Most teams that adopt generative ad tools point them straight at creative production and skip the upstream work. That is backwards. The highest leverage use of large language models in marketing is not generating the ad — it is understanding the audience before you decide what ad to generate. Here is how to do that work properly, with the specific prompts and analysis frameworks that have stood up to a year of practice.

The case for doing this work is simple. A merely competent ad to the right audience outperforms a brilliant ad to the wrong audience by a wide margin. Targeting has gotten harder since the privacy changes of 2021-2024 — you can no longer interest-target your way to a clean segment. The new edge is understanding your audience's language, fears, and aspirations deeply enough that your creative resonates regardless of where the platform algorithm sends it. AI tools, used carefully, can do meaningful audience research in hours that would otherwise take months.

What "audience research" actually means

Before tools, it helps to be clear about what good audience research produces. The output is not a persona document with stock photos and a fictional name. The output is a set of specific claims about the audience that creative can be built around. Good claims look like this:

• "They worry about whether the product is safe around their kids before they worry about price."
• "They have tried three competing products and were disappointed each time, so social proof has to address skepticism rather than enthusiasm."
• "They learn about new products primarily from one specific subreddit and one specific newsletter, both of which value technical honesty over polish."
• "They mostly buy on the second Wednesday of each month, which suggests budget-driven timing rather than need-driven timing."

None of these claims are demographic. None require interest-targeting data. All of them produce specific creative implications. That is the bar for useful research.

The five sources of raw input

Before you ask any AI to analyze anything, you need a corpus of raw text to analyze. The five sources that consistently produce useful research input:

1. Customer support transcripts. The most underused source of audience insight in most companies. Every refund email, every "how do I" question, every complaint is your audience telling you what matters to them in their own words. Gather six months of transcripts and treat them as primary research.
2. Product reviews — yours and competitors'. Amazon reviews, App Store reviews, Trustpilot, Google. Especially the three- and four-star reviews, which are where the genuinely informative critique lives.
3. Social media discussions. Subreddits, X conversations, TikTok comments, niche forums. Anywhere your audience talks unprompted about your category.
4. Sales call recordings. If you have any consultative sales motion, sales-call recordings are gold. The objections, the questions, the moments of relief — all useful signal.
5. Open-ended survey responses. Not multiple-choice survey data. The free-text questions where customers said anything they wanted. Most surveys produce these and most companies ignore the free-text answers.

For most teams, the bottleneck is not the analysis — it is gathering this corpus in the first place. Plan a week of data collection before you plan a week of analysis. The work is worth it.

The analysis prompts that work

Once you have the corpus, the analysis becomes an LLM exercise. Four prompt patterns produce different angles on the same data and you want to run all four.

Prompt 1: The vocabulary extraction. "Here are 200 customer reviews. List the 30 specific phrases customers use most often to describe the problem this product solves. Then list the 30 phrases they use to describe the product itself. For each phrase, note whether it appears in positive, negative, or mixed contexts."

This produces your audience's actual vocabulary. You will be surprised how often the words they use are different from the words you have been using in your creative. Match their vocabulary and engagement climbs noticeably.

Prompt 2: The objection map. "Here are 200 customer reviews and 50 support transcripts. List every objection or concern customers raise before purchase, ranked by frequency. For each objection, list the specific reassurances that, when present, seem to resolve it."

This becomes the basis for your ad creative's social proof. If "I was skeptical it would actually work" is the top objection, your ads should feature people overcoming that exact skepticism. Generic testimonials underperform specific objection-resolving testimonials by a wide margin.

Prompt 3: The competitor delta. "Here are 100 reviews of our product and 100 reviews of [competitor]. What do customers consistently say is better about our product? What do they say is worse? What do they wish either product had that neither does?"

This third question is the gold one. The "neither does" answer is where your next product feature lives and your next ad campaign can credibly position you. The fact that no one in the category provides X is your opening to be the one who does.

