How to ask AI for revenue ideas, and how Shoryaku AI does it
Ask ChatGPT or Gemini “what should we do to raise revenue?” and you usually get a correct answer. Keep posting on social media. Write more SEO articles. Try ads. None of that is wrong.
It is not wrong, and it still does not decide what to do tomorrow.
The reason is clear. The question is missing three things. There is no unit economics. You did not pass the current numbers. You did not make it write what the claim is based on. Add those three, and the answer that comes back changes a lot.
I am building Shoryaku AI ( https://shoryaku.net ), a marketing-support SaaS. The first half of this article is five practices you can use as they are in free ChatGPT or Gemini. The second half is what Shoryaku AI turned into a system so you do not have to do it every time.
Practice 1: pass “how much you can spend to acquire one customer” first
This is the one that matters most.
When you ask AI about acquisition, the budget talk usually ends at “up to ◯ × ¥10,000 a month.” What decides whether you can carry it out is not the monthly budget. It is how much you can spend to acquire one customer. Call that the CAC cap.
The calculation is not hard.
LTV = revenue per customer × gross margin × number of repeats
CAC cap = LTV ÷ 3
Even for the same “¥4,980 product,” this number changes by a factor of two or more.
| A ¥4,980-a-month service | A ¥4,980 one-time service | |
|---|---|---|
| Gross margin | 85% | 70% |
| Repeats | 16 months | 8 times |
| LTV | about ¥67,700 | about ¥27,900 |
| CAC cap | about ¥22,600 | about ¥9,300 |
The business on top can still recover if it spends ¥20,000 to acquire one customer. The business on the bottom loses money if it does the same thing. If you ask without saying which one you are, AI recommends the same ad action to both.
If you do not know the number of repeats, a rough figure is fine. Passing it and saying it is an estimate raises accuracy more reliably than not passing it.
The numbers for our business are as follows.
- Product: (product name)
- Price: (revenue per customer, and whether it is monthly or one-time)
- Gross margin: (if you do not know it, write that it is an estimate)
- Repeats: (months retained, or visits per year × number of years)
From these numbers, calculate LTV and the CAC cap (LTV ÷ 3).
Then propose actions. For any acquisition action that costs money,
always write "what it is expected to cost to acquire one customer"
and "whether it fits inside the CAC cap."
If it does not fit, do not recommend it. Write the reason.
If you are considering ads, also pass the cost per click you looked up in Keyword Planner, and it gets more specific. Cost per click ÷ (the share that reaches a signup) is what it costs to acquire one customer. At a cost per click of ¥420 and a signup rate of 0.45%, that is about ¥93,000 per customer. That is 4 times the CAC cap of ¥22,600 in the table above, so you can already see it cannot be adopted.
Practice 2: always make it line up another option for the same budget
AI leans toward agreeing with the action you asked about. Ask “should we run ads?” and the answer assumes you will.
The fast way to stop that is to make it write the option it did not adopt.
When you list an action that costs money, line up one free or low-cost
alternative that can be done with the same budget and the same effort,
and write why you choose this action instead of that one.
When “¥50,000 a month on ads” sits next to “4 pieces of content for the same ¥50,000 worth of time,” you can choose for the first time. A proposal that shows only one side has not been compared, so you cannot decide to adopt it.
Practice 3: make it split “does it disappear when you stop, or does it stay?”
Do not put an SEO article and an ad on the same “do this month” list. One starts working after 3 months and remains if you stop. The other works the next day and disappears if you stop. Put things of different kinds on the same ranked list, and the short-term one always comes out on top.
Attach these two to each action.
- Kind: cumulative asset (the effect remains if you stop) / flow (it disappears if you stop)
- Time until it works: immediate / within 1 month / 3 months or more
Then list, separately, what to do this week and what to set up for 3 months from now.
Once you require this, the ranking “it is free and takes 30 minutes, so it is the top priority” disappears. How easy it is to start is not a reason for priority.
Practice 4: make it label the basis
AI’s answers come out in the same confident voice whether the number was actually seen, whether it is a general rule, or whether it is a guess. You cannot tell them apart by looking, so make it write the distinction.
Label each piece of basis for a proposal with one of the following.
- Measured: based on numbers I passed
- Registered information: based on the business I described
- General principle: a general marketing principle
- Hypothesis: an inference with no support
- Insufficient data: the numbers needed to judge are not on hand
If the only basis for an action is "hypothesis," write it as something to test, not as something to carry out.
