AI search optimization is the work of making your website easy for AI systems to read, worth quoting, and possible to verify, so that when ChatGPT, Perplexity, Gemini or Google's AI Overviews answer a question in your field, your site is one of the sources behind the answer. It goes by several names (GEO, AEO, LLMO), but the names describe one discipline with three parts: access, content, and evidence.
Why does AI search need its own optimization?
Classic search returns a list of pages and lets the searcher choose. AI search returns an answer, already written, with a handful of sources behind it. That changes what winning means. In a ranked list, position three still gets read. In a composed answer, you are either one of the sources or you are absent, and absence is invisible: no impression, no click, no trace in your analytics.
It also changes what gets read. A person skims a page top to bottom. A language model retrieves passages: self-contained chunks of text that answer a specific question. A page can rank well in Google and still offer nothing an AI system can lift, because every claim leans on the paragraph before it and nothing stands alone.
Is this different from SEO?
It overlaps, but it is not the same job. The technical floor is shared: a crawlable site, clean HTML, working sitemap, fast pages. If your SEO house is in order, you are not starting from zero.
The differences sit above that floor. SEO optimizes pages to rank for keywords. AI search optimization optimizes passages to be quoted in answers. SEO measures positions and clicks. AI search visibility is measured by asking the engines your buyers' questions and recording which sources they cite. SEO earns links; AI search also weighs whether your claims can be corroborated elsewhere, because a model is more confident citing a source the rest of the web agrees with. We wrote a fuller comparison in our piece on GEO versus SEO; the short version is that one is a superset of neither, and doing one well does not automatically do the other.
Which AI search surfaces matter?
Four, today. Google's AI Overviews sit on top of normal results and lift passages from indexed pages, so ordinary Google indexing is the entry ticket. ChatGPT browses the web for current questions and cites what it reads. Perplexity is built entirely around cited answers and is the most transparent about its sources. Gemini and Copilot draw on Google's and Bing's indexes respectively, which makes Bing indexing suddenly relevant again.
They differ in mechanics but reward the same properties: pages that load without executing JavaScript, answers stated plainly near the top, and named sources behind factual claims. Optimize for the properties, not for one engine's quirks.
What does the work actually involve?
Three layers, in order.
Access. AI crawlers have to be able to fetch and parse your pages. That means checking robots.txt against the current list of AI user agents (GPTBot, ClaudeBot, PerplexityBot and the rest), serving real HTML rather than a JavaScript shell, and keeping key content out of PDFs and images. Many AI crawlers do not render JavaScript, so a site that only exists after a framework boots is partly invisible to them. This is one reason we build hand-written, framework-free sites: the HTML a crawler fetches is the page, nothing deferred.
Content. Each page should answer one question directly, in the first paragraph, in language that survives being quoted out of context. Follow-up questions become H2 sections. Vague positioning ("solutions for a changing world") gives a model nothing to work with; specific, checkable statements ("a booking flow that confirms in three steps") give it something to quote. This is the slowest layer and the one that compounds.
Evidence. Structured data (Organization, Article, FAQ schema) tells machines who is speaking. An llms.txt file summarises the site for AI systems in plain text. Consistent entity details across your site, registries and profiles let a model confirm you are who you say you are. None of this buys a citation on its own; it removes the reasons to doubt one.
How do you measure whether it works?
Directly: keep a fixed list of the questions your customers ask, put them to each engine monthly, and record which domains get cited. Movement on that list is the primary scoreboard. Secondarily: watch AI referral traffic in analytics (chatgpt.com, perplexity.ai and gemini.google.com show up as referrers) and watch AI crawler hits in your server logs. We run this loop on kiwastudio.com ourselves, and the pattern we see is consistent: citations follow within weeks of a page that answers one question cleanly, and never follow pages that answer none.
Expect modest referral volumes and treat them accordingly. The value of a citation is less the click and more the recommendation: the engine has already vouched for you by the time anyone arrives.
Where should you start?
In this order. First, spend an hour asking ChatGPT, Perplexity and Google the ten questions that matter to your business, and write down who gets cited. That is your baseline and your competitor list. Second, fix access: robots.txt, rendering, sitemap. It is the cheapest layer and it gates everything else. Third, rewrite your most important pages so each opens by answering its question in two or three quotable sentences. Fourth, add the evidence layer: schema, llms.txt, consistent entity details. Then re-run your baseline monthly and let the results direct the next round of writing.
None of this is exotic. It is mostly the discipline of saying what you mean, plainly, on pages a machine can read. The sites that struggle with AI search tend to struggle because they never quite said anything.
Findable by people. Citable by machines.
See how we run GEO