Prompt research is the practice of finding and prioritising the questions people put to AI assistants (the full, conversational prompts, not three-word keywords) so you can check whether your brand appears in the answers and fix the content that decides it. It complements keyword research rather than replacing it: keyword tools still show topics and question shapes, but they do not show how often people ask ChatGPT or Gemini, and a study of about 3,000 runs found AI answer lists almost never repeat. So the aim is a stable, prioritised prompt set you track by visibility percentage, not by position.
Key Takeaways
- Prompts are longer and more situational than keywords, but public, verified data on how much is thin. Search Engine Land's 12 October 2026 piece argues the shift with examples, not statistics.
- No tool shows true prompt volume. Semrush's Prompt Research report offers "AI Volume" estimates but does not document how they are calculated, per its own help page.
- Rankings in AI answers are unstable. SparkToro and Gumshoe found the chance of getting the same brand list twice was under 1 in 100; presence across many runs is the useful metric.
- Build a set of 40-100 prompts across personas, funnel stages and constraints, then score them for business value and evidence gaps.
- Source prompts from your own data first: sales calls, support tickets, Search Console, community threads.
- Prioritise decision-stage prompts where the answer has to name a brand or make a choice.
- Always re-run each prompt several times and log the date, tool and account state.
Why keyword research alone no longer covers it
Keyword tools were built around short queries with measurable volume. Two things break that for AI search.
First, many AI answers are built from several hidden searches, known as query fan-out. My query fan-out explainer covers the evidence: Seer Interactive found that 95% of Gemini 3's fan-out queries in its test had zero global search volume. The visible prompt may have no volume either.
Second, context is personal. In Becky Simms' Search Engine Land article (12 October 2026), the examples show assistants drawing on earlier conversations, so a prompt can omit context that changes the recommendation. She argues prompt tracking shows only one version of the answer. That is a reasoned argument with illustrative examples; she gives no figures, and I won't invent any either.
What people do use these tools for is documented at a high level. OpenAI's September 2025 usage study of about 1.5 million consumer conversations (an NBER working paper with an economist at Harvard) found "Asking" made up 49% of messages, "Doing" 40% and "Expressing" 11%. That is consumer ChatGPT overall, not buying prompts, so it only tells you that asking questions is the single largest behaviour.
What the evidence says about measuring prompts
Answers vary a lot
Search Engine Land reported a study by Rand Fishkin (SparkToro) and Patrick O'Donnell (Gumshoe.ai) in which about 600 volunteers ran 12 prompts through ChatGPT, Claude and Google's AI nearly 3,000 times. The chance of the same list twice was under 1 in 100, and the same order about 1 in 1,000. The authors concluded position tracking is essentially meaningless, while visibility percentage holds up better: some brands appeared in 60% to 90% of responses for a given intent. They acknowledge the method is imperfect and that open questions remain, such as how many runs are needed.
Different phrasings, similar brand sets
The same study found real user prompts were highly varied, yet returned similar brand sets when the underlying intent matched. For research, that means you are mapping intents, not chasing exact wording.
Volume figures are modelled
Semrush's Prompt Research report draws on a prompt database of over 317 million AI queries across Google AI Overviews, AI Mode, Gemini and ChatGPT, with daily rolling refreshes. Its "AI Volume" is described as an estimated number of searches, and the page does not explain how it is calculated. Useful for ranking topics against each other; not a number to promise a client.
The method: six steps
Step 1: Define the decisions, not the keywords
List 5-10 decisions your customers make where an AI answer could name a brand. For an edtech business that might be "choose a program," "check fees," "compare options," "assess outcomes." For SaaS: "shortlist tools," "compare pricing," "check integrations."
Step 2: Collect raw prompts from your own data
Lumen's guide to choosing AEO prompts lists six sources, and I agree with most of them (Lumen sells tools in this space, so take its product claims with salt):
- Sales and counselling calls. Listen to about 20 recent calls, flag recurring questions, and copy the wording.
- Support tickets and chat logs. Pre-sale questions especially.
- Search Console. Filter queries starting with how, what, which, best, should.
- Community threads. Reddit, Quora, niche forums for the phrasing people use pre-purchase.
- Staff interviews. Ask frontline staff what the site doesn't answer; Simms makes the same point.
- Keyword data, used only to spot question shapes such as "best X for Y."
Jenny Munn's prompt research guide adds a useful lens: flag high-intent modifiers (constraints like "under a budget," fit like "for small teams," risk like "that isn't too complicated," urgency like "out of the box"). She offers AEO services, and her "choice stress test" is a heuristic, not a validated method.
Step 3: Expand with a constraint matrix
Take each decision and multiply it by persona, funnel stage and constraint. Illustrative prompts for a hypothetical online course provider:
| Persona | Stage | Constraint | Example prompt |
|---|---|---|---|
| Working professional | Compare | No coding background | "Best online data analytics course for working professionals with no coding?" |
| Final-year student | Evaluate | Budget | "Affordable certification programs for data analysts in India?" |
| Parent | Verify | Trust | "How do I check if an online degree program is recognised?" |
| Career switcher | Decide | Time | "Can I switch to product management in six months?" |
These are illustrative, not observed prompts. Replace them with real ones from step 2.
