AI Audience Research: ChatGPT Prompts With Real Data

AI audience research means using an assistant to organize evidence about your audience, then checking that evidence before deciding what to create. Start with your posts, available analytics and the words people actually shared. Ask ChatGPT to separate what the sources show from what you still need to learn.

The useful result is a content decision you can explain: “These supplied comments ask about recording audio, so I’ll test a post showing my setup.” A plausible persona with an invented income or hidden fear does not establish who follows you.

Use the source-based method, copy the ChatGPT research prompt, or try the method with your connected Mysocial context.

Start With a Small Source Bundle

Choose one account, one question and one period. For example: “Which questions about recording a talking-head video appear in comments on my recent tutorials?” A social media manager should name the represented client account and keep its evidence separate from other clients.

You can begin manually. Copy permitted material into a document or spreadsheet, remove unnecessary names and personal details, and give each record a short source ID. Keep the original source available for your own checks. Share only information you have permission to use.

Source	What to include	What it can tell you
Your posts	Source ID, link, publication date, topic and caption or supplied transcript	What the post covered
Comment text	Comment ID, parent post ID, exact text and date if supplied	What those sampled comments said
Post metrics	Post ID, metric name, reported value and observation date if available	Performance within the metric’s stated scope
Platform demographics	Reported category, value, date and audience scope when available	The population and period that report actually measures
Voluntary feedback	Anonymized wording and its context, used with permission	What those respondents explicitly told you

Keep comment text and comment count separate. “12 comments” is a numeric metric; it does not reveal any of those comments’ questions. Similarly, a post’s views do not tell you the number of distinct people asking for a topic.

ChatGPT supports attached files, including spreadsheets and data exports. Provide the source material and the question you want answered; attaching a file does not remove the need to review the answer. OpenAI’s file and context guidance.

When a field is absent, write “not supplied.” Keep an observed zero as zero. If a comment’s date or a metric’s observation date is unavailable, say so rather than substituting the post’s publication date.

Copy This ChatGPT Audience Research Prompt

Paste this with your source bundle. Replace the account, question and period. It also works as a manual text prompt in Claude; no Mysocial connection is required for this route.

I create or manage content for [account and purpose]. My research question is [one question] for [period].

Use only the source material I supplied. First list which records you could read, their source IDs, dates and missing fields. Keep comment text separate from numeric comment counts, and post publication dates separate from metric observation dates.

Make an evidence table with: observed question or theme; exact supporting quote; source ID and date; number of distinct supporting comment records; unknowns; content hypothesis; and one small test. Count only supplied records and do not count repeated copies twice. Do not describe record counts as unique people or a share of my whole audience.

Quote only text present in the sources. Mark unsupported age, income, occupation, motives and other personal attributes unknown. Include comments that challenge a theme. Keep your interpretation separate from the observations.

Recommend one next post based on the evidence, explain what it would test, and say what later evidence would support or challenge the idea. If the inputs cannot answer my question, explain what is missing before recommending a topic.

The evidence table is a research aid you review. Check each quotation against its source and each count against the included records. An assistant can omit material or make an unsupported assumption even when the prompt asks it not to.

This method adapts the source-inventory and claim-review discipline in OpenAI’s research synthesis guide to audience comments. That guide covers research evidence; it is not a measured claim that AI improves audience research.

A Worked Example: Evidence to One Content Test

Fictional teaching example: every comment, date and number below is invented to show the method. This is not a Mysocial customer study, an audience benchmark or an actual research result.

A creator supplies two tutorial posts, labeled P1 and P2, and three anonymized comment records:

Record	Source and supplied date	Exact fictional comment
C1	P1; Oct 1, 2026	“What mic are you using?”
C2	P1; Oct 2, 2026	“Could you show how you remove background noise?”
C3	P2; Oct 3, 2026	“Can you show your mic settings in a noisy room?”

The separate metric sheet contains:

Post	Views	Numeric comment count	Metric observation date
P1	400	12	Oct 4, 2026
P2	Not supplied	Not supplied	Not supplied

P1’s 12 reported comments are not 12 comment texts in the bundle. Only C1 and C2 were supplied from P1; C3 was supplied from P2. P2’s missing metric fields do not erase C3’s supplied text.

