How to appear in ChatGPT and Google AI Overviews
Buyers now ask an assistant before they ask Google, and the assistant names two or three businesses. This is what decides whether yours is one of them. We wrote it as an agency that does this work daily, so we will not pretend it is guaranteed.
By Raoul (Alex) Müller, Founder ·
You appear in ChatGPT and Google AI Overviews by being the clearest, most consistent and most corroborated source on a question. In practice that means pages an AI crawler can reach, answers written so a machine can lift them without editing, business facts that never contradict themselves across the web, structured data that matches what a visitor sees, real reviews, and credible mentions on sites you do not own. No one can guarantee a citation.
What an assistant is actually doing when it names a business
It helps to be precise about the machinery, because most advice on this is written as though the model remembers you. It does not.
When someone asks ChatGPT for a plumber in Leeds or an SEO agency in London, the assistant runs one or more searches, fetches a handful of pages, reads them, and writes an answer grounded in what it just read. Likewise, Google’s AI Overviews works from Google’s index in much the same way. Perplexity, Gemini and Claude follow the same shape: retrieve, read, synthesise, cite.
That has three consequences. Your page has to be retrievable — reachable, indexed and relevant to the phrasing a buyer uses. It has to be liftable — the answer sitting in a self-contained passage a machine can quote without stitching six paragraphs together. And it has to be corroborated, because a model asked to name a trustworthy supplier leans on what independent sources agree about you.
Everything below is a way of improving one of those three properties. There is nothing else in the box.
The six things that decide whether you get quoted
Roughly in the order they are worth fixing. Most businesses lose on the first three and have never checked.
Pages a crawler can actually reach
If a page needs JavaScript to render its main content, sits behind a cookie wall, returns a soft 404, or robots.txt blocks it, no machine can quote it. Several AI crawlers execute scripts less capably than Googlebot, so content Google eventually renders may still stay invisible to them. Check robots.txt for blanket disallows of AI user agents that a plugin or a past developer added, view your key pages with scripts off, and confirm that search engines have indexed the important ones. It is unglamorous, and it is where most losses happen.
Answers a machine can lift without editing
A model quoting you wants a passage that survives being pulled out of the page. Put a direct, self-contained answer to the page’s core question within the first screen, in forty to eighty words, naming the subject rather than saying “it”. Then give the depth underneath, for the human. Headings should be the questions people actually ask, not clever labels. A machine quotes tables, short definitions and clean lists far more readily than a wall of prose, because their meaning survives extraction.
Business facts that never contradict themselves
Name, address, phone number, trading name, services, opening hours, founder, company number. If your website says one thing, your Google Business Profile another and an old directory a third, a model has to decide which version to trust, and the cheapest decision is to cite someone unambiguous instead. First, pick a canonical version of every fact and make the web agree with it. For example, where a business trades under more than one name, say so plainly — we do, with our own 102 Google reviews, filed under Digital Movement Australia, the same business under its original name.
Schema that matches what a visitor sees
Organization, LocalBusiness, Service, Article, FAQPage and Person markup give a machine an unambiguous version of what your page says. However, the rule that matters is that markup must describe visible content. Markup for reviews you do not display, prices that appear nowhere, or FAQs that are not on the page is a manipulation risk, not an advantage. Used properly, schema attaches a stable identity to your business: the same entity, described the same way, everywhere.
Real reviews, in the places assistants read
Ask any assistant to recommend a local supplier and watch what it cites: Google Business Profile, sometimes Trustpilot or a sector site, often a forum thread. Volume, recency and wording all feed the summary, which is why a steady trickle of genuine reviews beats a burst. Never buy reviews. Beyond the ethical and legal problem, purchased reviews read as generic, and generic text gives a model nothing specific to quote.
Credible mentions on sites you do not own
Corroboration is the signal you cannot fake cheaply. Trade press, local news, professional bodies, supplier and partner pages, real industry directories, podcast appearances, forum threads where someone recommends you unprompted. A model deciding who belongs looks for agreement between independent sources. For instance, one mention in a place your industry respects does more than fifty in places it does not.