Prompt 4: The unprompted association. "From these social media discussions, what activities, situations, or other products is our category most often mentioned alongside, when nobody is asking about us specifically? List the 20 most common associations."

This reveals the contextual frames your audience is actually thinking in. If meal-prep is mentioned alongside your protein powder, your creative should appear in meal-prep contexts. If late-night snacking is the association, that's a different creative direction entirely.

One critical methodological warning

LLMs hallucinate. If you ask "what do customers say about X" without providing the actual customer data, the LLM will confidently generate plausible-sounding but completely fabricated answers. This is the single most dangerous failure mode in this work.

The discipline: always paste the actual raw data into the prompt before asking for analysis. If your corpus is too big to paste, use a model with a long context window (the major ones now support hundreds of thousands of tokens) and chunk it deliberately. Never ask the model to "summarize what customers usually think" in the abstract. Always anchor on specific, present text.

A useful sanity check: every claim the LLM produces should be traceable to a specific quote in the input data. If you ask "show me the three customer quotes that support this claim," the model should produce them. If it cannot, the claim was likely invented.

From research to creative brief

The translation step from research to creative is where most teams fall down. They produce a wonderful 30-page audience document and then write the same generic brief they always wrote because the document was too long to actually integrate.

The discipline that works: compress every audience research project into a single page. Not a summary deck. One page. The page has four sections: the top three things this audience cares about (their words, not yours), the top three things they fear or are skeptical of, the top three contexts in which they encounter the category, and the top three competitive landscape facts.

That single page becomes the input to every creative brief for the next quarter. When you write a prompt for the ad generator, the prompt explicitly addresses one of the cares, fears, contexts, or competitive facts. If a prompt doesn't tie to anything on the page, you skip generating it — there is no audience-grounded reason for that creative to exist.

The research cadence

Audience research is not a one-time exercise. Audiences shift, competitors move, contexts change. We recommend a quarterly cadence: each quarter, collect a fresh corpus (you do not need to redo the historical analysis, just incorporate the new three months of input), re-run the four analysis prompts, update the one-pager. The quarterly refresh takes maybe a week of work and keeps your creative grounded in current rather than year-old understanding.

The teams that do this well notice the gradual shifts: a new objection appears that wasn't there six months ago, a competitor mention shifts from positive to mixed, a context emerges that didn't exist before. Catching these shifts early is the difference between creative that feels current and creative that feels six months stale.

Combining audience research with the creative platform

Where this gets compounding is when the audience one-pager directly feeds the prompt templates in your creative platform. Concretely: each of the three "cares" gets a prompt template, each of the three "fears" gets a prompt template, each of the three contexts gets a prompt template. That is nine pre-built templates that any team member can use to generate audience-grounded creative without having to redo the research thinking each time.

Over a quarter, those nine templates produce dozens of variants each. The ones that perform get tagged in the prompt library with the underlying audience claim they were testing. After a year, you have a database of "audience claim X is best expressed visually as Y" that becomes meaningful institutional knowledge.

This is the deeper version of the prompt library idea. Not just "prompts that worked" but "prompts that worked because they addressed this specific audience insight." The second layer is what makes it durable across staff changes and platform shifts.

What to do next week

If this is new to you, start small. Pick your highest-spend campaign. Pull six months of support transcripts and one month of competitor reviews into a single document. Run the four analysis prompts on that corpus. Write the one-page summary. Compare what the one-pager says to the current creative running in that campaign. The mismatches between the two are your immediate creative roadmap for the next quarter.

If the audit reveals serious mismatches — and it almost always does the first time — you have just identified the highest-leverage creative work your team can do. The teams we work with that adopt this practice consistently outperform teams that go straight from product brief to ad generation. Audience research is not the glamorous part of the workflow. It is the part that makes everything downstream actually work.

The tools to do this kind of research existed before LLMs but the labor cost was prohibitive — six months of qualitative researcher time per project. With current models, the labor cost drops to a week of one marketer's time per quarter. That is the shift that makes audience research finally affordable at the cadence it always needed.