AI-search strategy often starts with long, carefully constructed prompts invented by marketers. Two surveys reported in June 2026 offer a useful correction: many people still use short phrases or questions, while others add personal context such as location, budget or preferences.
That evidence supports a mixed research model. Traditional query analysis remains relevant, but it cannot represent every personalised or conversational path.
What the research covered
The analysis published by Search Engine Land described two studies conducted by Stella Rising:
- an August 2025 beauty-focused panel with 178 participants, including 113 active large-language-model users; and
- a January 2026 survey of 524 US adults who had used ChatGPT, Copilot or Gemini in the previous 30 days.
For the January panel, two-thirds of respondents said their prompts were 15 words or fewer, about 60% phrased them as questions, and 9% described using direct commands. In a shoe-shopping task, the median written prompt was eight words.
The article reports an approximate ±4.3% margin of error at a 95% confidence level for the January survey. The earlier panel was smaller and category-specific, so the authors describe it as directional.
These are useful findings, not a complete map of AI behaviour. They rely on self-reported or scenario responses from US panels and do not establish how every platform, country or buying journey works.
Why short prompts still matter
Short prompts can trigger live web retrieval, and Google's generative Search documentation says AI features may issue related “fan-out” searches to gather information. A brief starting query can therefore lead to a broader retrieval process.
This is one reason established SEO work remains valuable:
- clear pages can match a topic without repeating every query variation;
- descriptive internal links help discovery;
- strong service and comparison pages answer commercial intent; and
- crawlable, indexed content can be retrieved when a system needs current evidence.
Publishing a separate page for every synthetic prompt is neither necessary nor safe. Google warns against scaled pages created primarily to manipulate rankings or generative answers.
Why personal context changes the problem
A user may ask for an agency, then add a budget, location, industry, deadline or platform constraint in a follow-up. The eventual recommendation may depend on the conversation and on information the measurement tool cannot reproduce.
That makes a single “share of AI voice” number incomplete. Two users can enter the same words and receive different results because of platform, time, location, account context or conversation history.
The practical response is to measure a defined sample consistently and label it as a sample.
A defensible prompt-research method
1. Start with observed language
Collect questions from sales calls, contact forms, support conversations, internal site search and Google Search Console. Remove personal data and group the language by the decision it represents.
2. Separate three prompt sets
| Set | Purpose | Example source |
|---|---|---|
| Search-like prompts | Track short discovery and comparison demand | Search Console queries |
| Customer-language prompts | Reflect the constraints buyers actually mention | Anonymised sales questions |
| Scenario prompts | Test strategically important buyer contexts | Clearly documented synthetic personas |
Synthetic prompts are useful for coverage testing, but they should not be presented as actual customer demand.
3. Record the test conditions
Save the exact prompt, platform, model or product, date, country, sign-in state and whether the conversation was fresh. Note the cited sources and whether the answer retrieved current web content.
4. Evaluate accuracy and usefulness
Track whether the brand appears only after being named, whether factual statements are correct, which pages are cited, and whether the response would genuinely help the intended buyer. Repeat the sample before calling a movement a trend.
5. Turn gaps into better customer content
If important constraints are absent from the site, improve the relevant service or guide. Do not create pages that claim experience the business does not have. Useful additions might include process limits, service-area facts, decision criteria or sourced answers to recurring questions.
What this means for South African businesses
The US survey results should not be converted into local market percentages. They do support a practical hypothesis worth testing: South African buyers may combine short search language with local and commercial context.
Use first-party South African evidence to validate that hypothesis. For Symaxx, a small set of real questions about budget, city, platform, timeline and service fit is more valuable than hundreds of invented prompts.
The durable strategy is to answer genuine customer decisions clearly, keep those answers indexable, and measure how both classic and AI-assisted search surfaces use them.
Sources
- Search Engine Land: How real people actually prompt AI and what it means for GEO
- Google Search Central: Optimizing your website for generative AI features
- Google Search Central: Creating helpful, reliable, people-first content