# Audience Questions on Social Media: 120 Already Asked (https://www.socialcrawl.dev/blog/audience-questions-in-comments)

> We listed 120 audience questions on social media. 109 of those rows are comments. Sorted into who, what, why, how, and vs. Comment language is not a country.

On 17 October 2026 we clustered questions people had already written about retinol. The response listed 120 audience questions on social media. The window on the call was 17 October 2025 through 17 October 2026.

Of those 120 listed rows, other is 64/120, what is 37/120, vs is 12/120, how is 6/120, and who is 1/120. Why did not come back, so that bucket is 0/120. The field `total_questions` is 100. One hundred is the wrong denominator. Using 64/100 would describe a cap. The clusters list 120 rows, and shares against 100 do not sum to 1.

The pages that rank for this query are prompt listicles. [Bitly's 2026 catalogue](https://bitly.com/blog/social-media-engagement-questions/) offers prompts to copy into captions, polls, Stories, and newsletters. The [News Product Alliance](https://newsproduct.org/blog/101-questions-to-unlock-audience) page from its 2023 summit is the same job as survey items. This post counts questions people already wrote.

Comment language on a single post is not a country.

## What questions do followers ask in the comments?

The questions followers ask in comments are the rows below, counted from text people already wrote. SocialCrawl returned these clusters on one topic call. On retinol, what questions do followers ask comes back as 120 canonical rows.

Every retinol row has `count` 1, so a mention-weighted cut matches the row cut. The buckets add as 64 + 37 + 12 + 6 + 1 = 120.

| Intent | Listed rows |
| --- | --- |
| other | 64/120 |
| what | 37/120 |
| vs | 12/120 |
| how | 6/120 |
| who | 1/120 |
| why | 0/120 |

The why cluster was not returned. We did not impute a why row, so the listed share is 0/120. One-decimal renderings of the other fractions are 53.3%, 30.8%, 10%, 5%, and 0.8%. They sum to 99.9% because each fraction was rounded on its own. Quote the fraction.

`other` is residual intent. It is the largest bucket because clustering parked text that did not win who, what, why, how, or vs. The rows are still questions. One of them is quoted in the next section, spelling unchanged.

A pre-cluster kept count of 160 is not the row count. The response also lists 7 Reddit threads and 40 YouTube threads. Those thread counts are not comments read.

We classified each row from its first quote. YouTube comments are 98. Reddit comments are 11. Reddit posts are 5. YouTube watch-page rows with no author are 6. Naver is 0. Comments are 109 of 120, which is 98 + 11. The other 11 rows are post titles or video titles inside the same list. Those 11 are part of the 120 audience questions on social media this call listed. Treating the whole list as comments drops that split.

<KeyTakeaway>
Of 120 listed retinol rows, other is 64/120 and what is 37/120. vs is 12/120, how is 6/120, who is 1/120, and why was not returned (0/120). `total_questions` is 100, so 64/100 is the wrong quote for that mix.
</KeyTakeaway>

## Which of those questions are who, what, why, how, or vs?

Each label is an intent key. The text on a row is verbatim. The cluster text rule is `shortest_verbatim_member`, so a folded row shows the shortest wording, copied as written. `vs` is the comparison label. This pull also placed some review asks in that cluster.

Retinol who is a single row. Caps and spelling stay as returned.

> Hi. Except for the products from Joseon, the rest of the brands TEST on ANIMALS. Can you recommend Cruelty Free brands. Thank you!

This `other` row is still a question, spelling unchanged.

> Do you use sunscreen?

Three `vs` rows from that top five, copied as returned. The Celimax line and the rupee line compare products. The Kiehl line asks for a review and still says rwtinol.

> Celimax vs Medicube: which retinal booster actually wins for wrinkles? (pharmacist)

> ₹2000 Korean retinal vs ₹200 Indian Retinoid . Which one to use?

