How to Do Prompt Research (3 Methods)

How do you do prompt research? Start by finding prompts with real demand that your content does not cover, using a prompt discovery tool. Then measure the prompts you already appear for, using Google Search Console's generative AI performance report and Bing Webmaster Tools' AI Performance report. Prioritise by demand, publish against the gaps, then re-measure.

What is prompt research?

Prompt research is the practice of finding which prompts people ask AI engines, checking whether your content shows up in the answers, and deciding which ones are worth going after.

It is the AI-era counterpart to keyword research. Instead of a list of search terms with volumes, you end up with a list of conversational questions, an idea of how often each gets asked, and a view of whether you or a competitor gets cited.

One thing to clear up first, because the phrase collides with something else. Prompt engineering is the craft of writing better instructions to get better output from a model. Prompt research is a marketing job: working out what your customers ask, and whether your brand is in the answer. If you came here for the engineering kind, this is not that guide.

For the tooling side of this, see our prompt research solution page.

How is prompt research different from keyword research?

Keywords are what people type into a search box. Prompts are what people say to an assistant, and they are much longer.

Semrush analysed over 80 million lines of clickstream data and found ChatGPT prompts averaged 23 words when users were not in search mode, against 4.2 words when they were. That gap changes the whole exercise.

The practical consequence is that prompts sit upstream of keywords. Optimise a page for "best running shoes" and you can still be invisible for "which running shoes reduce knee pain if I have flat feet," even though the second person is the easier sale. The constraints in the prompt are the part your content has to match.

Keyword research is not obsolete. It still tells you how your market phrases things, and that phrasing is a good starting point for prompt discovery. It just stopped being the whole picture.

The two halves of prompt research

Every prompt research method answers one of two questions, and it helps to know which before you open a tool.

Discovery finds prompts you do not appear for. Measurement tells you how you are doing on prompts where you already show up. Most guides blur these together, which is why people open a reporting tool expecting to find new opportunities and come away frustrated.

Only the first method can show you something you are missing. The other three are records of your existing footprint, which makes them excellent for improvement and useless for discovery. You need both halves.

Method 1: Find the prompts you do not show up for yet

Start here, because this is the only step that can surface demand you have no idea exists.

A prompt discovery tool works from prompts people actually ask, rather than from your analytics. In OmniSEO's Prompt Explorer, the workflow runs:

1. Run your own domain. You see the prompts where you already surface, usually across more topics than expected.
2. Run a close competitor's domain. More useful than your own, because it exposes the ground they cover and you do not.
3. Run the topic, not a domain. Something like "tile flooring DIY" shows how real people phrase questions in that space, rather than how your team phrases them.
4. Read the content gap flag on each row. It tells you whether you currently cover that prompt.
5. Sort by prompt query volume. PQV estimates how often a prompt gets asked across AI engines, so high demand plus no coverage is your shortlist.

Open a prompt from the list and you can see the answer an engine returns and the sources it cites. Those cited pages are the most specific content brief you will get, because a competitor already wrote it for you.

Worth being straight about the boundary: the 7-day free trial gives you a snapshot of today with no setup or connections, and tracking those prompts over time across ten engines is a paid feature. If you only want to sanity-check one prompt first, our AI Prompt Volume Checker is free.

Method 2: Use Search Console's generative AI performance report

Google's report tells you which of your pages appear in its AI features. It does not tell you what anyone asked, and that limit is the most important thing to know before you open it.

Google launched the Search Generative AI performance reports on 3 June 2026. Per Google's own documentation, here is exactly what you get:

• Coverage: AI Overviews and AI Mode. Discover has a separate report, and Search Labs experiments are excluded.
• Metric: impressions, meaning how many times links to your site were shown inside a generative AI feature. There is no clicks metric.
• Dimensions: Dates, Pages, Countries, Devices. That is the complete list.
• No query dimension. You cannot see which searches or prompts triggered the impression.
• Availability: rolling out to a subset of properties, so it may not be in your account yet.

So how do you use it for prompt research? Two ways.

Confirm which pages earn AI visibility. Sort pages by impressions and you learn which of your content Google's AI features actually reach for. That tells you what format and depth is working, which you can copy on the gap pages from method 1.

Get query-shaped signal from the regular report instead. Your standard performance report still has queries. Filter it to phrases containing what, why, how, best, or versus, and you get a conversational subset of your real Google demand. It is not prompt data, but it is a reasonable proxy and it is free.

