Your Keyword Map Needs a Prompt Column. I Built One for Three Clients.

Keywords and prompts belong in the same spreadsheet. I tested the spreadsheet.

The planning model going around is simple enough to fit on an index card. Take your keywords. Rewrite each one as the full question a person would ask an assistant. Tag it top, middle, or bottom of funnel. Run it through ChatGPT and Perplexity. If no business gets named in the answer, drop it. Map what survives to the page that should own it.

The complaint going around is also simple. Teams are doing this in spreadsheets, by hand, and redoing it every month because the answers drift. Agencies with twenty clients cannot keep up. (I have an agency with more than twenty clients. I read that sentence twice.)

So I built the column for three clients on a Tuesday, a wrap shop, a dentist, and a kitchen remodeler, and I am going to show you the rows, because the rows are more useful than the model.

The prompts do not come from Search Console. Not for a local business.

The advice says Search Console is becoming a prompt source, because long conversational queries show up there now. For a big publisher, maybe. I checked five local client sites, 2,500 query rows across one 28-day window. Question-shaped queries: 75 rows, three percent. They carried 547 impressions out of 60,297, under one percent, and three clicks. Search Console is not going to hand a plumber a prompt list. It is going to hand him seventy-five rows of nearly the same thing.

And that is the useful part. The rows cluster. Thirteen phrasings of can you wrap a car with dents or scratches, 209 impressions, one click, position 8.4. Twenty phrasings of how long do windows last for the window company, 140 impressions, zero clicks, average position 81, which is a polite way of saying nobody is home. Two questions, thirty-three rows. Hundreds of prompts, a handful of intents. That is the whole clustering lesson and it was sitting in the export.

The real prompt source is the question report on the keyword itself, plus the phone. For kitchen remodel, the question universe leads with how much does a kitchen remodel cost at 4,400 searches a month, then how long does a kitchen remodel take at 720 with a difficulty of 21, then do you need a permit to remodel a kitchen at 320 with a difficulty of zero. For car wrap: how much to wrap a car at 9,900, how long do car wraps last at 1,900 with a difficulty of 10, how long does it take to wrap a car at 1,000 with a difficulty of two. Those are prompts. People already say them out loud to the front desk.

Eight prompts, two engines, and a filter that turned out to be real

I wrote eight prompts across the three clients, tagged them by funnel stage, and ran each through ChatGPT search and Perplexity. Google showed an AI Overview for six of the eight, so I read those too. Then I checked two things per answer: did any business get named, and was my client one of them.

Bottom of funnel, who does tuxedo wraps in dallas. The AI Overview cited the wrap shop. ChatGPT named it first. Perplexity called it the clearest match and pulled six of its twenty sources from the shop's own site. Owned. Bottom of funnel, best kitchen remodeler in lincoln ne. ChatGPT named six remodelers, all lifted from Angi. Perplexity named four others and cited the client's site without naming the client. Target. Bottom of funnel, best dentist in miami for teeth whitening. ChatGPT named six practices off directories. Perplexity named three. The client was cited, not named. Target.

Middle of funnel, how long does a kitchen remodel take in lincoln ne. The AI Overview cited the remodeler. Perplexity's answer opened with about 3 to 7 days on site if you are using the fast-turnaround model offered by some local remodelers. That is the client's entire pitch, quoted back, with the client's name removed. (Crazy to get plagiarized by a search engine and be grateful, I know.) The two how much prompts, teeth whitening in Miami and a kitchen in Utah County, produced price ranges. ChatGPT named two whitening studios on one of them. Perplexity named nobody on either.

Top of funnel, should I wrap my car or repaint it, and is a tuxedo wrap worth it on a truck. Advice, tables, tradeoffs, and not one business named by either engine. Perplexity did cite the wrap shop as a source for the tuxedo question, which is a different prize and I will come back to it.

The tally. Three bottom-of-funnel prompts, six answers, six of them named businesses. Five other prompts, ten answers, one of them named a business. The brand-mention filter is not a theory. It is a wall, and it sits exactly at the word who.

The column, and what each row tells you to build

The map has seven columns: keyword, prompt, stage, businesses named in the answer, client named, client cited, page that should own it. Four rows from mine, in plain language.

Tuxedo wrap dallas. Prompt: who does tuxedo wraps in dallas. Stage: bottom. Named: yes. Client named: yes. Client cited: yes. Owner page: the tuxedo service page that states the price. Verdict: defend it, do not touch it.

Kitchen remodel lincoln. Prompt: best kitchen remodeler in lincoln ne. Stage: bottom. Named: yes, six of them, all from a directory. Client named: no. Client cited: once. Owner page: the Lincoln location page, plus the directory profiles the engines actually read. Verdict: the name is the target, and the name lives on Angi and the BBB, not on the site.

Kitchen remodel timeline lincoln. Prompt: how long does a kitchen remodel take in lincoln ne. Stage: middle. Named: no. Client cited: yes, by the Overview and by Perplexity, which quoted the pitch. Owner page: the process page with the day count. Verdict: the citation is won. The name is not attached to it. Put the name in the sentence that gets quoted.

Wrap or repaint. Prompt: should I wrap my car or repaint it. Stage: top. Named: no, by anyone. Client cited: no. Owner page: none yet. Verdict: deprioritize for brand goals. Write it only if you want to be the source, not the recommendation.

Cluster, do not count. Then re-run it monthly and stop before you enjoy it.

Thirteen phrasings of the dent question is one page. Twenty phrasings of the window question is one page. Eight prompts across three clients collapsed into four intents: who is best, what does it cost, how long does it take, should I do it at all. That is the map. It is not four hundred rows. Anyone selling you four hundred rows is selling you the spreadsheet, not the answer.

On the three who prompts, the two engines agreed on one name for the wrap shop, two names for the remodelers, and zero names for the dentists. Same afternoon, same question, different shortlist. So the monthly re-run is real, not paranoia. Keep it to ten or twenty prompts per client, run it on the first of the month, and only move a row when the named list changes. The rest is decoration, and I have a whole other piece about decoration.

The clustering step is the one worth paying for

The manual part of this is not the prompts. It is the grouping, and the monthly grind of checking whether each page still covers the intent it was assigned. In the Semrush Content Toolkit, Topic Finder does the collapse for you: it takes a seed like kitchen remodel and returns the topics people actually search around it, with demand attached, which is the prompt column and the stage column generated instead of typed. SEO Writing Assistant then checks the draft against the topic it was assigned, so the page written for the timeline question actually answers the timeline question instead of drifting into the cost guide everyone else wrote.

The question numbers above came from the Questions filter in the keyword research side of the same platform, and the monthly re-run, whether the who prompts still name you, is what the AI Visibility features in Semrush One track without me copying answers into a sheet at 6 a.m.

The map is fine. It was missing a column.

The sub-searches an engine runs before it answers, and why most sites cover one of them, is in the fan-out field guide. What the same engines did with seven ranking pages last week is in the citation test.

Write the question the way the customer says it. Tag the stage. Ask two engines who they name. If the answer is nobody, it is not a brand prompt, it is a homework prompt. If the answer is six competitors from a directory, the work is on the directory. If the answer is your own pitch with your name missing, put your name in the sentence. One column. Four intents. A Tuesday.