Rethinking SEO for LLMs and Personalization

In this week’s episode of Voices of Search, we spoke with Venkata Pagadala, AI product manager for search at AT&T. Venkata has been publishing open research on what enterprise search teams overlook, and much of it returns to consumer behavior as the starting point, ahead of traditional keyword research.

During our discussion, we talked about how enterprise teams can turn their own data into a deeper understanding of consumers, why zero-volume queries often come from ready buyers, and what personalization means for how search work gets done.

Key Takeaways From This Episode:

• First-party data such as call recordings and site interactions shows customer needs that third-party keyword tools can’t, and competitors can’t copy it
• User-generated content like reviews and Reddit threads is where high-intent demand shows up, often with zero search volume
• Classifying millions of queries takes human industry knowledge, even when AI does the heavy lifting
• Search results and AI responses now differ from person to person, so a single prompt or keyword ranking says little
• AI makes creating output easy, which moves the bottleneck to quality control and a clear purpose
• Sharing research openly and asking why keeps the craft moving

Your Own Data Shows What Keyword Tools Miss

Venkata made an interesting point right off the bat, which is that Google is a middle player. It collects, classifies, and shows data. A company’s first-party data, such as call recordings and the touchpoints users have with its product, site, or platform, can be read without that middleman.

Competitors don’t have that data, and third-party tools like Ahrefs and Semrush, as useful as they are, only show so much.

“If it’s my own data, I can go as deep as I want, no filters,” he said.

The second step is thinking like a buyer instead of an SEO. When he shops for a microphone, he looks for comparison discussions on Reddit and honest blogs rather than sales blogs, then reads reviews to find the cons. 

“User-generated content is a gold mine,” he said, “because it comes straight from people who bought a product and were happy or unhappy with it.” 

His buyer’s checklist involves:

• Reviews and ratings, including the cons
• Response rates and how a company answers unhappy customers
• Comparison threads on Reddit and similar communities
• Wherever the target audience actually spends time, which shifts with age and product

He used Costco as an explicit example, since its return policies and customer service give him peace of mind when buying—and no keyword report captures that. 

“Many teams talk about these sources,” he said, “but few use them.” And so, his methodology involves asking how many strategies have actually pulled queries from YouTube comments, transcripts, or Facebook groups.

Classification Is Where AI Still Needs a Human

AI makes that consumer sentiment research possible at scale. Using APIs and MCP servers, a team can pull threads and comments in bulk.

“You can create slop at scale, or you can use AI to create meaningful content by understanding large data sets,” Venkata said.

The hard part is classification. Most topical research pulls 10,000 or 20,000 keywords. He analyzed 5.8 million queries and cleaned out 3 million that were noise. Doing that well takes real knowledge of the industry, including its sources, its big and small players, and even government sources, and even frontier models can’t classify a million queries on their own—which is something that AI can’t do.

His process starts by hand and widens from there:

• Review competitor navigation, sub-navigation, and footer links to find core topics across big, mid-size, and small players
• Use Wikipedia to build out subtopics
• Pull ranking data for competitors and your own site from two or three sources, then dedupe to find hidden gems
• Add internal URL structures and social signals

The result is creating a hierarchy. For instance, devices sit at the top. Brands like Apple and Samsung sit below that, then product lines like iPhone and MacBook, then individual models. 

Once that structure exists, queries from different sources can be pulled into it.

Zero-Volume Queries Hide Ready Buyers

Having this hierarchy also exposes the demand that typical keyword tools hide. Standard classification of an iPhone query stops at color, storage, and charging capacity. The questions Venkata finds on Reddit and similar platforms sound a whole lot different.

For example, someone with an iPhone 16 comments that their phone is too big while simultaneously asking for the best phone with long battery life. Another person, deep in the Apple ecosystem, likes the Samsung Fold and wants to know how to move their Apple data to Android.

“That is zero search volume, but this person is really trying to trade in this device and get a new device,” he said.

A team doesn’t need a page for every phrasing. It needs content for the underlying need, such as a size problem and a trade-in path to a newer model. People ask these questions on Reddit, Quora, Substack, and YouTube because they can’t ask them on a company’s website, which is why volume data never shows them.

Tyson called search volume a dirty KPI, directional at best after a decade of distortion from Keyword Planner to click data. He added that zero-volume queries make sense, since the industry wasn’t ready to meet needs this specific, like his wife’s search for the smallest iPhone after her iPhone 13 mini. 

