Instagram Competitor Analysis: Find Which Posts Worked for Any Account

1. Choose the accounts, and choose small ones

The hardest part. Ask your audience with a story sticker which accounts they follow, tap Instagram's similar-accounts button, and search the words your buyer would type. Keep 5 in your niche and 2 outside it, all small and fast-growing.

Why: the explore algorithm pushes up-and-comers to cold audiences, and that's the audience you want.

2. Tier 1: open Competitive Insights (free, in the app)

In your professional dashboard, open Competitive Insights. It tracks up to 10 public accounts over 3 months: posting volume, format split, follower growth and per-post engagement (Social Media Today covered the launch).

The best part is competitors' trial reels, the posts they're testing. Look for reels that appear twice or have an oddly low view count. No scraper returns this.

No button? You don't have the rollout yet, so go to tier 2. (Detail from heyDominik's walkthrough. App menus change, so check your own.)

3. Tier 2: 20 rows and one division

This is the whole idea. One row per post, 20 per account: account, permalink, plays, first caption line, length in seconds, and whether it ends with a comment keyword. Add a median per account, a ratio per post, and sort by ratio.

Use MEDIAN, never AVERAGE. One 100x post lifts the mean so far that every other winner scores below 1. Paid benchmarking tools report averages.

Rather paste rows into ChatGPT or Claude? Use the sheet prompt below.

account_median  =MEDIAN(FILTER($C$2:$C, $A$2:$A = A2))
post_ratio      =C2 / D2

# A account   B permalink   C plays   D account_median   E ratio
# then Data > Sort range > E descending
# never AVERAGE here: one 100x post buries every other winner

4. Read the top 20 and write one paragraph

For the top 20 by ratio, note the share ending with a comment keyword, the dominant format, the median length, and the repeating mechanic.

Mine, from the 422 posts in my own pull: 16 of the top 30 (53%) end with a comment keyword, the trigger behind comment to DM automation. Median length is 52 seconds, range 8 to 170, and only 12 of 30 fall in the 45 to 75 second band people quote, so length predicts little. 23 posts beat 5x and 105 beat 1.5x: expect about 5% real outliers. Recompute before you quote: my keyword share went from 77% to 53% when I ran the count.

python3 - <<'EOF'
import json, re, statistics as st
posts = json.load(open('posts.json'))
top = sorted(posts, key=lambda p: p.get('outlier_ratio') or 0, reverse=True)[:30]
kw = re.compile(r'\b(comment|drop|type|dm)\b\s+(the\s+word\s+)?\W{0,2}[A-Za-z]{3,15}', re.I)
n = sum(1 for p in top if kw.search(p.get('caption') or ''))
d = [p['duration_sec'] for p in top if p.get('duration_sec')]
print(f'comment keyword: {n}/30')
print(f'length: median {st.median(d):.0f}s, range {min(d):.0f}-{max(d):.0f}s')
print(f'in the 45-75s band: {sum(1 for x in d if 45 <= x <= 75)}/30')
EOF

# on my 422-post file this prints 16/30, median 52s (8-170s), 12/30 in band

5. Tier 3: scrape it when 10 accounts and 3 months run out

Apify's Instagram scraper takes profile URLs and returns plays, likes, comments, caption, timestamp, product type and video URL. Check the actor and price first. When I checked on 8 September 2026, the actor page read $0.0027 per result on the free tier, so my 9 accounts at 60 posts cost about $1.46. (apify~instagram-post-scraper is the per-post version.)

Send the token as a header from an environment variable. Apify's API docs accept either, and the query parameter leaks into shell history and proxy logs.

In Claude Code, add Apify's MCP server (a connector that lets Claude run the scraper) and run /mcp to sign in, as Claude Code's MCP docs describe. In ChatGPT, run the scrape in the Apify console and upload the JSON.

# verify the actor and today's price before spending anything
curl -s https://api.apify.com/v2/acts/apify~instagram-scraper \
  | python3 -c "import sys,json;d=json.load(sys.stdin)['data'];print(d['username'],d['name']);print(d['pricingInfos'][-1]['pricingModel'])"

export APIFY_TOKEN=...   # env only, never in a chat
cat > input.json <<'J'
{"directUrls": ["https://www.instagram.com/handle1/", "https://www.instagram.com/handle2/"],
 "resultsType": "posts", "resultsLimit": 60}
J
curl -s -X POST https://api.apify.com/v2/acts/apify~instagram-scraper/run-sync-get-dataset-items \
  -H "Authorization: Bearer $APIFY_TOKEN" \
  -H 'Content-Type: application/json' -d @input.json > posts.json

# Claude Code only
claude mcp add --transport http apify https://mcp.apify.com
export DEEPGRAM_API_KEY=...   # transcription, env only

6. Count rows per account before you trust a median

This cost me a week. In my own scrape, 7 of 9 accounts returned 60 posts and 2 returned exactly 1. A 1-row account's median equals its only post, so its outliers vanish and nothing errors.

Count every time, drop or rescrape short accounts, and make any agent show you the count first.

python3 - <<'EOF'
import json, collections
posts = json.load(open('posts.json'))
for creator, n in collections.Counter(p['creator'] for p in posts).items():
    flag = '  <-- partial, do not trust its median' if n < 20 else ''
    print(f'{creator:24s} {n:4d}{flag}')
EOF

7. Transcribe the reels so you can read the hook

The caption holds the call to action. The transcript holds the hook. Send each video URL to a transcription API such as Deepgram's pre-recorded audio endpoint and store the text with the post. Reading hooks this way pairs with writing hooks for reels.

Expect gaps: 349 of 422 posts came back with a transcript in my pull, since 26 were carousels and some videos were unreachable.

8. Tag one mechanic per post, most specific rule first

A keyword rule list reads caption plus transcript and assigns one tag: job killer, free alternative, dashboard reveal, hidden feature, curated number, contrarian, absurd analogy, leaked secret, urgency, platform tactics, story, other.

Order matters: first match wins, so put the most specific rules first and other last.

Why: you can filter to one mechanic and read 6 transcripts back to back.

9. Remix one post, all the way, on paper

Do this once before you build anything else.

Their post (property): you don't need cash to buy the land, then get a loan to build the house. That's a misconception. You can get one loan for the land and the construction together.

The structure: break a myth, give one shortcut, kill the objection, end on a DM ask, no face in the first frame.

My version: you don't need to post every day to grow. That's a misconception. You grow faster posting less and tracking smaller creators' outliers instead of copying the big ones.

Same skeleton, no shared sentence.