Step 4: Check the shape of the prompt
SeoNot's article on what replaces keyword research argues that question research replaces keyword research and that zero-volume questions are often the most valuable. It cites no data, and says nothing on how to monitor private AI conversations, so I treat it as a hypothesis. A more defensible test: does the AI answer have to name a brand or make a recommendation? If yes, it is decision-stage and worth tracking. If it can answer generically, it is informational.
Step 5: Score and prioritise
Use this sheet. Score each prompt 0-3 per column and sum.
| Column | 0 | 3 |
|---|---|---|
| Decision stage | Purely informational | Requires a recommendation |
| Business fit | We rarely serve this | Core offer |
| Evidence gap | We already own the answer | No page answers it |
| Can we credibly win? | Competitors dominate, we lack proof | We have first-party proof |
| Fan-out breadth | Narrow | Triggers many sub-questions |
Lumen suggests starting with roughly 100-200 candidates and trimming to prompts where you have "real credibility and a clear content opening." A final working set of 40-100 is workable for most small teams. Re-score quarterly.
Step 6: Test and log, with repeats
For each priority prompt:
- Run it at least three times on different days in each tool you care about.
- Record: date, tool, model if shown, logged-in state, location, whether you were cited, mentioned, or absent, and who was cited.
- Calculate visibility = runs where you appeared / total runs, per prompt cluster.
- Compare against two or three named competitors.
This follows the visibility-percentage approach from the SparkToro and Gumshoe work. For a deeper dive on share-of-voice maths see my definition post.
From prompts to content actions
Prompt research only pays off if it changes pages. Use this decision table.
| What you find | Action |
|---|---|
| You're absent, competitors cited | Add a direct answer section; check third-party sources cited |
| You're mentioned but facts are wrong | Fix your own page first, then external listings |
| Cited, but with old data | Update with dated facts; Lumen suggests refreshing content tied to high-value prompts every 60-90 days |
| Answer cites forums and review sites | Plan off-page work; see my off-page audit |
| Prompt is long and situational | Create situation-based content; see multi-part queries |
Seer Interactive's follow-up analysis is a good model of turning data into actions: it found Gemini searched for a company's awards and industries served, and without dedicated pages it used third-party directories. That analysis is one agency's data, one model and one week, but the lesson (give brand facts a home) is cheap to apply.
Google's guidance on AI features is a useful reality check: no extra requirements or special markup to appear in AI Overviews or AI Mode, just standard indexing, snippet eligibility and good content. Prompt research tells you what to write; it does not unlock a hidden ranking factor.
Tools: what to expect, honestly
- Semrush Prompt Research: estimated AI volume and example prompts across four platforms; method undisclosed.
- Ahrefs Brand Radar: I could not open a methodology page that explains prompt volume, so I can't characterise its accuracy. Ahrefs published a post titled "AI adjusted volume" on 17 August 2026 that I only saw as a title.
- DataForSEO: offers APIs to extract fan-out queries and enrich them with its own AI volume metric; its own article notes the metric is proprietary.
- Spreadsheet plus manual runs: slower, free, and good enough for a first pass. See my view on whether AI visibility tools are worth it.
Common mistakes
- Treating one run as a ranking.
- Tracking only prompts that contain your brand name.
- Copying prompts from a tool without checking real customer language.
- Optimising purely for informational prompts that never lead to a choice.
- Reporting estimated volumes as facts.
- Forgetting Search Console. See what Google's generative AI report shows for your own site.
FAQ
What is prompt research?
It is the process of finding, grouping and prioritising the questions and requests people give AI assistants about your category, so you can test whether your brand is cited and improve the content behind it.
Is prompt research replacing keyword research?
Not entirely. Keyword data still shows topics and question shapes, but it does not measure how often people ask AI tools, as Lumen's guide notes. Use both, and keep clear which numbers are estimates.
Where can I get real prompt data?
Your own sales calls, support logs, Search Console, community threads and staff knowledge are the most reliable. Tools such as Semrush estimate AI volume from a large prompt database, but the calculation is not documented.
How many prompts should I track?
Start with 100-200 candidates and cut to a working set of about 40-100 you can run regularly. Quality of selection matters more than count.
How often should I run each prompt?
At least three runs per prompt per tool before drawing conclusions, repeated monthly. SparkToro and Gumshoe found individual answers vary so much that single runs mislead.
Can I track my ranking position in ChatGPT?
You can record it, but the study found position is essentially random run to run. Track how often you appear across many runs.
Do I need paid tools?
No. A spreadsheet and disciplined manual runs can answer the first-order questions. Paid tools save time at scale but carry the vendor-method caveats above.
Does prompt research apply to B2B and edtech equally?
The method is the same: decisions, personas, constraints. B2B prompts often include integrations and pricing; edtech prompts include fees, recognition, outcomes and time commitments. Avoid unverifiable outcome claims in any answer content.
What should I do with prompts where I never appear?
Check whether a page answers the sub-question clearly, whether competitors are cited from third-party sources, and whether your facts are current. Then fix the gap and re-test later.
Work with me
If you'd like a prompt set built from your real customer language and a tracking sheet you can keep running, see my work and reach me through the contact form at younusfardeen.in. I bring 4+ years of marketing experience in SEO and AEO, and I'll be plain about what the data can and can't tell you.