Give AI answers something worth citing

What is known, and how well

Advice about AI visibility mixes three kinds of knowledge: what platforms document, what studies measured and what practitioners believe. They deserve different levels of trust. Before using any tactic, ask which of the three it rests on.

The rest of the chapter stays on the documented and the studied, and labels the rest. That is also the more useful stance for a business: tactics grounded in documentation or research are cheap to defend, and every tactic that merely improves the page for readers is a safe bet either way.

Contribute evidence, not echoes

Google’s guide to generative AI search says the most useful content has a unique point of view and is not commodity content. Its example of commodity content is a generic “7 Tips for First-Time Homebuyers”; non-commodity content offers an expert or experienced perspective, and Google advises against simply recycling what others have already said.^[2] A pattern follows for a business: publish the things only you can, such as a documented project, a worked example, a comparison with explicit criteria or a measured result.

Separate what you measured from what you believe. If you ran an experiment, say when, on what and how. If you did not, call it a review and keep your claims at that level. Honest scope is itself a mark of a source that is safe to quote. See the evidence and authorship chapter for the checks.

What the research found, with its limits

The most cited study is “GEO: Generative Engine Optimization” by Aggarwal and colleagues, accepted at KDD 2024. The authors built a benchmark of 10,000 queries across 25 domains, asked a language model to rewrite source pages using nine different tactics and measured how much each source was cited in the answers of a test generative engine, using a position-adjusted word count.^[3] In their overall results, adding quotations raised that score from 19.3 to 27.2, a 41% relative gain. Adding statistics (25.2) and citing sources (24.6) also helped, while keyword stuffing lowered it to 17.7.^[3]

Read the result with its conditions. The pages were rewritten by a model, the engine was one the authors built, with a second experiment on Perplexity.ai, and the metrics were their own. The effect varied by domain: for example, citing sources helped most on factual questions, while statistics helped most for questions about law, government and opinion. In a scenario where every page was optimized, citing sources raised visibility for sites ranked fifth in search results by 115.1% and lowered it for the top-ranked sites by 30.3%.^[3] The authors present that as a chance for smaller creators, and it is a simulation.

Build on the search foundation

For Google’s own features, the documentation is direct. AI Overviews and AI Mode draw on the same index and ranking systems as classic Search, and may use query fan-out, issuing several related searches across subtopics, so a page that covers the surrounding questions has more chances.^[1] Google’s guide to generative AI optimization says that optimizing for these features is optimizing for the search experience, and it lists what you can safely ignore.^[2]

Other services are different, and they document less. OpenAI says that for a site to be included in ChatGPT search results, its OAI-SearchBot crawler must not be blocked, and that this crawler is separate from GPTBot, which is used for training.^[5] Perplexity describes PerplexityBot as surfacing and linking sites in its search, and says it is not used for model training.^[6] There is no single submission switch. Allow access deliberately, as the crawler controls chapter explains, make sure pages work without JavaScript, and give those systems something distinctive to cite.

Make the contribution distinct

Before publishing, compare the draft with the references it uses. If removing the references would leave only a paraphrase of their documentation, the page needs more of you in it. Add a worked implementation, a decision table, a diagnostic process or a tested example. The platform facts belong to the platform; the method and the judgment belong to your page.

• A screenshot without context is decoration; add what the reader should notice.
• A chart without its dataset is an unsupported claim; link or describe the data.
• A code sample without its environment and expected result is hard to verify; state both.
• A recommendation without its conditions cannot be applied; give the situation it fits.

Be careful with the shortcuts. Google’s guide treats scaled, templated pages made mainly to manipulate rankings or AI responses as a violation of its scaled content abuse policy, whatever tool produced them.^[2] Twenty thin pages will not do what one strong page does.

Use citation observations to improve the source

If an AI answer cites your page, check what it cited and whether the answer represents the source fairly. A citation may support only a small statement and not the whole answer. Record the service, the question, the date and the cited URL so you can tell a repeating pattern from one observation. If the answer gets a detail wrong, look first at your page for an ambiguity: an outdated price in an old paragraph, a vague claim or two pages that contradict each other. Fix the source you control, then watch later results. A fix does not force another system to update at once.

When no citation appears, return to the fundamentals: is the page indexed and accessible, does it answer the question and does it contribute something that other pages do not? Use the citation audit chapter for a repeatable review, and avoid trading useful writing for tricks that the documentation says you can ignore.

Apply it to your website

• You know whether each AI-visibility tactic rests on documentation, a study or a hypothesis.
• Pages that matter contain evidence only you can supply: a documented project, a measured result, an explicit method.
• Statistics, quotations and citations are real, accurate and attributed to credible sources.
• Keyword stuffing, templated pages and inauthentic mentions are avoided.
• Important pages are indexed, accessible without JavaScript and eligible for a snippet.
• Crawler access for search and training is a deliberate decision, made per service.
• Citation observations are recorded with the service, question, date and cited URL, and used to fix ambiguity in your pages.