Playbook
LLM-friendly content is content that can be quoted without repair.
LLM-friendly content is a page an answer engine can lift a passage from without editing it first. By Arne Kellmann. Updated 2026-08-23. This page states the operating assumptions behind Blockquote's 34 checks — 29 deterministic, 5 LLM-judged — and turns them into a publishing brief.
In short
- Put the answer in the first sentence; delete the warm-up paragraph.
- Write headings as the questions a reader would type.
- Make every section true and complete when read on its own.
- Use structured data that matches what is visible, with honest dates.
- Prefer concrete numbers and named sources over adjectives.
What is LLM-friendly text?
LLM-friendly text answers directly, names entities explicitly, and keeps each section understandable when read in isolation. A model does not experience a page as a composed narrative first. It experiences a set of chunks — headings, passages and machine-readable hints — that must survive extraction. Those windows are finite: Blockquote's 5 LLM-judged checks, for example, read only the first 6,000 characters of body text, roughly the first 1,000 words of a page.
The practical test is simple: cut any one section out of the page, show it to someone who has not read the rest, and see whether it still states something true and complete. If the section only makes sense after three paragraphs of build-up, a retrieval system will either skip it or quote it wrongly.
How should a page open?
Open with a definition or a direct answer. The first sentence should be quotable on its own. The second should add the most important constraint, number or outcome. Only then should the page widen into nuance, examples or exceptions.
The common failure is the warm-up paragraph — two or three sentences explaining why the topic matters before saying anything about it. Answer engines rarely quote warm-up text, because it contains no claim to attribute. Move the answer to the first line and the warm-up becomes unnecessary. The payoff is measured: the Generative Engine Optimization study (arXiv 2311.09735, November 2023) tested content edits across a 10,000-query benchmark and found that adding statistics, quotations and source citations raised a page's visibility in generated answers by up to 40%.
Why do headings matter?
Headings define retrieval boundaries. Question-led H2s and H3s make it easier for a system to map a user query to a single self-contained answer block. Long stretches of unlabeled text force both reader and model into unnecessary inference.
Phrasing a heading as the question a reader would actually type also forces the section beneath it to answer that question and nothing else. A heading like "Background" permits any content; a heading like "How long does a scan take?" does not. Blockquote's structure checks — 30% of the total score — expect roughly one H2 or H3 per 400 words, and every H2 on this page is phrased as a question.
What should be machine-readable?
At minimum: organization identity, authorship, breadcrumbs where relevant, and FAQ or Article entities when they genuinely reflect the visible content. Structured data does not rescue weak writing, but it removes ambiguity about who published the page and what it represents.
Two constraints matter more than coverage. Structured data must describe what a reader can actually see on the page, and the dates in it must be true. Schema that contradicts the visible page is worse than no schema, because it gives a model a confident wrong answer about your content. The precedent is instructive: in August 2023 Google restricted FAQ rich results to authoritative government and health sites, yet FAQPage markup kept its second job of telling a model which visible question a passage answers. This page ships Article and FAQPage entities generated from the same source as the visible text, first published 2026-05-03.
What makes a page citable?
Claims concrete enough to be attributed. Named entities, specific constraints, real dates, version numbers, prices, counts, and source links placed next to the sentence they support. If a page only speaks in abstractions, a model has nothing stable to quote.
Original numbers are the strongest form of this. A page that reports something only its author could know — measurements, prices, results from its own work — gives an answer engine a reason to name the source rather than absorb the claim anonymously. Blockquote scores these habits with explicit thresholds: full marks require at least 1,200 words (800 words scores 0.7, 500 words 0.4), five external source links, and one emphasized key term per 200 words.
How is this different from classic SEO?
Classic SEO competes for a position in a list of links, where the click happens after the ranking. Citability competes for inclusion in a generated answer, where the source is named inside the text and there may be no list at all. Google's AI Overviews, launched at its I/O keynote on 14 May 2024, put that second surface in front of ordinary search traffic. The selection mechanics differ, so the checks differ.
The two overlap in places — crawlability, clean markup and useful content help both — and they do not conflict. Treat citability as a second scoring surface for the same page, not as a replacement for search ranking.
What does an AI crawler actually see?
Most AI crawlers read the raw HTML response and do not execute JavaScript. Content that only appears after a client-side framework renders it is invisible to them, and so is anything hidden behind a consent overlay that replaces the page body. The safe assumption is that a model sees your server-rendered HTML once, top to bottom, with a size limit. OpenAI alone documents three crawlers with different jobs: GPTBot gathers training data, OAI-SearchBot builds the search index, and ChatGPT-User fetches pages a user asks about.
That size limit has a practical consequence: crawlers truncate large documents, and anything after the cut does not exist for them. Put the claims you most want quoted early in the document, keep the markup lean so text is not buried under tag overhead, and never rely on a footer or an appendix to carry an important fact. Blockquote's own fetcher works the same way: it stops reading HTML at 3 MB and marks the scan partial.
When do lists and tables help?
Use a list when the content is genuinely enumerable — steps, criteria, options — and a table when two or more attributes vary across the same set of items. Both formats hand a model pre-segmented facts: each bullet or row is a claim with its own boundary, and that boundary survives extraction far better than the same facts woven through a paragraph.
The failure mode is decoration. A list of single-word bullets carries no claims, and a one-column table is a list wearing borders. If a bullet cannot be read as a standalone statement, fold it back into prose; formatting only helps when the visible structure reflects real structure in the information.
How long should a page be?
Long enough to cover its question completely, which in practice is rarely under a few hundred words. Thin pages give a retrieval system little to select from and little reason to trust the source. Substantive pages earn citations because they contain more distinct, quotable claims per question the page can plausibly answer.
Length is a byproduct, not a target. Padding a page with restatements adds words but no new claims, and models summarize repetition away. Extend a page by answering the adjacent questions a reader would ask next — each added section should contribute at least one fact that was not already on the page.
How do you verify that a change worked?
Rescan the page after publishing and compare the check results, not just the total. A total can stay flat while one category improves and another regresses, which is the case worth investigating.
Do not treat a single answer from a chatbot as evidence. Generated answers vary between runs, models and dates, so one citation proves nothing and one absence disproves nothing. The stable signal is whether the page now satisfies the conditions that make it quotable.
What does Blockquote measure?
A scan runs 34 checks in three weighted groups: schema at 40%, structure at 30% and citability at 30%. 29 of the checks are deterministic — a JSON-LD block is present or it is not, a heading hierarchy is measurable. The remaining 5 judge fact density, original data, direct paragraphs, pronoun clarity and quote extractability with Llama 4 Scout against a fixed rubric, and the report labels them so you can tell which is which.
Checks that do not apply are marked N/A and left out of the score entirely — FAQ markup is not counted against a page with no FAQ. That is why two pages with the same score can have different numbers of applicable checks.
Which sources does this playbook build on?
How do you check your own page?
The scan form runs all of the above as 34 checks and returns the failures with finished, copy-paste snippets — generated JSON-LD and content fixes rather than generic advice — so the shortest path from report to fix is pasting, publishing and rescanning.