How to Do Prompt Research for AI Search

A keyword tool reports what people typed into a search box in the past. It is a poor guide to how a buyer will word a question for an assistant next month. Fresher evidence sits inside your own walls, in recorded calls, the support inbox and public reviews, and the steps below turn it into roughly twenty tested prompts in a week.

Marketing leads and founders can run it for the price of their time and free access to the assistants buyers use. With no customer conversations on file, talk to five buyers first.

Step 1: Mine the conversations you already have

Start with the last ninety days of call notes or recordings, then the support inbox, then reviews of your company and of rivals. Capture each question in the buyer's own words and resist tidying. WRITER's AI visibility team begins the same way, recommending 20 to 30 raw buyer questions drawn from calls, tickets, win and loss interviews and similar sources, with no paraphrasing. That is a vendor's own account, so weigh it as such.

One row per question:
  quote exactly | channel it came from | date
  asker's job title | where they were in the purchase | did the deal close

You end up with raw questions labelled by role and stage, plus a view of the vocabulary buyers use for your category, which often differs from your own. Merge duplicates but keep their variants, mark questions from closed-won deals because they tie to revenue, note any that name a competitor, and remove personal or confidential details before the list leaves the call tool. Nothing here shows how common each question is across the market.

Step 2: Convert the questions into prompts and sort them

A prompt is a question phrased the way someone would type it into an assistant. Length matters, and the evidence points two ways. Pew Research Center found Google showed AI summaries for 8% of searches of one or two words and 53% of searches of ten words or more, favouring full questions. Semrush's clickstream study of ChatGPT in the second half of 2024 reported average prompts of 23 words without ChatGPT's search feature and 4.2 words with it. The study page gives no method, so read the gap as a caution that no single length suits every product.

Full sentence: "As a [job title] at a [kind of company], I need to [goal].
                 Which [category] tools are worth a look, and how do they
                 differ on [deciding factor]?"
Search-style:   "[category] for [use case]"

Aim for about twenty prompts in three to five themes. Tag each as branded, competitor-named or unbranded, and give the sensible ones both a long and a short form. At least half should be unbranded, since a prompt that names you flatters you; branded versus unbranded prompts explains the balance. Frame each theme as something your site should answer, cover early, middle and late stages, then lock the list and number the version. Gaetano DiNardi, quoted by AirOps, says the prompts that count most are the ones where someone is evaluating options, which argues for weighting those themes. Whether the prompts match real typing remains unproven at this point.

Step 3: Run everything and keep the receipts

Pick two assistants from those your buyers favour, for example ChatGPT and Perplexity, or Google AI Mode, and open clean sessions if you can. Every prompt gets three attempts, each answer pasted into the sheet, because answers vary and one pass misleads; why AI answers change every run covers the reasons.

Per run, record:
  prompt number | assistant | attempt 1-3 | date
  every brand mentioned | every URL cited
  were we named (Y/N) | were we cited (Y/N) | anything wrong about us

The output is mention and citation rates by theme, a catalogue of favoured sources and a list of factual errors about your brand. Sites cited repeatedly become off-site targets, errors get corrected at source where possible, and themes with no appearance mark content gaps. Citations also hint at what the system looked up, since Google says its AI features may use a query fan-out technique, issuing multiple related searches. What you will not learn is why you are missing, only where.

Step 4: Give every gap one assigned move

Each weak theme gets one action, chosen from three: a new page that opens with the answer, an upgrade to an existing page, or a mention on a source the assistants already cite. The seven-blocker audit helps with diagnosis. Book a re-run four weeks ahead.

Plan row: theme | gap found | chosen move (page, refresh or outreach)
          owner | due | date of re-run

Begin where won deals cluster, cap the work at three moves a month and re-run identical prompts. That gives a fair before and after, which shows whether presence moved without proving that an action caused it.

Reading the first results

Expect surprises. Prompts you assumed were about your product may return answers dominated by comparison sites, forums or analyst pages. Some answers will name a competitor you do not consider a rival. A few will describe your company with details that were true two years ago. Log each of these as a finding rather than an inconvenience, because they point at where the work is: correcting stale pages, writing the comparison page buyers expect, or earning coverage in the places assistants read.

Do not draw conclusions from a single theme. A cluster of three or four themes with the same pattern, such as a missing comparison page, is a stronger basis for action. Share the sheet with sales early. Reps recognise the questions at once and often add the ones you missed.

Our stance: keep short prompts in the set

Conventional advice says to write long conversational prompts because that is how people talk to assistants. The Pew figure backs that for Google summaries. The Semrush figure hints that when ChatGPT searches the web, the queries involved can be very short. A set built only from long questions would then measure a version of search your buyers may not use in every product.

We would therefore keep a few short, search-style prompts beside the long ones, label them and compare results by form. The price is extra runs and a bigger sheet, and the Semrush data has no stated method, so read it as direction. We would drop the short forms if your own analytics or buyer interviews showed that nobody in your market uses assistants that way.

The whole loop in one paragraph

Thirty verbatim questions go in; twenty prompts come out, grouped in three to five themes, half of them unbranded and some in short form. Each prompt runs three times in two assistants with every result logged. Three owned moves follow, and a re-run happens four weeks later. Renew the question list every quarter, run the set every week or two and review the plan monthly. For a longer version of the same idea, try the 30-day sales-call playbook, and see the GEO guide for context.

Where it breaks

Early-stage firms may hold only ten calls, in which case review sites and buyer interviews fill the gap. Personalised answers mean every run is merely a sample. Weekly prompt changes erase the trend. Names and company details should be removed from call notes before any of it enters a tool.