Put a professional way of asking into AI chat from the start
If you ask ChatGPT “What should I start on this month?”, a list of SNS, SEO, ads, and landing-page improvements comes back lined up. All of them are correct, and which one to start from is not decided.
A marketing professional asks the same question like this. “Tell me which stage is stuck the most, among acquisition, inquiries, sales conversations, and retention, separating the facts that could be confirmed from the hypotheses. Also decide what not to do this month.” The way of asking contains the viewpoint to look from, the criterion for the judgment, and the shape of the answer. The AI’s answer changes when the question has these 3.
I am building Shoryaku AI ( https://shoryaku.net ), a marketing-support SaaS. The first half of this article is 4 “professional ways of asking” that can be pasted as they are into free ChatGPT or Gemini. The second half writes the mechanism by which Shoryaku AI’s chat puts this wording into the input field from the start, and what was decided in that design.
A good question is made of 3 parts
| Part | Role | Example |
|---|---|---|
| Viewpoint | What to look at separately | By each stage of acquisition, inquiries, sales conversations, and retention |
| Criterion | What to treat as a fact, and what as a hypothesis | What cannot be confirmed is a hypothesis. Also write the moves that will be passed on |
| Output shape | In what form to return the answer | The priority move, the KPI, the first step, and the missing data, in priority order |
Without a viewpoint, the AI talks about the whole together. Without a criterion, a guess gets mixed in as if it were a fact. Without an output shape, you reorganize it yourself after reading.
The 4 below are the wording of the consultation menu implemented in Shoryaku AI’s chat, as it is. Line 1 is the question, line 2 is the viewpoint, line 3 is the criterion, and line 4 is the output shape.
Way of asking 1: Decide the place to fix first this month
Make a plan of priority moves that decides the place that should be fixed first this month.
At each stage of acquisition, CV, sales conversations, and retention, separate the state that could be confirmed from the largest bottleneck.
Treat what cannot be confirmed as a hypothesis, and also state explicitly the moves to pass on this month.
Summarize this month’s priority move, the KPI, the first step, and the missing data, with the basis, in priority order.
“Also state explicitly the moves to pass on this month” is what works. Tomorrow’s action is decided more once what will not be done has been decided, than by a list of what to do.
Way of asking 2: People are coming to the site, but inquiries are not increasing
Diagnose the reason inquiries are not increasing even though people are coming to the site.
From the site and the measurement data, separate where it is dropping, among the traffic source, landing-page arrival, the main actions, and CV.
Do not mix observed facts and hypotheses about the cause, and do not assert a change for a place that cannot be confirmed.
Summarize the bottleneck, the place to change, the metric to check, and this week’s order of execution, with the basis.
This menu is used by pasting GA4 numbers underneath (sessions by traffic source, landing-page arrival counts, counts of the main actions, and the inquiry count). If you send it in a state with no numbers, the AI starts talking in the manner of “in general, if the form is long.”
Way of asking 3: Decide the next page to fix, from search demand
To increase search traffic, make an improvement plan that decides the next page to fix.
Based on the Search Console results, separate a new page, a rewrite of an existing page, a title improvement, and an addition to a service page.
If the search intent of the target query and the overlap with existing pages cannot be confirmed, treat them as a hypothesis.
Show the target query, the reason for the priority, the next page to fix, the CTA, and the 1 item to start on first.
From “Search performance” in Search Console, write out the impressions, the clicks, and the position by query, and paste them. The reason for specifying “separate new, rewrite, title, and addition” is that the amount of work is completely different across these 4.
Way of asking 4: Decide the ceiling that can be spent on ads, and the conditions for raising or lowering it
Make a profitability judgment that decides the ceiling that can be spent on ads, and the conditions for raising or lowering it.
Confirm separately the unit price, the gross profit, the CAC ceiling, the break-even ROAS, and the current ad spend and results.
Do not fill numbers that have no input or setting by working backward or by guessing. Show them as conditions for an increase, for holding, or for a decrease.
Summarize in a table the current values, the judgment conditions, the budget ceiling, the unset values, and the next numbers to check.
