Discussion starter: an AI search question, posted with an answer from Niraj Raut to open the thread. If you have dealt with this on a site, reply with what you saw, especially where it differs.
- AI Search
No platform shares prompt volumes. How are you choosing which prompts to track: search query data, sales calls, customer interviews, or guesswork?
Disclosure: Niraj Raut, who posted this answer, runs the SEO consultancy linked at the end of it.
1 reply
-
adminAdmin
All four, but each for a different job, and guesswork only when it is labelled as a hypothesis. Sales calls and customer interviews tell you which questions decide a purchase and how buyers phrase them. Search query data tells you which topics have demonstrated demand in your market. Neither gives you prompt volume, so the aim shifts from “track the most-asked prompts” to “track the questions that shape buying decisions, in the language buyers use, and keep that set stable enough to trend”.
The biggest practical change is the unit of tracking. Track intents, each with several phrasings drawn from real sources, rather than single prompts you hope match what people type.
Stop trying to rebuild keyword volume for prompts
Keyword research trained us to rank targets by volume. That model breaks for AI answers for two reasons.
First, prompts are far less standardised than queries. In the SparkToro and Gumshoe research, 142 human-written prompts for the same headphone-buying need had a semantic similarity score of just 0.081, and the authors noted there were barely two that looked alike. Even perfect volume data for one exact prompt would describe a sliver of real behaviour.
Second, most AI use is not brand discovery. OpenAI’s September 2025 analysis of 1.5 million conversations found 49% of messages were “asking”, 40% “doing” and 11% “expressing”, and that practical guidance, seeking information and writing account for about three-quarters of conversations. The prompts where your brand could appear are a small commercial slice of that, so a volume-weighted prompt set would mostly track things that do not matter to you.
The better model is closer to message testing in market research: define the decisions buyers make, collect the language they use, and sample enough phrasings to estimate how often you appear when that decision is being researched.
What each source is good for, and how it misleads
Source Best used for Built-in bias Search Console queries (long, question-style, “best”, “vs”, “for [use case]”) Topics with proven demand, local terminology, modifiers Only shows queries where you already earned impressions, so it under-represents exactly the gaps you want to find. Search queries are also shorter than prompts Paid search terms and keyword tools Competitor and category queries you do not rank for Still query-shaped, not conversation-shaped Sales call transcripts and CRM notes Shortlist questions, objections, comparisons, deal-breakers Skews to late-stage buyers who reached sales, and to segments sales talks to most Customer interviews and onboarding surveys Direct evidence of whether and how buyers used AI tools, in their words Small samples and recall errors. People reconstruct what they asked Support tickets and site search Capability and policy questions, which make good accuracy checks Post-purchase or on-site, so weaker for discovery prompts Community threads and reviews Unprompted buyer phrasing and comparison framing Loud minorities and niche users Team hypotheses and LLM-generated variants Filling phrasing gaps, stress-testing wording Self-referential. Generated prompts tend to sound like the model, not like buyers The pattern in that table matters more than any single row: first-party qualitative sources give you the right questions, and search data tells you which of those questions have scale in your market. Neither alone is enough.
The partial volume signals that do exist
No AI platform publishes prompt volumes, but a few signals give you relative sizing:
- Panel-based estimates. Some vendors sell prompt volume data. Profound, for example, says it licenses conversations from multiple double-opt-in consumer panels and applies statistical modelling to correct demographic and geographic bias, with Australia among its covered markets. Treat this like clickstream-based keyword volume: useful for comparing topics, weak for absolute counts, and dependent on who joins the panel.
- Bing grounding queries. Bing Webmaster Tools’ AI Performance report shows a sample of grounding queries, which are the searches the AI system ran to retrieve sources, not the user’s prompt. They show what the retrieval layer looks for when your pages get cited.
- AI referral landing pages. GA4’s AI Assistant channel group shows which of your pages already receive assistant referrals. The topics behind those pages are evidence of real AI-driven interest.
