A video receives two thousand comments. “Where is the tutorial?”, “How much does it cost?”, and “How is this different from X?” all have many likes. The usual next move is to give those comments to AI, ask what users care about, and write the next post from its summary. The result is often broader, more generic content that does not resolve the actual obstacle to buying or using the product.
A comment section is not a poll. Likes may reward a joke, a shared opinion, or being early. Repeated low-like questions may show that new users are stuck at the same step. AI is useful here not because it chooses the topic, but because it turns unstructured conversation into evidence you can inspect: who is asking, what task they are trying to finish, what information is missing, and whether your product or existing content can answer it.
Define the question before importing comments
Each review should have one purpose: find next week’s tutorial topic, identify pre-purchase hesitation, test whether a feature is misunderstood, collect customer language for a landing page, or investigate a weak campaign. Different purposes require different filters. Asking AI to “summarise sentiment” across all of them guarantees a fuzzy conclusion.
For tutorial research, keep four types of comments: questions with a clear task; reports of a failure or blockage; questions about comparison, price or fit; and questions repeated by separate users. Leave emojis, giveaway replies, context-free requests, and unrelated arguments in the raw export, but do not let them dominate the model context.
Keep evidence with every comment
Export more than comment text. Include the text, date, likes, replies, source post topic, whether the author replied, parent comment, visible language or region, link, and an initial human note. Without the original post, AI cannot tell whether “this does not work” refers to price, performance or editing style. Without the parent comment, it cannot interpret “the second method.”
Add two columns: underlying question and fact to verify. The first restates the job in minimal language, such as “Can the free plan export without a watermark?” The second states what needs checking: the pricing page, help centre, product test, or support policy. This stops the team from turning a guess into a published answer.
Use AI as a labeler, not a content director
Work in batches of 100–200 comments and request structured output:
You are a user-research assistant. Using the comment and its context, assign one primary intent: operational help, feature understanding, price or purchase concern, comparison, usage obstacle, positive feedback, off-topic, or insufficient context. Extract the task the user is trying to complete and retain important limits in their wording. Do not infer identity, demand size, or purchase intent. For questions that can be checked with public information, name the fact to verify. Output comment ID, intent, task, key quote, fact to verify, and confidence.
“Do not infer” is the important constraint. “Is there a cheaper option?” does not prove the user will buy, and “why is this button missing?” does not automatically prove a product bug. The model can classify evidence; it should not turn one remark into a commercial conclusion.
Choose a topic from a cluster, not a single quote
After labelling, group repeats by task and condition. A topic is usually worth writing when it is asked by multiple independent people, has a specific answer that can be demonstrated, and connects honestly to your product, service, or existing material.
Questions such as “where is the template?”, “can I edit on my phone?”, and “my fonts changed after export” look separate. Together, they may reveal that the complete path from download to delivery is unclear. The next tutorial should cover device differences, import, font replacement, and delivery checks—not merely a template link. A single highly liked request for a free item is not automatically a content topic.
Use a short scoring system
Score every cluster from 0–3 for repetition, specificity, answerability, and business relevance. High-scoring clusters deserve content first. A cluster with high repetition but low answerability belongs with product or support, not in a piece of marketing copy. The score makes topic choices explainable rather than a debate about who was loudest.
Verify first, then let AI shape the tutorial
Using only the verified facts and user questions below, create a tutorial outline. Start with the real task the user is stuck on, not a generic trend statement. Each section should solve one step or choice, state conditions and limits, and distinguish alternatives. Do not add prices, functions, outcomes, or promises that are not provided. Finish with an actionable checklist and a list of questions that still need confirmation.
For current information about third-party tools or pricing, use Perplexity to locate candidate sources, then confirm the answer on the official page. Use ChatGPT for classification and outlines only after comments are anonymised. Order IDs, phone numbers, emails, private-message screenshots, and non-public customer details do not belong in the prompt.
A 40-minute weekly rhythm
- Export comments from three to five strong posts and remove obvious spam.
- Read twenty comments manually to set the goal and check the labels.
- Label in batches with AI, then sample five to ten records from each category.
- Build clusters, score them, and select one or two questions that are both verifiable and answerable.
- After publishing, see whether new comments reduce the original question or reveal a later-stage problem.
After several weeks, this becomes a more valuable topic library than any collection of viral templates. It records the jobs users are trying to do and the points where they stop. AI lowers the noise, leaving your team to verify facts, show the method, and resolve a real barrier.