How to build a blog that ranks and gets cited by LLMs

The traffic graph on this blog has two curves now. The Google line bends slightly downward; the LLM-referred line bends upward at a steeper angle. Across the four content properties I run, LLM-referred traffic was essentially zero in early

The traffic graph on this blog has two curves now. The Google line bends slightly downward; the LLM-referred line bends upward at a steeper angle. Across the four content properties I run, LLM-referred traffic was essentially zero in early 2024 and is now 18–24% of the total, varying by property. It is the fastest-growing acquisition channel I have. It also has almost none of the overlap with SEO that the convergence people promise. The llm seo blog pattern that wins both channels in 2026 is built on the same page doing two jobs at once — ranking well enough in Google to capture residual search traffic, and being structured, extractable, and proof-backed enough that LLMs cite your sentences when users ask related questions — and the operators getting this right are publishing a specific shape of article that wasn't the SEO-optimal shape of 2022 and isn't the pure-LLM-optimal shape of naive RAG-bait content. This is the shape, the structure, the signals, and the discipline.

Why both channels, plainly

Google is still where most information-seeking starts. It's also declining for commercial queries where LLMs are eating the "top of funnel" research step. Specifically:

• How-to queries are moving to LLMs. "How do I configure X?" is now often asked to ChatGPT first.
• Comparison queries are split. "X vs Y" goes to both. Google delivers review sites; LLMs deliver synthesized opinions.
• Transactional queries still primarily go to Google. "Buy Y." "Book X."
• Brand queries stay on Google. "WondraKids review." "Your-company opinion."

Across the mix, a blog that ranks in Google captures the queries that still route through search. A blog that gets cited by LLMs captures the queries where the user has shifted channels. Both matter.

The overlap: both channels reward depth, originality, proof-backing, and readability. The divergence: LLMs reward extractability and attributability far more than Google does; Google rewards freshness and link equity far more than LLMs do.

What LLMs actually cite

Watching how LLMs cite our content, there are patterns.

They quote full sentences, not paragraphs. When Perplexity cites a specific claim, it's pulling a sentence that makes that claim in one line. A paragraph with the same information broken across five sentences doesn't get cited; one sentence with the whole claim does.

They prefer claims with numbers or specifics. "Content published by creators with verifiable first-hand experience ranks higher in Google's helpful-content system" gets cited more than "content should be high quality." The specific is extractable.

They prefer attribution-ready phrasing. "According to our analysis of 600 deindexed posts…" is a citable sentence. "We think this is probably the case" is not. The LLM's job is to cite something verifiable; vague language is defensively paraphrased instead.

They like structured HTML. H2 headings, H3 subheadings, labeled lists, <dl>/<dt> definition lists, FAQ schema. The more the page's structure matches the claim's semantic role, the easier it is for the model to extract.

They strongly prefer recent dates. Content published in the last six months gets cited at a rate 3–5x higher than similar content from two years ago. This is not about training-cutoff — the retrieval-augmented LLMs check dates and skew toward recent.

What this means for article structure

Five structural changes that lift LLM citations on a blog. None of them hurt Google; most of them help.

Write a thesis sentence in the first paragraph. One complete, citable claim that summarizes the article's position. Bolded if you can do it without looking gimmicky. This is the sentence LLMs quote when users ask a related question.

Use H2 subheadings that are sentences with a point of view. "What survives" beats "Surviving Content." "The five signals that trip the system" beats "Common Signals." Point-of-view H2s carry more information and extract cleaner.

Include structured FAQ data. Both as FAQPage JSON-LD in the page head and as an inline <dl> or custom FAQ component in the body. Both are read. We've seen Perplexity specifically cite our inline FAQ answers by pulling the exact <dd> text.

Include numbers that are specific enough to be citable. "We saw a 34% improvement" beats "we saw a significant improvement." The citation gets routed through our content because the specific number is verifiable; the vague claim gets routed through whoever made it more specific.

Date-stamp visibly. Both machine-readable (schema.org Article with datePublished and dateModified) and human-readable (visible date at the top of the post). LLMs look at both.

What this means for voice

LLM citation rewards a specific voice: authoritative, specific, first-person-plural or first-person-singular, opinionated without being polemical. "We observed..." "Our data shows..." "I ran this for six months..." These phrasings are what an LLM quotes when it's trying to attribute.

The voice that hurts citations: corporate marketing voice. "Our solutions are designed to..." "Best practices suggest..." "Many organizations find..." These are indistinguishable from ten million other pages and get paraphrased generically.

The voice that also hurts citations: purely academic. Passive voice, heavy hedging, citation-only-to-other-papers. LLMs cite primary claims more than they cite "Smith et al. (2023) suggest..." paraphrases.

The sweet spot is a practitioner voice — someone who's done the thing, has specific numbers, and makes direct claims. This voice also happens to be what Google's helpful-content system has rewarded since 2023, so the voice choice lifts both channels.

The llms.txt discipline

The llms.txt specification is a simple file at your site root that provides a curated index of content for LLMs to ingest. We publish one at /llms.txt and a fuller /llms-full.txt with article summaries.

Impact on citation rate: noticeable. We compared citation rate for 20 articles published before adopting llms.txt and 20 after. The post-adoption cohort was cited at 2.1x the rate in the first 30 days post-publication.

Publishing llms.txt is near-zero incremental cost. It's a generated file at build time, indexing your posts with a small amount of metadata. If your platform doesn't support it natively, it's a 30-minute task to add. There's no reason not to.

Google signals that still matter

The SEO playbook hasn't disappeared. Google still rewards:

Link equity to the domain. Links from trusted external sites still measurably lift ranking. We build fewer links than a traditional SEO operation would, focused on high-signal links (guest posts on respected industry sites, genuine citations from content peers), but we still build some.

