Real vs Estimated Prompts: I Analyzed 100s of Real ChatGPT Queries - AI Search Visibility Blog | Insights & Data | Otterly.AI
Two questions come up every time I talk about AI search with marketing teams:
“How do I find real prompts from real people?”
And the skeptic’s follow-up:
“Why should I even bother optimizing for estimated prompts?”
Fair questions. So I ran an experiment to find out.
Study Methodology
About This Research: We analyzed hundreds of real prompts collected from market surveys. We asked SEO and marketing professionals what they would enter on ChatGPT when research products in our market category, and compared them against hundreds of estimated prompts generated through our own prompt research (with our AI Prompt Research tool) and Google Search Console data. The prompt data and research was collected for the United States as a key target market.
Key Findings (TL;DR)
71% longer prompts from real users than from estimated prompts
- Real users write 71% longer prompts than estimated/synthetic prompts (15.1 words vs 8.8 words on average).
- 78.9% of real prompts show tool-finding intent, compared to only 62.5% of estimated prompts.
- 52.1% of real prompts use personal pronouns (“I”, “my”, “me”) vs only 18.8% of estimated prompts.
- Real prompts are 3x more problem-oriented (21.1% vs 7.1%), suggesting users describe their challenges more often.
- Estimated prompts overweight “best” (commercial intent) while real prompts favor “what” and “I” (exploratory, personal).
Key Terms Explained
Real Prompts: Actual queries that real users have confirmed they enter into ChatGPT, collected through market surveys and user research.
Estimated Prompts: Queries that we researched with synthetic prompt generation, using data from Google Search Console and turning them into prompts via OtterlyAI.
Prompt Research: How People Use ChatGPT
As long as there’s no search console from OpenAI or any other direct access to their vast pool of prompt data, we won’t know for sure what people type on ChatGPT. But here’s what we know so far.
ChatGPT handles over 2 billion searches every single day. People use it for research and guidance across a wide range of topics.
According to the National Bureau of Economic Research (who teamed up with OpenAI), practical guidance and seeking information are two of the most common use cases on ChatGPT. Check out the full paper here.
But there’s one additional data point, I’d like to add here.
One more data point worth adding. Semrush conducted a large-scale study last year and published some interesting findings on query length. They found a massive difference in query length depending on whether users activated web search or not.
4.2 words longer average length of a prompt without web search activated
Users who activated web search typed an average of 4.2 words into the prompt interface. Users without web search averaged 23 words. That’s a 5x difference.
When users interact with ChatGPT without the SearchGPT feature, the prompts tend to be more detailed. During this period, we observed an average prompt length of 23 words, with some reaching up to 2,717 words.
People use ChatGPT for research and search, a lot. We don’t know the exact queries and prompts they enter, but we have enough signal to start making informed decisions.
How to Find Prompts, Both Real and Estimated
I wrote and spoke about this topic in length. I recommend checking out this article on all the different ways to find real and estimated prompts for ChatGPT tracking and optimization. But let me tell you how my current real-life process as a marketer at OtterlyAI looks like.
- Google Search Console: I regularly go through our queries to identify new and upcoming long-tail search queries that people might also put into ChatGPT.
- Market Surveys: I ran various surveys last year, asking real users in our market segments to give us the prompts they would enter on ChatGPT.
- Combining Both Lists: I merge both lists to track and improve my own brand and website visibility on ChatGPT.
My Prompt Experiment: Comparing Real vs Estimated Prompts
Then it hit me. I’m sitting on some interesting data. I have two buckets:
- Prompts I believe people enter on ChatGPT
- Prompts real people told me they would enter on ChatGPT
So I went on a journey to compare both lists and understand:
- Key differences and patterns
- Key similarities
- How far “off” I was with my estimated prompts
- How my brand would perform on AI search across both sets
Here’s what I found comparing hundreds of estimated prompts against hundreds of real prompts from real people.
At-a-Glance Comparison of Real vs Estimated Prompts
| Dimension | Real Prompts | Estimated Prompts | Strategic Implication |
|---|---|---|---|
| Avg Length | 15.1 words | 8.8 words | Real users need more context – create comprehensive guides |
| Personal Pronouns | 52.1% | 18.8% | Real users personalize – create “for you” content |
| Tool-Finding Intent | 78.9% | 62.5% | Real users hyper-focused on tools – prioritize comparison pages |
| Problem-Solving | 21.1% | 7.1% | Real users describe problems – create problem-first content |
| Conversational | 23.9% | 5.8% | Real users conversational – optimize for AI chat interfaces |
| “Track” vs “Monitor” | Prefer “track” | Even split | Use “track” in customer-facing content |
Here’s now a detailed breakdown of the comparison.
