I run a small AI code review tool. Over the last three months my site got 62 clicks from Google search. Not 62 thousand. 62.
Most launch postmortems are written by people who succeeded. This one is written by someone still in the trench, with the raw CSV exports open. Here is what three months of Search Console data actually taught me - including two wrong conclusions I almost shipped.
1. A report written from memory is fiction
I once wrote an SEO report from what I remembered of the dashboards. When I later sat down with the actual CSV exports, almost every number was wrong. Positions, impressions, which pages were "winning" - all of it. The conclusions happened to survive; the evidence behind them didn't.
The rule I follow now: no analysis without the export loaded. Memory smooths curves into stories.
2. "3 months vs 7 days" is a trap
My first report compared a 3-month average against the last 7 days and concluded "monotonic decline". Then I bucketed the daily data by week and saw something different: a floor in late September, then three weeks of recovery. The "decline" was rotation inside unequal windows, not a trend.
Worse, the rotation was directional. Pages I had recently consolidated into hubs grew up to 5x their baseline. Old template landing pages lost half their visibility in the same weeks. An aggregate chart hid the only interesting thing happening.
Equal windows or nothing.
3. Two pages moving toward each other are not necessarily related
After I merged two weak landing pages into a hub page, the hub jumped 24 positions and the old flagship page dropped 24 - the same week. Textbook keyword cannibalization, right? I was ready to write a post about it.
One screenshot killed the theory. Filtered by the money query, the hub had 0 impressions in 28 days. The old page served 75 of 77 impressions for that query. The hub had risen on its own narrow queries; the old page was being thinned by Google for being thin. Two independent processes that happened to cross.
Correlation between two moving lines is not a mechanism. Check who actually serves the query before you restructure anything.
4. One mention beat six months of on-page SEO
Over the summer I ran six rounds of title and description edits on head terms. Result: zero movement on any head position. The pages ranked where they ranked, and meta text was not the bottleneck.
Then one external mention happened in early August: 14 clicks in a single day on my brand query - more than half of all brand clicks for the entire quarter, on a day when normal brand traffic was about 2 searches.
Distribution beats on-page tweaks. It is not close.
5. Page one with zero clicks means your snippet lies
For one day, Google put one of my blog posts at the bottom of page one. 100 impressions that day. 0 clicks.
The post could hold the position - the content was good enough. The snippet answered a different question than the one people asked. That is fixable, but only if you know which query it ranked for. And I don't: 80% of that day's impressions came from anonymized queries that Search Console never shows, in the UI or the API.
The queries report is a shop window. The real inventory is behind it.
6. You don't own your brand query
Searching my own product name, foreign projects sharing the name outrank me. Same fight as on GitHub search, where dozens of repositories use the same name and mine is not in the top 10.
If brand demand is small, squatters win by default. The fix is unglamorous: consistent naming, organization schema, and enough real mentions that the brand query has something to attach to.
What I am doing about it
- Stopped snippet edits on head terms - proven dead end for a low-authority domain.
- Deepened the one page that actually serves the money query, instead of spreading effort across 30 templates.
- Moved the effort budget to links and mentions: awesome-lists, writeups, communities. The 14-click day came from distribution, so distribution is where the hours go.
No growth hack at the end of this post. The honest summary of three months: the site is small, the data is noisy, and the only signals that survived scrutiny were "make fewer, deeper pages" and "get mentioned".
The tool behind the post: review-action, an AI code review GitHub Action. Product site: code-critic.com.
If you have run a Search Console postmortem on your own launch, I would love to hear what surprised you - especially the conclusions you had to walk back.
Top comments (0)