How real people actually prompt AI — and what it means for GEO

Most group aren’t utilizing AI the measurement GEO discussions often assume. Two surveys of AI users conducted by Stella Rising recovered that galore prompts still look remarkably akin to accepted hunt queries. (Disclosure: I’m the VP of SEO astatine Stella Rising.)

One study focused connected a beauty-oriented user sheet successful August 2025, while the different surveyed a broader general-audience organization successful January 2026. Across some studies, prompts were short, often keyword-driven, and overmuch person to a Google hunt than the elaborate punctual templates celebrated successful AI trading circles.

At the aforesaid time, a increasing stock of users are adding individual context, specified arsenic their budget, location, profession, age, wellness concerns, aliases preferences. Those specifications springiness AI systems acold much accusation than a accepted hunt query ever could, creating a caller furniture of personalization that influences recommendations and marque visibility.

The mixed findings propose that GEO strategies request to relationship for some realities: Many AI searches still lucifer classical keyword queries, while the highest-value recommendations progressively look from prompts rich | pinch individual context. That’s wherever the opportunity — and the measurement situation — lies.

A batch of group are still typing for illustration it’s 2008

The biggest takeaway crossed some surveys is that the median AI personification is still throwing a keyword complete the wall and hoping for the best.

In the general-audience study from January:

• Two-thirds of respondents reported penning prompts of 15 words aliases fewer. 
• Only 12% wrote thing that would suffice arsenic a “real” punctual by the standards of an AI influencer thread. 
• About 60% phrased their queries arsenic questions, while only 9% gave a nonstop command. 

That mirrors what Pew Research has been seeing much broadly — 34% of each U.S. adults now usage ChatGPT, astir double the 2023 share, and 58% of adults nether 30 usage it.

When we ran a script task asking respondents to constitute the punctual they’d nonstop if they needed a caller brace of shoes, the median reply was 8 words. Real examples from the sheet included:

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

This lines up pinch Semrush’s clickstream data connected ChatGPT’s hunt mode, which shows the mean punctual magnitude is 4.2 to 8.7 words, fundamentally the aforesaid arsenic a Google query.

Longer, system prompts thin to look only erstwhile users are doing thing different than search, specified arsenic drafting, coding, aliases imaginative work.

For AEO and GEO work, that’s the portion to internalize. If you’re optimizing for prompts for illustration “Compare the apical 5 orthopedic-approved stepping shoes nether $150 for plantar fasciitis pinch 4.5+ prima ratings,” you’re optimizing for the incorrect distribution. 

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

The displacement betwixt the 2 surveys

In the August 2025 survey, we classified astir 50% of the free-text prompts arsenic “SEO-keyword-shaped,” meaning short, ambiguous, and brand-and-attribute-driven. By the clip the January 2026 study came back, that stock had dropped person to 30%. The remaining 70% had grown longer and much contextualized.

A fewer findings are worthy carrying forward:

• 24.5% of each prompts see the connection “best.” If you’re not appearing successful “best [category]” responses, you’re missing 1 of the highest-intent slots.
• 28% of prompts mention value aliases fund constraints. Users aren’t conscionable shopping. They’re shopping pinch a number successful their head.
• 16% of prompts are explicitly location-based. The “near me” query shape has successfully migrated from Google to LLMs.
• 32% of prompts see individual attributes (e.g., size, profession, wellness condition, life stage, etc.). This is the astir important number connected the page, and we’ll travel backmost to it.

On location specifically, the 16% fig lines up pinch what Local Falcon’s 2025 investigation showed for AI hunt overall: AI Overviews now look connected 92% of informational section queries, but only connected 15% of elemental local-pack queries. The intent is moving into LLMs faster than the proviso of optimized section contented for AI engines.

One caveat: These were 2 different surveys pinch 2 different audiences. The January 2026 general-audience sample was structurally much transactional than the August 2025 beauty-focused panel, which partially explains why less prompts looked for illustration keyword-style searches and much looked for illustration afloat requests. I wouldn’t over-index connected the “prompts are evolving” narrative, but I’d perfectly return the directional read.

The personification embedding furniture is wherever this gets interesting

The 32% fig (prompts containing existent individual context) is the astir under-discussed uncovering successful the dataset.

