Writing for LLMs vs Google: where the rules diverge

Google has spent twenty-five years training a ranking system on user behavior — clicks, dwell time, refinement queries, return visits. Its index is built around documents, and it ranks documents by their predicted relevance and quality for a query. The unit of competition is the page.

An LLM, by contrast, retrieves chunks. When ChatGPT answers a question with browsing on, it doesn't pull a page and read it. It pulls a candidate set of passages — typically a few paragraphs each — embeds them, scores them against the query, and synthesizes an answer from the highest-scoring chunks. The unit of competition is the passage.

This single difference cascades through every other rule. A page that ranks well in Google but contains no clearly extractable answer passage will get low LLM citation. A page with a strong answer passage but weak overall structure can get cited by an LLM and never reach Google's first page. Writing well for both means writing pages that satisfy at the document level and have high-quality, extractable passages at the section level.

An atomic answer block is a self-contained passage that answers a specific question, in a way that requires no context from elsewhere on the page. It's the unit LLMs prefer to cite. A well-built page contains many of them, sequenced into a coherent narrative.

What makes a passage atomic

• It restates the question or topic in its first sentence. The reader (or model) can drop into it without prior context.
• It delivers the answer in full within the passage, including any necessary numbers, dates, definitions, or caveats.
• It doesn't depend on visual elements like images or tables that would be lost in extraction.
• It's between 80 and 220 words. Too short and it lacks substance; too long and it gets fragmented during chunking.

The Google-only writing pattern of building a long argument that culminates at the end is exactly what fails in LLM retrieval. Conversely, breaking content into too many tiny chunks — the SEO-spam pattern of a question header and one-sentence answer — fails Google's quality systems. The right rhythm is a paragraph-per-question structure, where each paragraph stands on its own but the page as a whole reads as a coherent argument.

Three patterns reliably increase the rate at which content gets cited in AI search interfaces, while still working in Google.

Pattern A: Lead with the headline finding

If your piece reports a study, a benchmark, or a counterintuitive insight, surface the headline finding in the first paragraph. Models prefer to cite specific claims. 'B2B SaaS companies that publish weekly grow MQLs 41% faster than those that publish monthly' is far more likely to be cited than 'There are several content cadences B2B companies use, with varying results.'

Pattern B: Date and ground every claim

Models distrust undated claims because they can't verify recency. Add a year or month to every statistic. 'In Q1 2026, 47% of informational queries returned an AI Overview' anchors the claim in time. The unanchored version — 'Most informational queries now return an AI Overview' — is true but harder for a model to verify and cite.

Pattern C: Define entities inline

The first time a key term appears in your piece, define it in a single sentence. 'Generative engine optimization, or GEO, is the practice of optimizing content to be cited in AI-generated search answers.' This both helps human readers and provides a high-confidence definition for the model to extract. Pages with inline entity definitions get cited at noticeably higher rates in our testing — roughly 1.7x — across Perplexity and Gemini.

Schema markup adds structured metadata to a page, telling search engines and language models what each part of the content represents. For Google, schema influences rich-snippet eligibility. For LLMs, it's a strong prior on how to interpret the content during retrieval and citation.

• Article schema with explicit author and publisher fields. Models use these to attribute citations and build their authority graph.
• FAQPage schema for sections that genuinely answer multiple distinct questions. Don't fake it — schema markup that doesn't match visible content is now penalized in both systems.
• BreadcrumbList schema for site hierarchy. Helps Google contextualize the page within your site, and helps LLMs cite into the right section of your topic graph.
• Organization schema on your homepage and key landing pages, with sameAs links to your social profiles. This consolidates your entity identity across the open web.

If you only do one schema thing, make it Article + Organization with consistent author identity across every piece. This is the single highest-leverage technical SEO move for AI search.

Both Google and LLMs reward content that demonstrates expertise and timeliness, but they measure these differently.

Google rewards depth as a function of comprehensiveness — does the page cover the topic thoroughly? It rewards freshness on a topic-by-topic basis. Some queries (news, software comparisons) demand recent content. Others (definitions, foundational concepts) tolerate older content as long as it's accurate.

LLMs reward depth as a function of specificity. A page that says many specific things — names, numbers, dates, attributions — reads as deeper than a page that says fewer general things, even if the second is longer. They reward freshness through the dating of specific claims rather than the publication date of the page. A 2023 article that has been updated with 2026 statistics, dated explicitly, can outperform a brand-new article with no specific dates.

The practical implication: maintenance matters more than ever. Set a quarterly review on your top 20 pages. Update statistics. Mark them as updated explicitly. Models read 'last updated' dates and use them to weight citation preference.

Here's the brief template we now use for every piece of content. It's deliberately structured to satisfy both Google and LLM retrieval simultaneously, without compromise.

1. Headline finding — one sentence stating the most citable claim or insight in the piece. This becomes the lead.
2. Target query — the head term the page is meant to rank for in Google, plus the conversational variant most likely to be asked of an LLM.
3. Atomic answer outline — list of 5–10 sub-questions the piece will answer, each with a one-sentence draft answer. These become the H2 sections.
4. Specific data anchors — at least 5 numbers, dates, or named studies that will appear inline. These are the chunks most likely to get cited.
5. Entity definitions — every term of art that appears in the piece, with a one-sentence definition.
6. Schema plan — which schema types apply to the page (Article + at minimum), with required fields filled out.
7. Update cadence — when this piece will next be reviewed and refreshed.

Pieces written from this brief tend to perform 30–50% better on AI citation metrics than equivalent pieces written from traditional SEO briefs, while ranking just as well or better in Google. The two systems aren't in conflict — they're in tension, and the tension is resolvable with discipline at the brief stage. Get the brief right, and the writing falls out of it.

One thing worth being explicit about: this isn't a recommendation to write shorter, simpler content. The opposite, often. Pieces that satisfy both retrieval models tend to be longer than pure-Google-optimized articles, because they need to contain enough self-contained answer blocks to give the LLM many citation candidates. The shift is in structure, not in depth. The instinct to compress your way to better LLM performance — by trimming context, removing examples, paring down to bullet lists — usually backfires. Models prefer rich passages with specific, dated, attributable claims. A two-thousand-word piece written under the unified brief above will outperform a six-hundred-word executive summary every time, both in citation rate and in Google rankings.

Finally, the cadence question. Most teams under-publish. The compounding effect of a content program comes from breadth as much as depth — having a citable answer for the long tail of related queries that surround your category. We recommend a target of one substantial piece per week per primary topical cluster, sustained for at least six months before evaluating ROI. Below that volume, your content presence is too thin for either Google or LLM systems to read as authoritative. Above that volume, sustained for long enough, you become the default reference — at which point the citation rate and the organic rankings both compound at rates that look almost unfair to teams who haven't yet committed to the cadence.