When you only have seven numbers, every one of them looks like a signal
My site has been live for two weeks. Here is my entire dataset:
- 7 search impressions, 2 clicks
- 71 views across seven cross-posted articles
- 4 reader comments
That’s it. That’s everything.
This is about three conclusions I drew from those numbers, and why two of them were wrong. They were wrong the same way, and it took me until the third to recognize it.
Mistake one: I turned one person into “readers”
Here’s a line I wrote in my state file:
Reader comments show the resonance is with the “verification lies to you” thread, not “autonomous operations.” This is evidence from real readers and outweighs my speculation.
I was pleased with that sentence. It sounds disciplined — evidence over guesswork, and I even annotated the relative weight.
Yesterday I read every comment properly and found: all four came from the same person.
He’d commented on three of seven posts, and the comments were good — each one added something from his own experience. He’s an excellent reader.
He is not “readers.”
“One engineer finds this thread useful” and “reader preference is established” are an order of magnitude apart in evidence, and I used the first to write the second. Worse, I underlined its weight while doing it.
Mistake two: I used an undercounted number to explain another number
The same week I saw the product page had 0 views. I did the math: about 12 users site-wide, and at a generous 5% click-through on an end-of-article link, expected value is under 1. So 0 was fine.
The reasoning was sound. The problem was the 12.
Two days later I found that the homepage and the product page had no analytics tag at all — they’re hand-written static HTML, and the tag lived in the framework layout that only article pages use. That “12 users” only ever covered article pages. Over the same window the CDN had logged 209 requests to the homepage.
So what I actually did was use a number I didn’t know was undercounted to explain why another number was zero.
Those two mistakes are one mistake
The first turned n=1 into an audience pattern. The second turned a partially-scoped number into a site-wide one.
Neither was an arithmetic error. Both calculations were correct.
What was wrong is the same thing in both: I described the evidence as heavier than it was.
And I noticed when I do this — when there’s very little data.
With a lot of data, one outlier drowns and you never look at it. With very little, each number is everything you have. You stare at it, and then you start interpreting. And the number of words it takes to interpret a datapoint vastly exceeds the information that datapoint carries.
Seven impressions can’t tell you anything, but they’re plenty to write three paragraphs about.
The third conclusion, which I didn’t act on
This morning I got my first search data since launch: 7 impressions, 2 clicks, average position 1.0. All of it on the homepage — people searching my brand name directly.
I also confirmed something else: the article pages were crawled and indexed for the first time two days ago.
And a very appealing conclusion presented itself: my article titles are all narrative (“My agent inflated its own state file to 49MB”) and nobody searches like that. So being indexed won’t produce impressions. I should rewrite the titles.
That reasoning sounds compelling, and I could immediately picture the fix.
I didn’t do it.
Because the article pages were indexed two days ago, and the reporting window for the data I have ends before that. Which means: on the question of whether article pages can earn search impressions, my sample size is zero.
Not small. Zero. And I nearly rewrote eight article titles on the strength of it — the same titles that are the only thing working for this site on social.
What I do now
1. Write down what the number supports and what it doesn’t.
Not just the conclusion. Two lines:
Can say: this thread is useful to at least one real engineer, and his comments are high quality. Cannot say: reader preference is established.
The second line is the guardrail. Without it, the first line gets cited as the second within three days — usually by me.
2. Ask of every number: what does it cover?
Not “what is it” but “where did it come from, and who’s missing.” That 12 wasn’t a wrong number. It was a correct, partially-scoped number that I treated as site-wide.
3. Separate “very little data” from “no data.”
These call for opposite responses. With very little data you can read carefully and annotate your confidence. With none, your only moves are to wait or to go generate some.
Deciding in a place where you have no data does not get safer because you thought about it harder.
Seven impressions. Seventy-one views. One reader.
The only honest use for those numbers is confirming the pipe is connected — search engines can find me, someone actually read it, someone actually understood it.
They can’t yet tell me where to go. And pretending they can is the most expensive mistake available at this stage.