Why AI Search Is a Conversation, Not a Keyword

I want you to try a small experiment. Take the classic SEO keyword you’re proudest of — you know, the one that you fought your way to get to page one of Google.

Now, transform that keyword into a question a real customer would actually ask about that topic. Not the keyword. The messy, contextual, “here’s my situation” version of it. And put that question into the answer engines — ChatGPT, Google, Gemini, Perplexity, etc.

Notice anything? There’s a decent chance you’re not in the answer at all.

I ran this experiment myself. So, Content Marketing Institute (CMI) has ranked number one on Google for “what is content marketing” for as long as I can remember.

That page is far and away the most popular page on the entire website. So I typed “what are the best practices of content marketing” into Google. The AI Overview served up six best practices, citing (in order) Twilio, Harvard Business School, Adsmurai, and Adobe. CMI wasn’t mentioned.

But directly below that AI answer — below the sponsored results, too — there sat CMI. The No. 1 organic result. For that very query.

For thoroughness, I made the rounds. Gemini cited a similar group, but no CMI. ChatGPT answered confidently and cited no one at all (completely on brand). Claude actually included CMI midway through its list of best practices, but handed most of its citations to Neil Patel, 310 Creative, and MakeMEDIA. Four AI systems. Four different answers. Four different sets of winners.

Here’s my point. For 20 years, we’ve operated with one wonderfully reliable proxy for search visibility: the keyword ranking. We built entire measurement programs — entire careers — on it. Rank well, get found.

It was never a perfect proxy, but it was durable. And durability is what quietly turns a metric into an assumption.

The assumption I’m hearing more and more often these days goes like this: “We rank well in Google, so we’re probably fine in AI search.”

The research now says otherwise. And I mean emphatically otherwise.

The proxy is broken — and the data is unambiguous

Multiple independent studies over the past year have converged on the same finding. One analysis found that only 12% of URLs cited by ChatGPT, Perplexity, and Copilot rank in Google’s top 10 for the corresponding query, a figure a separate study of more than 18,000 queries independently confirmed. And more than a quarter of ChatGPT’s most-cited pages have zero organic search visibility. None.

Think about that weirdness for a second. You can be No. 1 on Google and invisible in the answer your buyer actually reads. You can also have no meaningful Google presence at all and still be the most-cited source in your category’s AI conversations.

The two games are being scored on different fields.

Now, there’s a nuance we have to acknowledge — because like everything in search, things are changing fast and the rule has exceptions. At the domain level, the correlation is actually pretty strong.

In other words: the authority we’ve spent years building still matters. If you’ve invested a decade in a credible domain, that investment isn’t wasted.

But here’s the catch. Which pages get cited, for which questions, follows entirely different logic. Your domain’s reputation transfers. Your rankings don’t. And since rankings are the thing most of us still watch week to week, our dashboards are keeping score for the old game while a new one is already underway.

Why the machinery produces different winners

There’s been plenty of talk about how “answer engine” use cases differ from keyword search. In the early days, the separation seemed to fall mostly along the question of whether you were looking for a discrete answer or a set of options.

If I’m searching for the definition of a word, I basically just want the definition. But if I’m searching for the “best Mexican restaurant in Los Angeles,” I want a range of options backed by some source I trust.

What’s been less well covered is something more fundamental: Traditional keyword search and AI answer engines are architected differently. They aren’t two interfaces to the same information retrieval. They’re two different machines. And the difference between those machines is exactly where our assumptions break.

Traditional search is a retrieval system. You type a query; the search engine ranks source pages against that query; you pick one.

The keyword (or key phrase) is the atomic unit of the whole transaction, which is why it became the atomic unit of our measurement.

AI answer engines are synthesis systems. And despite the popular framing, the queries that matter most aren’t the keyword-shaped ones.

People aren’t typing “best CRM software mid-market” into ChatGPT. They’re typing, “I run a 40-person sales team, and our pipeline reporting is a mess. What solutions should we look at?”

The system doesn’t rank pages against that question. It decomposes the question into many variations (researchers call this “query fan-out”), retrieves information against all of them, and then composes an answer, deciding along the way which brands to mention, in what order, with what tone, citing which sources.

So it’s the combination — different machinery and different use cases — that guarantees different winners. A page that never ranked for your precious keyword can quietly dominate the fan-out queries the AI actually ran.

