Aritra | Supaboard

I posted 20 replies on X yesterday as a growth experiment. Every reply logged in a database: which thread, what angle, views, likes, replies back. Day-one data: the reply I posted 1 minute after a thread went up got 27 views. The ones on day-old posts got 2. The parent post's trajectory matters more than what I write. Running this 4x a day for 30 days. Publishing the full dataset at the end.

Hmmm!

I thought I understood retention. Then I watched David Lieb’s YC talk. I was reading percentages without following the curves. In my latest video, I build a cohort dashboard with one prompt in Supaboard and explain what to look for beyond week-one retention.

171 replies in week one. 53% landed on threads that were already dead. Those 53% produced about 2% of the reach. Median: 8.5 views. Top decile: 100 to 400. The reply was never the variable. The parent thread was. Anyone getting better reach replying on slow threads?

700+ connectors is our own number and the least predictive thing on the page. Nobody churns over a missing connector. They churn because sales and finance define a closed deal differently, and the agent has to pick one. Has a connector count ever decided a tool for you?

Our BI Bench score is 96%. Best in the industry. Do the division. 4 wrong answers in 100. Ask 20 questions a week and one is wrong, with nothing on screen telling you which. Everyone markets the 96. Trust is decided by the 4. What should an AI analyst do when it is wrong?

Every AI analyst benchmark scores the query. None of them score the 2 to 4 weeks a Looker rollout burns before anyone asks the first question. Tableau and Power BI: days to weeks. The model was never the slow part. Which are you actually waiting on, the answer or the setup?

Week one on X: 5,900 impressions. Last seven days: 1,500. I post more now, not less. Output up. Reach down 75%. Easy read: the algorithm punishes volume. The one I avoid: week one was one good thread and the rest is filler. Reach fell while output rose. Which was it?

5,900 impressions bought 37 profile visits. Half of them followed. Profile converts at 50 percent. The click to get there converts at 0.6 percent. The advice is always write better posts. A good post never makes me wonder who wrote it. What makes you click a stranger's name?

Everyone shipped an AI analyst this week. The model was never the bottleneck. 700+ connectors in, what kills a question is finance and sales meaning different things by revenue. One definition is an opinion. Two is noise. Which do you want answering your board?

Three months in a row our search impressions went down. 17,578 a day in July. 13,372 in August. 11,934 in September so far. Clicks barely moved. 18.2 a day, then 15.2, then 14.8. That is 32% fewer impressions and 19% fewer clicks. Most of what we lost was never going to click anyway. In three days everyone posts a September update. Most of us will pick the window where the line goes up. That is not lying, it is choosing. Which number are you picking?

Two AI analyst launches crossed my timeline in five hours. Both demoed the same flow: connect the warehouse, ask a question, get a chart. Nobody demoed who is asking. Across 1,000+ companies the pattern is boring and consistent. The person with the question is the founder, the head of marketing, the ops lead. The person with warehouse access is not. So the question gets queued, reworded by someone else, and answered after it stopped mattering. An AI analyst behind the same access wall is a faster tool for the people who were never the bottleneck. If you are building one: are your weekly actives the data team, or the people who used to email the data team?

171 replies in week one. Median views per reply: 8.5. The top decile did 100 to 400 each. A tenth of the work carried nearly all the reach. The uncomfortable part is that the difference was not the reply. It was whether the thread was already moving when I got there. Volume was never the lever. Selection was. Where do you put your filter?

1,067,044 impressions on our site in the last 90 days. 733 clicks. 504 of those clicks were people typing our own name. Discovery is 229 clicks in three months. Branded search is a loop that feeds itself. Someone sees you on LinkedIn, Googles the name, clicks, and SEO takes credit for a decision made somewhere else. Strip the brand terms and what is left is what the content earned. Do you know your branded vs non branded click share without looking?

Everyone is cutting agent cost by routing cheap decisions to cheaper models. The biggest cost drop I have shipped had nothing to do with the model. One client was spending 4 hours, across multiple analysts, collating data before anyone could ask a question. That became 2 minutes. Analytics cost fell about 90%. The thinking stayed human. The spend was never in the fork. It was in everything that happened before the question was worth asking. Cheaper tokens per decision, or fewer decisions that need collating first. Which one is actually on your roadmap?

Agent count is the new headcount. One more agent does not fix a number on the P&L. It makes the org chart prettier. We connect data for 1,000+ companies. What breaks is almost never the model. It is that "active customer" is one number in Stripe and a different one in the CRM, and no agent is allowed to pick which one is real. A human settles that once and writes it down. Then the agents are useful. When your agents last disagreed, was the model wrong or the definition?

The reporting bottleneck is not the query. It is the reformatting. The analysis gets built once and rebuilt three times. Once for the deck. Once for the standup. Once for the board. Across 1,231 companies the shape repeats. The tool that wins is not the one that answers fastest. It is the one whose answer survives being moved somewhere else. Speed to insight is mostly solved. Speed to a slide someone reads is not. Where does your analysis actually die, in the query or in the slide?

Google put our site in front of 14,360 different search queries in the last 90 days. 123 of them produced a click. The advice is always "distribution is the game." Nobody mentions that distribution arrives as 14,360 rows nobody opens. Getting seen was never the hard part. Finding the 123 that mattered is. Most founders pick their next bet from memory, not from the list. When you chose what to build next, was it a number or a feeling?

When I first measured this account: 0.6% of the people who saw my posts clicked my name. About half of the ones who clicked followed. Every bit of growth advice I got was aimed at the first number. Wrong half. A 50% follow rate on profile visits means the writing works and the distribution does not. Tripling impressions just buys three times as many people scrolling past. Most founders have this shape somewhere in their revenue funnel and cannot see it, because the report puts the biggest number on top in the biggest font and buries the conversion step in a footnote. Which half of your funnel is actually broken: the one your dashboard shows first, or the one it makes you scroll for?

Connecting the data takes five minutes. Getting anyone to act on it takes a quarter. Looker quotes two to four weeks just to stand up. Everyone calls that the bottleneck. It was also how the org agreed the number was real. A number nobody watched get built is a number nobody defends in a meeting. Honest question for anyone who has rolled out reporting: did faster tooling make your team trust the numbers more, or less?

700 connectors is the easy part. Every source has its own idea of what a customer is. Stripe says subscription. The CRM says account. Support says email address. Point an agent at all three and you get three answers, each one correct inside its own source. The hard part was never retrieval. It was arbitration. Who owns the definition of "customer" where you work, and can you name them?