Stop thinking about keywords: how conversational search really works (2026)

The short answer: Google says the average question asked in AI Mode is three times longer than a normal search, and behind each one the machine quietly runs a string of searches you never see. That is query fan-out, and it is why AEO starts with conversations, not keywords. People haven't changed what they want to know. They've changed how they ask.

Think about the last thing you typed into Google for your business. Probably two or three words. Now think about the last thing you asked ChatGPT. Probably a paragraph, with your situation, what you'd already tried and what was worrying you.

Your customers have made exactly the same jump. In my experience, most marketing plans in this country are still built for the two-word version.

This article covers what changed, how Google's People Also Ask box saw it coming, what fan-out queries are and how they decide who gets named, how fast it is happening, and what to do about it this week. Every figure is sourced and dated, and vendor research is marked as such.

Watch the full video on YouTube. Listen to the podcast episode on Will AI Recommend You? on Buzzsprout. The cleaned full transcript is further down this page.

Why did keywords work for so long?

Keywords worked because search used to behave like a vending machine: a short phrase went in and ten blue links came out. A whole industry grew up around guessing which phrase to put in the slot.

Keyword research, search volume and "this phrase gets 1,900 searches a month, write a page for it" became the job. It worked, and anyone who has spent time in search, as I have for thirty years, ranked plenty of pages that way.

But there was always a crack in the model. Google has said since 2013 that around 15% of the searches it sees every day have never been seen before, and restated it in 2017 (Search Engine Land). Google's John Mueller has since said that whenever they recalculate it, it still comes out around 15% (Search Engine Journal).

No keyword tool contains those searches, because nobody had searched them yet. Even in the vending-machine era, the machine was being asked things nobody had planned for.

The drift towards longer searches isn't new either. In 2009, Hitwise reported that searches of eight or more words had grown about 20% in a year (MarketingCharts). What is new is the speed, and the fact that the machine now answers in full sentences.

What was People Also Ask really telling us?

People Also Ask was Google showing its working: a live map of the questions people ask next, published years before anyone was holding conversations with a search engine.

Google started testing the feature on 17 April 2015 (Common Denominator). By 2017, clicking one question loaded two or three more beneath it, and the list kept going for as long as you kept clicking (Search Engine Journal). Google's own demonstration started with bats and ended several clicks later at which kinds of bat live in Michigan.

Most businesses treated the box as clutter, or at best a list of blog titles. It was more important than that. It told you that when someone asks this, the next thing they usually ask is that, and then that. It was a conversation, laid out in advance.

Who answers People Also Ask now?

Google's AI does. People Also Ask used to be answered with a snippet lifted from a website, with a link to it. That was a route onto page one for sites that didn't hold the top position.

AlsoAsked analysed 19.2 million People Also Ask results in 2026. In the first week of September, 97% of the English-language answers it sampled were AI Overviews. A month earlier the figure was about 86%, and roughly 14 months earlier it was about 12% (Brafton, reporting AlsoAsked data, 18 September 2026).

When	Share of sampled PAA answers that are AI Overviews
About mid-2025	~12%
August 2026	~86%
First week of September 2026	97%

The questions are still there, so Google still believes people ask them. The answers are now written by the machine and stitched together from several sources.

Semrush found People Also Ask on about 90% of results pages that also carry an AI Overview (Semrush AI Overviews study, updated December 2025, vendor research). The modern Google page is an AI answer at the top and a stack of follow-up questions underneath. That is a chat interface wearing a search engine's clothes.

What is a fan-out query?

A fan-out query is one of the hidden searches an AI engine runs on your behalf: you ask one question, it breaks that question into many smaller ones, searches each separately and writes a single answer from what comes back.

Google described it at I/O 2025: AI Mode breaks your question into subtopics and issues a multitude of queries simultaneously on your behalf (Search Engine Journal). Its Search Central guide, published 15 May 2026, confirms that both AI Overviews and AI Mode use the technique.

Google isn't alone. ChatGPT, Gemini, Perplexity and Grok all fan out, at very different rates.

Engine	Hidden searches per prompt	Source
Gemini 3	10.7 on average (range 3 to 28)	Seer Interactive
Grok	6.8	Peec, May 2026
ChatGPT	2.1	Peec, May 2026
Perplexity	1.4	Peec, May 2026

Two findings matter most for anyone still buying keyword reports.

First, 95% of the hidden searches Gemini 3 generated had zero search volume (Seer Interactive, agency research). They don't exist in your keyword tool and they never will, because no human types them. The machine does.

Second, the hidden searches are getting more specific. Peec analysed more than 20 million ChatGPT fan-outs and found their average length roughly doubled, from about 6 words to about 12, between October 2025 and January 2026 (Peec, February 2026, vendor research). The machine is asking narrower questions, not broader ones, and broad, generic pages struggle to match them.

The machine is also searching more than once. After ChatGPT's default model changed to GPT-5.6 in August 2026, the share of prompts needing only one round of hidden searches fell from 94% to 43.5%, and the number of sources read per prompt roughly doubled from about 12 to about 24 (David Konitzny, via Peec, August 2026).

What does fan-out look like in practice?

Fan-out turns one simple question into the four or five things a sensible person would also want to know, then searches each one. Real fan-outs change from run to run, so treat these as typical patterns rather than fixed lists.

