Do GEO Visibility Tools Use Synthetic Prompts Instead of Real Buyer Questions?
Yes, most GEO visibility tools track synthetic prompts: questions written by you, by the vendor or by a model, rather than questions taken from real users' AI conversations. They have little choice, because AI companies do not publish what people ask. Synthetic prompts are fine when they are built from real buyer evidence and kept stable. They mislead when they are generated from keywords or your own marketing copy, in words no buyer would use, and then reported as if they were real demand.
Why don't GEO tools use real user questions?
Because AI companies do not publish them. The conversations people have with ChatGPT, Gemini, Claude and Perplexity are private to each company, and none of them publishes query logs the way search engines publish aggregated keyword data. What exists instead:
- Opt-in panels: some vendors pay consenting users to share their actual AI conversations; Profound reports more than 1.5 billion such conversations and Evertune more than 150 million prompts. These are real questions, but from a panel that may not match your buyers, and volumes are extrapolated from it.
- Clickstream data: browsing data from consenting users or extensions, used to estimate topics and volumes. Coverage is partial and methods are rarely disclosed.
- Search data: long, question-style queries in Search Console and keyword tools, which overlap with what people ask AI but are not the same.
- Your own data: sales calls, support tickets, site search, reviews and "how did you hear about us?" answers.
Any number a tool labels "prompt volume" is therefore modelled from a sample, not counted across all users; see whether AI prompt volume data is reliable.
Why do AI visibility studies use simulated prompts if real questions are unpredictable?
Real buyer questions are extremely varied. People describing the same need write very different prompts, with different lengths, details and conditions. So no fixed set can match what people actually type.
Simulation is still the right method, for the same reason market research uses questionnaires. A fixed set of prompts, each representing a real buyer need, lets you:
- Compare over time, because the same questions are asked each period.
- Compare brands fairly, because every brand faces the same questions.
- Repeat runs to measure how stable answers are.
- Cover needs deliberately, including valuable questions that are rare.
The aim is not to predict the exact words buyers use. It is to sample the needs they have, in wording close enough that engines retrieve the same kinds of sources. A prompt set is an instrument. Like any instrument it can be well or badly designed.
When do synthetic prompts mislead?
- Keyword-shaped prompts. "Best CRM software small business 2026" pasted in as a prompt reads like a search query, not a question, and retrieves different sources.
- Your own language. Prompts generated from your website repeat your positioning ("all-in-one revenue intelligence platform") and flatter your results.
- Branded prompts mixed in. Prompts with your name almost always return your name, inflating visibility.
- Improbable combinations. Generated prompts can pair conditions nobody combines, such as enterprise security requirements with a free-plan budget.
- Silent rotation. Prompts that change from period to period make trends meaningless.
- Modelled volume presented as fact, which makes some prompts look important for no real reason.
How can you tell if a prompt-monitoring tool is generating questions nobody asks?
Run five checks on the tool's prompt list:
- Provenance. Ask the vendor where each prompt came from: your input, their library, a model, or a data source. A good tool can tell you per prompt.
- The sales-call test. Take 20 questions buyers actually asked your team recently. Count how many have a close match in the tool's list. A low match rate means the list reflects the tool, not your market.
- The read-aloud test. Read ten prompts aloud. Would a buyer say them to a colleague? Keyword strings, stacked adjectives and marketing phrases fail.
- The branded share. Count prompts that include your brand or product names. More than a small share inflates results.
- The volume question. If prompts carry volumes, ask what data the volume is modelled from and what its error range is. A number with no method is a guess.
A tool that fails several checks can still be useful for running your own prompts, but its suggested prompts and headline scores should not drive decisions. For evaluating tools more broadly, see how to evaluate GEO tools before choosing one.
What does fixing a weak synthetic prompt look like?
Most weak prompts can be rescued by rewriting them from a real buyer's point of view. Some typical examples:
| Tool-generated prompt | Problem | Rewritten from buyer evidence |
|---|---|---|
| "Best HR software small business 2026" | Keyword string, no need stated | "What HR software works for a 30-person company that needs holiday tracking and UK payroll?" |
| "Top AI-powered revenue intelligence platforms" | Vendor language buyers rarely use | "What tools show which sales deals are likely to close, for a team of eight reps?" |
| "Is [your brand] the best project tool for agencies?" | Branded and leading | "What project management tool do small design agencies use to track client approvals?" |
| "Enterprise-grade free CRM with SOC 2 and unlimited users" | Improbable combination | "Which CRMs have a SOC 2 report and cost under $30 per user?" |
Each rewrite states one need, uses plain words, and carries one or two conditions a buyer actually mentioned. That is usually enough to make the prompt retrieve the sources real buyers would see.
How do you anchor synthetic prompts in real buyer language?
Keep the method synthetic but the language real:
- Start from evidence: sales notes, support tickets, reviews, site search and pre-sale chats.
- Write one prompt per need, in a buyer's words, including the conditions they mention.
- Validate a sample with people: ask a few customers or salespeople whether the prompt sounds like something a buyer would ask.
- Keep tool suggestions separate until reviewed, then promote the useful ones.
- Freeze and version the set, so trends reflect engines, not prompt changes.
Step-by-step prompt selection is covered in how to choose the right prompts to track. For B2B, see also which procurement buyer questions to use.
Can you get closer to real questions over time?
Yes, from your own channels:
- add "What did you ask ChatGPT or another AI tool when looking for a solution like ours?" to onboarding surveys or sales calls;
- review AI referral landing pages in analytics, which show which topics sent visitors;
- look at long, question-style queries in Search Console, which often mirror how people phrase AI prompts;
- read the questions people ask in reviews and communities about your category.
Each of these gives you real wording to test and refine your prompt set against.
What do people ask about synthetic prompts?
Are synthetic prompts worse than keyword tracking?
Not inherently. Keyword tracking also uses a chosen sample of queries. The quality depends on how well the sample represents real needs, in both cases.
Should I trust a tool's "top prompts" in my category?
Treat them as ideas to review, not as a ranking of real demand. Check them against your own buyer evidence.
Do real user prompts ever become available?
Yes, partly. Some vendors collect real prompts through opt-in consumer panels at large scale; check which plan includes them. They are samples from particular groups of users, not full logs, which AI companies do not publish, so check them against your own buyer evidence.
How many real buyer questions do I need to anchor a prompt set?
Twenty to fifty real questions from sales and support usually reveal the main needs and wording patterns for a single market.
Published by The Kovmere Team · updated Sep 28, 2026 · More resources