Gintarė Rimolaitytė, chief commercial officer at AI search visibility tool Trendos, recently said, “In traditional search, you optimize a page for a keyword. In AI search, you need the model to associate your brand with an entire topic. We track the same prompts daily across five engines, and the brands that surface consistently aren't the ones with the best keyword coverage. They're the ones AI connects to the broader conversation.”

The implication for marketers is now unavoidable. Unlike classic SEO, AI visibility will not be a position you can hold. It’s a distribution you have the opportunity to influence.

The measurement trap: Measuring by proxy

Understanding the difference between classic SEO and AEO/GEO is important because the marketing technology industry has largely responded to this shift the way it always responds to a technology shift: by pointing the old tools at the new thing and calling it close enough for rock and roll.

The first instinct of the measurement market has been to take a keyword list, send it to the LLMs via API, and pull back something that looks reassuringly like a rankings report. It feels familiar. That’s exactly the problem. It measures a behavior no actual user performs.

The answer AI search gives to a keyword can be completely different from the answer it gives to the real, conversational question. Which means keyword-based AI tracking can cheerfully and mistakenly report that you’re winning conversations you’re actually absent from.

I’ve come to think of this as the difference between proximity and proxy.

A proxy metric stands in for the thing you care about; a proximate metric gets as close as possible to the actual behavior. Keyword rankings were always a proxy for being “found,” but the proxy held because the keyword was the actual user behavior.

In AI search, it isn’t. The user behavior is the conversation. If your measurement doesn’t start from realistic conversations, then the prompts people genuinely ask, in the messy way they genuinely ask them, will measure the shadow of the thing rather than the thing.

This provides a practical buyer’s guide test for the rapidly crowding AI-visibility tool market. On whatever platform you evaluate, ask one question first: What does it actually send to the AI — keywords or our customers’ questions?

Everything else about the tool is downstream of that answer.

What good measurement actually looks like

Strip away the category noise, and a credible AI visibility practice comes down to answering four questions — continuously, not as a one-time audit, because as we just established, these systems change their answers constantly.

1. Are we present in the conversations that matter?

Not keywords. The conversations. Define the realistic prompts your buyers ask across the stages of their problem, and track whether you appear. The gaps are as valuable as the wins: every conversation where you’re absent is a content brief waiting to be written.

2. What share of those conversations do we hold?

Presence without proportion is trivial. The meaningful metric is share of voice against competitors who appear in the same conversations (and how that share trends over time). This is your new rank tracker, except the unit is the conversation and the scoreboard is probabilistic.

3. What’s the tone when we appear?

This is the one I prioritize. Being mentioned isn’t the goal; being represented well is. An AI system can name you as an option while framing you as “the expensive one,” “the legacy choice,” or “the one with the customer service challenges.” Technically, you were mentioned. Practically, you were positioned by a machine, at scale, in the moment of consideration.

Which is why I’d argue the real objective of this discipline isn’t share of voice at all. It’s what I’ve come to call “share of accurate representation.” And if I’m honest, accurate matters more to me than share. I’d rather appear in fewer conversations, represented truthfully, than dominate every conversation as a caricature.

4. Which sources are teaching AI what it believes about us?

This is the question that converts measurement into strategy. AI answers are built from citations, and research consistently shows that earned media and third-party mentions (not your website or blog) drive the bulk of them. When you can see which publications, communities, and databases the AI systems draw from in your category, you know precisely where next quarter’s content and PR effort should go.

What to do about all this

The good news: the cost of at least starting this discipline has collapsed to effectively zero.

A new generation of tools has emerged for exactly this work, and the better ones are built conversation-first rather than keyword-first. Trendos, for example — a platform that is gaining early traction with the SEO community behind Hostinger and NordVPN — tracks thousands of realistic, categorized prompts across the major AI models daily, rather than piping keyword lists through an API.

Whatever tool you choose, apply the test from above and you’ll quickly separate the conversation-native platforms from the keyword tools wearing new clothes.

Then run a simple 90-day play:

The gap in the data isn’t a reason to panic. It’s a reason to look harder.

Your rankings were never the point; being present, accurate, and trusted in the moment your buyer asks is the point.

The buyers haven’t stopped asking. They’ve just changed who they’re asking — and the brands that win the next five years are the ones measuring the conversation instead of the keyword.

Robert Rose shares how to use the Trendos platform, who it is and isn’t for, and how to get started with it in this CMI Tools of the Trade video.