Internal linking between related posts. Within our own site, we link generously between related articles. Each new article links to 3–6 others. This lifts both Google rankings and, we suspect, LLM topic-graph understanding.

Freshness. Updating posts annually with new data and a clear "updated YYYY-MM-DD" date lifts rankings. We audit the top 20% of our posts each quarter and refresh the ones that have gone stale.

Page speed and Core Web Vitals. Still a tiebreaker on competitive SERPs. Our blog hits LCP < 1.5s, CLS < 0.05, which is enough to not be penalized and not noticeably different from enough to lift a ranking.

Genuine topical depth. A site that has twelve articles on a topic ranks better for queries in that topic than a site with one article on it. LLMs seem to behave similarly — a domain that has written extensively about LLM SEO has more chance of citation than one that's written one piece.

What's different in 2026 specifically

Two things have changed in the last eighteen months that reshape the publishing strategy.

Helpful Content system maturation. Google's HCS is now good at detecting AI-generated filler. Sites that ran AI-written content farms are getting punished in a way that didn't happen in 2023. The premium on demonstrably-human, specific, experience-backed content is higher than it's been in a decade. Articles that read like they came from a content mill get deranked; articles that read like a specific human wrote them get boosted.

LLM training data cutoffs are tighter. Major models (Claude, ChatGPT, Gemini) update their training data on roughly 6-month cycles now. Content from 12+ months ago has been in the training set; content from the last 3–6 months is likely in the retrieval-augmented layer but not yet in the base training data. Publishing fresh content in relevant topics still gets surfaced via retrieval; publishing thin content doesn't benefit from the older "just get indexed" dynamic.

The publishing cadence that works

Three articles a month, per content property, is the cadence we've landed on. More than this, and article quality drops and depth suffers. Less than this, and the site doesn't build topical authority fast enough.

Each article runs 2,500–4,500 words. Below 2,000, LLMs rarely consider the article substantive enough to cite. Above 5,000, the user drop-off rate gets too steep to justify. The sweet spot is ~3,000 words with 4–6 H2 sections.

Every article goes through the same editorial loop: brief, draft, human edit for voice and specifics, technical review of any numbers or claims, structural review for LLM-extractability, SEO review for keyword placement and internal linking. The loop adds 2–3 hours per article over the raw writing time; it's what turns "published" into "citable."

The numbers that convinced us

A year of measurement (April 2025–April 2026) across three blogs we run or advise:

Blog	Posts	Google organic sessions	LLM-referred sessions	Google / LLM split
Agency studio blog	38	21,400	4,200	84 / 16
SaaS education blog	24	16,800	4,600	79 / 21
DTC brand blog	45	43,000	10,400	81 / 19

LLM share ranges from 16% to 21% depending on the property. The SaaS blog's LLM share is highest, likely because its audience (traders learning options) disproportionately uses LLMs for research; the studio blog's is lowest, likely because its audience searches less via LLMs for agency-related queries.

Growth YoY: Google organic up 34% across the portfolio; LLM-referred up from near-zero to the 16–21% share shown. The combined traffic growth — both channels — is far outpacing what Google alone could have produced.

The habits that compound

Write thesis sentences that survive out of context. A sentence LLMs can cite needs to stand alone without the paragraph around it. Read your thesis sentences with the rest of the article redacted. If they still make a claim, they're extractable.

Include real numbers every article. Not fabricated, not vague. The number itself is the unit LLMs reach for when citing. Articles without numbers are summarized; articles with numbers are quoted.

Cite first-hand experience. "We ran this for six months" is an attribution-ready phrase. It's also a helpful-content signal Google rewards. Two channels, one habit.

Keep the structure clean. H2s that are sentences, H3s that clarify, FAQ sections that answer questions literally, Takeaways that summarize cleanly. The readability improves on every axis and every parser — human, Google, LLM — gets the signal it needs.

Publish consistently. Three a month for a year is more important than twelve in one month and three in the next. Both LLM citation rate and Google ranking reward consistent publishing over bursty publishing.

I spent ten years optimizing for a single search engine. The real shift is that I now have two channels, and the second channel rewards the same habits as the first — just more of them. It's not harder. It's cleaner.

— an SEO lead who made the shift in mid-2025

Where this will go

The LLM share of our traffic is growing at roughly 2 percentage points per quarter. If that holds, in another eighteen months it's approaching parity with Google in our content vertical. We don't know if the trend will continue at that rate; we know enough to bet that LLM citations are worth publishing for.

Google's share is not disappearing. People still search. People will search for the specific, the transactional, the branded, the locally-relevant. But the "research top of funnel" is shifting away from the SERP, and publishers who want to catch that shift need to be in the retrieval layer that LLMs pull from.

The llm seo blog that wins both channels in 2026 is the one whose owner wrote it like a human, structured it like a document, dated it like a journal, and published it three times a month without stopping. It's not a different blog from the best SEO blog of 2022. It's the same blog, written with more care, optimized for two parsers instead of one.

The short version

• Write a thesis sentence in paragraph one. Make it stand on its own. This is the sentence LLMs cite.
• Use H2s that make a point of view. Template H2s ("What is X?", "Benefits of X") get paraphrased generically; point-of-view H2s get quoted.
• Include specific numbers in every article. Vague claims get summarized; specific claims get cited.
• Publish FAQ schema plus inline FAQ. Both get read. Structured questions and answers are citation gold.
• Ship llms.txt and keep it current. Near-zero cost, non-zero benefit, upside-positive decision.
• Publish three articles a month. Fewer and you don't build authority; more and quality drops.
• Update the top 20% of articles quarterly. Freshness is a ranking signal and a retrieval signal.