AI Search Optimisation: Separating Evidence From Assumption
AI Search Optimisation: Separating Evidence From Assumption
Ask five people in search what “optimising for AI search” involves and you’ll get five different answers. One will mention schema. Another will bring up llms.txt files. Someone will say GEO with complete confidence, and nobody will ask them to define it, because perhaps they’re not entirely sure themselves.
The terminology is a bit messy
Wikipedia’s own entry on the subject admits there’s no consensus definition as of early 2026, and that GEO, AEO, AIO, LLMO and GSO (ee-ay-ee-ay-oh) get used more or less interchangeably depending on who’s speaking or posting about it.
Then in May 2026, Google settled part of the argument itself. Its first official guidance on the topic says that from Search’s perspective, optimising for generative AI features is optimising for the search experience, and is therefore still SEO. It used the moment to squash a few methods agencies had been shouting about: llms.txt files, content “chunking,” and AI-specific schema markup all got called out as unnecessary for its own AI features.
Its worth flagging that this is Google talking about Google. ChatGPT, Perplexity and Claude run on different retrieval systems, and what earns visibility on one doesn’t automatically translate to another. Research tracking 680 million AI citations found only around 11% of domains get cited by both ChatGPT and Perplexity. So there isn’t really one thing called “AI search” to optimise for. There are several separate ecosystems, loosely connected at best, and each one has its own “habits”.
What the evidence actually says works
Set the vagueness aside and look at what’s been tested, and a clearer picture starts to form. Some of it runs against what’s been sold as best practice for the past couple of years.
Brand mentions matter more than backlinks. Ahrefs looked at 75,000 brands and found branded web mentions correlate with AI Overview visibility at roughly three times the strength of backlinks. A later expansion of the same study found YouTube mentions correlate even more strongly.
Schema isn’t all its cracked up to be. Ahrefs ran a controlled test on 1,885 pages that added structured data and compared them against a control group. The result for Google AI Overviews was a small negative effect. The movement for AI Mode and ChatGPT was negligable. Pages that get cited a lot do tend to have schema on them, but that’s because sites doing the content and authority work properly also tend to tick the technical boxes. The markup itself doesn’t seem to be doing much.
Specific, attributable content earns citations. Several other studies have found that adding statistics and quotations to a page lifted its visibility in generated answers. Original research and data-led pages earn citations at several times the rate of standard blog content.
Earned media does most of the heavy lifting. Muck Rack’s analysis of over 25 million links cited across ChatGPT, Claude and Gemini found earned media accounts for 84% of citations, with paid and advertorial content down at 0.3%. That figure’s held steady across three separate rounds of the same study since mid 2025.
How we approach GEO
Given all that, here’s how we approach GEO, and it’s working well for our clients.
We start with the prompts that matter commercially, not the ones that are easy to rank a citation count against, working out what a buyer actually types into ChatGPT or Gemini when they’re three quarters of the way to a decision, and building a governed list of those questions to track properly.
From there, we measure share of voice against named competitors on those questions, rather than counting citations in isolation. Share of voice on the questions that actually drive pipeline is the number that’s ultimately going to make a tangible difference.
The work behind that is mostly the work that’s always mattered, just aimed differently. Earned third party authority (digital PR, credible mentions, coverage). Clear entity signals help engines understand who you are and what you’re credible on. Fact dense, original content gives engines something specific to lift and cite, rather than something to skim past.
It fits Google’s “still SEO” position, because the fundamentals it leans on are the ones that have always mattered. It fits the earned media finding. It fits the schema result, because it’s about the substance of the content rather than the markup around it. And it fits the intent research, because it’s built around commercial prompts rather than informational visibility that never turns into anything.
Where it actually gets you
AI referral traffic is still tiny in comparison. It averages around 1% of total sessions across the sites according to multiple studies.
The visitors who do arrive convert well, though. and when you think about it, its quite obvious why. The AI’s already done the shortlisting. These visitors aren’t arriving to browse. They’re arriving to confirm a decision they’ve mostly already made, and probably been in a long chat with chat GPT about prior.
Treat GEO as a pre-qualification layer sitting in front of a small but valuable portion of your buyers, worth building for properly, as long as you’re building for the prompts your buyers actually type rather than chasing citations for their own sake.
If you want to know where you actually stand, forget the citation count. Run your category’s real buyer intent questions across the major engines and see who gets named. Including, sometimes uncomfortably, whether it’s you.
The AI Search Kickstarter will tell you, in a couple of weeks, not a couple of guesses. Worth booking a discovery call to see where you stand.
