When Search Console goes quiet

Act one: page-one rankings that were nothing

For weeks, the platform’s biggest “wins” in Search Console were queries like "bbca.jk" "dec 30, 2024" — an Indonesian bank ticker plus a date, in quotes, at position ~2, hundreds of impressions a month. No human searches like that. Quoted exact-phrase queries in that pattern are a machine — something programmatically checking whether specific figures appear in Google’s index — and every one of them recorded zero clicks. Once we classified queries by shape (quoted phrases, site: operators, pasted URLs), under 5% of the platform’s recorded impressions turned out to be human-shaped. We ended up de-indexing the foreign-ticker pages the bots were probing; they were spending crawl budget on an audience of scripts.

Act two: the collapse that wasn’t

Then impressions fell roughly 80% week over week. On a chart that looks like a penalty. In the warehouse it decomposed cleanly: the bot block was decaying, and the human-shaped remainder was roughly flat — a handful of impressions a day. A team managing by the top-line graph would have “fixed” something that wasn’t broken, churning titles and pages in response to crawlers losing interest. The honest reading was less dramatic: the site’s real organic baseline was near zero all along, and now the data finally said so.

Act three: the report goes dark

Last week the query report stopped entirely: day after day, zero query rows for the whole property, while site-level totals still showed a steady trickle of impressions. Our first assumption was a broken pipeline — the ingest job, the API grant, the database. All healthy. The cause is documented but easy to forget: Search Console omits rare queries from the query dimension for privacy. When a site’s remaining demand is a long tail of one-off searches, every query is “rare”, and the query report rounds the site down to nothing. The totals and the breakdown are different datasets, and at low volume they diverge completely. (The workaround: the page dimension is not anonymized the same way, so a page-only export still shows where impressions land — that’s the next column in our warehouse.)

What keeps the numbers honest

Every discipline here is boring, which is rather the point. A nightly job copies the full Search Console export into a Postgres table we own — the API only keeps 16 months, and you cannot classify bot patterns retroactively out of a summary chart. Queries are bucketed by shape before anyone reads them. And no single source is trusted alone: Search Console (organic only, two days behind), our self-hosted Umami analytics (first-party, but blockable), and the CDN’s server-side counts (complete, but bot-inflated) each lie in a different direction, so a claim has to survive all three. Paid campaigns land on dedicated tagged paths precisely so they can never be mistaken for organic progress.

The generalisable lesson is about instruments, not SEO. Any metric system read below its design volume — analytics, sales pipelines, error budgets — mostly reflects its own thresholds and filters. Whether the reader is an analyst or an AI agent, it needs to know how each instrument fails before it is allowed to draw conclusions from one.