Most people aren’t using AI the way GEO discussions often assume. Two surveys of AI users conducted by Stella Rising found that many prompts still look remarkably similar to traditional search queries. (Disclosure: I’m the VP of SEO at Stella Rising.)

One survey focused on a beauty-oriented consumer panel in August 2025, while the other surveyed a broader general-audience population in January 2026. Across both studies, prompts were short, often keyword-driven, and much closer to a Google search than the elaborate prompt templates popular in AI marketing circles.

At the same time, a growing share of users are adding personal context, such as their budget, location, profession, age, health concerns, or preferences. Those details give AI systems far more information than a traditional search query ever could, creating a new layer of personalization that influences recommendations and brand visibility.

The combined findings suggest that GEO strategies need to account for both realities: Many AI searches still resemble classic keyword queries, while the highest-value recommendations increasingly emerge from prompts rich with personal context. That’s where the opportunity — and the measurement challenge — lies.

A lot of people are still typing like it’s 2008

The biggest takeaway across both surveys is that the median AI user is still throwing a keyword over the wall and hoping for the best.

In the general-audience study from January:

• Two-thirds of respondents reported writing prompts of 15 words or fewer. 
• Only 12% wrote something that would qualify as a “real” prompt by the standards of an AI influencer thread. 
• About 60% phrased their queries as questions, while only 9% gave a direct command. 

That mirrors what Pew Research has been seeing more broadly — 34% of all U.S. adults now use ChatGPT, roughly double the 2023 share, and 58% of adults under 30 use it.

When we ran a scenario task asking respondents to write the prompt they’d send if they needed a new pair of shoes, the median answer was eight words. Real examples from the panel included:

• “Shoes nearby”
• “Tennis shoes”
• “Nike”
• “Ladies tennis shoes size 7 near me”
• “Best price for hiking shoes”

This lines up with Semrush’s clickstream data on ChatGPT’s search mode, which shows the average prompt length is 4.2 to 8.7 words, essentially the same as a Google query.

Longer, structured prompts tend to appear only when users are doing something other than search, such as drafting, coding, or creative work.

For AEO and GEO work, that’s the part to internalize. If you’re optimizing for prompts like “Compare the top five orthopedic-approved walking shoes under $150 for plantar fasciitis with 4.5+ star ratings,” you’re optimizing for the wrong distribution. 

Real prompts run 71% longer than the synthetic ones marketers tend to invent, but the median is still only 12 words, Otterly.AI’s analysis found.

The shift between the two surveys

In the August 2025 survey, we classified roughly 50% of the free-text prompts as “SEO-keyword-shaped,” meaning short, ambiguous, and brand-and-attribute-driven. By the time the January 2026 survey came back, that share had dropped closer to 30%. The remaining 70% had grown longer and more contextualized.

A few findings are worth carrying forward:

• 24.5% of all prompts include the word “best.” If you’re not appearing in “best [category]” responses, you’re missing one of the highest-intent slots.
• 28% of prompts mention price or budget constraints. Users aren’t just shopping. They’re shopping with a number in their head.
• 16% of prompts are explicitly location-based. The “near me” query pattern has successfully migrated from Google to LLMs.
• 32% of prompts include personal attributes (e.g., size, profession, health condition, life stage, etc.). This is the most important number on the page, and we’ll come back to it.

On location specifically, the 16% figure lines up with what Local Falcon’s 2025 research showed for AI search overall: AI Overviews now appear on 92% of informational local queries, but only on 15% of simple local-pack queries. The intent is moving into LLMs faster than the supply of optimized local content for AI engines.

One caveat: These were two different surveys with two different audiences. The January 2026 general-audience sample was structurally more transactional than the August 2025 beauty-focused panel, which partially explains why fewer prompts looked like keyword-style searches and more looked like full requests. I wouldn’t over-index on the “prompts are evolving” narrative, but I’d absolutely take the directional read.

