Open ChatGPT and type the question your best customer would type. Not your brand name, the question. Who should I use for commercial solar in Sydney. What’s the best ad management platform for an agency. Somebody gets named in that answer. If it isn’t you, it’s a competitor, and there’s no auction where you can bid your way in.

Do that before you read the rest of this, and write down who came back. That list is your competitive set now, and it’s often not the one in your board deck.

The mechanics of getting cited are knowable. Most of them are things a marketing manager can brief without a developer in the room.

How an AI answer picks its sources

Two things feed an AI answer. The first is what the model absorbed during training: broad, slow to change, and out of your hands this quarter. The second is retrieval. When you ask something specific or current, the assistant runs a search, opens a small number of pages, and writes its answer from those. Google’s AI Overviews work the same way on top of Google’s own index.

Retrieval is the part you can move. It also means the entry ticket is unglamorous: if your page isn’t among the handful of results that come back for the way the question was phrased, nothing else in this article matters. You can’t be quoted from a page the model never opened.

So the first job is the old job. Rank for the question. Then make the page easy to quote once it’s open. That relationship is the whole argument in our explainer on SEO versus GEO: one builds the authority, the other makes it usable.

Be one thing, and be it consistently

Models work in entities, not strings of text. Before anything gets cited, the system has to satisfy itself that the name in the question and the name on your page are the same organisation. That’s trivial for AGL. It’s hard for a business with three trading names, an about page last touched in 2019, a LinkedIn description that says something different, and a directory listing with the old address.

Fix the boring version first. One company name, spelled one way, in every place you appear. An about page that says plainly what you sell, who you sell it to, and where you operate. The same description on LinkedIn, your Google Business Profile and any industry directory you’re listed in. Where the details disagree, you don’t get a weaker citation. You get no citation, because the model can’t confirm you are who the page claims.

Answer the question before you warm up

Most business writing takes two paragraphs to clear its throat. That habit costs you here. A model looking for an answer will lift a passage that stands on its own, and a passage only stands on its own if it doesn’t lean on the sentence before it.

The practical version: write your subheading as the question a customer would type, then answer it in the next 40 to 60 words. Full sentences. No pronouns pointing back at the heading, no “as mentioned above”. Elaborate underneath for as long as you like. The first block is for the machine and the skim reader; everything after it is for the person who has decided you’re worth their time.

The same logic applies to lists and tables. A comparison table with real header cells gets parsed and quoted. The same comparison written as flowing prose across four paragraphs usually doesn’t.

Own a number nobody else has

Assistants cite whoever sits closest to the fact. If your page is where a figure originates, you get named. If you’re repeating a figure you read somewhere else, the original publisher gets named and you get nothing.

Original numbers are the cheapest advantage most businesses have and never spend. You already hold data nobody outside your company can produce: your own results, your pricing bands, what you see across your client base. Publish it with a date on it and a line on how it was measured.

Ours is a fair example. Pace’s site earned 116 Google impressions in the whole of February 2026, and 113,949 cumulative Google impressions by August, four months after the first new content went live. That’s Pace’s own Search Console data, dated, with the method written up in the Pace case study. Nobody else can source that number, which is exactly the point.

The machine-readable layer

Schema markup is the technical work here that pays for itself. Organization schema so your name, URL and profiles resolve to a single entity. Article schema with a real named author rather than “admin”. FAQPage markup where you genuinely have questions and answers, not bolted onto a sales page because a plugin offered.

Then check you aren’t blocking the crawlers that matter. GPTBot, ClaudeBot, PerplexityBot and Google-Extended are separate user agents from Googlebot, and plenty of sites block them by accident through a blanket robots.txt rule or a firewall setting nobody remembers switching on. It takes ten minutes to check.

We also publish a plain markdown copy of every page on this site at index.md, plus an llms.txt index at the root. Honest assessment: no assistant has committed to reading llms.txt, and I’d be sceptical of anyone selling it as a ranking factor. We run it because it’s cheap, and because a clean text version of a page leaves nothing to interpret for anything that does fetch the URL. If it turns out to do nothing, it cost us an afternoon.

Get mentioned where the models already read

Your own site is one input. The bigger one is everywhere else. Models learn who to trust from seeing a brand named repeatedly across sources they already weight heavily: trade publications, established directories, review sites, forums, and the reference sites those feed into.

A link buried in a sponsor footer does very little for this. A sentence that names your company, says what it does and quotes a person with a job title does a lot. That’s ordinary digital PR, judged on a different outcome.

Start from the answers you collected at the top of this page. Whoever got cited for your category questions is already in those sources. Find where they’re mentioned and go and earn a mention in the same places. It’s slow work and there’s no version of it you can buy in a week.

Measure it like you measure rankings

Write down 20 questions your customers actually ask. Run them through ChatGPT, Gemini, Perplexity and a Google search with AI Overviews showing, once a month, and record who gets named each time. Screenshot the answers, because they don’t stay put.

Two warnings. Answers vary between runs and between accounts, so treat any single result as a sample rather than a position. And query volume here is unknowable; none of the assistants report it, so anyone quoting you a search volume for AI answers is guessing. What you can measure is share: out of 20 questions, how many name you, and how that number moves over six months.

On the Google side there is a harder number available. AGL’s search programme with us produced 14.5k features in Google’s AI Overviews, alongside a 58% lift in average search position from the same engagement. Both came out of the same work described above rather than a separate AI project.

Most of this is the authority work search has always rewarded, arranged so a machine can use it. If you want the full version of how we run it, that’s generative engine optimisation. If you’d rather have the theory first, start with SEO vs GEO.

Either way, run the twenty questions. You’ll know within an hour how big your problem is.

How to get your brand cited in ChatGPT and AI Overviews