"How to fix a lawn": Google's own example

Google's guide to its generative AI features uses this question to explain the technique. "How to fix a lawn that's full of weeds" might fan out into "best herbicides for lawns", "remove weeds without chemicals" and "how to prevent weeds in lawn". An answer like that can draw on several kinds of page, such as a product comparison, a how-to and a short factual answer. A site that only targeted "lawn repair" covers one branch at best.

"Best headphones": a ChatGPT example

Peec's analysis of 5 million ChatGPT fan-outs used this one (Peec, May 2026). It becomes best wireless headphones, best noise-cancelling headphones for flights, best budget headphones and best headphones 2026.

Nobody typed "flights". The machine added the use case. ChatGPT also routinely adds words such as "best", "top", "reviews" and the current year, then merges the results. Pages that turn up across several of those searches score higher than pages that win only one (Peec).

A local service: the Doncaster accountant (illustrative)

Someone asks: "Which accountant near Doncaster is good for a small builder who's fallen behind on his VAT?"

Hidden search	What it's after	The page that wins it
accountants in Doncaster	Who exists locally	Business profile and directories
accountant for construction trades UK	Sector fit	A service page that names builders
help with late VAT returns small business	The actual problem	A plain-English explainer
HMRC penalty late VAT return	The fear behind the question	An answer page with real figures
[firm name] reviews	Trust	Reviews on third-party sites
accountant fees sole trader builder	Cost	Published price ranges

The firm that gets named has a page about builders, a plain answer on VAT arrears, published fees and reviews that agree with all of it. A firm that only ranks for "accountant Doncaster" may not appear at all.

A B2B decision: choosing a recruitment partner (illustrative)

Business buyers rarely ask once. Picture the operations director of a 120-person engineering manufacturer in Sheffield, with two maintenance engineer roles open for four months.

Turn	What they ask	Likely hidden searches	What gets a recruiter named
1	"Should we use a recruitment agency or keep trying ourselves?"	average time to hire maintenance engineer UK · cost of an unfilled engineering vacancy	Real time-to-fill and vacancy cost, in plain numbers
2	"Which agencies in South Yorkshire specialise in engineering and manufacturing?"	engineering recruitment agency Sheffield · [agency] reviews	The same name, location and specialism on your site, profiles and directories
3	"What do they charge, and is contingency or retained better for hard-to-fill roles?"	recruitment agency fees UK · contingency vs retained search	Published fee ranges and a straight answer on which model suits which role
4	"How do I know they can actually find maintenance engineers?"	[agency] maintenance engineer placements · recruitment case studies manufacturing	Named case studies with time to shortlist and retention
5	"What happens if the person leaves after a few weeks?"	recruitment rebate clause explained · free replacement guarantee	Rebate terms written out on the site, not buried in a PDF
6	"Out of those agencies, which would you pick for us and why?"	[agency A] vs [agency B] · [agency] Google reviews	Third-party reviews and mentions that agree with your own site

Turns 1, 3 and 5 never name an agency, yet they shape the shortlist. The recruiter whose content answered them is already trusted by turn 6. A keyword plan built on "engineering recruitment Sheffield" covers turn 2 and nothing else.

Why can you rank and still be ignored?

Because the AI judges your page against the hidden searches, not just the question your customer typed, a top-ten ranking no longer guarantees a place in the answer.

In July 2025, Ahrefs found that 76% of pages cited in Google's AI Overviews also ranked in the top ten for the same query. Its February 2026 update, covering 863,000 keywords and 4 million AI Overview URLs, put that figure at 38%. Around 37% of cited pages didn't rank in the top 100 at all. Ahrefs pointed to fan-out as part of the reason (Search Engine Journal, vendor research).

ChatGPT shows the same pattern. AirOps ran 16,851 queries through ChatGPT three times each and found that 32.9% of cited pages appeared only in results for a hidden sub-query, never for the prompt the person actually typed (Growth Memo, vendor research).

So it cuts both ways. You can be invisible for your customer's question and still be cited because you answered one they never typed. Or you can rank beautifully for your keyword and be skipped entirely. Ranking used to be a ticket into the answer. Now it's closer to a raffle ticket.

That matters because the answer is increasingly where people stop. Pew Research tracked 68,879 real Google searches by 900 US adults in March 2025. When an AI summary appeared, people clicked a traditional result on 8% of visits, against 15% when there wasn't one. They clicked a source link inside the summary on just 1% of visits (The Register, reporting Pew). Being named inside the answer is increasingly the whole interaction.

How fast is search becoming a conversation?

Fast. Google's AI Mode passed a billion monthly users in its first year, and its queries more than doubled every quarter.

What Google says

Google published a year-one review of AI Mode on 19 May 2026 (Google). The headlines:

• AI Mode passed a billion monthly active users globally, and its queries more than doubled every quarter since launch.
• The average AI Mode search is triple the length of a traditional search.
• More than one in six US searches now use voice or images, with image searches growing over 40% a month.
• Planning queries grew 80% faster than AI Mode overall over six months, and brainstorming queries 30% faster since launch. Searches starting "where should I" and "ideas for" are growing.

Some of the fastest-growing questions aren't facts. They're plans and decisions, which is exactly where a business wants to be named.

Longer questions get the AI's answer

Pew Research found that only 8% of one- or two-word searches triggered an AI summary, against 53% of searches of ten words or more (The Decoder, reporting Pew). The more your customer talks like a human, the more likely they get the machine's answer instead of a list of links.

The ad money is already moving

Search Engine Land's analysis of one agency's Google Ads accounts to August 2026 (also summarised by Optimixed) shows the same shift in paid search. It is one agency's data, not Google-wide.