- Search Console’s generative AI reports show AI Overview and AI Mode impressions by page, but no queries. Pages with high AI impressions tell you which topics Google’s AI features already surface you for.
How I would build the prompt set
- Map the decisions, not the keywords. List the intents a buyer moves through: problem framing, category discovery, alternatives to a named competitor, fit for a specific use case, pricing and switching, and brand validation (“is X good for Y”).
- Harvest real phrasing for each intent from at least two independent sources, such as a sales transcript and a Search Console query, or an interview and a community thread.
- Score each intent on business value (deal size or margin), evidence of demand (how many independent sources it appears in), likelihood of AI use (multi-constraint research questions go to AI more than single-fact lookups, in my judgement), and whether better content or third-party coverage could plausibly move the result.
- Write three to five phrasing variants per intent, keeping real buyer wording, constraints and context. A prompt like “payroll software” is a keyword. “Payroll software for a café with casual staff that handles STP” is how people talk to an assistant.
- Split branded from unbranded. Brand-validation prompts almost always mention you, so they inflate any combined rate. Report them separately.
- Version the set. Keep a stable core that changes only with a documented version bump, plus a smaller rotating exploratory set for new hypotheses.
Hypothetical example:
A payroll software company selling to Australian small businesses reviews 30 sales calls and finds the same three questions keep deciding deals: handling casual and award-based staff, switching providers mid-financial-year, and Single Touch Payroll reporting. Search Console confirms steady impressions for switching and STP queries, but shows almost nothing for hospitality-specific terms, where the site has no content. Onboarding survey answers show several customers asked ChatGPT to compare payroll tools for cafés. The core set becomes around 12 intents with four phrasings each, weighted towards the hospitality and switching intents, because those carry the highest deal value and appear in three sources. Generic “best payroll software” prompts stay in the set but get less weight.
Cost check:
Variants multiply fast. 50 intents with 4 phrasings across 3 platforms, run 3 times weekly, is 1,800 answers a week. Start with fewer intents and more phrasings rather than the reverse, because buyers phrase the same need in very different ways and one phrasing can misrepresent the whole intent.
Selection bias is the risk nobody reports
Whoever chooses the prompts controls the result. A set built from pages you already rank for, or from phrasing your marketing team uses, will make visibility look better than it is. A set built from competitor positioning will do the opposite. Two safeguards help: document the source of every prompt in the set, and include intents where you expect to be weak. If your tracked visibility is high everywhere, the prompt set is probably describing your content, not your market.
A 60-day check that your set reflects real buyers
Add one open question to onboarding or post-purchase surveys: “If you used an AI assistant while researching, what did you ask it?” After 60 days, map each answer to your intent list.
- Metric: coverage rate, the share of reported questions that map to a tracked intent
- Target: 70% or more of reported questions map to an existing intent (a working threshold, not a benchmark). Well below that, the set reflects your assumptions more than your buyers
- Secondary signal: the share of respondents who used AI at all, which tells you how much tracking budget the channel deserves
- Next step: add unmatched intents that recur to the exploratory set, and promote them to the core only at a planned version change
What would change the approach: if few customers report using AI tools, shrink the prompt set and put the effort into accuracy checks on brand questions. If a panel source shows large demand for a topic none of your first-party sources mention, treat it as a hypothesis worth a few exploratory prompts, not an instruction.
My position: use sales calls and interviews to choose the questions, search data to size and phrase them, panel estimates for relative weighting where available, and guesswork only in a labelled exploratory tier. Then document every source, because the prompt list is the method, and the method is what makes the numbers defensible.
Need help with this on your own site?
Niraj Raut works with businesses in Nepal, Australia, the UK, Europe and the US on technical SEO, ecommerce SEO, local SEO and AI search.
Add a reply Cancel reply
Replies are open to approved members. Log in or apply to join.