A ChatGPT Ads Strategy Built From Real Prompts

Last verified 9 October 2026. ChatGPT Ads Manager is still in beta, so options described here may change.

A ChatGPT ads strategy starts with the prompts your buyers type, because ChatGPT matches ads to conversations instead of keywords. Get the prompts right and the context hints, ad copy and images follow from them. Get them wrong and even a good ad lands in conversations where nobody is ready to buy.

We learned this with our own money. Our first two Search Agency campaigns spent Rp 813,413 for 171 clicks, and the campaign built around a clear audience earned a click-through rate of 3.69%, almost double the 1.95% of the one built around a vague agency pitch. The strategy we use now came out of that test. If you haven't set up an account yet, start with the setup guide, and for reading reports after launch, see part two.

Why a ChatGPT ads strategy starts with prompts

ChatGPT ads have no keyword list. OpenAI says the system picks ads using the context of the current conversation, your landing page, your title and copy, and the context hints you write. Eligible ads then compete in a relevance-weighted, second-price auction, so an ad that fits the conversation can beat a bigger bid.

That makes the prompt the unit of planning. A keyword like "SEO agency Jakarta" is three words. A prompt is a full situation, such as a startup founder asking how to grow organic traffic in six months on a small budget. The situation tells you who is asking, what they are worried about and what would make them click, which is everything an ad needs.

It's tempting to skip this step and write one ad for the whole business. Our AI Search campaign did exactly that, and it paid Rp 8,459 per click against Rp 3,183 for the more focused bootcamp campaign.

How to find the prompts your buyers ask with StoryMint

We use StoryMint for prompt research. It's the AI visibility platform built by Search Agency's founder, so treat this as a tool we know well and also have a stake in. The workflow below works in any tool that tracks AI answers, and you can do a slower version by hand by asking ChatGPT your buyers' questions and writing down what comes back. Every screenshot in this section comes from our own Search Agency account.

1. Start from personas, not products

Prompts come from people, so the first step is a set of buyer personas. StoryMint builds them from your domain and market and lists each persona's pain points, which you then pick from when you research prompts or check visibility. Our AI Search campaign has six, from a budget-conscious marketing manager to a growth-focused business owner.

Each persona also comes with a list of messages to avoid, which is equally useful for ads. The traditional marketer moving to digital, for example, should never see technical jargon or instant-results promises. Those lists become the guardrails for your ad copy later.

2. Turn a persona into real prompts

Next, open Prompt Research under the AI Search menu, choose a persona and run Find Prompts. StoryMint analyzes that persona's pain points and goals and returns the questions they are likely to type, grouped by the seed query each one came from. For our growth-focused business owner it returned 15 prompts from 5 seed queries.

Read the list with a buyer's eye before you use it. "Berapa harga jasa SEO?" (how much do SEO services cost) and "Apa jasa SEO terbaik di Indonesia?" (which SEO service is best in Indonesia) come from someone shopping for an agency. "Berapa gaji SEO?" (how much does an SEO earn) came from the same seed but is a job seeker's question, and an agency ad in that conversation would waste budget. Prompts in English, such as "How to measure ROI with AI?", sit closer to research than to buying.

3. Expand the best prompts with query fan-out

AI assistants break one question into several smaller searches before they answer. Every prompt in Prompt Research has a "Use in Query Fan-Out" button, and the fan-out analysis shows those sub-queries and checks each one against your site. We ran it on "Berapa harga jasa SEO?" and got 30 sub-queries.

Nineteen of the 30 were covered by an existing page on search.agency, and 11 were gaps with no matching page. The gaps include "faktor penentu harga jasa SEO" (what decides SEO pricing), "model penetapan harga jasa SEO" (SEO pricing models) and "biaya per jam jasa SEO" (hourly SEO fees). Each gap is two things at once. It's an ad angle, since a pricing-curious buyer clicks on an ad that explains pricing, and it's a page we haven't written yet.

4. Check which prompts already mention you

The last step is LLM Visibility, which asks ChatGPT, Google Gemini, Perplexity and Google AI Mode the same question and records whether each one mentions your brand and cites your website. On the prompt "Recommend agile SEO tactics for a new tech startup aiming for aggressive market penetration in 6 months", all four mentioned Search Agency and none of them cited our URL.

Sort prompts into three ad roles

Not every prompt deserves ad budget. We sort each one into one of three roles, based on what the visibility check shows and whether the person is close to a decision. We call this the prompt-to-ad map.

Prompts where AI already names you are good places to reinforce. Across five startup prompts StoryMint has checked for us, Search Agency was mentioned in the answers while our website was cited 0% of the time. An ad under that answer gives the reader a direct link that the answer itself doesn't provide.

Prompts where AI names a competitor or nobody are your highest priority for ads, because paid placement is the only way to appear there until your content earns a mention. Pure research prompts, like "Apa itu SEO dalam AI?" (what is SEO in AI), carry little buying intent. Those belong to content and Topic Ownership Strategy work instead of ad spend.

