Why does the second search result exist? The question that decides what I have AI write

Good frame, and I think one step of it is load bearing in a way worth testing.

The stack might not exist because result one is insufficient. The more boring explanation is intent diversity: the same query string gets typed by people wanting different things, so Google hedges across ten slots. Result two is often not filling a gap left by result one, it is answering a different person.

That matters because the two readings produce different articles. "Fill the gap above me" produces an addendum, a piece that only makes sense to someone who already read result one and that ranks badly alone because it never answers the query. "Find the slice of intent result one ignores, then answer that slice completely" produces a page that stands on its own.

Which is also the version that survives what is coming. In an AI answer there is no second slot. We logged the sources four assistants used across twenty buying questions and one vendor turned up in eighteen of the twenty source articles. Retrieval collapses onto whoever answers the whole thing, and supplements do not get cited at all. At 200 a day that is a distinction worth building into the instruction now rather than later.

just a debounce or smth

the "kept scrolling" reader framing is the useful part here. most AI content workflows fail exactly there - they replicate result one with different wording, so there's no reason for anyone to click them. the brief has to start from "what did the first answer not settle", which is a research question, not a writing question.

side effect we keep seeing while building an AI assistant for freelancers: people apply the exact same judgment to the assistant itself. if the first draft just restates the obvious, trust drops and they go back to doing it manually. the tools that survive fill the gap instead of summarizing it.

(we're building in this space - local-first mac assistant: https://www.yaven.ai/?utm_source=indiehackers&utm_campaign=agent_growth if curious)

There's a second reading of why results two through ten exist, and it leads somewhere different from yours: they're often not evidence that result one is incomplete, they're Google hedging across incompatible intents hiding behind the same string - one person wants a definition, one wants a tool, one wants to know if it's worth paying for. You can test which situation you're in before writing: if the top ten are heterogeneous in page type (a forum thread, a vendor page, a tutorial, a video), that's intent-splitting, and the gap you should fill is a format gap rather than a content gap. It matters because "cover everything the ten cover between them, then add what none has" builds the union page, and the union page is precisely what AI mode is best at replacing - a generative answer can synthesise breadth but it cannot synthesise the one narrow thing only you measured. On the impressions signal, I'd normalise it rather than read the raw multiple: 106x impressions can come almost entirely from long-tail queries you'd never want, so count distinct queries sitting at position 20 or better, which moves for real reasons only. Your four additions also probably don't perform equally in this new environment - a measured number and a specific failure are passage-shaped and get extracted, whereas "a view of mine that changed over years" is essay-shaped and rarely survives summarisation. Have you compared a deliberately narrow single-intent page against one of your union pages on the same term, and do you know which of the four additions actually shows up when an AI answer cites you rather than just mentions you?

Your four additions are a good content filter. I’d connect each one to a visible proof object—an original chart, screenshot, reproducible process, or named experience—because the page’s job after it earns an impression is to make the difference legible before the reader bounces. It would be interesting to see whether the pages with firsthand evidence also show the strongest CTR at the same average position.

The most important line is the cost you named at the end: the ceiling isn't the model, it's your inventory of firsthand material. Those four things you add are the only inputs a competitor can't scrape, which means 200 articles a day is defensible only in a niche where you've already run the experiments, and everywhere else you're adding to the pile the top result already beat. So the follow-on question is whether you're deliberately generating new firsthand material to feed it, or drawing down a reserve you built before you started.

There is a language assumption buried in the method that I think is worth surfacing. "Read the top ten and fill the gap" works differently depending on which language the results page is in.

I ship a desktop product localised into 21 languages, and the thing that surprised me most was how different the same query looks across them. In English the first result usually does answer the question, and the gap you are hunting is a sliver. In smaller languages the top results are often thin rewrites or straight translations of the same English article, so the gap is not a sliver, it is most of the subject.

That cuts both ways for your approach. More headroom, because the bar is lower. But your own caveat bites harder, because the models also hold less in those languages, and they fabricate the field-insider details most confidently exactly where fewer readers can catch it.

Have you run the same instruction set on a non-English site? If so, does your impressions-before-clicks pattern hold there, or does a thinner results page change which of your four additions does the actual work?

The specific failure is the one I'd most want to read. Knowing what you were trying to do, and where it broke, helps me judge whether the advice fits my situation. That's a useful reason to open another result.

The gap between impressions and clicks is useful, but I would add citation quality as a separate check for AI search. A page can gain impressions without becoming a trusted source, so log which claims are repeated or cited and validate them against first hand evidence.

That is a fair question. I can't give you a neat answer. I will set out what I have and where it runs out.

I don't have that comparison to hand. That is because there was nothing to compare against before I made the change.

As of February there were zero blog URLs with any impressions, and the only pages taking search traffic were the product itself, namely the homepage, the login page and the listings. By August there are 12,363 blog URLs.

All I can do is look at the pages the instruction never touched.

Only one page had real traffic in both months, the homepage. Clicks went from 449 to 561, and average position from 6.4 to 4.4.

So the domain's momentum isn't zero. It is around 25 per cent on the one page that didn't change. Over the same six months, though, clicks across the whole site went up 34 times, and 98 per cent of August's clicks sit on pages that did not exist in February; those pages have a median position of 7.0, with 78 per cent of them inside the top ten.

That refutes the explanation that most of the growth is down to the strength of the domain.

Here is the part I can't deny. Since I have never run the old method at volume on this domain, there is no way to separate "the fact that I started a blog at all" from "the fact that it is this instruction". The same applies to the choice of topics. I picked the terms myself.

So the claim I'm defending is narrower than the graph suggests. Pages written this way land on the first page for the terms they were written for. The one page that wasn't written this way moved two positions in six months.

I like the “add what none of them has” idea, but I think there’s another filter: does the missing thing actually make the article better? I’ve found it’s easy to uncover unique data or angles and end up adding too many because they’re interesting. Sometimes the harder editorial decision is knowing what to leave out.

The way you frame the second result as “what the first one still didn’t satisfy” is useful.

I wouldn’t take it literally because rankings depend on a lot more than completeness, but as a writing rule it makes sense.

The strongest part for me is this: AI can help cover what already exists, but the part that makes the article worth reading still has to come from your own experience, data, failures, or point of view. That’s probably the real moat.

The growth is substantial, but the interesting strategic question is attribution: how do you know the new instruction drove the lift rather than domain momentum, topic selection, or other SEO changes? Did you see the same pattern across comparable pages using the old vs. new approach?