How do I optimize my website for AI search?

AI search optimisation is described as a new discipline more often than it deserves. Most of it is SEO you should already be doing. What is genuinely new is three additions — and the order you do them in determines whether the rest of the work counts for anything.

What actually changes

Classic SEO optimises for a position in a list of results. A person sees ten options and picks one. AI search optimises for inclusion in a single generated answer, and for the three to five sources footnoted beneath it.

The overlap is large because assistants are still built on the open web. They retrieve candidate pages using the same signals search engines always have — indexability, relevance, authority — and a page that already ranks well is a strong candidate.

SEO ends when a human clicks. AI optimisation ends when a model decides your sentence is the best available answer and reproduces it with your name attached.

That difference in finish line produces three additions.

Addition 1: retrievability

The most consequential and least discussed change in technical SEO, and the first thing to fix because everything else depends on it.

Googlebot renders JavaScript in a headless Chrome engine. AI crawlers do not. An analysis by Vercel and MERJ covering more than 500 million GPTBot fetches recorded zero JavaScript execution. GPTBot downloaded JavaScript files roughly 11.5% of the time and never ran them, and the same held for ClaudeBot, PerplexityBot, Meta’s crawler and ByteDance’s. They fetch raw HTML under tight timeouts, take what is there, and move on.

What to do:

• Disable JavaScript and load your key pages. If the answer is not there, move that content to server-rendered HTML.
• Check robots.txt and your CDN or WAF for AI crawler blocks. The block is usually at the CDN, inherited from a bot-management preset nobody chose.
• Check server response times. Tight crawler timeouts mean a slow server can produce no crawl at all, silently.

This is binary. Either your content is readable or it is not, and no amount of editorial work compensates for failing it.

Addition 2: extractability

Once you are retrievable, the question becomes whether the model can use what it found.

Answer first, context second. Each section should open with a direct answer in one or two sentences, then add nuance. Models lift passages, not pages, and a self-contained chunk is far easier to quote than one that builds across half a page.

Question-led headings. Phrase them the way a person asks an assistant, because that is what retrieval matches against.

One complete answer per page. A page covering six topics answers none of them cleanly enough to be preferred. Depth on a single question beats breadth across many, which inverts a decade of content-hub advice.

Specific, attributable claims. This is the highest-leverage editorial change available. The Princeton and IIT Delhi GEO study tested optimisation methods across roughly 10,000 queries and found Statistics Addition, Quotation Addition and Cite Sources were the three strongest, delivering roughly 30–40 percent relative visibility improvements. Vague claims get paraphrased and lose attribution; a dated, sourced number gets lifted with your name on it.

Structured data that matches. Article, FAQPage, Organization. Mismatch with the visible text is worse than nothing.

Addition 3: corroboration

The biggest mindset shift from link-era SEO, and the one that takes longest.

Models weigh what the rest of the web says about you more heavily than what you say about yourself, and unlinked mentions count. Reviews, forum threads, roundups, comparison posts and news coverage all shape what a model believes about your brand — with no hyperlink required.

Where SEO chased backlinks, AI optimisation chases presence: a consistent, credible footprint across the places both people and models read. Reddit is cited disproportionately often, which is why forum visibility has become part of the discipline rather than a growth hack.

Consistency matters more than volume. Ten sources describing you identically is stronger evidence than fifty describing you five different ways, because inconsistency gives a model nothing stable to state. In practice that makes a surprising amount of this work editorial housekeeping — the same category description, the same product names, the same positioning, everywhere.

The most actionable input is the citation list itself. When you run your category’s questions through the assistants, the URLs they cite are a ranked map of the sources that influence your category. Being accurately described on those specific domains is worth more than a hundred generic mentions.

The SEO foundation still applies

None of the above replaces the basics, and skipping them makes the additions worthless.

You cannot be cited from a page that was never retrieved, and retrieval depends on being indexed, relevant, internally linked and reasonably authoritative. A site with no organic presence is not in the candidate set an assistant chooses from, so the fix there is foundational SEO rather than GEO tooling.

This is why the honest sequencing is: fix retrievability first because it is binary and cheap, keep doing SEO because it gates everything, then layer extractability and corroboration on top. Teams that abandoned SEO to chase AI visibility have generally ended up weaker at both.

A thirty-day plan

1. Days 1–2. Load key pages without JavaScript. Check robots.txt and CDN for crawler blocks. Fix whatever fails.
2. Days 3–5. Write twenty real buyer questions. Run them through ChatGPT, Gemini and Perplexity. Record who is named and what is cited. This is your baseline.
3. Days 6–10. Write one plain sentence stating your category, audience and alternatives. Put it on the homepage and in Organization schema.
4. Days 11–20. Take the five prompts where competitors are named and you are not. Rewrite or create one page per prompt, answer-first, with one specific number nobody else has.
5. Days 21–30. Work the citation map: get accurately described on the specific third-party domains the assistants already cite in your category.
6. Day 30 onward. Re-scan weekly and read the trend, never a single run.

Five expensive mistakes

Abandoning SEO to chase AI visibility. Retrieval runs on organic signals. Teams that reallocated their entire budget away from SEO have generally ended up weaker at both, because they removed the foundation the new work stands on.

Optimising before measuring. Without a baseline you cannot tell whether anything worked, and generated answers vary enough between runs that post-hoc impressions are worthless. Twenty prompts, run twice, stored — before touching anything.

Judging on Perplexity alone. It reflects changes within days, which makes it a useful early indicator and a misleading measure of overall position. A programme that looks successful in Perplexity and has not moved ChatGPT at all is common and worth knowing about.

Publishing volume. Fifty thin pages produce fifty things nobody quotes. The unit that earns citations is one question answered completely, not coverage across many answered partially.

Treating schema as a shortcut. Structured data supports understanding; it does not manufacture authority. Adding FAQPage markup to a page nothing can read changes nothing at all.

What good looks like after ninety days

A realistic outcome for a site starting from zero: retrievability fixed and verified, a fixed prompt set with three months of stored history, mention rate up meaningfully on the constrained questions you can genuinely win, one or two pages being cited by name, and a clear list of the third-party domains where you still need to be described accurately.

What is not realistic in ninety days: being named on broad category questions dominated by market leaders, or a step change in ChatGPT specifically, which lags the others by the longest margin.

Measuring it

There is no position to track, so track four things per prompt and per engine: whether your brand was mentioned, whether your URL was cited, which competitors appeared, and your share of the answer space across the whole prompt set.

Store the raw answer text. This is the part most teams skip and it is the part that makes the measurement usable — generated answers vary between identical runs, so a score without stored evidence cannot be audited, trended honestly, or defended when a stakeholder disputes it.

Corroborate with analytics, where referral traffic from ChatGPT, Perplexity and AI Overviews is increasingly visible as its own channel. Our guide to measuring AI referral traffic covers the attribution side.

Sources

• Vercel and MERJ — analysis of 500M+ GPTBot fetches finding zero JavaScript execution; GPTBot downloaded JS files ~11.5% of the time without running them; same behaviour for ClaudeBot and PerplexityBot.
• Aggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande — “GEO: Generative Engine Optimization”, ACM SIGKDD 2024 (arXiv:2311.09735); ~30–40% relative visibility improvement from statistics, quotations and cited sources.
• Google — 2.5 billion monthly active users for AI Overviews and 1 billion for AI Mode, stated 19 May 2026.

Frequently asked questions