> Thank you. Please review Kiehl’s and No. 7 rwtinol

Turning those rows into [social listening leads](/blog/social-listening-leads-api) is a different post.

The second topic is 에어프라이어. Retinol rows are YouTube and Reddit. Air-fryer rows are Naver Q&A 60, YouTube comments 48, and 1 YouTube watch-page row with no author. The surfaces differ, and the split is not an English audience versus a Korean audience.

That response lists 109 rows. `total_questions` is 100. The `count` fields sum to 120, because a few wordings were folded. The fractions below use the 109 listed rows. Mention shares, `count` divided by 120, are supporting only.

| Intent | Listed rows | Mention share |
| --- | --- | --- |
| other | 40/109 | 41/120 |
| what | 27/109 | 36/120 |
| vs | 20/109 | 21/120 |
| how | 19/109 | 19/120 |
| why | 2/109 | 2/120 |
| who | 1/109 | 1/120 |

Listed rows add to 109 (40 + 27 + 20 + 19 + 2 + 1). Mention counts add to 120 (41 + 36 + 21 + 19 + 2 + 1). The same clusters over `total_questions` are 40/100, 27/100, 20/100, 19/100, 2/100, and 1/100, which sum to 109/100.

The only wording with `count` above 2 is 에어프라이어 추천 (what, count 9). Gloss: air fryer recommendation. Three rows sit at count 2. 에어프라이어로 고구마 굽기, 최적의 시간은 얼마인가요? is what. Gloss: best time to bake a sweet potato in an air fryer. 에어프라이어 vs 오븐, 고구마 굽기에는 뭐가 더 좋을까요? is vs. Gloss: air fryer versus oven for a sweet potato. 코다리조림, 에어프라이어 활용 가능할까요? is other. Gloss: can an air fryer handle braised dried pollack. The who row is 이 에어프라이어 브랜드 알려주세요. Gloss: tell me this air fryer brand.

We left `other` rows in `other` even when they look like what or how. A larger how or vs fraction on this topic is not a language effect. Only 48 of 109 air-fryer rows are YouTube comments. The 120 audience questions on social media from the retinol call are a separate list. Adding 120 to 109 would invent a corpus this pull did not cluster.

## How mixed is the social media comment language, and is that a country?

The comment language mix here is one post at a time. A social media comment language share is the language of labelled comments on that post, not a country.

Share equals counted divided by labelled comments, and labelled equals n minus abstained. The abstained count is part of the same statement as the share. counted/n is a different ratio. Each handle below is one post (`posts=1`).

| Handle | Language | Counted / labelled | n | Abstained | Share | Interval |
| --- | --- | --- | --- | --- | --- | --- |
| youtube mkbhd | en | 100/100 | 100 | 0 | 1 | 0.9737 to 1 |
| tiktok charlidamelio | en | 35/35 | 50 | 15 | 1 | 0.9282 to 1 |
| tiktok bellapoarch | en | 32/39 | 50 | 11 | 0.8205 | 0.6998 to 0.8997 |
| youtube MrBeast | en | 70/86 | 100 | 14 | 0.814 | 0.7357 to 0.873 |
| youtube BLACKPINK | en | 53/78 | 100 | 22 | 0.6795 | 0.5878 to 0.7591 |
| youtube BLACKPINK | ko | 14/78 | 100 | 22 | 0.1795 | 0.1192 to 0.2613 |

youtube mkbhd English share is 1, from 100 of 100 labelled comments, n=100, abstained 0, interval 0.9737 to 1. tiktok charlidamelio English share is also 1, from 35 of 35 labelled comments, n=50, abstained 15, interval 0.9282 to 1. That 1 is not 50 of 50.

On the BLACKPINK YouTube post, English share is 0.6795, from 53 of 78 labelled comments, n=100, abstained 22, interval 0.5878 to 0.7591. Korean on that same post is 14 of 78 labelled comments, share 0.1795, interval 0.1192 to 0.2613, with the same n of 100 and the same 22 abstentions. That Korean share is the language of comments on one YouTube post, not evidence the audience is in Korea.

