A Website Built for AI Search: How to Appear in ChatGPT and Google AI Overviews
Search has stopped being a list of blue links. Google shows AI Overviews, ChatGPT answers in prose and names its sources, and people increasingly get what they need without visiting a single website. On 2026 figures, AI Overviews cut clicks on the top organic result by roughly a third, and on queries that trigger an AI answer users click through to a traditional result only about 8% of the time.
For a business this changes the objective. The fight used to be for position. Now it is for the mention. Ranking first and never appearing in the AI answer is now worse than ranking fifth and being cited.
How AI optimisation differs from ordinary SEO
| Parameter | Classic SEO | AI search optimisation |
|---|---|---|
| Objective | Position in results | Citation frequency in answers |
| Unit of optimisation | A page per query | A passage per question |
| What matters in the text | Keywords, length | Direct answers, structure, facts |
| Primary trust signal | Backlinks | Consistency of mentions across sources |
| Success metric | Traffic | Presence in answers and branded demand |
How a model decides whom to cite
Simplified, the mechanism works like this: the system breaks a question into sub-questions, retrieves passages that answer them directly, and assembles a response. The winner is not the longest text but the one from which a complete thought is easiest to extract.
Three practical consequences follow:
- A paragraph must stand on its own. If it only makes sense after the three before it, it will not be used.
- The answer goes first, the reasoning after. The classic “build-up, then the point” structure loses here.
- Specifics beat adjectives. A number, a timeframe, a name gets quoted; “individual approach” does not.
What to do with site structure
Questions as headings
Subheadings phrased as the questions clients actually ask give the model a ready anchor. Not “Our Services” but “How much does a corporate website cost”.
A short answer directly under the heading
Two or three sentences that fully answer the question. Elaboration follows. That passage is what ends up in the answer.
Tables and lists
Structured data is parsed more reliably than solid prose. Comparisons, stages and criteria are better presented as a table.
Schema.org markup
FAQPage, Article, Organization, Product. It is not a guarantee of citation, but it is how you tell a machine which part is the question, which is the answer, and which is the company name.
The technical minimum
- Content in HTML, not only in JavaScript. Some crawlers do not execute scripts. If text appears only after rendering, the page is empty to them.
- Speed. Slow pages are crawled less often and less thoroughly.
- Access for AI crawlers. Decide deliberately in robots.txt who is allowed: GPTBot, PerplexityBot, ClaudeBot, Google-Extended. Blocking reduces risk but also removes you from answers.
- Clean heading structure. One h1, sequential h2 and h3 with no skipped levels.
- Update dates. Freshness influences source selection in fast-moving topics.
Off-site mentions matter more than you think
A model does not form its picture of a company from one website. It reconciles what it sees in many places: directories, industry platforms, discussions, profiles, publications.
In practice this means the company name, description, city and contact details must match everywhere. If three sources describe you differently, the model either picks one at random or leaves you out entirely. Consistency functions as a trust signal here.
How to measure the result
Standard analytics will not show this: a visit from an AI answer often looks like direct traffic. What to do instead:
- Regularly put twenty to thirty questions you should be found for to the models, and record whether you are mentioned and in what context.
- Watch branded search demand. Growth in direct brand queries while traffic stays flat is a classic sign that you are being cited.
- Look for referrals from chatbot domains where they are passed.
- Ask new clients how they found you. “ChatGPT recommended you” has stopped being exotic.
What not to do
- Generate hundreds of pages covering every phrasing. Models collapse duplicates well and ignore thin content.
- Hide text behind tabs and accordions unnecessarily. Anything absent from the initial source risks being ignored.
- Write for word count. A long text without facts gives nothing to quote.
- Block every crawler “just in case”. That should be a deliberate decision, not a default.
Where to start this week
Take the five questions clients ask most and give each its own block: the question as a heading, a short answer beneath it, detail below, facts and figures inside. Add FAQ markup. Verify the text is visible in the page source.
This is not a one-off campaign but a change in how you write. A site that is easy to extract an answer from also wins in classic search, where the same passages feed rich snippets.