A reviewed analysis could distinguish:

• Observation: C1 and C3 ask about recording equipment or settings; C2 asks about editing background noise. There are three supplied comment records. Their IDs alone do not establish three distinct people.
• Evidence: preserve each exact quote and its source/date. “Recording setup” is the analyst’s theme label, not wording shared by every commenter.
• Unknown: the rest of the comments, silent viewers’ preferences, demographic information and P2’s metrics. The sum of known views is 400 across one measured post; the two-post total is unknown.
• Hypothesis: a recording-setup tutorial may answer the questions in C1 and C3. A separate editing tutorial may answer C2.
• Small test: publish one recording-setup tutorial, explain the equipment and settings you actually used, and invite viewers to identify the step that remains unclear.

“Two of these three comment records ask about setup” describes this supplied sample. It does not mean two-thirds of the audience want that topic. Nor do the invented 400 views prove that the original topic caused demand.

Use Your Available Mysocial Context

If your posts are already connected to Mysocial, your assistant can begin with the evidence available in the workspace you select. Connect Mysocial to ChatGPT or Claude, choose the relevant account and period, and review what the assistant can actually read before it summarizes anything.

Access matters: Free can relay available posts and captions, post metrics, follower history and measured platform demographics. Availability still depends on the channel and its data. Stored comment text, transcripts and generated intelligence require the relevant paid access and content that has been collected or processed. A numeric comment count on Free does not unlock the comments themselves.

Mysocial separates platform-reported demographics from model-inferred audience segments and psychology. Treat an inferred explanation as a hypothesis to review. It does not verify a person’s income, motives or private circumstances.

Research a question with your available sources

@Mysocial MCP

Edit your prompt Use my selected Mysocial workspace and own posts for [account], [topic] and [period]. First show available sources and access gaps. Use comment text only if stored and accessible; counts are not text. List observed questions with exact quotes, source links and dates; keep missing fields unknown. Separate hypotheses from findings. Suggest one content test and what would challenge it. Read existing data only.

Review and send in your AI app.

Setup & access

Connect and select Mysocial before sending. Connect Mysocial →

Connect an eligible channel and select Mysocial in ChatGPT or Claude. Available posts, metrics and measured demographics can be read on Free. Stored comment text, transcripts and generated intelligence need relevant paid access and available processed content.

Result: An evidence table, explicit unknowns and one content hypothesis to review and test.

The buttons open your AI app with a draft prompt. Connect and select Mysocial there, review the prompt, then send it yourself; copy and paste if needed. If the needed comment text is unavailable, use the permitted manual source bundle above or resolve access first. The prompt asks for existing readable evidence. Review any request for fresh collection or paid processing separately; do not assume it is part of this research step.

Review the Result Before Planning Content

Follow the evidence back to its source:

1. Check coverage. Did the assistant read all supplied records? Which dates, posts or fields are missing?
2. Check quotations and counts. Confirm the wording and source IDs. Remove duplicate records; keep distinct records separate from distinct people.
3. Look for disagreement. A comment asking for a beginner explanation and one requesting an advanced example may call for different posts.
4. Challenge the interpretation. Could a question reflect an unclear sentence in one tutorial rather than a broad audience interest?
5. Choose one test. State what you will make and what response would cause you to revisit the idea.

OpenAI advises checking important claims, quotes, current sources and unsupported assumptions before relying on an answer. Its review guidance.

If you use a persona to brief a writer, make it a working summary of the supplied evidence. Include “unknown” where support is missing. Do not fill gaps from someone’s name, accent, profile photo or an assistant’s guess.

Measure the Next Post, Then Update the Hypothesis

Record the post you tested, its publication date and the question it was meant to answer. After a comparable observation period, inspect the metrics your platform actually provides and the text of relevant new comments. Keep platform, format, metric definition and observation period consistent when comparing posts.

A useful follow-up prompt is:

Here is my original hypothesis and its source table. Here is the test post and the new dated evidence. Separate what changed from what remains unknown. Which new comments support or challenge the idea? Compare only supplied, compatible metrics. Do not attribute a performance change to the topic alone. Recommend whether to repeat, revise or investigate further.

A strong performance result can still have other explanations, such as timing, distribution or a different hook. A quiet comment section can leave the question unresolved. Ask your audience directly when the missing input matters.

Repeat the research when new evidence or a new decision warrants it. You do not need to rebuild a complete persona on a fixed timetable.

When the evidence supports a topic, add the hypothesis and its sources to your content calendar. Use the social media content guide to turn it into a draft, then return to the audience evidence after publishing. The point is to keep learning from what people actually say and do.