When you try this, there are more “hypothesis” labels than you expected. Knowing that is itself useful. Before you spend time on a proposal that is all hypothesis, you can go get the missing numbers first.
Practice 5: leave room to say “we cannot judge yet”
AI answers whenever it is asked. Show it an ad with only 30 clicks and ask “should this be stopped?” and it will write reasons to stop. With 30 clicks, you cannot tell whether it is good or bad.
If you hand it four answer types, this problem mostly goes away.
Do not turn an action into something to carry out when the data needed to judge it is missing.
Split each action into these four.
- Carry out: it can be judged from the numbers that were passed
- Check: a fact in the admin screen has to be confirmed first
- Watch: there is not enough data. Collect numbers first
- Skip: it conflicts with the business direction or the current state
Once “watch” and “check” are available, the AI no longer has to force a conclusion. I think this is the instruction that raises the quality of the answer the most.
Doing this every time is honestly tiring
All five work. What you find when you actually do them is that you paste the same things again every time.
- The business numbers (price, gross margin, repeats)
- The current measured values (traffic, ad clicks and cost, search impressions)
- The current site (where a button is, and where there is none)
- Actions you received before and decided “this does not fit”
If you are going to do that every week, it is faster to make it a system. I built Shoryaku AI to take that on.
What Shoryaku AI does
At the center of Shoryaku AI is “improvement recommendations,” which produce marketing actions to carry out from this week through this month. The five practices above are applied every time, without a person having to instruct it.
Data is pulled automatically once you connect. Connect Google Analytics, Search Console, Google Ads, Meta ads, and Instagram, and you do not paste it again after that.
LTV and the CAC cap are calculated automatically. Choose a revenue model (one-time purchase / monthly or annual subscription / repeat visits / work ordered each time) and the rest is calculated. Input is optional. If it is blank, defaults are applied from the industry and the customer type, and they are passed to the AI with it stated that they are estimates. If ads are being considered, the cost to acquire one customer is also back-calculated from keyword cost per click and passed along.
And the actions that come back are not shown on screen as they are. This is the biggest difference from doing it by hand.
| Check | When the condition is not met |
|---|---|
| Does a paid action have an acquisition-cost estimate and an alternative? | That action is not shown |
| Is an action over the CAC cap sitting at the top priority? | Priority is lowered by one step |
| Is the basis measured, or a hypothesis? | An action centered on hypotheses has its confidence lowered, and its priority lowered |
| Is there enough data to judge? | It is moved from “carry out” to “watch” or “check” |
Confidence is not the AI’s self-report. It is recalculated by a fixed rule from the kinds of basis that were saved and from the amount of data. Even if the AI marks something “top priority,” it drops to runner-up if the basis is mostly hypothesis, and it is skipped if it contradicts the actual site.
What remains comes out as one card per action.
| Item | Contents |
|---|---|
| Priority and timing | Top priority / runner-up / consider, this week / next week / this month |
| Kind | Cumulative asset / flow, and time until it works (immediate / within 1 month / 3 months or more) |
| Money | Expected cost, expected CAC (within the CAC cap / over it / for a test) |
| KPI | The metric, the current value and the target, and how it is measured |
| Steps | 3 to 8 steps. Each step has a time required |
| Basis | Labeled measured / registered information / principle / hypothesis / insufficient data |
| Premises and risks | What has to be true to carry it out, failure causes, when to stop |
| Alternatives considered | Options not adopted, and why |
For an action you carried out, the numbers are compared 14 days later and 28 days later. Here too the wording is “An improvement was observed. The cause is not established.” Other things happen in the same period, so it cannot be asserted.
It stops at the proposal. Execution is not automated. It does not change ad bids or create campaigns, so the last decision stays with a person.
To try it
There is no cap on how many times improvement recommendations can be generated (within remaining credits). What is capped is how many times actions can be created or rebuilt: 10 a month on Light, 20 a month on Standard, and 25 a month on Pro. The acquisition-cost check and the confidence calculation described here run the same way on every plan.
The 14-day free trial can be started with no card at the start.
If you want to look at numbers first from a CSV you already have, there is also a CSV sales diagnosis that does not require signup. That one is aggregation without AI, and it is explained in another article.