Write clear answers, then measure AI citations

CH.01

Make the answer understandable in context

A useful passage states its subject, answers a real question and supplies enough context to assess the answer. Clear structure helps readers; it does not guarantee selection by an AI system.

Check whether someone arriving at a subsection can understand the claim without guessing what “it” refers to. Name the product, method or population when ambiguity matters. Keep qualifications beside the claim, and link to the fuller explanation when the answer depends on earlier material.

Public AI search services do not expose one common extraction pipeline. Avoid assuming that every engine reads only isolated paragraphs, ignores the rest of the page or copies a sentence verbatim. An accurate, self-contained explanation is a writing choice; citation remains an outcome to observe.

CH.02

What the original GEO experiment measured

Aggarwal and colleagues’ KDD 2024 paper used a 10,000-query benchmark with a 1,000-query test split. Its main system supplied Google’s top five source texts to GPT-3.5-turbo. These are experimental conditions, not a specification for today’s public search products.

The paper’s Table 1 reports best improvements of 41% for Position-Adjusted Word Count and 28% for Subjective Impression. They are relative changes in two visibility metrics, not fixed effects assigned to statistics, external links and quotations, nor percentage-point changes in citation frequency.

CH.03

Give each paragraph a clear job

Use a direct answer when the reader needs one, then provide the explanation and evidence needed to understand it. Keep a narrative when the story or sequence is part of the task.

Read headings and opening sentences together. They should explain what the page covers and where to go next. Do not force every paragraph into a definition, repeat nouns mechanically or strip away a qualification just to make a sentence easy to quote.

Remove repeated introductions and claims that add no information. Keep examples, exceptions and longer explanations when they help someone use the answer correctly. The goal is a complete answer with less effort for the reader.

CH.04

Put important information where readers need it

Opening with the main point can reduce the work of finding it. It is a useful editorial choice, not a universal rule that the first 30% of every page controls citation.

Google’s AI Search guidance says there is no ideal page length and no need to split content into tiny chunks for AI systems. Start with the reader’s task: a short practical question may need a brief answer; an explanation may need definitions, a worked example and exceptions.

A 40–60-word opening can be a drafting exercise for English content, not a template to impose on every language or topic. Use only the length needed to orient the reader. Place essential limitations beside the answer and let the rest of the page supply depth.

CH.05

Measure changes without calling every comparison an A/B test

A live page’s before-and-after citation record is observational. Retrieval, indexing, model behavior, competitors and demand can change at the same time. A cleaner paragraph may be worthwhile even when citations do not immediately change.

Choose a relevant prompt set and record the surface, language, location, account state, dates, page version and cited URLs. A starter set of 50–100 prompts is a planning choice, not a statistical guarantee. Keep failed requests and missing observations separate from complete answers without your URL. Use a consistent denominator and repeat schedule.

A controlled test needs a suitable assignment method and comparable conditions. A system you operate can test supplied versions under fixed retrieval settings; that result remains specific to that system. On public search, report the observed change and competing explanations. Do not publish duplicate pages merely to manufacture an “A/B” comparison or treat an API response as the consumer interface.

CH.06

Finish with a short editorial check

Check the material claims and the main reading path before publication. Revisit a passed check when the relevant content changes or a specific problem appears.

Use this as a practical worklist. Mark an item as unresolved when evidence is missing; use “not applicable” when the page has no reason to include it. A page does not need a statistic, expert quote and chart just to satisfy a template.