Nearly one-third of users are willingly handing LLMs accusation that nary Google query would usually carry, specified arsenic their size, job, training plan, surviving situation, aliases kids’ ages. We spot prompts successful the information like:

• “What shoes would you urge for regular opinionated astatine work?”
• “Find maine a cost-effective brace of moving shoes that I tin bid connected Amazon. My size is men’s 10.”
• “Please show maine the apical 5 shoes for wide feet successful a size 8 for women that are comfortable, stylish, nether $120, and that younger group won’t make nosy of for a Gen X personification for illustration me.”

That past 1 unsocial packs successful gender, ft width, size, budget, style intent, generational identity, and a existent societal anxiety. No accepted hunt query was ever going to aboveground each of that.

This is the personification embedding furniture astatine work. When personification interacts pinch ChatGPT aliases Gemini repeatedly, the exemplary builds a floor plan of who they are, which progressively persists done memory. The personification is, successful effect, training the adjunct connected themselves. Once that spot is established, they extremity penning aboveground queries and commencement penning requests that presume the adjunct knows them.

That displacement has 2 implications for really brands should deliberation astir visibility:

• The prompts that thrust acquisition decisions are often not the ones that show up successful a SERP aliases keyword tool. A existent Gen X female asking astir wide-fit, $120, “won’t-get-made-fun-of” sneakers will ne'er look arsenic a tracked SERP keyword. But that’s the punctual that decides whether your footwear ends up successful the proposal set.
• The worth of a marque citation increases substantially erstwhile it appears wrong a context-rich prompt. If the exemplary is already factoring successful personification attributes, the brands it surfaces are prefiltered for relevance. That’s a overmuch higher-quality belief than a generic bluish link.

Where synthetic prompts fresh — and wherever they don’t

A communal maneuver successful GEO punctual investigation is to conception synthetic personas (“I’m a 38-year-old merchandise head training for a half marathon successful Boston who prefers brands focused connected sustainability…”) and past usage those personas to stress-test which brands an LLM surfaces nether different scenarios. There’s existent merit to the approach. If the personification embedding furniture is doing the dense lifting successful the answer, the only measurement to simulate the reply is to simulate the user.

But synthetic prompts don’t seizure everything. Real prompts are messy, layered, and influenced by caller speech history, persistent memory, and signals the exemplary has picked up complete weeks of use. You tin trade a 50-word persona and still miss the nuance of a personification who has been talking to ChatGPT astir their day, preferences, and family for 6 months.

Instead, usage synthetic prompts to representation the personas your marque needs to beryllium visible to, but don’t dainty the resulting visibility scores arsenic crushed truth. Combine them pinch existent punctual information wherever possible. That tin mean customer interviews, societal hunt patterns, support tickets, aliases regex pulls of question-shaped queries from Google Search Console.

What to really track

This people leads to the adjacent question: Should you way SEO keywords successful your AI visibility level if one-third of existent prompts look for illustration SEO keywords?

The reply is yes, pinch 1 filter.

Across the past quarter, our squad has seen web retrieval rates connected tracked prompts climb sharply. On respective customer accounts, much than 90% of monitored prompts now trigger unrecorded web hunt wrong ChatGPT aliases Google’s AI Mode. 

When that happens, the LLM is efficaciously moving a real-time SERP and synthesizing the result. That intends the short, keyword-shaped prompts we identified — astir 30% of the full — are still very overmuch successful play. They behave for illustration AI-flavored Google queries and should beryllium tracked accordingly.

The select is this: Don’t discarded search slots connected prompts that are axenic caput position aliases single-brand queries. Those are apt to beryllium answered from exemplary weights aliases short canned responses alternatively than retrieval, and they won’t springiness you a useful publication connected visibility.

Here’s a applicable setup:

• A synthetic-persona punctual group that exercises the personification embedding layer, mapped to the personas your marque really needs to win. Use this to aboveground which competitors a exemplary defaults to nether different personification conditions.
• A real-prompt group originated from question-shaped GSC queries, customer sheet inputs, and regex-extracted “who/what/where/can/should” patterns. These are the short, retrieval-triggering prompts astir users still write.
• A mini qualitative room of messy, context-rich existent prompts pulled from the type of activity we did for the study. Use it to sanity-check whether your contented really answers the mobility the personification is asking, not the mobility your keyword instrumentality says they’re asking.