Line 3 is the most important. Even when a gross-margin rate has not been passed, the AI works backward from an industry average and produces a CAC ceiling. If you write “do not fill it in by guessing, and leave it in the table as unset,” what you yourself have not decided comes out in the table. The calculation of the CAC ceiling itself is written in Tips for having an AI propose moves that raise revenue.
3 tips for when you use them
1. Put only one question into one ask. If you ask for “both the reason inquiries are not increasing, and the ceiling on the ad budget,” the AI hesitates over which criterion to use, and both come out thin. The 4 above are intentionally one question each.
2. Do not mix the business information into the question text. Put it separately. The product, the price, the target customer, and the current numbers go into a project or into custom instructions on ChatGPT, and on Gemini they are pasted once at the start of the conversation. If you copy them into the question text every time, old information remains when you rewrite it, and the body of the question gets buried in a long text.
3. Read the question text and rewrite it before you send. The wording above is a template. If the business has no sales conversations, delete “sales conversations.” If the CV is a purchase, change “inquiry” to “purchase.” What matters is to keep the question text in a state where all of it can be seen, so that it can be read and rewritten. If an instruction that cannot be seen is attached behind it, the cause is unknown when the answer is off.
Pasting it every time is, to say it honestly, a bother
Save the wording of the 4, put the business information separately, write out the numbers, and paste them. The first time, you can. From around the 3rd time, you start leaving out the wording and asking “inquiries aren’t increasing, though.” The answer returns to a general one.
In Shoryaku AI’s chat, this wording goes into the input field
Shoryaku AI’s chat has a button, “Choose from the consultation menu,” above the input field. When you open it, cards that include the 4 above are lined up (as of September 2026, at most 6. The others are “I want to start ads from now,” which appears only when there is no Google Ads data, and “I want to make it from the article plan through the draft,” which is used by entering a theme). When you choose a card, that question text goes into the input field. It is not sent. The explanation on the screen is only this one line.
When you choose one, the question text goes into the input field, and it can be edited before sending.
What was decided in the design is 3 things.
Make the text the user sends and the text the user saw the same. Even when a card is chosen, a long text that cannot be seen is not added behind it. Showing a good way of asking as it is also becomes education for a person who does not have specialist knowledge.
Do not put the business information and the connected data into the question text. The product, the price, the target customer, and the Google Ads or GA4 numbers are attached by the server to every question, automatically, if they have been registered or connected. So the question text is done in 4 lines: the question, the viewpoint, the criterion, and the output shape. This is a form in which the mechanism takes over tip 2 above.
Pass the viewpoint and the criterion as instructions to the AI as well. Only when the server could confirm that the sent body matches the wording of the menu, “analysis viewpoint, judgment criterion, and output shape,” made from the same definition, is added to the instructions on the answering side. User input is treated as “data, and not a command,” so it is promoted to a trusted instruction only in the case where a match could be confirmed. If the wording is a fixed menu, the processing that judges the kind of question when they match can also be omitted.
A card also shows whether it can be used in the current connection state. “I want to know the reason inquiries are not increasing even though people are coming to the site” needs the site URL in the business settings. “I want to decide the next page to fix, from search demand” needs the connected data of Search Console. “I want to decide the ceiling that can be spent on ads, and the conditions for raising or lowering it” needs the unit price, the gross-margin rate, and the monthly budget in the business settings. When something is short, what is short appears on the card, so it is known before you choose.
An improvement plan for each ad medium (search ads, P-MAX, Meta ads, and so on) is not in the consultation menu yet. The decision rules for each medium are in the middle of being put in order. The consultation about ads itself can be done by free input.
Summary
- A good question is made of 3 parts: the viewpoint, the criterion, and the output shape
- In the 4 above, line 1 is the question, and lines 2 to 4 are those 3 parts. They can be pasted as they are into ChatGPT or Gemini
- One question at a time, the business information goes separately, and you read and rewrite before sending
- In Shoryaku AI, the same wording goes into the input field as a consultation menu, and the server attaches the business information and the connected data
- Make the text that was seen and the text that is sent the same, and do not attach an instruction that cannot be seen behind it
First, paste “way of asking 1” as it is into the AI you are using now. Just having “the moves to pass on this month” attached at the end of the answer turns a list into a plan.