Three ways to give AI chat your own numbers

If you ask ChatGPT "what caused CPA to rise last month," you do not get a decent answer. It does not have your numbers.

What comes back is something you would find in a textbook: "a rise in bids, an increase in competitors, seasonal factors, and so on are possible." It is not wrong, but reading it does not decide what you do tomorrow.

There are 3 ways to pass the numbers. The effort, the situation each one fits, and the risk are all different.

I am building Shoryaku AI ( https://shoryaku.net ), a marketing-support SaaS. The first half is the comparison of the 3, and what you must always do before you upload a CSV. The second half is about Shoryaku AI's chat.

There are 3 ways to pass them

The easiest is pasting a CSV, and the most dangerous is the same method. I will start there.

What you must always do before you upload a CSV

If you have ever exported an order CSV from an online shop, you know that it contains names, addresses, phone numbers, and email addresses.

If you upload that to an AI service as it is, you have handed your customers' personal information to an outside service. Even for a business of 1 person, that is not a comfortable story. At a company, it may touch the internal rules. If the data was entrusted to you by a business partner, it can also be a breach of contract.

The procedure is 3 steps

1. Delete the columns that can identify a person.

Name, address, phone number, email address, customer ID, order number. Delete "customer ID" and "order number" too. On their own they mean nothing, but if you match them against other data you have, they can be turned back into a person.

2. Do not pass the line items as they are. Aggregate first.

Sales by day, sales by product, counts. A pivot table in a spreadsheet is enough. Turn 5,000 rows of order lines into an aggregated table of about 30 rows.

3. Paste the aggregated table directly into the chat box.

If it is a few dozen rows, you can paste it as it is, without attaching a file. Not leaving an attached file behind makes the cleanup easier too.

Passing it after you aggregate also makes the answer better

This is not only for safety.

If you make an AI add up several thousand rows, it gets it wrong. And it does not tell you that it got it wrong. Pass the totals and averages as numbers you calculated yourself. Let the AI do only "how to read those numbers." With that split, the answer is more trustworthy.

How to ask after you paste

Handing over the numbers is still not enough. A generality still comes back. Add 3 instructions.

The following are our aggregated figures for (period).

(paste the aggregated table)

Conditions:
- Use only the numbers you were given. Do not fill in missing items by guessing.
  Write "this cannot be told from this data."
- For each claim, always state whether the basis is "measured" (it is in the numbers you were given),
  "inference" (an interpretation derived from the numbers), or "hypothesis" (there is no supporting evidence)
- When you compare, align the latest 28 days with the 28 days before that

The 2nd instruction is the one that works best.

An AI's answer comes out in the same confident style for the actual numbers, the generalities, and the things it thought up. If you make it attach labels, you can see which part of that text is actually based on the data. When you try it, there should be more "hypothesis" labels than you expected.

If you connect with MCP

If you ask the same thing repeatedly, connecting is faster than pasting every time. Google Ads and Google Analytics have official MCP servers, and you can have an AI tool on your machine read the data directly.

The weight of the setup, and the concrete steps, are in another article.

Even then, there are things you still explain every time

Even after the problem of passing the numbers is solved, 1 more thing remains.

• What you sell, at what price, and to whom
• How much you can spend in a month
• What you tried before that did not work

Without these premises, the answer goes back to a generality. If you pass only the numbers and it ends at "let's A/B test the ad copy" or "let's improve the landing page," that is because the AI does not know your business.

You can write them into custom instructions or a project feature, but the state of the business changes, so you have to keep updating them. Quietly, that does not continue.

Shoryaku AI's chat

It is built so that the 3 things in the first half are done with no setup.

If you are already connected, you only ask the question. You do not need a CSV or a screenshot. It goes and fetches the numbers it needs from whichever of Google Analytics, Search Console, Google Ads, Meta Ads, and Instagram are connected.

It fetches only when it needs to. Questions are classified into 17 kinds, and 3 of those (a question that is out of scope, small talk, and a question that needs confirmation) are answered on the spot and stop there. It does not call the ads API for "think of Instagram post ideas," and it does not fetch outside data for a general marketing question. It is built so you are not kept waiting on an unrelated question.

Business information goes in every time. The offering, the price, the revenue model, the target, and the monthly budget. On top of that, it also passes moves you previously rated as not a fit, so the same proposal is less likely to be repeated.

The answer attaches a label of "measured," "inference," or "hypothesis" to each claim. The instruction the first half recommended you write every time is built in here.

And writing "measured" alone does not let it through. Each such claim is required to attach at least 1 reference to actual data, and an answer that contains a reference that does not exist is rejected. This is to prevent a state where only the label looks solid and there is nothing inside.

When the data could not be fetched, the answer itself is not turned into an error. It states what could not be fetched, and still answers. Even if a connection is broken, it answers as far as it can with the information it has on hand.

You can also attach a CSV

There is also a feature to attach a CSV to the chat, but what is sent is not the file.

It aggregates inside the browser, and then sends only the few dozen aggregated rows. The line items, and the rows from before the personal-information columns were dropped, are not sent to the server. It is the shape of "delete the personal-information columns, aggregate, and then pass it," from the first half, done with no manual work (the same mechanism as CSV sales diagnosis).

About the cost

You can use it from Lite at ¥980 a month. Chat does not consume the analysis count, so you can ask without watching the count (within the remaining credits).

It is not something you cancel ChatGPT or Claude to switch to. For writing, looking things up, and writing code, a general-purpose AI is a better fit. The price is set on the premise that you bring only the questions that involve your own numbers here, and keep using what you use now for everything else.

So that they are easy to use together, you can also pull the same data from an AI tool on your machine via MCP.

What to check before you use it

• Only 5 channels are supported. Yahoo! Ads, TikTok, and LINE are not supported
• If you do not enter the business information, the accuracy does not come out. Input is optional, but the more you enter, the more the answer changes
• It is read-only. You cannot pause an ad or change a budget from the chat
• It is not a replacement for a general-purpose AI chat. It is built for marketing questions

To try it

The 14-day free trial can be tried with no card registration at the start.

How to choose a plan is in another article. The overall picture of the connections is summarized here.

• Marketing
• AI
• Product
• SaaS