this post.

What happened to voice search

Voice didn’t disappear. People still ask their phones for directions and their speakers for the weather. What changed is who answers.

Google announced in March 2025 that Gemini would replace Google Assistant, and the removal of Assistant from Android phones began on September 4, 2026, rolling out over several weeks, according to reporting in the Khaleej Times. Smart speakers and Google TV devices weren’t part of that phase. So the spoken question on a phone now goes to a generative assistant that writes an answer, and it doesn’t read out the top result the way the old voice-search advice assumed.

That matters for the advice we used to give. “Optimize for voice search” used to mean writing a short, snippet-shaped answer near the top of the page so Assistant would read it aloud. Today the assistant reads several sources and writes its own answer. You’re not competing to be read out. You’re competing to be one of the sources it trusts enough to use, and, if you’re lucky, to cite.

Where the long tail went

Two things Google has said in public explain most of it.

First, the questions got longer. When Google started testing AI Mode in March 2025, Robby Stein, Google’s vice president of product for Search, told TechCrunch that people were asking questions about twice the length of traditional searches. That’s the long tail, right there, just not in the classic search box.

Second, the engine writes long-tail queries of its own. Google calls it query fan-out. In a video interview reported in July 2025 by Search Engine Journal, Stein described AI Mode taking one question, running several related searches in the background (including some on angles the user never mentioned), and combining the results into one answer. Deep Search can run far more of these over several minutes.

Put those together and you get the strange Search Console data at the top of this post. The person asks one long question. The engine turns it into a dozen narrower ones. Your page gets an impression for a query no human typed, and the click (if there ever was going to be one) goes nowhere, because the answer was assembled before anyone reached your site.

What we’d do now

Here’s how I’d change the advice today.

The last row is the one I’d push hardest. Keyword tools are slow to show conversational, multi-part questions, because nobody searches the same twelve-word question twice. Your client’s inbox is full of them, and that’s where I’d start a page’s question list, before any tool’s “questions” report.

If you want the longer version of this approach, our post on answer engine optimization goes into page structure, and the post on search intent optimization covers how to tell what each query actually wants.

Where long-tail keywords still pay

This next bit is the part of the 2024 post I’d keep almost word for word, because it held up.

When people are ready to buy, they type. They type model numbers, part numbers, sizes, and the name of their suburb. “Bosch GSB 18V-55 chuck replacement” and “emergency plumber Parramatta Sunday” are long-tail keywords, and they’re more valuable than ever, because they’re the queries an AI answer is least likely to finish for the searcher. The person wants a specific page, a specific product, or a phone number.

So for eCommerce product and category pages, local service pages, and niche B2B products with real model names, the classic long-tail work still pays. Get the exact names, numbers, and places on the page, in the title, and in the structured data. That part of SEO hasn’t changed much since 2010 (well, apart from the structured data).

Checklist

• Pull your own Search Console queries for your top pages and look for machine-shaped queries you didn’t expect.
• Merge thin pages that each target one long-tail variation into one page per topic.
• Write each H2 or H3 as a plain question, and answer it in the first two sentences under the heading.
• Build the question list from real customer emails and support tickets, not only from keyword tools.
• Keep exact model numbers, part names, and locations on product and local pages.
• Remove any statistic you can’t source, including old ones.

FAQ

If your own Search Console has started showing queries like the ones at the top of this post, I’d like to see them. Post a couple in the comments and I’ll add the interesting ones here, with credit.

A correction to our 2024 version

Updated October 2026. The first version of this post, published in December 2024, included a table estimating voice search adoption by age group from 2020 to 2024, with values like “likely 70%+.” Those figures weren’t taken from any study, and they shouldn’t have been presented as data. We’ve removed the table, and we haven’t replaced it with anyone else’s estimates.

If you used those numbers in a report or a client presentation, we’re sorry. Please treat them as unreliable. Everything we cite in this version links to its source, and if you spot something we’ve got wrong, the comments are open.