Query length	Share of impressions a year earlier	Share of impressions, Aug 2026
1 to 2 words	42%	24%
3 to 4 words	33%	48%

Searches of seven or more words went from 1% to 4% of conversions over the same period. Advertisers are paying for this shift whether or not they've noticed it.

It's happening in the UK, and not just among AI users

RealityMine, a Manchester research firm, tracked behavioural panels in the UK and US (RealityMine, 23 April 2026). In the UK, the average Google search grew from 21 to 26 characters in a year. Searches over 30 characters rose 24%. Google searches per person rose 26% against May 2025, and more searches now start with who, what, where, when and why.

The important finding is who is doing it. The pattern was the same for heavy ChatGPT users, light users and people who don't use ChatGPT at all. Google's own AI answers have taught everyone that a full question gets a better result.

RealityMine also found the mix of search intent barely moved between May 2025 and March 2026. People aren't searching for different things. They're saying them differently.

How many follow-up questions do people ask?

No platform publishes an average number of follow-ups per person. The best available figures show most sessions are still short, but follow-ups are among the fastest-growing behaviours Google reports.

Measure	Figure	Source
Follow-up growth in AI Mode	Up more than 40% a month on average in the US since launch	Google, May 2026, via PPC Land
Queries per AI Mode session	2 to 3, against 5 or more in a traditional Google session	Semrush, via AuthorityTech (vendor)
Messages per ChatGPT conversation	1.7 on average; a third open with a question	WebFX, 13,252 conversations, November 2025 (agency)
Prompts needing one round of hidden searches	Down from 94% to 43.5% with ChatGPT's GPT-5.6	David Konitzny, via Peec, August 2026

The honest reading is that most people still ask one or two things. Buyers keep going. A commercial decision, such as choosing a recruiter or an accountant, sits at the long end of that curve.

Every follow-up triggers another round of hidden searches. And the machine now asks itself follow-up questions even when the person doesn't. So a two-turn conversation can easily mean twenty or more searches you'll never see in a report.

Is Google right that this is still just SEO?

Google is half right. It's right about the tactics people are selling, and wrong to suggest the unit of search hasn't changed.

On 15 May 2026, Google Search Central published its first guide to optimising for generative AI features. It says optimising for AI search is still SEO. It warns that creating pages for every variation of how people might search, fan-out queries included, can breach its scaled content abuse policy. It also says you don't need special AI markup, chunked content or an "AI-friendly" writing style (TechWyse summary).

On the tactics, we agree. A page for every fan-out query is spam, and anyone selling it is selling hopeful activity with a new label. The research backs that up too: AirOps found pages that covered every sub-question of a topic were cited less than pages covering a solid chunk of them (Passionfruit, summarising AirOps). Breadth is not the same as usefulness.

It's also worth being straight about the size of the channel. Referrals from AI assistants are still a small slice of total website traffic. Conductor put them at about 1% of visits across ten US industries in 2025, growing roughly 1% a month (Conductor). Semrush's 2026 channel study puts AI traffic lower still, at 0.14% of the total, though it grew 66% in 2025 (Semrush). It's fast-growing and the visitors are high quality, but it isn't a replacement for everything you already do.

Where we disagree is the idea that nothing has changed. A plan built on a keyword list starts from the wrong unit. The keyword is no longer the unit of search. The conversation is. AEO is building the marketing for that conversation, and AEO leads to SEO, not the other way round. Done properly, the traditional rankings follow as a by-product.

What should you do this week?

Start from the questions your customers actually ask, in their own words, and build a small number of pages that hold the whole conversation.

1. Collect real questions, not keywords. Record your sales calls. Ask your customer service team what they hear every week. Write each question down exactly as it was said. Not "forklift hire" but "can I hire a forklift for a weekend without an operator's licence?" That's how they talk, and now it's how they search.
2. Use People Also Ask as a free map. Search your main service and keep clicking the questions. You're looking at Google's own guess at the conversation. Then ask ChatGPT, Gemini and Google AI Mode the same question the long way, and notice what they go off and look up.
3. Build one page per conversation, not fifty pages per keyword. Answer the main question in the first few lines. Then answer each follow-up under a heading that is the real question, with a complete answer directly beneath it. Cover the questions that matter, not every possible variation.
4. Publish what only you know. Prices, process, timescales, real case studies with numbers, and the answer to the awkward question competitors avoid. The machine has read everyone's version of the basic answer. What earns the citation is the part only you can give it.
5. Make the outside world agree with you. The final turn of most buying conversations is a trust check. Your name, location, specialism and claims need to say the same thing on your site, your Google Business Profile, directories and reviews.

Does your business come up?

I've been writing about answer engines since 4 October 2019, 1,153 days before ChatGPT launched. I got parts of it wrong: I expected us all to be talking to speakers in the kitchen by now, and we aren't, quite. The direction was right. Search was always heading towards a conversation. It has simply arrived faster than anyone planned for.

So try this today. Open ChatGPT or Google AI Mode and ask the question your best customer would ask before they ever found you. Ask it properly, the long way, with their situation and their worry in it.

Does your business come up? If it doesn't, who does, and what did they publish that you haven't?

What does Simon say in the video? (full transcript)

Simon Young, Question.Marketing. Recorded October 2026. Timestamps follow the published edit (33:44). Lightly cleaned: fillers and false starts removed, British spelling, wording otherwise as spoken.

Why should you stop thinking about keywords?