What the gap looks like in real data
One Semrush comparison, for “dentist London”, shows how tightly AI visibility tracks corroboration.
| Semrush metric | D3 Dental | Bupa Dental |
|---|---|---|
| Authority score | 14 | 71 |
| AI visibility | 0 | 81 |
| Referring domains | 168 | 21,500 |
| Organic visits per month | — | 2.4 million |
Zero against eighty-one is not a rounding difference. AI answers about dental care quote one of these organisations constantly; the other stays effectively absent, despite being a real practice with a real website and 168 sites linking to it.
Be careful what this proves, though. This is one comparison, and the causation runs both ways: Google names Bupa because independent sources corroborate it, and those sources partly corroborate it because Google names it. However, that does not mean a single-site practice can never earn a citation. It does mean that on a broad question like “dentist London”, the assistant reaches for the source the rest of the web already agrees on.
So the lesson for a smaller business is not to fight that fight. Instead, pick the questions where you can plausibly be the best-corroborated answer — narrower services, specific towns, particular buyer situations — and win those. We took the larger example apart in our breakdown of how a UK health brand earns citations at that scale.
Nobody can guarantee you a citation
This needs saying plainly, because the market is filling with people who will not.
AI answers are not a ranking list. The same question, asked twice, can produce different sources. Ask it through two different accounts, in two different locations, with two different histories, and it can produce different sources again. Providers also retune models and change how they ground answers without notice. As a result, a business can appear on Monday, vanish on Wednesday and return on Friday without changing a word on its website.
There is also no submission form, no paid placement and no file you can upload that puts you in the answer. We publish an llms.txt file on this site because it is a courteous way to state our facts in one place, not because it buys a citation. It does not.
So the honest promise is a probability, not a place. Do the six things above properly and you become a likelier source on the questions you care about. That is what we sell and what we report on. Any agency guaranteeing a ChatGPT citation is either misunderstanding the technology or hoping you do.
A fixed set of buyer questions, tested monthly across Google AI Overviews, AI Mode, ChatGPT, Gemini, Claude and Perplexity, with a record of whether you appeared, who appeared instead, and which sources the answer used. Movement in that record is the measurable outcome; a guarantee is not.
Five things that are being sold and do not work
- Publishing hundreds of thin answer pages. Volume without substance gives a model nothing worth quoting and dilutes the pages that were working.
- Schema types invented for the occasion. If it is not in the schema.org vocabulary, it is decoration.
- Paying for mentions on sites nobody in your industry reads. Corroboration counts only when the sources are independent and credible.
- Rewriting a whole site in robot-friendly prose. Pages that irritate humans lose the rankings that feed retrieval in the first place.
- Chasing a single prompt result. One flattering screenshot proves nothing; a tracked prompt set over months proves something.
The pattern behind all five is the same: an attempt to signal quality without producing it. However, the systems you are optimising for look past exactly that, because their commercial value depends on giving people answers that hold up.
What does work overlaps heavily with ordinary technical and content SEO: fix what cannot be crawled, write the answer a buyer actually needs, keep every fact about your business identical everywhere, and earn a few mentions a sceptical person would respect.
How to tell whether any of it is working
Three habits, none expensive, that turn this from a feeling into a record.
Fix a prompt set and never change it
First, write down twenty to thirty questions a real buyer would type, in their words. Then run the same list every month across the assistants that matter. Record whether you appeared, who did, and which pages the assistant cited. The value is in the consistency, so resist the urge to improve the list.
Watch the sources, not just yourself
The pages cited instead of yours are a free brief. If three answers keep quoting a comparison table, a pricing page or a forum thread, that is the format the question wants. Build the version you can honestly be best at.
Ask new enquiries how they found you
Assistants pass very little referrer data, so analytics under-reports them badly. One extra field on the enquiry form, recorded against each deal in a CRM that keeps the source attached, will tell you more this quarter than any tool.
Find out where you stand today
Tell us the questions your buyers ask. We will run them, show you who appears instead of you, and say what we would change first. AI search optimisation is our own specialism, and the one thing on this site we test on ourselves every month.
Get your AI visibility checked