What actually changed
Search used to end at a list of links. Increasingly it ends at an answer, with two or three sources cited beneath it. That collapses ten results into a shortlist of three, and being fourth is now identical to being fortieth.
The mechanics differ from ranking, too. An AI engine is not matching a keyword to a page — it is retrieving what it can find about an entity, checking whether the sources agree, and generating a recommendation. So being consistently described across the web matters as much as any single page, and being mentioned somewhere trusted can beat ranking first on your own site.
What we can evidence
This is a field with a great deal of confident theorising and very little published data, so here is ours. Across the owned property portfolio: 302,037 occasions where Bing Copilot cited a property in an AI answer, and 147,783 appearances inside Google AI features. Both figures come from platform exports published on the case studies, with the window, filter and methodology visible on the screenshot.
Equally important is what we cannot evidence. ChatGPT, Gemini and Perplexity publish no per-answer export, so we do not claim citation counts for them. On our own case studies those surfaces are explicitly marked not audited rather than implied — and any agency quoting you a precise ChatGPT citation figure should be asked where the export came from.
What the work consists of
The uncomfortable truth is that most of it is SEO done properly. There is no separate AI channel to buy.
- Entity clarity. Consistent name, description, location and service list everywhere you appear, so an LLM resolves you to one business rather than three ambiguous mentions
- Structured data. Organization, LocalBusiness, Service, FAQ and Product markup, so machines read facts instead of inferring them
- Answer-shaped content. A direct answer in the first two sentences, then the detail — because that first block is what gets extracted
- Citation sources. Presence on the directories, listicles and publications that AI engines actually quote in your category
- Crawlability for AI agents. GPTBot, ClaudeBot, PerplexityBot and Google-Extended mostly do not execute JavaScript. A site that hides its copy behind hydration is invisible to them
- Share-of-voice tracking. Running the prompts your buyers run, on a schedule, and recording who gets named
Why this matters more in Sri Lanka, not less
It is tempting to file AI search under “later” in a smaller market. That gets it backwards. Sri Lankan search volumes are small enough that an AI assistant has fewer strong local sources to choose from — so the businesses that establish themselves as the consistent, well-structured answer now will be difficult to displace later.
There is also a language dimension. Ask an AI assistant to recommend a supplier in Sri Lanka and it will draw on English-language sources almost exclusively, because that is what exists. Being the best-structured English source about your category here is an unusually cheap position to take.
What nobody can promise
Nobody controls whether an LLM cites you, any more than anyone controls Google’s rankings — and the AI surfaces change faster than search ever did. Anyone selling guaranteed ChatGPT placement is selling something they cannot deliver.
What is reliable is the underlying work, because entity clarity, structured data and being cited by trusted sources have been good SEO for a decade and are now good GEO as well. If AI search stalls entirely, none of it is wasted.