Method 3: Use Bing Webmaster Tools and Microsoft Clarity

Microsoft gives you the closest thing to query-level AI data, plus the only view of what AI traffic does after it lands. Two tools, two jobs.

Bing Webmaster Tools AI Performance

Bing's AI Performance report covers Microsoft Copilot, AI-generated summaries in Bing, and select partner integrations. It went into public preview in February 2026 with:

• Total Citations, how often your content appeared as a source in AI answers.
• Average Cited Pages, the average number of unique pages cited per day.
• Grounding queries, the key phrases the AI used when retrieving the content it cited.
• Page-level citation activity and a visibility trend over time.

In June 2026 Bing added four more, all in preview: Intents classifies grounding queries as Informational, Commercial, Navigational, Research and more. Topics clusters related queries into themes. Citation Share shows what percentage of citations for a given grounding query went to your site. Compare tracks change over time.

Here is the part almost everyone gets wrong. A grounding query is not a user prompt. It is the phrase the AI searched in order to find sources, which is downstream of whatever the person actually typed. Bing also says the grounding query data is a sample, and that Citation Share is observational rather than a ranking, does not expose competitor domains, and carries no quality score.

Read that way, it is genuinely useful: it tells you what retrieval phrasing your pages win on, and Intents tells you whether that is commercial or informational. Just do not mistake it for a list of your customers' prompts.

Microsoft Clarity

Clarity answers what the other tools cannot: what people actually do once they arrive from an AI assistant.

It has two AI channel groups. AIPlatform captures sessions from an organic link inside an AI chat, and PaidAIPlatform captures clicks on paid ads inside one. Supported sources are ChatGPT, Copilot, Gemini, Claude and Perplexity.

Find both under the Referrer card, on the Channels tab.

From there you can filter to AI traffic, compare its behaviour against other channels, attribute conversions, and watch session recordings for that segment specifically.

One limitation to know: Clarity counts direct use of those platforms on their own sites, like chatgpt.com and copilot.microsoft.com. AI features embedded in other products, including search engines, are not included. So your AI Overview traffic is not in this number.

The reason it is worth the effort: in Clarity's study of more than 1,200 publisher and news sites, AI-driven traffic grew 155.6% over eight months while search grew 24.0%, and sign-up click-through from LLM referrals ran at 1.66% against 0.15% from search. The honest caveat is in the same study: AI referrals were still under 1% of overall traffic, and the sample skews toward publishers, so treat the conversion gap as directional rather than a forecast for your B2B funnel.

For the GA4 version of this measurement, see our guide to tracking AI referral traffic.

Put the three together into one workflow

Run discovery monthly and measurement fortnightly. That cadence is enough to catch movement without drowning in noise.

1. Discover. Run your domain, two competitors, and your two biggest topics. Shortlist gaps with high PQV.
2. Publish or update against the top three gaps, matching the structure of pages already cited for those prompts.
3. Confirm in Search Console that the new pages start picking up impressions in AI Overviews and AI Mode.
4. Check Bing for new grounding queries pointing at those pages, and whether Citation Share is moving.
5. Check Clarity for AI sessions and conversions on them, which is the only step that connects any of this to revenue.
6. Re-run discovery and repeat.

If you want the monitoring half automated rather than manual, that is what prompt tracking does.

What prompt research still cannot tell you

No tool available today shows you a log of the literal prompts your customers typed. It is worth saying plainly, because a lot of writing on this topic implies otherwise.

What you actually get is a set of good approximations. Bing shows sampled retrieval queries, Google shows pages without queries, and Clarity shows sessions from standalone AI sites only.

Discovery tools estimate volume rather than measuring it directly, the same way keyword volume has always been estimated.

That is enough to work with, and it is a long way from the "we are effectively blind" position that was accurate eighteen months ago. Just build your plan on ranges and directions, not on decimal places.

Your first prompt research session

Here is the homework, and it takes about forty minutes.

Run your own domain and your closest competitor through a discovery tool, and write down the three highest-volume prompts where they appear and you do not. Then open Search Console's generative AI report and Bing's AI Performance report, and note which of your existing pages already earn AI visibility. The overlap between those two lists tells you what kind of content works for you, and where to point it next.

So which is your team actually short on right now: prompts you have never seen, or pages that deserve to be cited and are not?