I opened my LinkedIn notifications to inspect the comment program I started on 11 August.
The repository audit records show about 780 completed comments or handled posts through 7 September. Some early records count handled posts instead of read-back comments, so I treat that as an operational estimate, not a perfect analytics total.
The notification sample was much smaller. It showed impression counts for four recent comments:
| Comment subject | Impressions |
|---|---|
| A ten-layer production system | 377 |
| A free vLLM and Dynamo academy | 449 |
| Long-horizon OCR on messy scans | 3,066 |
| A single $100,000 bug bounty | 12,033 |
| Total | 15,925 |
One comment produced 75.6% of all measured impressions in this sample.
That does not prove a universal rule. Four notifications are not a clean experiment. LinkedIn selected which milestones to show, and each source post had a different audience.
It does show me what to inspect.
The best comment had a precise gap
The $100,000 bounty post said one vulnerability submission earned six figures.
My comment did not restate the achievement. It asked what made that one finding so expensive:
Six figures for a single submission usually means one specific class of bug done right. What made the finding so expensive is something I’d personally find worth knowing.
That comment received 13 reactions and 12,033 impressions. The author replied with his full technical write-up.
The comment worked because it named the missing piece. The post gave the result. The comment asked for the mechanism.
That made the reply useful to the author, to me, and to anyone reading the thread.
Specific curiosity beat polite curiosity
The two smallest comments ended with a form of what I would love to know.
They were not empty. Each named a technical area. But the questions were broad:
- Which system layer fails most often?
- How hands-on are the exercises?
The OCR comment was sharper. It asked where one-shot OCR fails on messy scans. That comment reached 3,066 impressions and received a reply about possible project work.
The stronger questions have three parts:
- A concrete claim from the post.
- A boundary, failure mode, or mechanism that the post leaves open.
- A question the author can answer from direct experience.
The weaker pattern has only curiosity. It sounds pleasant, but it gives the author less to work with.
Reach was not the only useful signal
The notifications also contained replies about Traefik failures, agent verification, stale retrieval chunks, and access controls. One author suggested connecting about projects after the OCR comment.
Those replies matter more than a view with no next step.
I now separate three outcomes:
- Distribution: comment impressions and reactions.
- Learning: a reply that adds technical detail.
- Relationship: a reply, profile visit, connection, or project conversation.
A comment can succeed in one category and fail in another. One large impression count should not hide that difference.
The volume is ahead of the measurement
About 780 handled items in four weeks is a lot of activity. Four notification counts are almost no measurement.
That is the uncomfortable lesson.
The program recorded inputs carefully: post type, author, draft, reaction, approval, and completion. It did not record enough outcomes per comment. This makes it easy to optimize the writing process while guessing about the result.
The next batch should capture a small result table after 48 hours and again after seven days:
| Field | Why it matters |
|---|---|
| Comment impressions | Distribution |
| Reactions | Lightweight agreement |
| Author reply | Conversation |
| Useful technical detail | Learning |
| Profile views after batch | Intent |
| Connection or project follow-up | Relationship |
I do not need every metric for every comment forever. I need enough paired data to compare question shapes and post types.
What I will change
- Use fewer broad
what I would love to knowendings. - Ask about a mechanism, limit, or failure mode.
- Track replies and relationships beside impressions.
- Measure batches before changing the writing rules.
- Keep notification numbers labeled as a selected sample.
The 12,033 number caught my attention. The reply with the technical article was the better result.