[00:01] Hey, everybody. Simon here from Question.Marketing, and I wanna talk to you today about try to stop thinking about keywords. Think, think, think more about how people are going to be speaking, how we're gonna ask questions of LLMs, and how search is evolving in 2026, and this is the latter part. We're into Q4 of 2026. We're looking now at how queries fan out and, I want you to understand how people are now searching and use that in your marketing, which is gonna be hugely important. Probably one of the most important changes in how search works is happening right now, and the opportunity is huge.

How are people searching now that they can ask questions?

[00:47] So as we move forward, we're gonna go into a world where, you know, we used to type for example, accountant Doncaster. You would type in a phrase or a keyword, and we're now moving into a world where people are prompting, obviously. We're asking questions of LLMs. And you might then, in that example, somebody might ask, "Which accountant near Doncaster is good for a small builder who's fallen behind on his VAT?" Now, one of the findings from Semrush was that queries are now have gone from maybe, you know, three or four words when we were working a keyword universe or a key phrase universe, to now 23 words on average, where people are asking questions.

So we now need to think, how does this affect us?

Why have 15% of Google searches never been seen before?

[01:50] One thing that is really important to understand, and I don't think many people actually realise this, before AI, way before AI, all the way through search, people have on average been typing in things that aren't answered, things that are not answered, right? So 15% of the time, and this is according to Google, every single day, there are searches going into Google that have never been asked before, right? No keyword tool can reference them or has use of them because they've never been typed in.

This is how the human mind works, right? We need to understand that if somebody hasn't searched something yet, how can we optimise for it? And if, you know, if you are an SEO listening to this or you're dealing with an SEO company who are still talking about keywords, trust me, they are getting it wrong. We need to move to a query-based, prompt-based, question-based way of optimising.

What was People Also Ask telling us?

[03:01] So if we go back in time a little bit let's look back to April 2015. Google started testing a system called People Also Ask, and hopefully we all understand how this works. When you search something, a little way down the page you will see People Also Asked. Now, our business is Question.Marketing. The reason it's Question.Marketing is because it is about the questions that people are asking, and I want you to seriously think about how you optimise for the questions. Now, based on the previous slide that I showed you, think about also the fact that 15% of the queries that are going into search and LLM have never, ever been typed in before.

So that might sound ridiculous, but you know, there's 15% of the world that are putting things in a different way. So purely rephrasing things, wording them differently, asking different questions, and that is every single day. So think how that compounds, right? 2017 each click then started to load a lot more questions, and the box never ends. If you've ever tried this, if you go to the People Also Ask section and just keep clicking and clicking and clicking it, open and close the questions and answer sections of Google, it produces hundreds and hundreds and hundreds of queries, and it's starting to then populate what people are asking around a subject and how a query is fanning out.

Who writes the People Also Ask answers now?

[04:41] So now we're at 2025. If we, well, look at mid-2025, only 12% of the People Also Asked answers are in AI Overviews. So I think this is really crucial because we're starting to see the shift, and this is where fundamentally traditional SEO is finished, just because it's not about keywords anymore, and it is now about how we optimise content to answer many, many queries, and that's what I'm trying to or will hope to explain to you in some examples in this video as we go on.

August 2026, jump forward, 86% of those People Also Asked answers are now in AI Overview. So we're now in a significant position where answers really do matter. So it's not just about phrases. It's so obvious that people try and optimise for a phrase. You know, in our example earlier, find me an accountant near Doncaster, or who's the best accountant in Doncaster? Or which Doncaster accountant should I use? Or is there an accountant in Doncaster that could help me with my VAT?

You can now see that there could be thousands of iterations of that, even that just one basic phrase, and now once you start going into billions of different searches that happen all the time, you can hopefully understand how we can talk about 15% of the searches that go into these AI platforms have never even been asked before. So how does an AI deal with that? Think about it. Now, September 2026 and things are moving really quickly, 97% of the sampled answers are AI written. So I think we can see that Google has started mapping this conversation years and years and years before anybody was even beginning to have a conversation with it.

So they see the end game. You've got to work with how this is moving. The box that showcased your content is now written by Google's AI, right? So if we move from 12% of the People Also Asked answers being used about, well, just over a year ago, 97% of that is getting used. So 19.2 million sampled English People Also Asked answers in the first week of September 2026 alone.

Why is search now a conversation?

[07:27] So 90% of results pages with an AI Overview also show People Also Are Asked in an AI answer on top. Right? Remember that. So follow-up questions it's a conversation. Think of search as a conversation now, not as a keyword or a key phrase, right? We're not, we're not trying to come top for a keyword or a key phrase, right? This is the whole of SEO ending right now just because it's nothing to do with keywords anymore. It's to do with conversation, and it's to do with answers and questions and how people discern what they're looking for.

How do they move through that map of questions and answers that's already there? So one question, and this is what I'm trying to get across to you guys, right?

What does fan-out look like for a local accountant?

[08:23] One question may fan out into many, many different answers, right? So in the example, if somebody asks which accountant near Doncaster is good for a small builder who's fallen behind with his VAT, right? Only a very small search volume, right? So don't start looking at... In the past, if somebody typed in accountant Doncaster, and they just started clicking links, that's a very general search. If somebody ends up on a website, they might produce an enquiry, they might be a decent customer. They may or may not. But once somebody qualifies selves, we have to understand that behind that are a whole bunch of queries that AI is using to map who to cite and recommend.

