How AI Search Engines Actually Choose What to Cite
Jillian Oco, CMO

AI assistants like ChatGPT, Gemini, and Google's AI Overviews decide what to cite largely the same way traditional search does, by rewarding pages that already rank well, state facts plainly, and get referenced elsewhere. Schema markup, llms.txt files, and AI-specific tricks make little measurable difference. What moves citation odds is content built to be lifted and quoted, not content built to game a new algorithm.
Why GEO readiness belongs on your own marketing checklist
Most MSPs treat their own website as a lead form with a logo on top, not a channel competing for space inside an AI answer. That's a mistake. Prospective clients increasingly ask an assistant to compare managed service providers, summarize what a security stack should include, or explain what managed detection and response means before they ever type a query into Google. If your site isn't structured to be a source for those answers, a competitor's is. The research below comes from controlled tests, not vendor guesswork, and it points to a few specific, checkable things worth doing instead of chasing every new AI SEO tactic that shows up in your inbox.
Does adding schema markup get your pages cited by AI search?
No, not on the evidence so far. Ahrefs tested 1,885 pages that added JSON-LD schema between August 2025 and March 2026, matching each one against control pages with similar citation history that never added schema, and found a 4.6 percent decline in citations from Google AI Overviews in the 30 days afterward, according to Ahrefs' analysis of schema markup and AI citations. The same test found a 2.4 percent uptick in Google's AI Mode and a 2.2 percent uptick in ChatGPT, changes the study attributes to random variation rather than the schema itself, according to Ahrefs' test.
| Platform | Citation change after adding schema |
|---|---|
| Google AI Overviews | -4.6% |
| Google AI Mode | +2.4% |
| ChatGPT | +2.2% |
Structured data still has a job. It supports eligibility for conventional rich results and keeps your pages readable by every crawler, not just AI ones. It just isn't the lever that gets a page pulled into an AI-generated answer. If a vendor pitches a package built around AI-specific schema or a proprietary markup standard, treat that pitch with the same skepticism you'd apply to a vendor promising a guaranteed Google ranking. The honest answer is that nobody, including the platforms themselves, has found a markup shortcut yet.
How often does a click actually happen after an AI summary runs?
Rarely, and that changes what ranking is worth. Pew Research Center tracked the browsing of 900 U.S. adults for a full month in 2025 and found an AI-generated summary appeared on 18 percent of the Google searches in that study, according to Pew Research Center's analysis of search behavior and AI summaries. Visits where a summary appeared led to a click on a traditional search result only 8 percent of the time, compared with 15 percent of visits where no summary showed up, and just 1 percent of visits to a summary included a click on a link inside the summary itself, per the same Pew research. For an MSP, the win isn't necessarily a click. It's being the source the assistant names when a prospect asks what a managed backup plan should cost or which compliance frameworks apply to their industry.
What actually improves citation odds, according to the research
A 2023 Princeton-led study that tested generative-engine optimization techniques across a benchmark of search queries found that some techniques improved a source's visibility in AI-generated answers by up to 40 percent, according to the GEO: Generative Engine Optimization paper. The paper's authors also note that the effectiveness of these techniques varies across domains, according to the same GEO paper. The consistent thread across the research is structural. Content that states a clear claim, supports it with a specific fact, and answers a question directly tends to get pulled into AI answers more often than content written primarily to rank in a conventional search results list.
In practice, that means rewriting the kind of page most MSP sites already have. A service page that opens with a vague mission statement before ever saying what the service includes gives an assistant nothing clean to quote. A service page that opens by naming what the service covers, who it's for, and what a client gets gives the assistant a sentence it can lift directly. The same applies to pricing and process pages. An assistant summarizing "how do managed IT services get priced" will favor a page that states its own pricing structure plainly over one that buries it behind a contact form, because the plain statement is the one it can safely repeat without guessing at what you meant.
Mentions and citations are not the same win
Showing up in an AI answer and getting linked from it are two different outcomes, and most brands only get one. A study of 12,000 AI responses collected across Google AI Mode, Google AI Overviews, Gemini, and GPT-5.4-mini found that only 23.1 percent of brand mentions in those answers were also backed by a citation in the same response, according to Buzzstream's study of AI mentions versus citations. Looking at it from the other direction, 69.9 percent of citations in that same study named the brand in the surrounding response text, and roughly 80 percent of citations pointed to pages the brand didn't own, according to the same Buzzstream research. That last figure matters most for an MSP's own marketing. Press coverage, review sites, and partner pages you don't control are carrying most of the citation weight, which means digital PR and third-party placement belong in a GEO plan, not just on-site content work. A listing on a reputable MSP directory, a quoted comment in a trade publication, or a case study hosted on a technology partner's site can end up doing more for citation odds than another round of edits to your own homepage. Budget and attention should follow that evidence rather than stay locked on the website alone.
Does Google require special markup for its AI features?
No. Google's own generative AI optimization guide states that sites don't need new machine-readable files, AI-specific markup, or an llms.txt file to appear in Search's generative AI features, and that standard SEO fundamentals still apply because those features run on Google's core ranking systems, according to Google's generative AI optimization guide. That guidance lines up with what Ahrefs measured independently. The lever is still the content itself: clear claims, direct answers, and facts a model can lift without distorting them.
Turning GEO readiness into a recurring service, not a one-time audit
None of this is a one-time fix. Answer engines change how they retrieve and weigh sources on their own schedule, which means a GEO audit done once in January is stale by summer. The platforms themselves keep shifting too. Google added generative AI performance reporting to Search Console the same year it published the guidance cited above, which means the data needed to track whether a page is actually showing up in AI features is new enough that most agencies haven't built a workflow around it yet. That's an opening for an MSP that already owns the client relationship. A website health check, structure cleanup, and recurring content review can be packaged the same way you already package patch management or backup monitoring, as a flat monthly line item instead of a one-off project. ActiScore runs that audit against a client's site under your own brand and flags the structural issues the research above actually ties to citation, not a generic SEO score that looks good in a slide deck. If you're deciding where GEO readiness fits next to the tools you already resell, the stack builder walks through how it pairs with what's already in your catalog, which sits alongside the rest of the Actiforge product lineup available to white-label.
Treat AI-search visibility as infrastructure, not a marketing trend. The clients asking you to manage their security stack and their backups will eventually ask why their own site never shows up when a prospect asks an assistant for a recommendation. Having an answer, and a packaged way to fix it, is worth more than chasing the next GEO hack. See the full stack.
