Three things decide whether an AI tool names your business in an answer: a page that gives the answer plainly and early, a real named person behind it whose experience holds up, and other sites the tool already trusts saying similar things about you. The technical side – schema, structured data, the various files people now recommend adding – helps a machine read your site. On the current evidence, it isn’t what gets you named.
That’s the short version. The rest of this is the practical one, in the order we’d work through it.
This is the follow-on to our guide about what GEO is and how AI search works. If the whole idea is new to you, start there. If you know the theory and want to know what to do about it on Monday morning, keep reading.
Start with the things that do make a difference
There’s a lot of advice about AI search, but a fair amount of it doesn’t survive rigorous testing, interrogation or research. You’ll read that adding a particular schema type multiplies your citations by some exact-sounding amount, or that a new file in your site’s root directory is now essential. Some of this advice is delivered with a confidence that simply isn’t warranted.
The findings that do stand up to scrutiny are duller and more useful. Independent testing fails to find a reliable link between technical markup and whether an AI tool cites a page. What it points at instead is the content: whether it’s clear, whether it’s specific, and whether other credible sources say the same thing about you.
So we start with the content and the mentions, and treat the technical work as support. If you take one thing from this page, take the priority order of all of this stuff, and that’s as shown below, with the best actions first.
Answer the question early, and plainly
AI tools that search the web while they answer a question or search query tend to judge a page on how well its opening section responds to the question at hand. They read the top of the page, take the part that answers it, and rewrite that into their reply. A page that leads with the answer is easy to lift. A page that clears its throat for six paragraphs isn’t.
So the answer goes first, and the detail sits underneath for the reader who wants it. That’s close to the reverse of how most marketing copy is built, where the useful bit arrives after the scene-setting, and it takes some unlearning.
On a recent project we rebuilt a set of pages around this idea. Each one opened with a short answer block, two or three sentences responding to the exact question someone would ask, followed by a plain comparison table of the options involved. The thinking was simple enough. Give the machine, and the person, an answer in a form that can be quoted without rewriting it, and you become the easiest source to reach for.
Write things a machine can quote without guessing
Leading with the answer is half of it. The other half is writing sentences that can be pulled out of the page and still make sense in their own right.
Be specific. AI tools favour real figures, named things and clear claims over hedged prose. “We have extensive experience in complex payroll” gives an AI assistant nothing it can use. “We run weekly payroll for construction firms, including CIS deductions and subcontractor verification” can be lifted straight into an answer. It’s also more useful to the human reading it.
Keep each point self-contained. If a claim only makes sense once you’ve read the three paragraphs above it, it’s hard to extract cleanly. Say the whole thing in one place, even if that means repeating a few words you’ve already used.
Use structure where it earns its place. A comparison table, a one-line definition, a clear sequence of steps: all are easy to read and reuse. But there’s a point where content optimised for machines stops working for people – all bullets and no sentences for instance – and a page nobody wants to read earns no links and no trust. Write for the person first. Extractability comes from clarity, not instead of it.
Frame some headings as the questions people ask. “Do I need to register for VAT?” is more AI-friendly than “VAT registration considerations”, because it mirrors how someone types a question into an AI tool and puts the matching answer directly underneath.
Get mentioned where AI tools are already looking
This is the lever most often ignored, and frequently the strongest one. A large share of what AI tools cite is coverage on other people’s sites rather than a brand’s own pages. AI assistants are more persuaded by other people describing you than by you describing yourself, which is a fair instinct on their part.
The work behind this is old-fashioned and slow. Get quoted in the trade titles your sector reads. Get listed accurately in the directories and membership bodies your buyers check, and audit the listings you already have, because a good number of them will still have an old address on them. Make sure that when someone credible writes about your field, you’re one of the names they reach for.
That’s as much about digital PR and relationship-building as SEO, and it doesn’t happen in a fortnight. It’s also the hardest part to fake, which is exactly why it counts for so much.
Put a real, named person behind the content
AI systems are built to favour content that shows real experience and authority. For a business, that’s less about any single tag and more about substance: was this written by someone who does this work day in, day out, and can anyone check that?
So name your authors, and give them bios that explain why they’re an authority. Let the expertise show in the writing itself: in the detail, learned opinion and caveats that can only come from having done the thing. Content assembled to fill a page reads as content assembled to fill a page – both to people and, increasingly, machines.
What about schema and llms.txt?
To be frank this is where we part company with a lot of the advice out there.
Schema markup (the structured data in your page’s code that tells search engines what your content is) is useful, and you should use it. It powers rich results in classic search, it keeps your business details consistent across everything that reads them, and Microsoft has said it helps its systems understand content. Filled in properly, with real detail rather than the bare minimum, it’s worth having.
What it doesn’t appear to do, on the current evidence, is reliably earn you a citation in an AI answer. Independent testing has repeatedly found no dependable link between adding markup and getting cited, and tools reading a page live often skip the code and take their meaning from the visible text. Google’s guidance on its AI features says plainly that they need no special schema and no new files to work.
So our position is straightforward. Implement solid schema for the reasons it’s always been worth doing, but don’t expect it to be the thing that gets ChatGPT to name you, and be wary of anyone who promises it will.
The same goes for llms.txt, a plain text file some now recommend adding to describe your site to AI tools. It won’t hurt you. But adoption among the tools that do the citing is thin, and Google doesn’t use it. It isn’t the lever. Spend the effort on the content and the mentions first.
How do you know if it’s working?
AI visibility needs to be measured differently from rankings, and the simplest check costs nothing.
Ask the AI tools the questions your customers would ask, and see whether you come up and how you’re described. Do it across a few of the different AI assistants, because they don’t agree with each other. Then watch your analytics for referral traffic from AI tools (which is usually tiny at first) and keep an eye on people searching your brand name directly – since some will see you in an answer and go and look you up.
Record where you’re starting from before you change anything, even if the honest answer is zero. That zero is what lets you prove the work later.
Where we’d start
Get the order right and a lot of this looks after itself. Answer a question directly, write with enough specificity that a sentence can stand on its own, and earn mentions on sites people trust. That’s where the evidence points and where the effort pays.
Do the technical work as well, schema included, but do it for what it’s good for, and don’t let anyone sell you the tags as the silver bullet, because there isn’t one. There is only clear, credible content that a machine can read and other people vouch for, which is – unsurprisingly – the same thing that has always earned trust online.
Whether all of this will still be true a year from now is a fair question. The tools are moving quickly and the advice will move with them. It’s something to keep an eye on, and we will.
If you want a hand working out where your site stands on any of this, our SEO and GEO service page sets out how we approach it.