The user embedding layer is where this gets interesting

The 32% figure (prompts containing real personal context) is the most under-discussed finding in the dataset.

Nearly one-third of users are willingly handing LLMs information that no Google query would normally carry, such as their size, job, training plan, living situation, or kids’ ages. We see prompts in the data like:

• “What shoes would you recommend for daily standing at work?”
• “Find me a cost-effective pair of running shoes that I can order on Amazon. My size is men’s 10.”
• “Please tell me the top five shoes for wide feet in a size eight for women that are comfortable, stylish, under $120, and that younger people won’t make fun of for a Gen X person like me.”

That last one alone packs in gender, foot width, size, budget, style intent, generational identity, and a real social anxiety. No traditional search query was ever going to surface all of that.

This is the user embedding layer at work. When someone interacts with ChatGPT or Gemini repeatedly, the model builds a profile of who they are, which increasingly persists through memory. The user is, in effect, training the assistant on themselves. Once that trust is established, they stop writing surface queries and start writing requests that assume the assistant knows them.

That shift has two implications for how brands should think about visibility:

• The prompts that drive purchase decisions are often not the ones that show up in a SERP or keyword tool. A real Gen X woman asking about wide-fit, $120, “won’t-get-made-fun-of” sneakers will never appear as a tracked SERP keyword. But that’s the prompt that decides whether your shoe ends up in the recommendation set.
• The value of a brand citation increases substantially when it appears inside a context-rich prompt. If the model is already factoring in user attributes, the brands it surfaces are prefiltered for relevance. That’s a much higher-quality impression than a generic blue link.

Where synthetic prompts fit — and where they don’t

A common tactic in GEO prompt research is to construct synthetic personas (“I’m a 38-year-old product manager training for a half marathon in Boston who prefers brands focused on sustainability…”) and then use those personas to stress-test which brands an LLM surfaces under different scenarios. There’s real merit to the approach. If the user embedding layer is doing the heavy lifting in the answer, the only way to simulate the answer is to simulate the user.

But synthetic prompts don’t capture everything. Real prompts are messy, layered, and influenced by recent conversation history, persistent memory, and signals the model has picked up over weeks of use. You can craft a 50-word persona and still miss the nuance of a user who has been talking to ChatGPT about their day, preferences, and family for 6 months.

Instead, use synthetic prompts to map the personas your brand needs to be visible to, but don’t treat the resulting visibility scores as ground truth. Combine them with real prompt data wherever possible. That can mean customer interviews, social search patterns, support tickets, or regex pulls of question-shaped queries from Google Search Console.

What to actually track

This naturally leads to the next question: Should you track SEO keywords in your AI visibility platform if one-third of real prompts look like SEO keywords?

The answer is yes, with one filter.

Across the last quarter, our team has seen web retrieval rates on tracked prompts climb sharply. On several client accounts, more than 90% of monitored prompts now trigger live web search inside ChatGPT or Google’s AI Mode. 

When that happens, the LLM is effectively running a real-time SERP and synthesizing the result. That means the short, keyword-shaped prompts we identified — roughly 30% of the total — are still very much in play. They behave like AI-flavored Google queries and should be tracked accordingly.

The filter is this: Don’t waste tracking slots on prompts that are pure head terms or single-brand queries. Those are likely to be answered from model weights or short canned responses rather than retrieval, and they won’t give you a useful read on visibility.

Here’s a practical setup:

• A synthetic-persona prompt set that exercises the user embedding layer, mapped to the personas your brand actually needs to win. Use this to surface which competitors a model defaults to under different user conditions.
• A real-prompt set sourced from question-shaped GSC queries, customer panel inputs, and regex-extracted “who/what/where/can/should” patterns. These are the short, retrieval-triggering prompts most users still write.
• A small qualitative library of messy, context-rich real prompts pulled from the type of work we did for the study. Use it to sanity-check whether your content actually answers the question the user is asking, not the question your keyword tool says they’re asking.

