The worst hour of my month used to be podcast prep. Two hours a show, actually, researching guests and writing questions, and I dreaded it enough that I put it off until the last possible minute. What finally fixed it was not working faster. It was building a system that did the research so I could stop guessing at what would make a good episode.
Content calendars have the same shape of problem, and most social teams have not fixed it yet. You block out a Thursday afternoon, you open the planner, and you and two colleagues start pitching ideas at each other. Somebody suggests a myth-busting post. Somebody else says we should do a behind-the-scenes. By four o’clock, you have twelve tiles filled in and no evidence behind any of them.
Then the posts go out, and the comments come in, and the comments are full of questions you did not answer. Your community manager replies to the same one nine times that month. Nobody writes that down anywhere, because answering it is a support task and planning content is a marketing task, and those two things live in different tools with different owners.
So you brainstorm again next month.
The list you need is already written
Brooke Sellas runs social care for global brands at B Squared Media, and when she demonstrated her workflow on the first episode of The AI Social Playbook, she did something that reframed the whole exercise for me. She took a month of one client’s comments, DMs, and ad replies, and asked her AI which questions people kept repeating.
What came back was not a list of topics. It was a ranked list of actual questions, in the words customers used, sorted by how often they came up. Does it work in my automatic self-cleaning litter box? How much does this actually cost? How does the subscription work? Does it clump? Does it track outside the box? How does it compare to other brands? Does it actually control odor?
Brooke laughed when the first one appeared, because her team answers it constantly. Nobody on that account had ever thought to count it.
Now look at the sixth question. Customers were asking how the product compares to the competition, and they were naming the competitors themselves inside those comments. Brooke’s read was immediate: build a comparison landing page that includes every brand the customers named, and pull traffic from those competitors. Her buyers ran the keyword research for her and never sent an invoice.
How to run it yourself this afternoon
The mechanics are unglamorous, which is why they work.
Start with the export. Go into your social tool or the native channel analytics, find the inbox or messages report, and pull one month of conversations into a CSV. You want the message text, the platform, whether it was an organic comment or an ad comment, and, if your tool provides it, a permalink back to the live conversation. Brooke’s file has those permalinks, and they matter later.
Then write the context before you write the ask. This is the part people skip, and skipping it is why so much AI output reads like nothing. Brooke’s block names the category, the product, the buyer, and the date range: you are analyzing customer conversations from a kitty litter brand’s social media channels, the brand sells to discerning cat owners, and the data includes comments, DMs, and replies from July 2026.
Only after that does she give the instruction. Categorize these conversations by theme, and for each theme report how often it appears, the emotional tone, and what the customer is actually asking for or trying to accomplish.
That final clause is the one to steal. Emotional tone tells you somebody is annoyed. What they are trying to accomplish tells you what to publish.
When the themes come back, run the second pass. Ask for the top questions people asked that month and how you might create content that answers them. In Brooke’s demo, that returned the ranked FAQ list, and then a summary that paired each question with a suggested format: a compatibility graphic for the litter box question, a cost breakdown carousel for pricing, a short demo video or GIF where seeing it settles the argument.
Read that last part again, because it is the piece most people will miss. The output is not just what to talk about. It is what shape the answer should take, derived from the kind of question being asked. A yes-or-no compatibility question wants a graphic. A pricing question wants a breakdown. You are not choosing formats by what your designer has time for anymore.
Where it breaks
I have to be straight with you about the limits, because I hit the first one myself during the recording.
This workflow needs repetition to work. Brooke’s client generated 1,902 incoming messages in one month. My own accounts do not come close, and neither do most of yours. If you pull thirty comments, you will get thirty questions and no ranking, because a frequency sort needs frequency. Brooke’s advice is to tell your AI what you actually have and ask what a meaningful sample looks like for the question you are asking. It will work with a small set, and it will also tell you when the set cannot support the conclusion you wanted.
Second, verify before you publish. Brooke calls her audit the discount double-check: ask the model to show you the verbatim comments it is flagging, alongside its own summary, with the row references. Her export has permalinks, so confirming a flagged comment takes one click. If yours does not, that check gets slower and you should still do it. AI reads sarcasm as sincerity, and a content calendar built on a misread joke is a bad month.
Third, the model will try to help too much. When Brooke pulled up the verbatim complaints, it offered to draft replies to those customers. She said no and kept going. That is a small discipline that saves an afternoon.
One more thing worth naming. Brooke ran her whole demo in a plain chat rather than her social listening project, on purpose, so the project’s stored assumptions would not shape a live analysis. That sits in real tension with her own rule about loading up on context. Every sentence of context you write is also a steer. I do not think there is a clean answer, and I would rather tell you that than pretend there is.
The principle underneath the tactic
Zero-click is not a trend to react to. People stopped clicking the blue link, so the answer has to live where the question was asked. Everyone in marketing agrees on that by now, and almost nobody has changed how they decide what to publish.
What Brooke demonstrated is the missing half. If the answer belongs on the platform, then the question has to come from the platform too. Your comment inbox is not a support queue that happens to be noisy. It is the only place your buyers tell you, unprompted and in their own words, what is standing between them and a purchase.
You already own that data. You are probably deleting it every quarter.
When the next tool arrives, and it will, ask it the same question I would ask this one: does this help me find out what I do not know, or does it help me confirm what I already believe? Brooke said the valuable insights are the ones that surprise you, and she is right. I tell my AI to challenge my assumptions for exactly that reason, because I make plenty of them and I would rather be corrected on a Thursday than in a quarterly review.
Pull last month’s comments. Ask what people kept asking. Then go look at how much of your current calendar answers a question nobody sent you.
