Why Specific Content Gets Cited by AI (And How to Prove It With Your Own Data)

"If you write an insightful blog post on a specific enough topic, and people link to it, you have a real chance at influencing everyone's LLM output in a year or so."

That single Bluesky post from Dan Abramov, the React core contributor behind Overreacted, set off one of the more useful SEO conversations of the summer. Tyler Gaw replied that he'd watched it happen to his own writing: "I've seen a couple of mine, not even that insightful, just specific, get pulled into them and used within like 6 months." Google's John Mueller weighed in with five words of endorsement: "Make more insightful & useful stuff." Search Engine Journal's Roger Montti covered the whole exchange on July 16.

The claim is worth taking seriously because it matches how these systems select sources. It's also worth treating as a hypothesis rather than a slogan, because the evidence behind it is a handful of anecdotes. The good news: this is one of the few pieces of AI search advice you can actually test on your own site, with your own data, in about a quarter.

Why narrow pages win citations

Start with what an answer engine does when someone asks it a question. It converts the question into a search, retrieves candidate pages, pulls the passages that seem to resolve the question, and cites a small number of sources. Not ten blue links. Usually two to five citations per answer.

Each step of that pipeline favors the specific page.

Query match. A post titled "Why useEffect fires twice in React 18 strict mode" is a near-perfect match for the one question it answers. A post titled "React hooks best practices" is a mediocre match for fifty questions. Retrieval scores the first page higher for its question every time, and AI answers are assembled one question at a time.

Competition. The broad topic puts you against documentation, Wikipedia, and every content-marketing blog that ever chased the head term. The narrow topic might have three genuine candidate pages on the whole internet. You'd rather be one of three than one of thirty thousand.

Extraction. Models cite what they can quote. A focused page states its answer in the first hundred words, and everything after supports it. A broad page buries eleven answers across 4,000 words, and each is entangled with the rest. Passage-level retrieval slices pages into chunks; the focused page basically is one clean chunk. We've covered the difference between being retrieved and being cited before: getting fetched is step one, and being quotable is what converts the fetch.

Notice what's absent from the pipeline: domain authority in the classic sense. It still helps, and links still matter (Abramov's claim includes "and people link to it"). But the selection step runs on passages, which is why Tyler Gaw's "not even that insightful, just specific" posts made it in. Specific beat authoritative. That's a genuinely different game from 2019 SEO, and it's winnable by small sites.

Specificity is measurable, not a vibe

Advice like "write focused content" usually dies in the gap between publishing and any observable result. If citations take 6-12 months to accumulate, most teams give up before the signal arrives, or worse, keep publishing broad content because nobody can prove it's underperforming.

The crawl layer closes that gap. Before any AI system cites a page, its crawlers have to fetch that page, and fetches are observable on your server today. Which URLs does GPTBot actually read? Does ClaudeBot return to a post after you update it? Did PerplexityBot find the new piece within a week, or has it never seen anything but your homepage? Every one of those questions has a factual answer sitting in your logs, and almost nobody looks, partly because GA4 and other JavaScript analytics can't see bots at all.

Crawl activity is a leading indicator, not a guarantee. A fetched page can still lose the citation to a better passage elsewhere, which is why the crawl-to-referral ratio is worth tracking alongside raw crawls. But the ordering is fixed: no crawl, no citation. A specific post that draws repeat AI crawler visits in its first month is in the pipeline. A post the bots skip for eight weeks is not going to appear in anyone's answer, and you just learned that in weeks instead of a year.

The specific-enough checklist

Specificity has a working definition you can apply before publishing. A post is specific enough when:

• The title is a question one real person would ask, phrased the way they'd ask it, and the post answers only that one. If the title needs "and" to describe the contents, it's two posts.
• The first paragraph contains the answer, not the wind-up. Someone who reads nothing else should leave resolved.
• It contains at least five facts nobody else has published together: your numbers, your measurements, your failure case, named tools with versions, dates. Passages get cited for containing information, and recycled information is already attributed to someone else.
• Every H2 is a sub-question of the main question. If a heading would fit equally well in a different article, cut it.
• An expert would bookmark it, precisely because it's the only decent page on that narrow thing. "Only decent page on X" is the whole strategy compressed into a phrase.
• You can name the query it should win. One query. If you can't, the retrieval system can't either.

Run your last five posts against that list. In my experience most content marketing fails item one immediately: titles describe territories ("A complete guide to email deliverability") rather than answering questions ("Why Gmail throttles your sends at 2 p.m."). Territory pages made sense when Google rewarded comprehensiveness. Passage retrieval rewards resolution.

Where llms.txt fits

A pile of narrow pages creates a discovery problem: crawlers have to find the right one among hundreds. That's the problem llms.txt exists to solve. It's a curated markdown index at a stable URL listing your pages with one-line descriptions, so an AI system can map a question to the URL that answers it without crawling everything you've ever published.

Specific content and llms.txt compound each other. An index line like "Why Gmail throttles bulk sends in the afternoon" routes a crawler straight to a page that can win its query. An index full of entries like "Blog" and "Solutions" routes nowhere. If you maintain the file by hand, write the descriptions as the questions each page answers; if a tool generates it, make sure your page titles already do that work.

How to run the experiment on your own site

Here's the protocol, sized for one quarter:

1. Collect five real questions your customers ask that have no dedicated page anywhere obvious. Sales calls, support tickets, and Reddit threads in your niche are the richest sources.
2. Publish one narrow post per question, passing the checklist above. Keep a recent broad post as your control.
3. Register the pages in your sitemap and llms.txt the day they ship, so discovery lag doesn't muddy the test.
4. Watch per-URL AI crawler activity weekly. You're looking for two signals: time-to-first-fetch for each bot, and repeat visits. A page GPTBot fetches once and abandons is being deprioritized; a page it re-reads after each update is being tracked.
5. At 90 days, compare. Crawl counts, distinct AI bots seen, revisit rates, and any AI referral traffic, narrow posts vs the control. Then spot-check by asking ChatGPT and Perplexity the five questions and seeing who gets cited.

Expect the retrieval clock to move first: live answer engines can cite a crawled page within days or weeks. The training clock, the one Abramov was talking about, runs 6-12 months, and you'll have long since decided whether the strategy works from the crawl data alone.

The teams that win AI search over the next couple of years will be the ones that measured which of their pages AI systems actually read, then fed that loop every month. Guessing right about specificity was never required. If you want the measurement side handled, the live demo shows exactly what per-bot, per-page crawl data looks like on a real site.