Search stopped being a list of links some time ago. A growing share of questions now get answered directly on the page, or inside a chat interface that names a handful of sources and moves on.
That changes the goal. You are no longer trying to occupy position three. You are trying to be one of the few sources a model chooses to quote.
This guide is a practical sequence for doing that, written for UK businesses rather than for enterprise SEO teams. It assumes you have a normal website and no dedicated technical resource.
How does AI search actually pick its sources?
Traditional search ranks documents. AI search retrieves passages, then writes an answer from them, then cites where the passages came from.
That distinction drives everything else. A page can rank well and never be cited, because its useful content is buried in the eleventh paragraph behind a personal anecdote and a definition of the industry.
The retrieval step usually happens in three stages. The system reformulates your question into several related queries, gathers candidate passages from its index and from live web results, then selects the few it can most confidently use.
What this means in practice. Your competition is not the page ranking above you. It is the specific paragraph that answers the question more cleanly than yours does.
Google has published its own guidance on this and it is worth reading in full: the AI features and your website guide and the succeeding in AI search post both make the same core point, which is that there is no separate AI ranking system to optimise for.
The platforms differ in how they source. AI Overviews and AI Mode work from Google’s index. ChatGPT combines what is in its training data with live retrieval through its own crawler, documented in OpenAI’s bot documentation. Perplexity leans heavily on live retrieval and cites more visibly than either.
Step 1: Make sure AI crawlers can reach your content
This is the step that quietly disqualifies businesses, and it takes ten minutes to check.
Open your robots.txt file and read it properly. Look for blocks on GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot, Google-Extended and CCBot, because a blanket disallow added at some point in the past will keep you out of every system that respects it.
Decide deliberately rather than by inheritance. Blocking training crawlers while allowing search crawlers is a legitimate position, but blocking everything and then asking why you are not cited is not.
Then check that your content is visible without JavaScript. If the main text of your page is rendered client side, some retrieval systems will see an empty shell, so view the raw HTML source and confirm your key paragraphs are actually in it.
Step 2: Structure content so it can be extracted
Write so that any single section can be lifted out and still make sense on its own.
That means self-contained sections, no unresolved pronouns at the start of a paragraph, and no answers that depend on a sentence three screens earlier. Assume the reader arrived at that heading and nowhere else.
Practical rules that make a measurable difference:
- Keep paragraphs to three sentences or fewer, so a passage boundary lands cleanly
- Use question-led H2s that match how people actually ask, not clever headings
- Put the direct answer immediately under the heading, then the nuance
- Use a table wherever you are comparing more than two things on more than two dimensions
- Define terms explicitly the first time, in a sentence that starts with the term itself
Step 3: Answer the question in the first 60 words
The opening block is the single highest-leverage part of any page. It is what gets lifted verbatim.
Write it as a standalone answer that would satisfy someone who reads nothing else. Include the specific detail, because vague answers are unusable to a system that has to commit to a claim.
Before:
In today’s fast-moving business environment, understanding the cost of call answering can be a complex undertaking. There are many factors at play, and every business is different. In this article we will explore the various considerations that go into pricing.
After:
UK call answering services typically cost between 75p and £1.50 per minute, £1 to £3 per call, or £50 to £300 a month on a bundled plan. Out-of-hours cover usually carries a surcharge. An AI receptionist is normally a flat monthly fee with no per-call charge.
The second version contains numbers, ranges and a comparison. It is quotable, checkable and useful. The first version says nothing at all, and no retrieval system can do anything with it.
Step 4: Strengthen your entity and brand mentions off-site
Models corroborate. If your site is the only place a claim appears, it is treated as a claim rather than a fact.
The practical work here is unglamorous. Make sure your business name, address, phone number and description are identical everywhere they appear, so that a model reading three sources sees one consistent entity rather than three similar ones.
Then work on being mentioned in places you do not control: industry directories, trade press, supplier and partner sites, local press, podcasts, and any credible roundup in your sector. Unlinked mentions still count here, which is a meaningful difference from classic link building.
Get the basics right too. A complete Google Business Profile, a Companies House record that matches your trading name, and consistent LinkedIn details all feed the same picture.
Step 5: Add structured data
Structured data does not buy you a citation. It removes ambiguity, which is a different and still worthwhile thing.
Three types cover most business sites, all documented at schema.org:
- Organization on the site as a whole, establishing name, logo, contact details and social profiles
- Article or BlogPosting on content, establishing author, publish date and last-updated date
- FAQPage on any page with a genuine question and answer section
Validate what you add with Google’s Rich Results Test and confirm there are no errors. Markup that does not match the visible content on the page is worse than no markup at all.
Step 6: Earn third-party citations
For a large class of queries, the sources a model quotes are not vendor sites at all. They are roundups, comparisons and listicles written by someone else.
Ask a model the questions your buyers ask, then look at what it cites. If the answer to “best X in Manchester” is drawn from three directory pages and a local blog, then getting accurately listed on those four pages is worth more than any change to your homepage.
This is the highest-leverage step for most small businesses and the one nearly everyone skips. It is also slow, which is why it is worth starting first.
Step 7: Measure and iterate
You cannot manage this without a baseline, and the baseline does not require buying anything.
Build a fixed set of 15 to 20 prompts spanning the buying journey, from “what is X” through “best X for Y” to “X vs Y”. Run them across the platforms that matter to you, record whether you are mentioned, whether you are cited with a link, and what is said about you.
Run the same set monthly. The same prompt returns different answers on different runs, so check each prompt a few times and record how often you appear rather than whether you appeared once.
On the analytics side, watch referral traffic from AI platforms separately in your analytics, and watch impressions and clicks by query type in Search Console.
What does not work?
A few things circulate widely and are worth dismissing quickly.
Keyword density. Retrieval works on meaning, not on term frequency. Repeating a phrase does not make a passage more retrievable, it makes it worse to read.
Paying for placement. There is no mechanism to buy inclusion in AI Overviews or in a ChatGPT answer. Advertising placed alongside AI features is advertising.
Publishing more of the same. Volume without differentiation produces content that no model needs to cite, because the point is already covered by a source it trusts more.
Chasing every new file convention. Proposals like llms.txt are cheap to adopt and currently marginal. They are not a shortcut past crawlability, extractability and credibility.
Treating GEO, AEO and AI search optimisation as three disciplines. They are three names for the same work. Pick one term, use it consistently, and spend the energy on the work instead. For what it is worth, US-published guidance usually says optimise for AI search; we use optimise, and the practice is identical.
Where to start if you only have a day
Check robots.txt and confirm nothing you care about is blocked. Rewrite the opening 60 words of your five most important pages into direct answers. Add FAQPage markup to the two pages that already have real questions on them.
Then run a baseline prompt set and diarise a repeat for a month later. That is a day’s work and it puts you ahead of most of your competitors, who have done none of it.