So if somebody asks which accountant near Doncaster is good for a small builder who's fallen behind his VAT, that's probably not been typed in very much at all. It's hundred percent definitely not been optimised for, 'cause I can imagine there's not somebody who's produced a page that is specifically aimed at that phrase and for somebody who has fallen behind with their VAT. If somebody has, I'd be amazed, and well done you if you did. But anyway, behind that, the AI will be searching accountants in Doncaster, accountant for construction trades, help with late VAT returns, HMRC penalty for a late VAT return, firm name and then reviews, and accountant fees sole trader builder.

So hopefully we can now see there are many questions being then queried by the LLM in the background. So if somebody's telling you they want to optimise you for just accountants in Doncaster, that's not gonna work, right? It will be general, and it will not be specific to the person's needs. So therefore and it's not how people type it in. They won't converse with LLMs in that way, right? So hopefully you can now see that building up a set of questions and queries behind an actual query that somebody's typing in is what actually matters. So this will then produce a result of maybe one or two decent recommendations from an LLM.

What do Google and ChatGPT fan-out examples show?

[10:52] Other examples for example, if someone... Google's own example was how to fix a lawn, right? And in that, they named herbicides for lawn weeds, best grass types for your climate, how to prevent weeds, lawn watering schedule, right? So these are all things that matter to the first query. Another one from ChatGPT, and this was from Peec's research. Somebody searches best headphones. We would look for example, the AI might be looking for the best wireless headphones, best noise-cancelling for flights, best budget headphones, best headphones 2026.

The last one is quite interesting because we are seeing that quite a lot of queries that include a date or statistics do get referenced quite often. So what you need to think about is how do I come into the main search based on all of the content that somebody might ask a question around?

Why are the questions that decide the answer missing from keyword tools?

[11:57] Most of the questions that decide the answer are not in your keyword tool. So somebody's optimising you for best accountant in Doncaster, all those phrases, you're definitely not optimising for or even telling people, telling the LLM that you're optimising for that. You don't have the content built behind it to do that. And you're most certainly not measuring whether you come up for those foundational phrases and keywords, right? The other things that matter to build the search query to give the answer.

Google or Gemini. So Gemini will look for roughly 10 searches per prompt. 95% of those probably had zero search volume. So that's really important. We're looking for the fact that, you know, these are phrases and keywords and things that are people asking. They're They don't have volume in traditional search. Between six and 12 words per ChatGPT are hidden search, and 32.9%, 33% of ChatGPT-cited pages surfaced only via a hidden subquery.

So meaning that when, that subquery, the content matched it and was cited, that then brought forward a page that could not possibly have necessarily been optimised to come up for a particular phrase or keyword. So it's the fan-out that matters. You have to understand, and I don't think I don't think very many SEO agencies understand what I'm talking about here, and maybe I'm educating some of them to understand how to build content properly behind the scenes that matches the query, okay? It's not, it's not rocket science.

This is, this is a process that has to go through to understand how to build content in this new LLM search universe.

Why is a top 10 ranking now a raffle ticket?

[14:09] Ranking in the top 10, now buys you effectively a raffle ticket. It doesn't get you into the answers. In mid-2025, if you look back, roughly 75% of pages would produce or would be cited in those, AI Overviews. Now, in February 2026, and I'm recording this video in October 2026, if we look back to February, that had already reduced roughly by half. So 38% of pages that were previously in the top 10 now being cited.

So you are not, or your content as it stands, if you have spent years, and you may well have spent years optimising your content to appear for certain keywords or phrases, the foundation that that is built on is quicksand, right? It is disappearing. Half of the overlap between AI citation and Google or traditional search positions has now disappeared in less than a year. And I think this is set to shift completely to the point where the only thing that matters is the questions, the answers, and the citation in the fan-out query.

You've got to build the content around how you're stacking content to answer a question to be cited. This isn't a new thing, right? It's definitely not

How fast is AI Mode growing?

[15:53] A new thing, right? There's a billion-plus searches now in monthly in AI Mode. And that's only, you know, really in Google, that's properly only a year after they've really moved into providing AI Mode to people. AI Mode queries have more than doubled every quarter, so it's exponential. It's just a speed thing. This is just growing so quickly. This the searches that come out are three times the length of a traditional search on average. So people are beginning to talk to these LLMs.

We are going to see a shift from typing in phrases and keywords to asking questions. We are gonna ask questions and want to learn more, right? And one in six, this is crucial as well, one in six searches in the US now are voice or image search, right? So we're moving very, very quickly away from what you would call traditional search. I've heard people talking about, you know, the web disappearing, you know, moving to an AI LLM assistant universe. That is happening, right? I got called a crackpot or you know, was I on mushrooms about a month ago for saying, you know, there is a possibility that the traditional web or search disappears.

I think I'm gonna be proven right. We are moving to a world where the web, yes, it exists, but it's more like a repository or a library, and it's not where we end up, you know, we don't end up on people's websites. We answer our queries within the LLM. We probably buy there. We book meetings there. We converse there. Our AI assistants talk to each other, talk book meetings, et cetera, et cetera. Why would you spend your time or moreover waste your time searching through websites when you can get the answer given to you?

Why do longer questions get the machine's answer?

[18:10] The longer the question, now listen to this carefully, the longer the question, the more likely the machine answers it for you, right? So searches that triggered an AI summary by length, one to two words, only 8%, and 10-plus words, 53%, right? And this is only set to increase again. So 8% versus 15%, a click will click a normal result with an AI summary versus one without one, and 1% will click a source link inside the AI summary.

So this is why we're starting to see people not click through to sites, and this is again set to increase.

