Building a client prompt set from ranking keywords

Every new engagement starts with the same blank page: which prompts do we track? Ask the client and you get their product names, which is the one set of prompts they already win. Ask a keyword tool and you get terms people type, which do not always look like what people ask an AI assistant. Both are incomplete starting points.

The most reliable shortcut we have found is to start from what the client's site already ranks for in Google, then translate and prune. It is not the final prompt set, but it gets you to a credible first version in an afternoon instead of a week.

Why ranking keywords are a good seed

A keyword the site already ranks for is a topic where the domain has some demonstrated relevance. It is also a topic someone actually searches. Using those as seeds avoids two classic mistakes: tracking prompts nobody asks, and tracking only the brand terms the client already owns.

Competitor keywords are just as useful. If a rival ranks for a topic the client does not, that topic shows up as a candidate prompt, and the first tracked answers will show whether AI engines treat the gap the same way Google does.

One honest limit. Google rankings and AI answers do not map one to one. A keyword the client ranks first for may produce an AI answer that names three other brands. That is exactly why it is worth tracking, but we never present a keyword rank as a prediction of an AI answer.

Turning keywords into prompts by hand

The translation step is where the work is. Keywords are short and fragmentary. Prompts are questions, comparisons, and requests. We apply four moves.

Add the intent. "crm for dentists" becomes "what is the best CRM for a small dental practice" and "how does X compare with Y for dental clinics." One keyword often yields two or three prompts with different intent: discovery, comparison, and decision.

Add the buyer. If the client sells to different roles, we write the prompt as each role would ask it. That feeds persona work later.

Add the market. Language and location change AI answers. We decide which country or city each prompt belongs to, and we keep markets in separate monitors so results do not mix.

Cut the rest. We drop navigational terms, near duplicates, and anything where a good answer would not name a brand at all. A tracked prompt should be one where being named is worth something.

For a first set we aim for a manageable number, enough to cover the main topics and intents without becoming a pile nobody reads. Quality of coverage matters more than count.

The tooling shortcut

Doing the lookup by hand means pulling ranking data from a keyword tool, exporting it, and cleaning it. The translation step stays manual in any case, but the lookup does not need to be.

According to the Promptwatch changelog, Prompt Explorer gained a Ranking keywords tab in September 2026. You enter a page or domain, Promptwatch looks up the Google keywords it ranks for, and turns them into prompt suggestions you can add to a monitor. It sits beside the other sources in Prompt Explorer: Search Console, People Also Ask, seed queries, and website pages. Because it accepts any domain, we also run it on a competitor's domain to see which topics they cover that the client does not. The changelog says "a page or domain," so we do not assume it respects restrictions on competitor data beyond that.

A second, smaller change helps the first week. Since September 21, creating a monitor suggests a set of starting prompts before you save, so a new monitor starts collecting answers instead of sitting empty. We treat those as scaffolding. They are a prompt to think, not a prompt set.

Our workflow for a new client

1. Pull the client's own ranking keywords by page and group them by topic.
2. Pull one or two competitors' ranking keywords and mark topics where the client has no coverage.
3. Review the suggestions with the client. We cut anything off-strategy and add the questions their sales team hears.
4. Rewrite for intent and role. Each kept keyword becomes one to three prompts.
5. Tag and topic them. Organising prompts into topics and tags from the start makes dashboards readable later.
6. Split by market. A monitor's language and location are fixed once it is created, per the changelog, so we plan the monitors before adding prompts.
7. Freeze the version. We record the date and the prompt count, so later changes show as changes.

That last step is the one we see skipped most. A prompt set that quietly grows invalidates trend lines, so we version it like any other asset.

Where Promptwatch fits

Promptwatch is the platform we run client programs on, and the Ranking keywords tab is the reason the first version of a prompt set no longer takes a week. The lookup, the suggestion, and the move into a monitor happen in one place. The judgement on what a client's buyers actually ask stays with us and the client, and we do not hand that to a generated list. Details of plans and limits are at promptwatch.com. Prompt counts differ by plan, with 50 on Essential, 150 on Professional, and 350 on Business, while every agency plan has unlimited prompts, so the plan shapes how wide you can start.

What to expect from the first month

Expect to rewrite a good share of the prompts after the first two weeks of data. Some will return answers with no brands at all. Some will turn out to be near duplicates. A few buyer questions will appear in sales calls that your set missed. That is normal. The aim of the first version is not to be right. It is to be good enough to learn from, and documented well enough that you know what changed when you change it.