Language & Structure Comparison
Prompt Length Distribution
Real Prompts:
- Average: 15.1 words
- Median: 12 words
- Range: 2-86 words
- Pattern: High variance, many verbose explanations
Estimated Prompts:
- Average: 8.8 words
- Median: 7 words
- Range: 2-86 words
- Pattern: More consistent, SEO-optimized brevity
Insight: Real users write 71% longer prompts on average, suggesting they’re less certain about what to search for and need to provide more context. This has major implications for keyword targeting and content structuring.
| Type | Real Prompts | Estimated Prompts |
| Questions (contains ? ) | 22.5% | 18.4% |
| Statements | 77.5% | 81.6% |
Starting Word Patterns
We analyzed the starting words of both real and estimated prompts. Quite fascinating to see those differences.
Starting Word Patterns
Real Prompts – Top 10:
- “what” – 15.5%
- “how” – 11.3%
- “i” – 11.3%
- “best” – 11.3%
- “find” – 8.5%
- “ai” – 7.0%
- “tools” – 4.2%
- Others – various
Estimated Prompts – Top 10:
- “best” – 18.8%
- “how” – 14.9%
- “ai” – 13.6%
- “what” – 12.3%
- “top” – 2.9%
- “i” – 2.6%
- “find” – 2.3%
- “tools” – 1.9%
Critical Difference: Real prompts start with “what” and “i” more frequently (exploratory, personal), while estimated prompts heavily favor “best” (commercial intent). This suggests estimated prompts overweight bottom-funnel behavior.
Intent Comparison between Real and Estimated Prompts
Key Intent Gaps:
1. Tool Finding Intent (78.9% → 62.5%)
- Real users overwhelmingly want to find specific tools
- Estimated prompts underweight this by 16 percentage points
- Implication: Content should heavily emphasize tool comparison, feature listings, and “which tool” content
2. Problem Solving Intent (21.1% → 7.1%)
- Real users express problems (“help me,” “I need,” “how can I”)
- Estimated prompts are 3x less problem-oriented
- Implication: Missing significant demand for troubleshooting content and use-case guidance
3. Monitoring Intent (50.7% → 58.6%)
- Estimated prompts correctly identify monitoring as core need
- However, may be overemphasizing pure “monitoring” at expense of exploratory research
Term Frequency Comparison
These are the terms used in both data sets. Keep in mind that they are highly specific to my industry.
| Term | Real Prompts | Estimated Prompts | Delta |
|---|---|---|---|
| track | 29.6% | 27.8% | -1.7% |
| monitor | 15.5% | 25.6% | +10.1% |
| tool | 63.4% | 40.5% | -22.9% |
| brand | 53.5% | 36.6% | -17.0% |
| visibility | 35.2% | 33.0% | -2.2% |
| chatgpt | 11.3% | 12.0% | +0.7% |
| ai search | 52.1% | 33.0% | -19.1% |
| geo | 5.6% | 2.3% | -3.4% |
| optimization | 22.5% | 17.8% | -4.7% |
| best | 29.6% | 27.2% | -2.4% |
How Does My AI Search Performance Stack Up?
Among the most important questions: does my brand performance differ between real and estimated prompts?
The answer is, not really. Our brand performance was slightly better for estimated prompts, likely because we’ve been optimizing for those queries over the last few months. All in all, the brand ranking for us vs. our competitors looked quite similar. The same vendors performed at comparable levels across both sets.
| Our Own Brand Performance on AI Search |
|---|
| Brand Visibility – Real Prompts | Brand Visibility – Estimated Prompts | Brand Ranking – Real Prompts | Brand Ranking – Estimated Prompts | |
| OtterlyAI | 43% | 50% | 3 | 2 |
| Competitor 1 | 66% | 58% | 1 | 1 |
| Competitor 2 | 52% | 48% | 2 | 3 |
| Competitor 3 | 42% | 38% | 4 | 4 |
Here are two screenshots from OtterlyAI – side by side:
My Recommendation
Take action. Start using the data sources you already have. Conducting real market surveys will help you tailor your strategies even further toward the audiences you serve.
P.S. If you’ve been saying “we don’t know what people do on ChatGPT“… that’s no longer an excuse. You have enough data to get started on your AI search performance today.
Frequently Asked Questions (FAQ)
How do I find real prompts from real people?
You can find real prompts through market surveys asking your target audience directly, analyzing Google Search Console for long-tail queries, and combining multiple data sources to build a comprehensive prompt list for AI search optimization.
Why are estimated prompts worth optimizing for?
While real prompts provide the most accurate picture of user behavior, estimated prompts still capture valuable search intent patterns. The key is understanding the gaps between them – real prompts are 71% longer and 3x more problem-oriented, so optimize your content strategy accordingly.
How long are real ChatGPT prompts?
Real ChatGPT prompts average 15.1 words with a median of 12 words and can range from 2 to 86 words. This is significantly longer than estimated prompts, which average only 8.8 words. When users don’t use web search, prompts can reach up to 2,717 words.