At that point, you’re not conscionable search AI visibility. You’re search it crossed the afloat spectrum of really existent users get to your content, from a three-word “good stepping shoes” query to a 40-word “I’m a 60-year-old pinch plantar fasciitis…” request.

---

What the broader information tells america astir AI search

A fistful of further findings from the January 2026 study thief explicate why these punctual patterns matter.

Users progressively spot AI recommendations

Up to 68% of users spot ChatGPT’s recommendations much than Google’s, pinch astir citing detail, deficiency of ads, and personalization arsenic the reasons. 

• NIM’s investigation has recovered ChatGPT often produces more efficient, meticulous user decisions than Google.
• A January 2026 study reported that trust successful Google weakens arsenic AI usage rises. 
• Yext’s 2026 investigation puts ChatGPT’s stock of “top-quality source” responses astatine 35% among dense AI users.

AI hunt is becoming a regular habit

Half of progressive AI users usage these devices regular aliases respective times per time to complete tasks they utilized to do connected Google.

• MCP reported 37% of consumers now commencement searches pinch an AI instrumentality alternatively of Google.
• OpenAI’s February 2026 numbers put ChatGPT’s play progressive users astatine 900 million — much than double a twelvemonth earlier.

Citations still thrust traffic

85% of users click done to cited sources astatine slightest immoderate of the time; 21.9% ever do. The mention is not the extremity of the funnel. 

• Conductor’s 2026 benchmarks showed AI referral postulation up 357% year-over-year. 
• Semrush reported outbound referrals from ChatGPT up 206% successful 2025. 
• Emarketed saw AI-referred visitors converting astatine 4.4x the complaint of modular organic. 
• Volume is still mini (Conductor pegs it astatine astir 1.08% of full traffic), but it punches good supra its weight class.

Voice whitethorn yet beryllium having its moment

34% of users are now utilizing sound chat regular aliases much often. This is the first dataset I’ve seen that really delivers connected the “voice hunt will matter” committedness we’ve been proceeding for a decade. 

It’s worthy pairing each of this pinch Ahrefs’ latest AI Overviews CTR research: The beingness of an AI Overview correlates pinch a 58% little clickthrough complaint for the top-ranking page. The postulation that does travel done is qualified. The postulation that doesn’t is gone

AI hunt is settling into a richer, much personalized form. The intent stack is the aforesaid 1 Google has ever served. What’s caller is the embedding furniture and the search demands it entails. That creates a clear group of priorities for SEO and GEO teams.

What changes — and what doesn’t

Here are 3 things you tin do pinch this accusation if you’re an SEO lead, contented lead, aliases strategy lead:

• Audit your prompt-tracking setup: If it’s each synthetic prompts aliases each keyword-shaped prompts, you’re missing half the picture. Build the layered model outlined above.
• Map your contented to the personification embedding layer: For your apical categories, database the personas (e.g., age, life stage, profession, condition, budget) astir apt to transportation existent prompts into AI search. Then cheque whether your PDPs, blog content, and FAQs really reply those people’s questions.
• Don’t wantonness the SEO-keyword work: Roughly one-third of existent prompts still look for illustration classical hunt queries. With web retrieval moving astatine 90%+ connected galore of the prompts we monitor, the spread betwixt an SEO keyword and an AI punctual is narrower than the GEO sermon implies.

The behaviour alteration is real. The sophistication of AI thought leaders’ prompting is partially present and partially oversold. Most group are still doing Google-style searches. They’re conscionable searching wrong an interface that knows much astir them.

If that’s wherever the assemblage is, that’s wherever we person to optimize.

Methodology

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

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

The January 2026 study surveyed 524 progressive LLM users via Centiment, defined arsenic having utilized ChatGPT, Copilot, aliases Gemini successful the erstwhile 30 days, pinch a separator of correction of astir ±4.3% astatine the 95% assurance level. 

Given its smaller size and category-specific composition, the August 2025 sheet should beryllium viewed arsenic directional alternatively than statistically typical of the broader U.S. AI personification population.