Specific, narrow blog posts get cited by AI more than broad ones. The retrieval mechanics behind it, and how to prove it with your own AI bot crawl data.
Quick answer
Narrow, specific posts get cited by AI models more readily than broad ones because retrieval matches queries to passages: a page about exactly one thing matches cleanly, faces few competing sources, and hands the model a quotable answer. Practitioners report the effect showing up in 6-12 months for training influence and much faster for live retrieval. You do not have to take it on faith. Publish narrow posts, then watch per-URL AI crawler activity for 60-90 days. Pages the bots fetch and revisit are in the citation pipeline; pages they skip need a sharper focus.
"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.
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.
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.
Specificity has a working definition you can apply before publishing. A post is specific enough when:
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.
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.
Here's the protocol, sized for one quarter:
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.
Written by Crawlytics Team. Crawlytics tracks AI bots, generates llms.txt, and powers WebMCP commerce, all from one snippet on any stack. See how it works →
The evidence so far is anecdotal but consistent, and the mechanism holds up. Dan Abramov reported Claude referencing his year-old long-form posts on narrow topics, Tyler Gaw saw specific posts pulled into AI outputs within roughly six months, and Google's John Mueller endorsed the underlying advice. Mechanically it makes sense: retrieval systems match queries to passages, and a page about exactly one thing produces a strong match with little competition, while a broad roundup competes against thousands of similar pages and buries its answers mid-page. No large-scale public study has quantified the effect yet, which is why testing it on your own site with crawl data matters.
Two different clocks run at once. Retrieval-based citations, where ChatGPT, Perplexity, or Copilot searches the live web and cites what it finds, can appear within days to weeks of your page being crawled and indexed. Training influence, where a model has absorbed your content and reflects it without searching, runs far slower: Tyler Gaw observed roughly 6 months and Dan Abramov estimated about a year. Watching AI crawler activity on a URL tells you the pipeline is working long before either kind of citation shows up.
The target is different. Niche SEO targets low-competition keywords to win a ranking slot on a search results page, where a broad page can still win with enough authority. AI answers do not have ten slots; an engine typically cites a handful of sources per answer, and it selects the passage that most directly resolves the question. That selection step rewards pages whose entire content is the answer, punishes pages where the answer is one section among twenty, and cares less about domain authority than classic Google rankings do. Specificity is the unit of competition, not just a keyword tactic.
Check server or edge logs for AI crawler User-Agents, or use a bot analytics tool that does it for you. JavaScript-based analytics like GA4 miss most AI bots because crawlers do not execute scripts. Crawlytics identifies bots like GPTBot, ClaudeBot, PerplexityBot, and ChatGPT-User by User-Agent and shows per-page, per-bot, per-day activity, so you can watch whether a new post starts drawing crawls in its first weeks and whether the bots come back after each update.
That is exactly the case it serves best. An llms.txt file is a curated markdown index at a stable URL that tells AI systems which pages matter and what each covers. A catalog of focused, single-question pages with clear one-line descriptions gives a crawler an efficient map from question to URL, which is far more useful than a list of vague category pages. It is a discovery aid rather than a citation guarantee: it improves the odds the right page gets fetched, and the page content still has to earn the citation.
This page is part of Crawlytics.app. View all pages: llms.txt · llms-full.txt