Where is the ad money going?

[19:02] Ad money is already following these longer search queries, right? So share of Google Ads impressions by query length. If we look at one to two words, a year earlier, 42%, now 24%, and longer queries three to four words a year earlier, 33%, and now in August 2026, 50- nearly 50%, 48%. So conversions from seven or more words has moved from 1% to 4%, and this is evidence from Search Engine Land analysis of Google Ads data in August 2026, right, via Optimixed.

This data shows that people are moving. Again, you cannot be surprised about this, that people are moving to asking questions and researching within their query.

Are UK searches getting longer?

[20:05] UK searches are getting longer even for people who never used ChatGPT. So characters in the average UK Google search in a year has gone from 21 to 26, so not a huge shift, but you can see that it's moving. UK searches over 30 characters are up 24%, and Google searches per person in the UK versus May 2025 up 26%.

So people aren't searching for different things. They're saying things differently. They're asking more questions. They want more depth in their search.

How many follow-up questions do people ask?

[20:47] So most people will still ask only once or twice, right? Buyers will keep diving in, right? So we are seeing a slight divergence between the seriousness of a query, if we put it that way. What we, what we're seeing is a shift in, you know, you are getting your answers more quickly, so you only have to maybe ask once or twice, and I don't know that people are in the mood or they haven't necessarily learnt quite yet to dive deeper and deeper on more general topics.

That's actually good because it's, you know, you get the answer straight away, you can take action on it, or if you're just browsing, you're not, you're not wasting any time, right? You don't have to go and dive through four, five, six websites. You don't have to go and check reviews. You don't have to make your own mind up. The AI can do that for you, right? People are making sort of two or three queries per AI Mode session versus five in traditional Google. So again, people are getting answers more quickly. 1.7 messages in the average ChatGPT conversation, so seeing that backs up our one or two people ask metric.

But we are seeing 94% down to 43% of prompts needing one round of hidden searches. So we are seeing the fact that there are many, many more hidden searches behind content that is getting served to you or served to the people that are potentially trying to find your business.

What wins each question when choosing a recruitment agency?

[22:36] So I've, I built a little example. Now, if you're listening on my podcast, I'll try and explain clearly what's going on. So I've built a little example. So if a recruitment agency, so somebody's looking, a business is looking for a recruitment agency and somebody's thinking about, "Should we use an agency or shouldn't we?" And the question they ask, "Should we use an agency or keep trying ourselves?" That could be a question, right? Which agencies in South Yorkshire, Yorkshire, I went a bit West Country then.

Which agencies in South Yorkshire specialise in engineering, right? Consistent name and specialisms win. What do they charge? Contingency fees or a retainer? What would win that? Publishing your fees, right? So you would have to have your fees published because if your competitors aren't, one of the sub-queries would pick up your fees, and you may then be referenced in the answer.

How do I know they can find a maintenance engineer? So how does, how do we know that that particular recruitment business can find me the person I'm looking for? And what will win that query? A named case study with times and dates on it. So if you can demonstrate that you can find maintenance engineers in a particular area That's how the content would be picked up and you would then be referenced. So think below the query that's getting asked.

What if the person leaves after a few weeks? Quite a quite a regular question that would come up, you know, if you wanted to deal with a recruitment agency. And what would win that? Publishing your rebate terms on your site. And which recruitment agency would you pick for us and why? And what wins it? Reviews that agree with your site. So external reviews on an external site would then pick you up. So what we can see is it's not about these top level queries. You know, somebody can ask, "Should we use an agency or keep trying ourselves, looking, looking for staff ourselves?" Now, you know, if you, if you wanna win those queries that's quite difficult if you don't build the content below it.

That's what I'm trying to get across here to you. Think below, think more deeply into what people are searching and how, you can satisfy the query, and it's not about optimising for the phrase in the first place. Hopefully, I'm making myself clear.

Is Google right that this is still just SEO?

[25:44] So Google, you know, like to caveat this, Google says it is all just SEO at as usual, right? Google says we're still in an SEO world. And we are to some extent, right? 'Cause we haven't fully gone into the universe where we are gonna end up. So we are, we are not fully into a full AI Mode where we end up talking to LLMs, talking to our AI assistants like Muse Charm, et cetera, and it gives us the answer and we have a proper conversation with it, right? But when we're having those conversations, think about the knowledge that it needs to have behind it.

It's not about a webpage that says, "This is the best recruitment company in Doncaster," because somebody created a page saying they are the best recruitment company in Doncaster. It's about all of the fan-out queries that you need to have below it, right? But Google would say a page per fan-out variation is spam, and that is true. It's We're not going back to the good old days where I go and build 50 different web pages, just to cover all the individual scenarios, right? Google would call that content abuse, and pages covering every sub-question will get cited less, right?

You need to cover a solid base and foundation and be cited externally as well. One, that's one of the, obviously, one of the most important things that are gonna happen, right? AI referrals are still only, and I say still only, around 1% of traffic for most sites. But again, we have said that an, a massive chunk of the web search that is happening still, even when answered by LLM or AI query that's just cutting the dead wood, the traffic that nobody wanted anyway. You know, it's traffic that's not gonna convert. So all of that, if you get rid of, I don't know, 75% of the traffic just because it gets rid of all the people that would have landed on your site and never made an enquiry, or you weren't a good fit or just bounced, if we get rid of all of that and then we say 1%, that's 1% of 25%, right?

So we're now at four or 4%, right? It's only set to increase.

