What a Google Quality Hit Actually Looks Like: Our Own Search Console Data

On June 23, 2026, this site earned 662 impressions in Google Search. On June 25, it earned 12.

Not a dip. Not a wobble. A 98% collapse in 48 hours, and two and a half months later, it has not recovered. This post is the full post-mortem, with our real Search Console numbers, because almost every “traffic drop recovery” article you’ll find is written by an agency about an anonymous client, with the embarrassing details sanded off. Ours are all here.

If your impressions just fell off a cliff, this is what the pattern looks like from the inside, what we ruled out, what the data says caused it, and the recovery plan we’re actually running. It’s a diagnosis-and-plan post, not a victory lap. We’ll publish the follow-up when we have recovery data, whichever way it goes.

The shape of the collapse

Here’s the daily impression data from our Search Console export (web search, last 3 months):

Period	Daily impressions	Avg. position	What was happening
June 7–14	141 → 570 → rising	31–56	Google testing the site broadly
June 15–23	612–901/day (peak: 901 on June 17)	48–55	Maximum visibility
June 24	113	48	The drop begins
June 25	12	41	Floor
June 25 – Sept 6	1–25/day, usually under 10	erratic (2–94)	Flatline, 2.5 months and counting

Three things about this shape matter for anyone pattern-matching their own drop.

It was a step function, not a slide. Seasonal decline, content decay, and lost rankings all look like slopes. A one-day, roughly 98% impressions collapse means a switch flipped: a site-level classification, a manual action, or a technical block. Nothing else moves that fast.

Positions went erratic, not uniformly worse. After June 25 our average position bounces between 2 and 94 day to day. That’s not “we rank worse.” That’s “we barely appear at all,” so a handful of stray impressions swing the average wildly. Small denominators, noisy averages.

The rise predicted the fall. Look at the first two weeks: impressions climbed fast while average position sat around 45–55 and CTR sat near zero. Google was testing hundreds of our pages on page 4 through 6 and finding that nobody ever chose us. In hindsight, that was the audition. We failed it.

What we ruled out first

Before blaming an algorithm, we checked the boring explanations. You should too, in this order:

1. Indexing blocks. Homepage returns HTTP 200. Meta robots says index,follow. robots.txt allows search crawlers (we use Cloudflare’s Content Signals to restrict AI training, but search=yes is explicit). Not a technical block.
2. Site down during the window. No outage on or around June 24–25.
3. Security or hack. No injected pages, no strange queries in the export. The query list is all on-topic.

That leaves Google-side revaluation. One check we’d tell every reader to run before anything else: Search Console → Security & Manual Actions → Manual actions. A drop this severe and this fast is consistent with either an algorithmic reclassification or a manual action, and the recovery paths differ completely. Do not skip this check.

We ran it. Clean: “No issues detected.” No manual action, no manual reviewer flagged this site. That points the diagnosis firmly at an algorithmic, site-level quality reclassification, the kind our own scaled-content footprint would trigger without any human ever looking at the site. It also means the fix is entirely in our hands: no reconsideration request to file, no waiting on a review. Prune the footprint, fix the URL hygiene, and let the next crawl and reassessment do their work.

What the data says actually caused it

This is the uncomfortable part, and it’s the reason this post exists. When we audited what Google had been showing during those two high-visibility weeks, the pattern was impossible to miss.

The pages that earned the most impressions earned zero clicks, and they were permutations. Our top pages by impressions:

Page	Impressions	Clicks	Avg. position
/blog/ai-mode-seo-checking-software/	1,154	0	71
/copilot-seo/	758	0	69
/tools/keyword-density/	601	0	75
/blog/ai-seo-tools-visibility-optimization/	493	0	56
/blog/ai-translators/	471	0	62

The query report is worse. Whole families of near-identical queries: “ai mode seo checker,” “ai mode seo checkers,” “ai mode seo checking tool,” “ai mode seo checking tools,” “ai mode seo checking software,” “ai mode seo tracking software.” Each with 40–150 impressions, all at position 60–90, all with zero clicks. Then the same family again with “copilot” swapped in.

That’s not a keyword strategy. That’s a footprint. Google’s systems are explicitly built to detect content produced at scale to cover keyword permutations, and our own export reads like a training example for that classifier.

Meanwhile, the pages that earned actual clicks were the opposite kind of content:

Page	Clicks	Avg. position
/blog/google-trends-ai-fanout-data/ (original data analysis)	8	~7
Homepage	6	13
/blog/best-seo-mcp-servers-github/ (hands-on tooling)	2	11

A set of specific, genuinely useful pages also held top-3 positions on small volume: NAP consistency for AI local SEO (position 1.9), our AI citation rate research (1.9), defensive SEO for brand narrative (1.8), AI SEO for startups (2.4).

The verdict the data forces on us: the site could rank. Google’s classic systems liked our specific, first-hand content. But the permutation layer stapled on top of it poisoned the whole domain. The timing (our test window ran through late June, weeks after the May 2026 core update began rolling out) is consistent with a site-level quality reclassification, though Google publishes rollout windows, not per-site verdicts, so we’re naming the pattern, not the exact system.

The recovery plan we’re running

We’re treating this as a rebuild, not a tweak. Four workstreams, in order.

1. Prune before publish. The site sheds its permutation pages: the [engine] seo [checker/software/tool/tracker] family and the thin [niche] AI SEO templates that never earned a click. Consolidate the few with genuine substance, remove or noindex the rest. Adding good content on top of a bad footprint doesn’t work; the site-level classification applies to everything, including the new stuff.

2. Fix URL hygiene. Our export shows the same posts collecting impressions on multiple URL variants: with and without trailing slashes, plus stray http and www versions. Every one of those splits signals. Site-wide 301s to a single canonical form, self-referencing canonicals, one URL per post.

3. Publish only what the data endorsed. Original data analysis, hands-on tooling walkthroughs, and direct answers to specific questions. Those three content types earned every single one of our clicks. Capped at 2–3 posts a week, because publishing velocity is itself part of the scaled-content footprint. Our notes on structuring posts for AI citation describe the format we’re standardizing on.

4. Measure the recovery honestly. Weekly log of impressions, indexed page count, and the new generative AI performance report baseline. If nothing moves within a full core-update cycle after the prune, we’ll say so in the follow-up.

What we’d tell you to do with your own cliff

• Date the drop precisely (daily export, not the default chart). A step function means classification, action, or block; a slope means competition or decay. The fix depends entirely on which one you have.
• Check manual actions before theorizing. Thirty seconds, completely different playbook.
• Rule out technical causes with curl, not memory. Status code, meta robots, robots.txt, canonical. Verified on the live site, today.
• Audit your zero-click impressions. Pages that got shown hundreds of times and chosen never are Google telling you what it tested and rejected. That list is your prune list.
• Find what still earns clicks. Whatever survived is your rebuild blueprint. Ours was small, specific, and data-driven. Yours will be whatever your audience actually chose when given the option.

This is part 1: diagnosis and plan, published while still in the hole. Part 2 will show the same charts after the prune, whatever they look like.