On tiktok khaby.lame, 24 of 49 comments were labelled English, which is 24 of 35 labelled, n=49, abstained 14. `share` is null, `too_few` is true, and the interval is 0.5473 to 0.7975. On the TikTok handle nike, 21 of 50 comments were labelled French, which is 21 of 32 labelled, n=50, abstained 18. `share` is null, `too_few` is true, and the interval is 0.5109 to 0.7772. nike is the public handle that was queried. French is the largest labelled language on that post, and the response did not return a share for it.

French on khaby.lame does have a share, 0.2, from 7 of 35 labelled comments, n=49, abstained 14. That is not 7/49. English on TikTok nike has share 0.1875, from 6 of 32 labelled comments, n=50, abstained 18. That is not 6/50. Both clear the headline rule. Neither is in the table above.

Instagram natgeo English share is 1, from 12 of 12 labelled comments, n=15, abstained 3, interval 0.816 to 1, and `too_few` is false. Fifteen comments are too thin to lead.

The eight posts together contain 514 comments and 417 labelled comments. Abstained comments sum to 97, and 514 minus 97 is 417. Those totals are the sample size, not a comment language mix. We did not average the shares.

Divergence was false on all eight calls. TikTok follower samples are n=150, under 200. YouTube and Instagram returned `audience` and `follower_locations` as null. TikTok audience samples on these calls are 1,300, 1,298, 1,299, and 1,299. Those samples are not a country mix.

A social media comment language code is a language code. [W3C, following BCP 47](https://www.w3.org/International/questions/qa-choosing-language-tags), says the first subtag always designates a language, and a region subtag is optional. The [YouTube comment resource](https://developers.google.com/youtube/data/reference/rest/v3/comments) and the [Instagram IG Comment node](https://developers.facebook.com/documentation/instagram-platform/instagram-graph-api/reference/ig-comment) (updated 21 January 2025) have no country field. Per-comment sentiment is a different post, the [sentiment analysis API](/blog/comment-sentiment-analysis-api).

<CTA variant="hard" />

## How do you reproduce this social media audience research?

We pulled this social media audience research on 17 October 2026. The calls finished between 07:14 and 07:18 UTC.

Both topic calls returned HTTP 200, `coverage` 1, `partial_failure` false, and no warnings. `GET /v1/prism/audience-questions` with topic retinol and timeframe year charged 30 credits. The same path with topic 에어프라이어 charged 30 credits. Both windows ran from 2025-10-17 through 2026-10-17. Both returned `total_questions` 100. We would have swapped the topic if a call had listed fewer than 10 questions. We did not run a backup topic.

Methodology on both calls was `audience-questions-v3`. Clustering method was `intent_choice`, version `intent-choice/1`. The intent keys are who, what, why, how, vs, and other. `confidence_floor` was 0.5. Dedupe was canonicalized surface form. The cluster text rule was `shortest_verbatim_member`.

The table below is the pre-cluster candidate count.

| Topic | Texts seen | Judged | Kept | Kept by judge | Q&A-site rows | Heuristic fallback |
| --- | --- | --- | --- | --- | --- | --- |
| retinol | 580 | 580 | 160 | 160 | 0 | 0 |
| 에어프라이어 | 255 | 255 | 141 | 70 | 71 | 0 |

160 is not the 120 retinol rows. 141 is not the 109 air-fryer rows. The 71 Q&A-site rows are not the 60 classified Naver pages. Retinol lists 7 Reddit threads and 40 YouTube threads. The air-fryer call lists 20 YouTube threads and 100 Naver threads. Those thread counts are not comments read. The fetch legs returned HTTP 200.