What should you do this week?

[28:18] So keywords, look, I can't stress it enough, keywords are no longer the unit of search or value, right? The conversation is what you have to be optimising for. So a little bit of exercise, something you can do this week that's gonna make a difference for you as a business.

Start with questions, right? Not keywords. I think I've been clear about that throughout this whole presentation. Record your sales calls, right? Write down everything that a customer says in their own words. I've been telling people to do this for years, right? And we're all, maybe we're all a little bit lazy, but, I have, in a recent example, taken one of my clients, I've taken all of their phone calls for the last six months, downloaded them all. I took out everything that was less than five minutes long. The remaining calls, we transcribed them all, put them into Claude, and some amazing insights came out of it, right?

Not just things where we could learn from which salespeople were working, but all of the questions that people were asking, right? And that is hugely valuable. So then we can build content around that to answer those questions on topic, right? And not trying to be based on keywords or phrases, but based on actual things that your actual customers are saying, right? I can't stress it enough. Record your calls. Start with the questions, right? It's not about the keywords anymore.

How do you use People Also Ask as a map?

[30:00] Use the People Also Asked free map pack, right? So within that, you've got People Also Asked. Use that as a template. So go in to Google and ask it the questions. Look for the People Also Asked section. Download every single one of those questions and then draw a line through the ones which aren't relevant to your business, and you'll be left with possibly hundreds that are relevant, right? And you can use that, go and answer those questions in your own words, record that, produce it into pieces of content, publish it, right, and share it, and then you've got a very good chance of covering off a lot of this ground that we've been talking about in this video.

Stick to one page for a conversation, right? Put your answer first and then follow up follow up under it in a heading that is another question. So structure your page correctly, but keep those questions on topic. Stick to a theme. Don't go wandering off into massive sub-queries. Keep it succinct and accurate, and accuracy is one of the huge things, right?

Why should you publish what only you know?

[31:15] And most importantly, say things that only you know, right? If you can get stuff, and I say to people, "Get it out of your head onto paper," right? That would be the old school way of doing it. Get it out of, out of your head and onto paper. Well, now we're talking about digitising your knowledge and talk about your prices, your process, things people can do with your products that, you know, they might find as an advantage or some way to use it that hadn't necessarily been published. Put down your numbers and your experience and just give something to the knowledge base. And as we discovered earlier, maybe this was a light bulb moment for an awful lot of you.

I can certainly say it was for me when I first found this out. 15% of things that are getting typed into the web on a daily basis have never been typed in before. So think about that for a minute. You can't optimise for all of that, but you can optimise for the sub-queries that come below it.

How do you check whether AI recommends you?

[32:23] So ask it the long way, right? The way your best customer would ask the question. Some people of us, you know, we all can be very guilty of being too close to our own business and thinking like, well, not putting ourselves in the customer's shoes. So ask the questions. Go and ask ChatGPT and Gemini and Claude and Perplexity and ask them all the questions that you think that you should come up for and have a look, right? Do you come up? Are you cited at all? And if not, work out why and who is coming up and why they're coming up ahead of you.

Some of the sources that I quoted today are in the blog. I will publish them in my transcript, but I think I've really tried to get to the bottom of this and show you how, you need to think more broadly beyond keywords and phrases, and I think that's the lesson for today. Just make sure you are understanding what those sub-queries are below a particular one, and if you have to, go back into the video, look at the examples I've given, and hopefully that's useful and I'll see you in the next one. Cheers guys.

Where do the figures in the video come from?

Figures Simon cites on camera, with the receipt behind each one. Times follow the published edit. Vendor and agency research is marked as such in the article above.

Time	Figure as cited	Receipt
02:12	Around 15% of the searches Google sees every day have never been seen before. This is a Google Search figure.	Search Engine Land; Search Engine Journal
03:06	People Also Ask testing began in April 2015; by 2017 each click loaded more questions	Common Denominator; Search Engine Journal
04:47	Share of sampled People Also Ask answers that are AI Overviews: about 12% (mid-2025), about 86% (August 2026), 97% (first week of September 2026), from 19.2 million results	Brafton, reporting AlsoAsked
07:28	People Also Ask on about 90% of results pages that carry an AI Overview	Semrush AI Overviews study (vendor)
11:19	"Best headphones" ChatGPT fan-out example	Peec (vendor)
12:38	Gemini runs roughly 10 hidden searches per prompt (10.7 average for Gemini 3); 95% had zero search volume	Seer Interactive (agency)
13:00	ChatGPT fan-outs roughly doubled in length, from about 6 to about 12 words	Peec (vendor)
13:06	32.9% of ChatGPT-cited pages surfaced only through a hidden sub-query	Growth Memo, reporting AirOps (vendor)
14:24	Pages cited in AI Overviews that also rank in the top 10: about 76% (July 2025) down to 38% (February 2026). Simon rounds the first figure to 75%	Search Engine Journal, reporting Ahrefs (vendor)
15:56	AI Mode passed a billion monthly active users and its queries more than doubled every quarter. Google's figure is users, not searches	Google
16:28	The average AI Mode search is triple the length of a traditional search	Google
16:47	More than one in six US searches now use voice or images	Google
18:20	AI summaries appeared for 8% of one or two word searches against 53% of searches of ten words or more	The Decoder, reporting Pew Research Center
18:36	People clicked a traditional result on 8% of visits with an AI summary against 15% without; 1% clicked a source link in the summary	The Decoder; The Register, reporting Pew Research Center
19:07	Share of Google Ads impressions: one to two word queries 42% to 24%; three to four word queries 33% to 48%; seven or more words 1% to 4% of conversions	Optimixed, reporting Search Engine Land
20:16	UK average Google search grew from 21 to 26 characters; searches over 30 characters up 24%; searches per person up 26% against May 2025	RealityMine
21:47	2 to 3 queries per AI Mode session against 5 or more in traditional Google	AuthorityTech, reporting Semrush (vendor)
21:58	1.7 messages in the average ChatGPT conversation	WebFX (agency)
22:12	Prompts needing only one round of hidden searches fell from 94% to 43.5% after GPT-5.6	Peec (vendor)
26:52	Google warns that a page for every fan-out variation can breach its scaled content abuse policy; pages covering every sub-question were cited less	TechWyse; Passionfruit, reporting AirOps

