Our entire website got three clicks from Google last month. Not three thousand. Three. This is the post where we open Search Console, read the numbers out loud, and admit we built a pile of pages before doing the ten minutes of thinking that would have told us how they’d do.

We are publishing our own ugly numbers because that is the whole point of these notes.

The pattern we kept running

Most of our apps are, underneath, a structured dataset plus a private place to track your own copy of it. A while back we started putting the dataset half on the open web as free reference pages (maintenance schedules for over a thousand vehicles, plant and pet care timelines, and so on). We wrote about why in One Dataset, Two Products: do the structural work once, let it serve both the app and the web.

It felt like it was working. So when we shipped a line of on-device AI apps, we ran the same play for them: 28 new pages: grammar rules for Better Writing, travel phrases in fourteen languages for Pocket Lingo, journaling prompts for Daily Vibe. Same idea, same confidence. Publish the data, funnel the search traffic into the app.

Then we finally read the analytics properly, and the confidence did not survive contact.

What Search Console actually showed

Ninety days, the whole domain:

  • 1,115 impressions. 3 clicks. Average position 37.

Position 37 is the fourth page of Google. It means our pages are indexed and are being shown, yet they rank far too low for anyone to ever reach them. We are a young domain with almost no authority, and it shows in one number.

But the shape underneath that number is the interesting part:

  • The pages that get shown are the maintenance ones. “2015 Honda CR-V maintenance schedule.” “Aston Martin Vantage.” Specific, low-drama demand. The pages rank on page two or three, so they collect impressions and almost no clicks. Impressions without authority is a waiting room, not a result.
  • The 28 AI pages have zero impressions. Now, in fairness: they are only a few days old, and nothing ranks in a few days. So this is not a post-mortem. It is worse than that: the realization that we can already predict these pages will struggle, and that we could have predicted it before building them.

The three-question test we skipped

Here is the ten-minute check we should have run on the queries first. Not “is there search volume”; every keyword tool says yes. The questions that matter:

  1. Is it underserved? Someone has to be able to out-rank. “2015 CR-V maintenance schedule” is served by a handful of thin, ad-choked pages, so a small plain page can win. “Affect vs effect” is served by Grammarly, Merriam-Webster, and ten thousand others who have owned it for a decade. We cannot out-authority them; we shouldn’t try.
  2. Do we have an enumerable moat? Maintenance is make × model × year: thousands of specific pages, each one different, each backed by data we actually have. At the time of this report, our app-derived grammar content was eleven core rules. The web guide has since expanded to 41 rules (the same 11 app-core rules plus 30 site-authored lessons and examples), but the original lesson still holds: a modest set of familiar grammar topics is not the same moat as thousands of genuinely distinct records.
  3. Does the page win the click, or does the answer win it? This is the 2026 one, and it’s the killer. Ask Google “how do you say thank you in Japanese” and the AI Overview says arigatō right there. No click needed, ever. A maintenance schedule is a table you scan and save; the answer box can’t replace it, so the click survives. For a phrasebook page, the click is already gone before we rank.

Maintenance passes all three. The AI content fails two of them. We didn’t need ninety days of data to know that; we needed to ask the questions before we opened the editor.

So where do the AI apps get found?

Not Google. That’s the conclusion, and it’s freeing rather than grim. The on-device AI apps have their “wow” moment inside the app (Ask Monie turning “coffee 4.50” into a logged expense, Pocket Lingo translating with no signal), and moments like that belong on the App Store and in short video, where showing beats telling. We were trying to win a search battle for apps whose whole pitch is something a search result can’t show.

The maintenance content is the opposite: nobody makes a TikTok about a service interval, but they will type it into Google at 11pm before a road trip. Different apps, different front doors. We had them backwards.

The lesson

Score the queries before you build the pages. Underserved, enumerable, click-surviving. If a topic misses two of the three, the pages are craft with nowhere to go, and the channel for that product is somewhere else entirely.

We still like the maintenance pages, and we’ll keep them. But the next move there is not more pages. It is authority: the slow work of a few real links and time, so the impressions we already earn turn into clicks. And the AI apps go where they were always going to win.

We shipped the wrong way to feel why. Again. That’s what these notes are for.

More soon. We’ll keep them honest, three clicks and all.