We collected real posts to learn what "good" looks like. Here is what the data can and cannot say.

Ask an AI "is this post good?" and it will answer from its general sense of the internet. That sense is fluent and mostly average. For a judge to be useful, it should compare a piece with real posts from the kind of account the client actually is.

So we collected real posts. This is what we gathered, what it showed, and the limits I want to be clear about.

Why we did it

Our first scorer judged posts against a rubric. The rubric was reasonable but it was our opinion. A better test: how does this piece compare with posts that real audiences actually reacted to, from accounts of a similar size and kind?

We needed that for the judge, and for the writers too. A post written as a brand cannot borrow the shape of a founder's personal story.

How we did it

All data was read from LinkedIn through my own logged-in Chrome, read-only. Numbers are as displayed (reactions, comments, reposts). We recorded followers so we could compute an engagement rate, (reactions + comments) divided by followers, and sort accounts into size tiers.

We built four sets over two days:

The seeds loaded into the system now hold 121 entries for Mukika, 106 for Mabbit and 55 for IIPTA.

For IIPTA, a niche with very small engagement, we collected 13 posts by searching LinkedIn on topics like patent agent exams and trademark in India. The median was under 20 reactions. That is a finding, not a failure.

An honest failure first

Our first attempt at IIPTA benchmarks produced zero verified entries. Search tools returned titles and links but not post text or numbers. LinkedIn and Instagram fetches were refused. One direct fetch from the device was blocked as bot-check circumvention and I did not retry it. We added nothing we could not verify. An empty file is better than an invented one.

What the data showed

The first 20 posts (hook-level patterns, overlapping counts):

• Personal story or vulnerability opener: 11 of 20. It included the best engagement per follower: a 6,000-follower account with 355 comments.
• Concrete number in the hook: 7.
• Contrarian or bold claim in line one: 6.
• Milestone, announcement or "comment to get it": 5.
• Skimmable listicle: 3.
• Named-brand teardown: 2. That is our category, and it was nearly absent.

Mid-tier creators (36 posts):

• Personal or underdog stories beat analysis. One creator's ex-boss open letter reached 5.4% engagement; his recent B2B analysis posts were under 0.1%.
• Community and event recaps with tagged people reached 1.6% to 2.5% for a 6,900-follower account.
• A short quotable line from a well-known marketer got 1.1%; a career question from a 2,300-follower account got 2.9%, against 0.4% to 0.8% for her statute explainers.
• Accounts of 80,000 to 90,000 followers sat at 0.3% to 2.8%. For a small account, 1% to 2% is a good target. A large account gets about 0.7%.

IIPTA warning: statute explainers at 2,000 to 23,000 followers got 2 to 18 reactions (0.03% to 0.8%). The 7% to 14% results came from employer-network award posts, not education. Link-out posts from firms stayed under 0.1%.

Brands and media (64 brand posts):

• First-person personal stories: 3 of 64 posts (5%), and all three were employee posts the brand reposted.
• Humour and shared pain out-drew product posts about 3 to 5 times. A HubSpot in-joke post got 1,174 reactions. A Semrush study post got 80.
• Reports and studies got the least reach.
• Media pages get tiny engagement: one outlet had 25 reactions on 805,000 followers.
• India media runs wire stories with numbers in the title. US media runs "how and why brand X did Y" cases and puns.

The implication we drew: when a client speaks as a brand, benchmark against brand and media rows. Use creator rows only for hook style, never for story structure. That is why each client now has a "Speaks as" setting: brand, founder or both.

The limits, and I mean them

• Small and not random. Many mid-tier "top posts" come from featured sections and their age is unknown.
• Engagement is as displayed. Saves and impressions are not visible, so educational posts are understated.
• Post permalinks were not available. We can find a post by its excerpt, not its URL.
• The "why it worked" notes are my inference.
• LinkedIn only. Zero Instagram entries. Instagram analytics are hidden at post level.
• No board game, workshop or edu-game brands. The first-pass doc listed this gap.
• Excerpts are opening lines only.

How it shaped YuktiBOS

• Posts are scored against benchmarks matched to size tier and "speaks as".
• A later check showed the benchmark-driven parts of the judge (engagement, r = 0.55 with readers; craft, 0.66) tracked human opinion better than the conversion and truth parts. The benchmark idea held up; the rest needed rebuilding.
• The next build plan stores a benchmark-set version on each scorecard, so old pieces can be rescored against newer sets. I have not confirmed it is live.
• Instagram benchmarks and edu-brand posts are still a gap.

Numbers at a glance