All guides

Getting cited by AI search

Marketing7 min read

Ask ChatGPT or Perplexity a question in your category and watch what it cites. Some of those sources will be obvious. Some will be a competitor’s page from 2023 that ranks nowhere in Google, quoted at length because it happened to contain the one statistic the model needed.

That second case is the interesting one, and there is now research explaining it.

The research

A team from Princeton and IIT Delhi published “GEO: Generative Engine Optimization” at KDD in 2024. They built GEO-bench, a benchmark of 10,000 queries spanning 25 domains, then tested nine different content changes to see which ones made a generative engine more likely to surface and quote a given source.

The headline result is that optimisation works. Their methods improved visibility by up to 40% in generative responses. What matters more is which methods did the work.

Adding direct quotations from relevant sources performed best, with a relative improvement of 30 to 40% on their position-adjusted word count metric. Adding statistics and adding citations to credible sources followed closely behind. Improving fluency and plain readability produced gains of 15 to 30%.

Keyword stuffing produced almost nothing. The authors were direct about it: methods carried over from traditional search offer little to no improvement in generative engines.

That is the finding worth sitting with. Two decades of SEO habit built around keyword density transfers badly to a system that reads for meaning and quotes for evidence.

Why quotes and numbers win

A language model assembling an answer needs material it can attribute. A page saying “response times matter for conversion” gives it nothing to work with. A page saying “Harvard Business Review found firms contacting leads within an hour were seven times more likely to qualify them” gives it a sentence it can lift, attribute, and stand behind.

The model is looking for citable units. Specific figures, named sources, direct quotes, dates. Prose that hedges everything into generality is unquotable, so it goes uncited even when it is entirely correct.

This has an unglamorous implication for content teams. The work is research, not writing. Finding the real number and sourcing it properly takes longer than producing another 1,200 words of competent generic advice, and it is the part that determines whether anything cites you.

Make the page readable by machines

None of this matters if the assistant cannot fetch your page.

Check your robots.txt for what you allow. Crawlers including GPTBot, ClaudeBot, PerplexityBot and Google-Extended each respect robots directives, and plenty of sites block them by accident through a broad disallow rule inherited from an old configuration. Decide deliberately whether you want to be in these systems. Blocking them is a legitimate choice. Blocking them without realising is not.

Server-rendered HTML helps considerably. Content that only appears after client-side JavaScript executes is unreliable for crawlers that do not run a full browser, which is one of several reasons we build content sites the way we describe in why Astro is a great framework for simple content sites.

Structured data gives the machine an unambiguous version of facts your prose states loosely, and we cover that separately in structured data and rich results.

Structure for extraction

Write so a single section answers a single question completely. An assistant pulling an answer wants a self-contained passage, and a paragraph that depends on three earlier paragraphs for context is hard to lift cleanly.

Put the answer near the top of the section, then support it. Use headings that read like questions people actually ask. Keep the useful number in the same sentence as the claim it supports rather than three sentences later, because proximity is what makes a passage quotable as a unit.

Date your content and say when it was updated. Models and the humans checking their output both weight recency, and an undated page about a fast-moving topic is a page nobody wants to stake an answer on.

What to measure

Track your own citations. Ask the assistants your customers use the ten questions your customers actually ask, record which sources come back, and repeat monthly. It is manual and unglamorous. It also beats every tool currently promising to automate it.

Watch referral traffic from assistant domains in your analytics, with the caveat that most AI citations produce no click at all. Somebody reading your statistic inside a chat response never visits your site. That is a real change to how content marketing pays back, and pretending otherwise leads to disappointment when the traffic graph stays flat while brand mentions climb.

The honest position

This field is about two years old and much of what is written about it is confident beyond the evidence. The Princeton study is the substantial piece of published work, it tested a specific set of engines at a specific time, and generative engines change fast.

What survives that uncertainty is unremarkable: cite real sources, include real numbers, write clearly, let machines fetch your pages. That advice was defensible before any of these systems existed, which is a decent sign it will still be defensible when they change again.

The same scepticism applies to putting these models inside your own business, and we go through where that works in where AI earns its keep.

If you want help working out which questions your customers ask assistants, and whether anything of yours comes back, that is part of how we approach marketing. Get in touch and we will take a look.

References