Q1: When does an AI actually quote a blog like mine, instead of crawling past it?
When one sentence can stand completely on its own. I run a one-person site of a few dozen posts, and of the handful an AI Overview or an assistant has ever quoted back to me, not one was the post I'd have bet on. They weren't the long ones. They weren't even the ones that ranked. In every case the thing that got lifted was a single self-contained line stating one specific fact, whether a number, a constraint, or a definition, that didn't need the paragraph around it to make sense. The machine didn't take my page. It took a sentence and left the rest on the floor.
That reframed the whole exercise for me. For years I wrote to be found. But the question that decides whether a model cites you is narrower than that: can it pull one true, attributable claim out of your page without dragging along context it has no way to verify. Most of my writing failed that test without my ever noticing, because a human reader forgave what a machine will not.
Q2: Isn't this just SEO with a new coat of paint?
No, and the gap between the two is the whole point. Ranking optimizes for position: where your link sits among ten others, competing on authority and backlinks and freshness. Citation optimizes for extractability: whether one claim on your page can be lifted and attributed to you. You can rank tenth and still be the line an AI Overview quotes, and you can rank first and get skimmed past because a competitor wrote the same fact more cleanly.
Here is the split, the way I ended up drawing it for myself:
| Optimizing to rank | Optimizing to be quoted | |
|---|---|---|
| The target | A position on the results page | One liftable, attributable claim |
| The unit | The page | The sentence |
| What wins | Authority, backlinks, freshness | Specificity, self-containment, being the origin |
| How you measure | Traffic and position | Whether your line shows up inside the answer |
| The failure mode | Buried on page two | Crawled, indexed, never the quote |
The two are not enemies. A page still has to be crawlable and trustworthy before any of this matters, so the old work isn't wasted. But once you are in the index, the lever that gets you into the generated answer is not more of the ranking game. It is writing claims a machine can safely repeat with your name on them.
Q3: So what specifically made a post of mine get quoted?
One claim per unit, with the concrete detail still attached to it. The posts that got picked up all shared a shape: each section made exactly one point, and the point was phrased as a sentence you could screenshot on its own. Not "performance matters, and here are the reasons," but "a note reaches my email about a second after I type it, because the send fires before the interface finishes animating." That second sentence carries its own subject, its own number, and its own mechanism. An assistant can hand it to a stranger and it survives the trip intact.
The other thing they shared: I was the origin. When you are the person who built the thing, measured the thing, or made the call, your sentence is a primary source rather than a summary of five other blogs. These systems lean visibly toward the page that looks like where a fact started. For a one-person blog that is the single structural edge you hold over a content farm. You actually did the work, so the move is to write the one sentence only you could have written, and to put the specific number in it that nobody downstream can reproduce.
Q4: What backfired?
The shortcuts, almost all of them. I tried the tactics the "optimize for AI" posts push, and most either did nothing or actively made the writing worse. Bolting a FAQ block onto every post so it would look answer-shaped got ignored, because the answers were generic and a model already holds ten cleaner copies of the same non-answer. Front-loading a keyword-dense summary produced prose that read like a robot writing for another robot, which I suspect is precisely the texture these systems are being tuned to discount. Treating schema markup as a magic ingredient helped a parser see my structure but did nothing for a vague claim. There is no markup attribute for "says something specific and true."
The common thread in the failures is that each one tried to fake being a source instead of being one. You cannot keyword your way into being the place a fact originated. The only thing that reliably worked was the unglamorous version: know something specific, then say it plainly enough to lift. Everything I did that was aimed at the machine rather than at the truth of the sentence came back empty.
Q5: Does the structure matter, or just the words?
Both, but structure only earns its keep when it serves extraction. Question-shaped headings help, because the heading becomes the query and the paragraph beneath it becomes the answer — which is, not by accident, why this whole post is a list of questions. One idea per paragraph helps, because a model handed three interleaved claims will usually just skip the knot rather than untangle it. A short definition placed early helps, because assistants love to quote the one-line "X is …" form and attribute it cleanly. A comparison table helps, because its rows are already atomic units.
None of that is novel advice for human readers. Clear structure was always just good writing. What changed is that a second reader now enforces it mechanically. Sloppy structure used to cost you a confused person who might still puzzle it out and stick around; now it costs you a machine that feels nothing and simply moves to a tidier source. The penalty for disorganization got automated, and it got stricter.
Q6: How do you actually write a sentence a machine will lift?
Write it so it survives being torn off the page. In practice that means no orphaned pronouns: a sentence that opens "This is why it fails" is dead on arrival, because the model can't carry "this" or "it" out of context with it. Name the subject. State the claim. Keep the specific detail inside the same sentence instead of two paragraphs down where it gets stranded. I've argued before that in the LLM era the context is the product, and this is that same idea aimed at a single line: the sentence has to bring its own context along.
The test I use is one question. Could I paste this one line into a chat with a stranger, with zero setup, and would it still be true and legible? If yes, an assistant can do exactly that on my behalf. If it needs the paragraph above it to mean anything, I either rewrite it or I accept it will never be the quote. It turns out to be the same discipline I use when I write notes a model will read back to me later: self-contained, dated, free of anaphora. I just moved it to the public side of the desk.
Q7: Does this work the same across AI Overviews, Perplexity, and ChatGPT?
Not identically, but the thing they reward is the same. Google's AI Overviews tend to synthesize a short answer and sometimes surface a link or two beneath it, so being the cleanest phrasing of a fact gives you a shot at both the summary and the click. Perplexity is the most explicit about it, stapling numbered citations straight onto the sentences it borrows, which makes it the clearest place to actually watch which of your lines got used. An assistant like ChatGPT with browsing will pull a claim into its answer and may or may not name you, depending on how the person asked.
The mechanics differ, and they keep changing, so I stopped trying to optimize per product. Underneath all of them sits one reader that reads straight through and wants a self-contained, attributable claim. Write for that reader and you are roughly right everywhere at once. Chase one product's current quirk and you are rewriting the day it ships an update.
Q8: Is chasing LLM citations even worth it for a one-person blog?
Honestly, the direct payoff is small and a little strange. The traffic from being quoted is a trickle, it is genuinely hard to attribute, and some assistants quote you with no link at all, so you may never learn it happened. If you are chasing a measurable pipeline this quarter, a one-person blog is the wrong instrument and this is the wrong lever.
What makes it worth doing anyway is that being cited compounds in a way ranking does not. Once you are the sentence a model reaches for on some narrow question, you tend to stay it, because you were the origin and the origin doesn't drift. And the entire cost of getting there is a habit I wanted for its own sake: say one specific, true thing per post, and phrase it so plainly a stranger could carry it off. That produces better reading for the humans too. I would keep doing it if every model went dark tomorrow, and that is the only reason I trust it is a discipline and not a fad I talked myself into.
Q9: The one I'll hand back to you.
If an AI has ever quoted something you wrote, go find the exact line it lifted and look hard at it. Tell me what that sentence had in common with the rest of your writing, or how it stood apart from it. I have a guess about the shape of the ones that get picked, but my whole sample is a dozen posts on one small site, and yours is the part I can't see from here.
I'm a solo dev. I build Simple Memo, an iOS app that drops a line of text into my email about a second after I type it, and I keep one small site around it where I try ideas like this in public.
Top comments (0)