How to Get Cited in AI Search: What the Evidence Shows

What does the actual research say?

There is far more marketing than science here, so it is worth separating the two.

The one genuinely controlled study most people cite is GEO: Generative Engine Optimization by Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, presented at KDD 2024. The team built a benchmark of roughly 10,000 queries, applied nine different content modifications to source documents, and measured how often each modified document was cited in the generated answer.

What worked:

Modification	Effect on visibility
Adding citations to sources	Strong positive
Adding direct quotations	Strong positive
Adding statistics	Strong positive — combined lift reported up to 40%
Improving fluency and authoritative tone	Modest positive
Keyword stuffing	Neutral to negative

The pattern is coherent and, on reflection, obvious: a generative engine needs something concrete to attribute. A paragraph containing a specific number, a dated event, or a quoted authority gives the model a discrete factual unit it can lift and cite. A paragraph of confident generalities gives it nothing to point at, so it summarises the idea without naming you.

Note the shape of that finding. It is not a formatting trick. It says: have specific, verifiable things to say, and show your sources. Which is what good writing has always meant.

Why doesn't ranking first guarantee a citation?

Because AI answers are not built by reading your result page. As covered in how AI Overviews work, Google uses query fan-out: it decomposes your question into several sub-queries, runs them in parallel, and pulls the best passages from across all those result sets.

So the competition is not "the top ten for this keyword." It is "the best passage for each of a dozen sub-questions you never explicitly targeted." A page ranking fourth for the head term can be cited because it happens to contain the cleanest paragraph answering a sub-query — and a page ranking first can be skipped because its relevant section is three paragraphs of preamble followed by a hedge.

This is why the measured overlap between top-ten rankings and AI citations has fallen so sharply, from roughly three-quarters to under half in analyses through 2026. Ranking still helps — you have to be in the index and retrievable — but it is a qualifier, not the decision.

What about brand mentions?

Multiple 2026 vendor analyses report the same finding: unlinked brand mentions across the web correlate more strongly with AI citation than backlink metrics or on-page optimisation. Figures floating around put branded web mentions at roughly three times the correlation strength of backlinks.

Treat those specific numbers with real caution. They come from commercial studies with undisclosed samples and no controls, and they are correlational — well-known brands are mentioned more and are cited more, and both may follow from being genuinely well-regarded rather than one causing the other.

That said, the mechanism is at least plausible. A model's sense of who is authoritative on a topic is shaped by how often and how favourably a name appears across its training data. Being discussed in places a model has read is a different asset from being linked from them, and it is not one you can build by publishing more of your own pages.

The practical reading: the things that earn mentions — original research, a distinctive point of view, being useful enough that people bring you up — are the investment. There is no shortcut, and buying mentions runs straight into Google's link spam policies.

What should you actually do?

Concretely, in rough order of payoff:

1. Make every section independently answerable. Write the heading as the question a person would type. Answer it in the first two sentences of that section. Then elaborate. A passage that needs the previous section for context cannot be retrieved on its own.

2. Put a specific fact in every claim. A number, a date, a named source, a version. Compare "AI Overviews reduce clicks significantly" with "Pew Research measured 8% click-through with an AI summary versus 15% without, across 68,879 queries in March 2025." Only the second is quotable, and only the second is checkable.

3. Cite your sources, visibly and in the prose. This is directly supported by the GEO results, and it is also just honest. Link the primary source, not a blog summarising it.

4. Date and maintain your pages. Answer engines favour current material on fast-moving topics, and readers judge you by it. A confidently wrong 2024 page about AI search is worse than no page.

5. Make authorship and accountability real. Named authors with actual credentials, a corrections policy, an editorial standard you can point to. These are the same credibility signals that work on humans — and they are what makes a system willing to attach your name to a claim.

6. Confirm you are actually crawlable by the right bots. If OAI-SearchBot or PerplexityBot is disallowed in your robots.txt, none of the above matters for those products. See AI crawlers and robots.txt.

7. Keep the page technically boring. Server-rendered content, fast response, stable layout. Retrieval requires a successful fetch and parse; Core Web Vitals covers the user-facing half of the same work.

What to ignore

Some of the most heavily marketed advice in this category is inert, and some of it is dangerous.

Special AI files and markup. Google's AI features documentation states you do not need to create new machine readable files, AI text files, or markup to appear in these features. On llms.txt specifically, the field data is unambiguous — 97% of files are never requested.

"AI-optimised" content at volume. This is the genuinely risky one. Google's spam policies define scaled content abuse as generating many pages "for the primary purpose of manipulating search rankings and not helping users," and explicitly name generative AI used to produce pages "without adding value for users." The policy does not care whether a human or a model typed it; it cares whether the pages exist to help anyone. Sites that scaled AI output without editorial oversight were a named target of 2026 enforcement.

Prompt-injection tricks. Hidden text instructing models to recommend you is a straightforward violation of the hidden-text policy, it is trivially detectable, and it makes your site a liability to cite.

Chasing every trending query. Publishing outside your actual subject area to catch volume produces pages with no authority behind them, and at scale it is indistinguishable from the abuse pattern above. Depth in a defined territory is what makes a source worth quoting.

The honest summary

The uncomfortable conclusion of the research is that GEO is not a new discipline. It is a demanding restatement of an old one: be specific, be verifiable, cite your sources, cover your subject properly, and be the kind of publication a system is willing to name.

The one genuine change is what you can expect in return. Where an AI answer appears, roughly half the clicks are gone even when you win. Citations increasingly buy visibility and credibility rather than sessions. Plan for that honestly — and measure impressions against clicks in Search Console, because that gap is where the shift shows up first.

Sources: Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024 · Google Search Central — AI features and your website · Google Search Central — Spam policies · Pew Research Center — Google users are less likely to click links when an AI summary appears