Language calls used `GET /v1/prism/audience-language` with platform set to tiktok, youtube, or instagram, the handle, and `posts=1`. Eight calls succeeded, each with `posts_opened` 1, and none were retried. The handles were charlidamelio, khaby.lame, bellapoarch, nike, MrBeast, mkbhd, BLACKPINK, and natgeo. Instagram nike, Instagram blackpinkofficial, zachking, veritasium, and nasa were not called. The run stopped at eight successes.

At `posts=1` the charge was 8 credits on TikTok, 2 on YouTube, and 6 on Instagram. Four TikTok calls, three YouTube calls, and natgeo come to 32 + 6 + 6 = 44 language credits. Topic calls are 30 + 30. The total deducted is 104. The balance moved from 14,234 to 14,130. A follow-up balance read used 0 credits and still showed 14,130. Leg credits of 12 and 7 inside the question payloads were not added to 104.

A reader with a SocialCrawl key sends those requests to our data API and can rerun this count of audience questions on social media. Divide each cluster's row count by the listed rows on that response, 120 for retinol or 109 for 에어프라이어. Dividing by `total_questions` does not sum to 1. Quote `share` only when `too_few` is false and `share` is not null. On this pull that value is counted divided by labelled comments, n minus abstained.

Two topics and eight posts are the whole of this social media audience research. YouTube n=100 is the sample returned, not every comment on the video. `share` ignores abstentions, so the abstained count stays next to the share. We did not re-judge labels or impute anything. No call failed, and nothing was refunded.

<CTA variant="explorer-mid" />

## Frequently asked questions

### What intent mix shows up when audience questions are clustered?

On the retinol call from 17 October 2026, the audience questions on social media clusters list 120 rows and `total_questions` is 100. The listed mix is other 64/120, what 37/120, vs 12/120, how 6/120, and who 1/120. Why was not returned, so that listed share is 0/120. Quoting 64/100 uses the cap as the denominator, and those shares do not sum to 1. The air-fryer call lists 109 rows, with other 40/109, what 27/109, vs 20/109, how 19/109, why 2/109, and who 1/109, drawn from Naver Q&A 60, YouTube comments 48, and 1 YouTube watch-page row with no author. That is a different surface mix, not a Korean-audience comparison, and the two calls are separate lists. `other` is residual intent. On retinol it includes "Do you use sunscreen?"

### Is comment language the same thing as audience country?

No. A share is counted divided by labelled comments on one post. On the BLACKPINK YouTube post, Korean share is 0.1795, which is 14 of 78 labelled comments, with n=100 and 22 abstained, interval 0.1192 to 0.2613. English on that post is 53 of 78 labelled comments, share 0.6795, interval 0.5878 to 0.7591, with the same n and the same 22 abstentions. That Korean share is the language of comments on one YouTube post, not evidence the audience is in Korea. Divergence was not flagged on any of the eight calls. YouTube and Instagram `audience` and `follower_locations` are null. TikTok follower samples are n=150, under 200.

### How can someone else rerun this comment study?

Send `GET /v1/prism/audience-questions` with topic retinol and timeframe year, and the same path with topic 에어프라이어. Then send `GET /v1/prism/audience-language` with platform set to tiktok, youtube, or instagram, a public handle, and `posts=1`. This pull ran on 17 October 2026. Both topic windows were 2025-10-17 through 2026-10-17. Each topic call cost 30 credits. At `posts=1`, TikTok cost 8, YouTube cost 2, and Instagram cost 6, and the language calls summed to 44. With the topic calls, 30 + 30 + 44 = 104. Divide cluster rows by the listed rows on that call, 120 or 109, not by `total_questions`. Publish `share` only when `too_few` is false and `share` is not null. That share is counted divided by labelled comments, which is n minus abstained.