At that point, you’re not just tracking AI visibility. You’re tracking it across the full spectrum of how real users get to your content, from a three-word “good walking shoes” query to a 40-word “I’m a 60-year-old with plantar fasciitis…” request.

---

What the broader data tells us about AI search

A handful of additional findings from the January 2026 survey help explain why these prompt patterns matter.

Users increasingly trust AI recommendations

Up to 68% of users trust ChatGPT’s recommendations more than Google’s, with most citing detail, lack of ads, and personalization as the reasons. 

• NIM’s research has found ChatGPT often produces more efficient, accurate consumer decisions than Google.
• A January 2026 study reported that trust in Google weakens as AI usage rises. 
• Yext’s 2026 research puts ChatGPT’s share of “top-quality source” responses at 35% among heavy AI users.

AI search is becoming a daily habit

Half of active AI users use these tools daily or several times per day to complete tasks they used to do on Google.

• Search Engine Land reported 37% of consumers now start searches with an AI tool instead of Google.
• OpenAI’s February 2026 numbers put ChatGPT’s weekly active users at 900 million — more than double a year earlier.

Citations still drive traffic

85% of users click through to cited sources at least some of the time; 21.9% always do. The mention is not the end of the funnel. 

• Conductor’s 2026 benchmarks showed AI referral traffic up 357% year-over-year. 
• Semrush reported outbound referrals from ChatGPT up 206% in 2025. 
• Emarketed saw AI-referred visitors converting at 4.4x the rate of standard organic. 
• Volume is still small (Conductor pegs it at around 1.08% of total traffic), but it punches well above its weight class.

Voice may finally be having its moment

34% of users are now using voice chat daily or more often. This is the first dataset I’ve seen that actually delivers on the “voice search will matter” promise we’ve been hearing for a decade. 

It’s worth pairing all of this with Ahrefs’ latest AI Overviews CTR research: The presence of an AI Overview correlates with a 58% lower clickthrough rate for the top-ranking page. The traffic that does come through is qualified. The traffic that doesn’t is gone

AI search is settling into a richer, more personalized form. The intent stack is the same one Google has always served. What’s new is the embedding layer and the tracking demands it entails. That creates a clear set of priorities for SEO and GEO teams.

What changes — and what doesn’t

Here are three things you can do with this information if you’re an SEO lead, content lead, or strategy lead:

• Audit your prompt-tracking setup: If it’s all synthetic prompts or all keyword-shaped prompts, you’re missing half the picture. Build the layered framework outlined above.
• Map your content to the user embedding layer: For your top categories, list the personas (e.g., age, life stage, profession, condition, budget) most likely to carry real prompts into AI search. Then check whether your PDPs, blog content, and FAQs actually answer those people’s questions.
• Don’t abandon the SEO-keyword work: Roughly one-third of real prompts still look like classic search queries. With web retrieval running at 90%+ on many of the prompts we monitor, the gap between an SEO keyword and an AI prompt is narrower than the GEO discourse implies.

The behavior change is real. The sophistication of AI thought leaders’ prompting is partly here and partly oversold. Most people are still doing Google-style searches. They’re just searching inside an interface that knows more about them.

If that’s where the audience is, that’s where we have to optimize.

Methodology

Both studies referenced in this article were conducted by the Stella Rising team. You can read it in “New Data: How Consumers Use LLMs for Search in 2026 (And What It Means for GEO).” 

The August 2025 study surveyed 178 members of Stella’s Glimmer Insights community, 113 of whom were active LLM users. 

The January 2026 study surveyed 524 active LLM users via Centiment, defined as having used ChatGPT, Copilot, or Gemini in the previous 30 days, with a margin of error of approximately ±4.3% at the 95% confidence level. 

Given its smaller size and category-specific composition, the August 2025 panel should be viewed as directional rather than statistically representative of the broader U.S. AI user population.