Notes on figures as spoken:

• 01:25: Semrush's 23-word figure is for ChatGPT prompts made without web search, against about 3 to 4 words for a Google search, from 2024 clickstream data (Semrush). It does not measure how all queries changed over time, and Semrush's April 2026 update found those prompts had shortened to about 13.5 words.
• 03:41, 06:18 and 32:04: Simon applies the 15% figure to "search and LLM", "AI platforms" and "the web". The published figure is for Google Search only.
• 10:57: Google's own lawn example lists herbicides, weeding without chemicals and weed prevention (Google). Grass types and watering are Simon's additions.
• 11:41: no study shows that queries containing a date or statistics get cited more. The closest evidence is that ChatGPT adds the current year to some of its hidden searches (Peec).
• 15:56: Google's figure is a billion monthly active users, not searches.
• 27:31: AI referrals were about 1% of traffic on Conductor's 2025 data (Conductor) and 0.14% on Semrush's 2026 data (Semrush).
• 28:11: Simon's point is that if AI referrals hold steady while the rest of the traffic falls to about a quarter, their share of what remains rises about fourfold. It is a hypothetical, not a statistic.

---

FAQ

What is a fan-out query?

A fan-out query is one of the hidden searches an AI engine runs on your behalf. You ask one question, it breaks that question into many smaller ones, searches each separately and writes a single answer from what comes back. Google has described AI Mode doing this (Search Engine Journal), and ChatGPT, Gemini, Perplexity and Grok all fan out at different rates (Peec).

Why can't my keyword tool see these searches?

Because nobody typed them. Google has long said around 15% of the searches it sees every day have never been seen before (Search Engine Journal), and 95% of the hidden searches Gemini 3 generated had zero search volume (Seer Interactive, agency research). The machine writes these searches, so no keyword report contains them.

Does ranking in the top 10 still get you cited in AI Overviews?

Less than it did. Ahrefs found 76% of pages cited in AI Overviews ranked in the top ten in July 2025, and 38% in February 2026 (Search Engine Journal, vendor research). Simon calls a top 10 ranking a raffle ticket: it gets you in the draw, not into the answer.

Should I create a page for every fan-out query?

No. Google's guidance warns that pages for every variation of how people search can breach its scaled content abuse policy (TechWyse summary), and AirOps found pages covering every sub-question were cited less than pages covering a solid chunk of them (Passionfruit). Build one page per conversation instead.

What should a business do this week?

Start from the questions your customers actually ask, in their own words. Record your sales calls, use People Also Ask as a free map, build one page per conversation with the answer first and each follow-up under a question heading, publish what only you know, and make your site, profiles and reviews say the same thing.

How can I check whether AI recommends my business?

Open ChatGPT, Gemini, Claude, Perplexity or Google AI Mode and ask the question your best customer would ask, the long way, with their situation and their worry in it. If you do not come up, look at who does and what they published that you have not.

---

Sources

• Google, How AI Mode is changing the way people search in the U.S., 19 May 2026
• Google Search Central, guide to optimising for generative AI features, 15 May 2026 (updated July 2026); summary via TechWyse
• Semrush ChatGPT clickstream study, February 2025
• Conductor AEO/GEO benchmarks report, November 2025, and Semrush traffic channel mix study, April 2026
• Search Engine Land and Search Engine Journal on Google's 15% new-queries figure
• Search Engine Journal on fan-out in AI Mode
• Common Denominator and Search Engine Journal on the history of People Also Ask
• AlsoAsked data via Brafton, 18 September 2026
• Semrush AI Overviews study, updated December 2025
• Pew Research Center via The Decoder and The Register, July 2025
• Seer Interactive, Gemini 3 fan-out research
• Peec: fan-out length, fan-outs by engine, what is a fan-out, GPT-5.6 retrieval changes
• AirOps via Growth Memo and Passionfruit
• Ahrefs via Search Engine Journal, February 2026
• Search Engine Land Google Ads analysis, also via Optimixed, September 2026
• RealityMine, 23 April 2026
• Google follow-up data via PPC Land; Semrush session data via AuthorityTech; WebFX, November 2025
• Hitwise via MarketingCharts, 2009

---

Episode: "Stop thinking about keywords: how conversational search really works", Simon Young, Question.Marketing, recorded and cut 6 October 2026. Duration 33:44. Series: Will AI Recommend You? About this page: the article is Simon's own blog post ("Your customers stopped using keywords. Has your marketing?", 6 October 2026), with links restored from its source list. The transcript is built from the edited master with ElevenLabs Scribe ASR, cleaned for fillers and false starts, with brand names corrected and British spelling applied. Film and audio remain the source of truth.