Turn prompts into context hints

Context hints are where your prompt research enters Ads Manager. OpenAI's guidance on context hints asks for a clear, natural phrase focused on one idea, describing what you offer, who it helps and when it's useful. Hints guide matching, but they don't act as targeting rules.

Our first AI Search hints were a comma-separated keyword list. Rewritten from the persona and fan-out work, they read like this.

• "Startup founders in Indonesia who need a 6-month SEO and AI search roadmap on a limited budget"
• "Marketing managers who must show clear ROI from SEO to their boss"
• "Business owners who want their company recommended when customers ask ChatGPT"
• "Business owners in Indonesia comparing SEO agency prices and what drives the cost"

Each hint maps to one persona and one ad group. Mixing all of them into a single ad group repeats the original mistake, because one ad can't speak to a founder and a marketing manager at the same time.

How to write ChatGPT ad copy that earns the click

Good ChatGPT ad copy picks up where the answer stops. The person has read a helpful reply, so the ad has to offer the next step, such as a plan, a tool or a person who can do the work. OpenAI's creative guidance sets the rules plainly. The title should spell out the value with informative details over broad marketing language, and the copy should add new information instead of repeating the title.

Length decides whether the message survives. OpenAI's bulk upload checklist recommends 16 to 24 characters for titles and 32 to 48 for copy. Our first description ran to 77 characters and was cut off mid-word in the preview.

We also run titles through StoryMint's headline analyzer before launch. Our original title, "AI Search Agency", scored 6.5 out of 10 with a C grade and the verdict "Clear but lacks uniqueness". A persona-specific alternative, "AI Search Audit for SMEs", scored 7.5 with a B, and the tool suggested adding a concrete benefit. Below is how the persona work turns into ads, each written for one persona and inside the recommended lengths.

The persona's avoid list keeps copy honest. OpenAI's ad policies prohibit ads that "exaggerate results", which rules out lines like guaranteed rankings or instant citations. That matches what our personas said they distrust anyway.

How to choose ChatGPT ad images

The image is the largest visual in a ChatGPT ad, and our own results show how much it matters. Inside the bootcamp ad group, two ads ran side by side. The ad with a dark "SEO Fighter Bootcamp 2026" image earned a 4.24% click-through rate across 2,356 impressions, while the speaker-photo ad earned 2.22% across 900. The AI Search ad, which used our logo, managed 1.95%.

The two bootcamp ads also had different titles, so the image isn't the only explanation. The pattern still fits OpenAI's image guidance, which asks for simple, relevant visuals that match the title and copy, and warns against abstract or cluttered ones. An image that names the program tells the reader what they'll get before they read a word. A logo tells them who is selling, which matters less in the second after an answer. A few practical rules come out of this, and none of them need a designer.

• Use square PNG or JPG files of at least 256 by 256 pixels, as Ads Manager asks.
• Show the product or program itself, such as a course name, a product shot or the result of the service.
• Test a logo-only image against a product image before assuming your brand mark is enough.
• Fix your page's social sharing image, since the "Add new ad" option prefills ads from your website metadata.

That last point catches many advertisers. OpenAI says the prefill uses existing website metadata for the image, title and description, and doesn't generate new imagery. If your landing page's share image is a generic stock photo, that is what your suggested ad will show.

How to optimize ChatGPT ads with one test at a time

ChatGPT ads optimization works best as a planned series of single-variable tests. Our first two campaigns differed in location, offer, hints, copy and image all at once, which is why we can't say which one caused the gap in results. A clean plan changes one thing per round and keeps everything else fixed.

Give each round at least a week. OpenAI's reporting shows clicks within about 15 minutes, but spend lags by seven to eight hours and conversions can take 24 to 48 hours, and a daily budget is treated as an average across the week. Our own test ran for about two days, which was enough to compare click-through rates and far too short to judge conversions, since 171 clicks produced zero recorded registrations.

Set conversions up properly before round four. That means OpenAI's pixel on the page, a conversion event for the action that matters, and UTM parameters so the traffic also shows up in GA4. Part two covers what to check when an ad isn't serving or clicks don't convert.

Use ad results to guide your AI Search content

Ad data tells you which prompts and angles your buyers respond to, and that is exactly what your organic content needs to cover. When an angle wins as an ad, it deserves a page that answers the same need in depth, so AI assistants can cite you inside the answer as well as below it. The 11 pricing gaps from our fan-out run are a good example. If an ad about SEO pricing wins, the next pages to write explain pricing models, hourly fees and what drives the cost, so the answer itself can cite us.

Run the loop monthly. Re-check visibility on your priority prompts, move prompts between the three ad roles as your organic presence grows, and cut ad spend where the answer already names and links to you. Over time the ads cover the gaps while AI search measurement tracks whether your content is closing them. If you'd like that loop run for your brand, our AI Search team builds the prompt map, the ad tests and the content plan together, so paid and organic work from the same research.