What Our Automated GSC Miner Actually Found

A weekly script that reads Search Console and emails an authoring queue sounds like automation theater until you see what it returns. This is the field report for ours: what the miner is, dated first outputs across our 3 production builds, and the one diagnosis rule that has earned permanent residence in the pipeline [our data].

What is a GSC miner and why build one?

It is a scheduled pipeline that pulls a property's Search Console queries, extracts the question-shaped and underserved ones, scores them against the existing page inventory, and emails a ranked authoring queue — every week, without anyone remembering to look [our data]. Search Console earns the job because it is first-party, free, and the only demand data that describes your site rather than a tool vendor's panel.

It is also, honestly, a proxy. Search Console has no AI Overview dimension — Google's AI-features documentation covers eligibility and snippet controls, not per-query AI reporting — so question-shaped query mining approximates AI-shaped demand rather than measuring citations. The approximation methods we use on top of GSC are detailed in AI Overview impressions in Search Console.

How fast did Search Console return mineable data?

Faster than the folk wisdom says — day 4 on a domain that had existed for day 4. The fleet record, dated from each build's git history [our data]:

One caveat on comparing rows: the migration build inherited a domain with years of search history, so its day-6 figure measures the pipeline's speed, not a cold start. The greenfield row is the clean test of how fast a new domain generates usable signal.

The greenfield number is the one worth repeating: a brand-new domain generated actionable query data inside its first week. Waiting for "enough data" would have cost the 5 pages that shipped from it.

What did the first automated outputs look like?

Small, specific, and mostly not new pages. The insurance miner's first weekly report (2026-08-19) produced 1 comparison page — a head-to-head the query data showed demand for — and 3 snippet fixes on existing pages [our data]. Fixes outnumbered new pages 3 to 1 that week; one week on one build, reported as exactly that.

The split matters more than the counts. A naive miner treats every underserved query as a brief for a new page; the useful version distinguishes build-new (impressions with no matching page) from fix-existing (impressions, a matching page, and clicks that never come). The second category is where the diagnosis rule below does its work.

Just as important is what the record does not contain. We have no traffic or revenue outcome to attach to any miner-produced page — the pages are days old, and our record proves pipeline latency and output shape, nothing more [our data]. A field report that claimed more than that would be manufacturing the same unverifiable numbers this niche launders constantly; when the miner-produced pages have months of history, their outcomes will be reported with dates.

What is the position-3-zero-clicks rule?

Our triage heuristic for the fix-existing pile: when a query accrues impressions with the page holding roughly position 3 and zero clicks, the miner flags it as a snippet problem — the title, description, or extractable answer mismatches what the query is asking — rather than a ranking problem or a content gap needing a new page [our data]. The documented fix in our record is exactly that shape: retitle and rewrite the answer block to the question actually being asked.

There is a second suspect worth naming without blaming: Seer Interactive measured organic CTR down 61–70% when an AI Overview is present on a query (2025, Seer Interactive) — a good position with no clicks is also what click absorption looks like from inside Search Console. We state the rule's evidence tier plainly: GSC cannot show whether an AI Overview sat on any specific query, so per-row causation is unobservable, and our own record attributes the pattern to snippet mismatch, not to AI surfaces. Either way the cheap action is the same — rewrite the answer block — and its diagnostic logic is the same one worked through in our AI Overview approximation method.

Should you build an automated miner?

Probably not first, and we say that as the people who built 3. A monthly manual export answers the same questions at single-site scale: pull queries, isolate the question-shaped ones, score them against your pages. The full method — the exact extraction and gap-scoring steps our automation runs — is written up as a manual workflow in how to mine Search Console for AI-shaped queries, and it requires nothing but the GSC export button and a spreadsheet.

Automation earned its keep for us at 3-brand scale, where it closes the loop that makes a content system self-feeding — the property described in the authority engine. The honest state of the cadence: the miner is scheduled weekly on all 3 brands, and each produced its first automated report on 2026-08-19 — configured everywhere, 1 run deep [our data]. The honest sequencing: publish a citable library first, mine second, automate third. A miner pointed at a site with nothing worth ranking returns an empty queue, which is why the pages come before the pipeline in the operator's guide.