The point is, people have need-based questions that high-volume keywords aren’t uncovering, and that’s what we need to be searching for when we want to target the LLMs head-on.

Personalization Makes Volume a Weak Guide

Venkata sees personalization as the reason this shift matters. Within months, he expects search results to differ significantly from person to person, and AI responses already do.

In his view, a marketer’s job is to reach consumers wherever they are. That meant newspapers and billboards, then websites and apps, and voice agents are next. Search volume did a good job of collecting and categorizing demand, he maintains, but the industry went too far too fast—and now he questions how anyone wins in AI by starting from search volume.

Search used to be close to one-to-one with a single engine. Now attention splits across five or so platforms, which makes any one ranking an incomplete picture. The alternative he points to is personas, query fan-outs, and prompt tracking, with a caution about what a tracked prompt proves.

“A single prompt should not be a definition of whether we are ranking or not,” he said.

Tyson tied that back to results themselves. As systems add a context layer on each user, the variation in results will grow, which rewards content built around personalization over content aimed at a single high-volume term.

Think of AI as a Sports Car that Someone Still Has to Drive

Accessible AI lets people without coding skills build dashboards, mockups, and analysis that once needed developers. Tyson noted that this narrows the old divide between technical and content-focused SEOs. Venkata’s advice is to understand what you’re asking before you build anything.

“AI is like a sports car,” he said. “You need to understand where to drive and where not to drive.”

In other words, you need to know the purpose of what you’re creating, whether it fills a real gap, and whether it can work at enterprise scale. A home-built agent can write anything, he said, but the question is whether it works and how it differs from what others have. The skill is less about creating and more about understanding a problem and using AI to solve it—just like knowing how to handle a sports car on top of where you’re headed.

Tyson pointed to where that leaves teams. Production is now table stakes, so the bottleneck becomes quality control. Venkata put the stakes bluntly—leaders can run any transcript through an AI to check whether the person talking makes sense, so you have to be ready to stand behind the output.

His preferred way to earn leadership’s trust is a mockup. Analyze 100, 10,000, or a million reviews, name the three biggest pain points, and build a mockup that solves them using existing site components and content. 

Two habits keep the workflow repeatable:

• Document each solved problem in a tool like Obsidian or Notion so an AI agent can run the classification next time
• Dictate instead of typing, using a tool like Whisper Flow, since people talk faster than they type

“The storytelling days are gone,” he said. “The story now comes through the mockup.”

SEO Is a Craft Built on Asking Why

Venkata still calls the work SEO, treating AEO and GEO as facets of one platform-independent craft with search at its core. Tyson agreed, describing it as a craft that has always been about data retrieval systems and how to stay relevant inside them. Venkata added that personal agents may become the next audience.

“What keeps the craft moving,” he said, “is the habit of asking why.”

He credits Bill Slawski, known for his writing on patents and entities, as an inspiration, since he followed his own curiosity instead of money or endorsements. Bill started posting heavily when almost no one wanted to hear about AI. According to Venkata, in the GPT-3.5 days, no one trusted him, and sharing publicly anyway led to speaking opportunities, papers, and plenty of corrections from people who knew more.

That curiosity is also why he shares his research publicly. He posts on LinkedIn and his own website without asking for emails, and he sees it as a chance to show that SEO takes research and understanding.

“Education should never be restricted by money,” Venkata stated.

Tyson framed that as a rising tide for the industry, where one person sharing what they learn leaves the whole community better off.

SEO Now Starts With the Consumer, Not the Keywords

Every thread in this conversation traces back to one change, which is that search results and AI answers vary by person, so a keyword’s volume says little about who is asking or why. The teams that stay useful start with the consumer, using first-party data and the places where people talk. They let AI process the volume, apply human judgment to classify and check it, and share what they learn.

Venkata’s approach keeps the craft the same while changing where the evidence comes from. Teams that understand the people behind the queries will find demand that no keyword tool reports.

Voices of Search is a daily SEO and content marketing podcast hosted by Jordan Koene and Tyson Stockton. The show delivers actionable strategies and data-driven insights to help marketers navigate the ever-evolving world of search engine optimization and content marketing. New episodes air weekly, covering everything from technical SEO to AI discovery, featuring industry leaders and practitioners sharing real-world frameworks and proven tactics.

Subscribe to Voices of Search on Apple Podcasts, Spotify, or your favorite podcast platform. Follow Previsible on LinkedIn for updates and subscribe to the VOS YouTube channel for video episodes and clips.