Shoryaku AI

A prompt for landing-page improvements that tests one element at a time

Marketing11 min

Paste the body of a landing page into an AI and ask it to “produce improvements,” and you get ten or more ideas: the headline, the CTA, the color, the form, the price, case studies, and more. If you change all of them, the page changes, and you can no longer tell what affected the result.

What you need when you improve a landing page with AI is not a volume of copy options. Pick the single largest bottleneck right now, and pair one change with one test.

This article covers a prompt that passes landing-page information and measured numbers to free ChatGPT or Gemini, and narrows where to try first: the first view, the claim, the CTA, or the form.

The page body alone cannot set priority

Even on the same landing page, a person who arrived from an ad and a person who arrived from a search for the company name expect different things. An AI that has only seen the text does not know what intent arrived, where people left, or what counts as a result.

Before you ask for improvements, pass at least the following.

  • Who it is for: who the page is aimed at
  • Source and intent: ad text, search terms, a social post, a referral, and so on
  • Purpose of the page: purchase, booking, inquiry, a request for materials, and so on
  • The current page: URL, headlines, body, CTA, form fields
  • Measured numbers: sessions, CTA clicks, form starts, completions, CVR
  • Constraints: a price you cannot change, legal wording, production time, and so on

If there are no measured numbers, do not have the AI predict an improvement. Have it output “hypotheses about problems you can check on the page.” Do not treat a CVR improvement as a fact when you cannot tell whether traffic is low or conversion inside the page is weak.

Match the visitor’s expectation to the page’s answer

Google’s explanation of ad quality lists the usefulness and relevance of the ad and the landing page as parts of ad quality. How to check landing-page results covers the landing-page report and how easy the page is to use on mobile.

This is not only about ads. The first screen of the landing page has to answer, right away, the expectation the person had before they arrived.

Expectation before the visit: I want to know about ad analysis for a small shop
The page's first answer: who it is for / what you will learn / what to do next
Primary CTA: Try the analysis for free

If there is traffic and signups are weak, look in this order first.

Place What to check Number to look at mainly
First view Can you tell who it is for, the problem, the benefit, and the primary CTA? Bounce and the next action, primary CTA clicks
Claim and proof Does it answer the expectation from before the visit, and is there a reason to choose it? Scroll, clicks on each CTA
CTA Is the action, and what happens after it, specific? CTA click rate
Form Are there more fields than needed, or a place where input gets stuck? Start rate, completion rate, where people leave

Do not change the whole page at once. Pick the first place from the numbers and from what you observe.

Make the change and the hypothesis a pair

For example, the following change is not one test.

Change the headline, make the CTA yellow, and shorten the form

Even if the result improves, you cannot tell whether to keep the headline, the color, or the form. Pair the change and the reason, like this.

Observation: there is traffic from ads, but few clicks on the primary CTA
Hypothesis: the headline in the first view does not answer the ad's claim
Change: change only the headline, to a sentence that makes the result the ad promised clear
Metric to watch: primary CTA click rate, and the final conversion
Keep: CTA position and color, the form, the price, the ad settings

It is possible for only primary CTA clicks to rise while the final conversion falls. So check the final result when you can, not only the click rate in the middle.

An input template to pass to ChatGPT or Gemini

If the AI cannot read the URL, paste the page body too. If you want it to look at a problem that is specific to mobile, also pass a screenshot, the screen width, and the display order.

Purpose of the landing page:
- Final conversion:
- Who it is for:
- Product and price:

Traffic:
- Source:
- Ad text, search terms, post text:
- What the person who arrived is expecting:

Current landing page:
- URL:
- Page title:
- Headline and body in the first view:
- Primary CTA wording and where it goes:
- Main sections and headings:
- Information published as proof:
- Form fields:
- What looks off on mobile:

Measured numbers (the same period):
- Period:
- Sessions:
- Primary CTA clicks:
- Form starts:
- Form completions:
- Conversions:
- Differences by source and by device:

Recent changes:
- Changes to the landing page, ads, price, or measurement:

Constraints:
- Wording that cannot change for legal or brand reasons:
- How much production time can be used this time:

Write “unknown” in any field where you do not have a measured number. 0 and unknown are different. This keeps the AI from reading a blank as 0.

A landing-page prompt that narrows to one element

After the input template, paste the following instructions.

You help design tests for landing-page improvements.
Using only the information that was passed, narrow the first change to try to one element.

Rules:
- Do not invent numbers, customer quotes, results, or industry averages that are not in the input
- Do not guarantee an effect. Write improvement ideas as hypotheses
- If there are no measured numbers, do not assert that CVR is low or high
- Check whether the expectation before the visit matches the landing page's headline and CTA
- From the first view, the claim and proof, the CTA, and the form, pick one place to check first
- In one test, change mainly one element
- Also state the elements you are not changing
- When you can, watch the final conversion, not only the click rate
- Do not propose unsupported claims such as a count of companies using it, a satisfaction rate, scarcity, or urgency
- Do not remove paths that are needed, such as legal, privacy, or contact

Output:
1. What the person who arrived expected, and the answer the landing page currently gives
2. Facts that could be confirmed
3. At most 3 hypotheses for the bottleneck
4. The one hypothesis to test first, and why it was chosen
5. The one element to change. Current version and changed version side by side
6. Elements not changed this time
7. Primary metric, supporting metrics, and the final conversion
8. Comparison period, and the condition for reviewing
9. Information that is missing for the decision, and how to get it

After the output, look at whether it can be tested, not at how striking it is

Even if the AI’s change looks attractive, fix it before you run it unless these five are in place.

  1. Which observation the hypothesis was chosen from
  2. Whether the element being changed is narrowed to one
  3. Whether the parts that stay the same are written down
  4. Whether you can check the primary metric and the final conversion
  5. What result means keep it, revert it, or check something else

When there is little data, do not call a winner over a short period. Keep “cannot judge” as a result too. If you try several ideas, try them one at a time in order, or split the comparison in an environment with enough traffic.

How to look back at the numbers after an action is also explained in How to review a marketing move with AI without asserting cause.

Do not recollect the site information and the numbers every time

You can use this method with a free AI, but you have to recollect the page body, headlines, CTA, form, traffic source, and measured numbers every time you change something.

Shoryaku AI uses, as material for analysis, what it could confirm from the registered site — page title, headlines, CTA, form, and similar information — together with the data you connected. If there is traffic and the conversion step is weak, it looks first at the first view, the CTA, and the match with search intent, ahead of more acquisition. A proposal includes the target page, the element to change, the reason, the KPI, and the period to check. If measured numbers or page information are missing, that is treated as a gap.

If traffic reaches the landing page from Google Ads, combining this with A prompt for investigating a worse Google Ads CPA sorts out whether to check the ad side or the landing-page side first.

Summary

When you ask AI for landing-page improvements, look at whether they are narrowed to a unit you can test, not at how many ideas there are.

  1. Pass who it is for, the traffic source, and the purpose of the page
  2. Pass the CTA, the form, and measured numbers, not only the page body
  3. Set the expectation from before the visit next to the page’s answer
  4. Choose a hypothesis for the bottleneck
  5. Change one main element at a time
  6. Check the intermediate metric and the final conversion
  7. Record it when you cannot judge, too

First, put the first view of the landing page next to the ad text or the search term from just before someone arrived. If the two do not answer the same expectation, you can narrow the first hypothesis to test by a lot.