### What does a "vs" question mean in the cluster?

`vs` is the comparison label. On retinol it is 12/120 listed rows. One row, copied as returned, is "Celimax vs Medicube: which retinal booster actually wins for wrinkles? (pharmacist)". The same cluster also holds "Thank you. Please review Kiehl’s and No. 7 rwtinol", including the spelling rwtinol. On the air-fryer call, vs is 20/109 listed rows, including "에어프라이어 vs 오븐, 고구마 굽기에는 뭐가 더 좋을까요?" at count 2. The label holds comparisons and some review asks together.

<RelatedPosts currentSlug="audience-questions-in-comments" tags={["development", "learning"]} locale="en" />

<FaqJsonLd
  faqs={[
    {
      question: "What intent mix shows up when audience questions are clustered?",
      answer:
        "On the retinol call from 17 October 2026, the audience questions on social media clusters list 120 rows and total_questions is 100. The listed mix is other 64/120, what 37/120, vs 12/120, how 6/120, and who 1/120. Why was not returned, so that listed share is 0/120. Quoting 64/100 uses the cap as the denominator, and those shares do not sum to 1. The air-fryer call lists 109 rows, with other 40/109, what 27/109, vs 20/109, how 19/109, why 2/109, and who 1/109, drawn from Naver Q&A 60, YouTube comments 48, and 1 YouTube watch-page row with no author. That is a different surface mix, not a Korean-audience comparison, and the two calls are separate lists. other is residual intent. On retinol it includes \"Do you use sunscreen?\"",
    },
    {
      question: "Is comment language the same thing as audience country?",
      answer:
        "No. A share is counted divided by labelled comments on one post. On the BLACKPINK YouTube post, Korean share is 0.1795, which is 14 of 78 labelled comments, with n=100 and 22 abstained, interval 0.1192 to 0.2613. English on that post is 53 of 78 labelled comments, share 0.6795, interval 0.5878 to 0.7591, with the same n and the same 22 abstentions. That Korean share is the language of comments on one YouTube post, not evidence the audience is in Korea. Divergence was not flagged on any of the eight calls. YouTube and Instagram audience and follower_locations are null. TikTok follower samples are n=150, under 200.",
    },
    {
      question: "How can someone else rerun this comment study?",
      answer:
        "Send GET /v1/prism/audience-questions with topic retinol and timeframe year, and the same path with topic 에어프라이어. Then send GET /v1/prism/audience-language with platform set to tiktok, youtube, or instagram, a public handle, and posts=1. This pull ran on 17 October 2026. Both topic windows were 2025-10-17 through 2026-10-17. Each topic call cost 30 credits. At posts=1, TikTok cost 8, YouTube cost 2, and Instagram cost 6, and the language calls summed to 44. With the topic calls, 30 + 30 + 44 = 104. Divide cluster rows by the listed rows on that call, 120 or 109, not by total_questions. Publish share only when too_few is false and share is not null. That share is counted divided by labelled comments, which is n minus abstained.",
    },
    {
      question: "What does a \"vs\" question mean in the cluster?",
      answer:
        "vs is the comparison label. On retinol it is 12/120 listed rows. One row, copied as returned, is \"Celimax vs Medicube: which retinal booster actually wins for wrinkles? (pharmacist)\". The same cluster also holds \"Thank you. Please review Kiehl’s and No. 7 rwtinol\", including the spelling rwtinol. On the air-fryer call, vs is 20/109 listed rows, including \"에어프라이어 vs 오븐, 고구마 굽기에는 뭐가 더 좋을까요?\" at count 2. The label holds comparisons and some review asks together.",
    },
  ]}
/>

<DatasetJsonLd
  name="Audience questions in comments, pulled 17 October 2026"
  description="On 17 October 2026 we listed 120 retinol rows. 109 of those rows are comments. Comment language is not a country."
  url="https://www.socialcrawl.dev/blog/audience-questions-in-comments"
  dateModified="2026-10-17"
  variableMeasured={[
    "retinol_listed_rows",
    "retinol_comment_rows",
    "intent_share_of_listed_rows",
    "comment_language_share",
    "labelled_comments",
  ]}
/>