A prospective client told us recently that they'd stopped searching for suppliers. They just ask an assistant, read the three names it comes back with, and start there. No blue links, no comparison of ten tabs. It sounded like a small change in habit. It's actually a change in who gets to be on the shortlist.

So: how do you end up in that answer? There's a lot of noise about this — new acronyms every quarter, tools promising to get you 'ranked in ChatGPT'. Underneath the noise, the mechanics are knowable, and most of what works is stuff a small team can do without buying anything.

What is AI search, really?

AI search is when an assistant answers your question in prose instead of listing links — and cites a handful of pages it drew from. ChatGPT, Perplexity, Claude, Google's AI Overviews and Bing Copilot all do a version of this. Behind the scenes they run searches, fetch a few promising pages, read them, and write an answer grounded in what they found.

That last part is the important bit for you. Most of the time the assistant is not reciting something it memorised during training — it's reading live pages at the moment of the question, the same way a fast research assistant would. If your page is fetchable, readable, and clearly answers the question, it can be cited even if it was published last week. (The related mechanic — why models invent answers when they haven't fetched anything — is the subject of our post on AI hallucination.)

Does being cited by an AI actually bring customers?

Fewer visits, better ones. An AI answer resolves the easy questions without a click, so informational traffic drops. But someone who arrives after an assistant named you as a fit has already been pre-qualified — they've read a summary of what you do and clicked anyway. In our own inbound, those conversations start further along than search traffic ever did.

The strategic point is narrower than 'AI is the new SEO'. It's that being absent from these answers is invisible in your analytics. You won't see a drop attributed to it. You'll just quietly stop being considered, and the first sign will be fewer inbound conversations from people who describe you accurately before you've said a word.

How do AI assistants decide what to cite?

No one outside those companies has the full picture, and the systems change constantly. But the patterns are consistent enough across platforms to plan around, and they're less mysterious than the tooling market wants you to believe:

• It has to be fetchable. If your robots file blocks the AI crawler, or the content only appears after JavaScript runs, or it sits behind a login or a cookie wall, you're not a candidate. This disqualifies more sites than anything else on this list.
• It has to contain the actual answer, in text. Not implied by a diagram, not buried in a video, not spread across five paragraphs of preamble. Assistants extract passages; give them a passage worth extracting.
• Specifics beat adjectives. A number, a date, a version, a named process, a real constraint. 'Fast, reliable, world-class' is unquotable — there's nothing in it to cite.
• Corroboration matters. Assistants are more confident naming a company that appears consistently across places it doesn't control — a directory, a marketplace listing, a forum thread, a client's own site, a press mention — than one that only describes itself on its own domain.
• Freshness helps on anything time-sensitive. A dated page that's clearly current gets picked over an undated one that might be from 2019.

Step one: check that the AI crawlers can read your site

This is the fastest, least glamorous win. Open your own site's robots file — it lives at yourdomain.com/robots.txt — and look for rules that block the AI user agents. Plenty of sites blocked them in 2024 over training-data concerns, and nobody has revisited the decision since. Blocking a training crawler is a legitimate choice; blocking the search crawler means you can't be cited.

The names worth knowing: GPTBot, OAI-SearchBot and ChatGPT-User (OpenAI), ClaudeBot (Anthropic), PerplexityBot and Perplexity-User (Perplexity), Google-Extended (Google's AI surfaces, separate from ordinary Googlebot), Applebot-Extended (Apple Intelligence), and Bingbot, which feeds several assistants at once.

Two more technical checks, both common failure points. First, view your page with JavaScript disabled — if the main content vanishes, some crawlers see an empty page. Server-rendered HTML is the safe default. Second, make sure your pages are actually reachable from a sitemap and from ordinary navigation; orphaned pages get crawled last, if ever.

Step two: write pages that can be quoted

The single highest-leverage change is structural: answer the question near the top, in a self-contained paragraph, before you explain, qualify, or tell the origin story. Forty to sixty words that would still make sense if someone lifted them out of the page — because that's exactly what happens.

Then keep the structure honest all the way down. Write headings as the questions people actually ask, not as clever labels. Put one idea under each heading. Use lists for things that are genuinely lists — steps, criteria, options — and prose for reasoning. If a comparison is the point, lay it out as a comparison instead of describing it in paragraphs.

Every article in this journal is built that way, which is not a coincidence — it's the reason we can point at it as an example. Headings are questions, each section opens with a direct answer, and the underlying markup emits the full article text as structured data so an assistant reading the page programmatically gets the whole thing rather than a teaser.

Step three: be specific about who you are

Assistants assemble a picture of your company from everything they can read, and they're much more willing to name an entity they can describe confidently. That means the boring facts need to be stated plainly and identically everywhere: legal and trading name, what you do, who you do it for, where you operate, who works there, how to reach you.

Two things carry disproportionate weight here. Named authors with real credentials and links to their other work — an assistant assessing whether to trust a claim looks for a person behind it. And structured data: schema.org markup for your organisation, your articles, your people and your services, which turns 'this text mentions a company' into 'this is the company, here is its name, its logo, its profiles, its authors'. It's plumbing, it's invisible to readers, and it's one of the few things you fully control.

The fastest way to find the gaps: ask three different assistants what your company does, and read the answers as a customer would. Whatever's wrong or missing tells you exactly which page to write next. Do it quarterly.

Does llms.txt matter?

Short answer: it's cheap, it isn't a ranking lever, and you should not spend a week on it. llms.txt is a proposed file — a plain-text map of your important pages, meant to help an AI find its way around your site. Adoption grew sharply through 2026, but the picture on the receiving end is mixed.

• Perplexity has said it fetches the file and uses it to prioritise which pages to read.
• Anthropic recommends it in its own guidance for making documentation agent-readable, and publishes one.
• OpenAI's crawler documentation doesn't mention it; their stated mechanism for crawler control is robots.txt.
• Google stated in June 2026 that llms.txt is not required for Google Search.
• Independent analyses through 2026 found the overwhelming majority of published llms.txt files receive no AI requests at all.

Our read: it's genuinely useful if you publish documentation an AI agent might need to navigate, and roughly neutral otherwise. Ship one in an afternoon if it's easy. Don't mistake it for the work — a well-structured page that a crawler can reach does far more.

Step four: get mentioned where you don't control the page

This is the hard part, and it's the part that separates companies that get named from companies that merely have good websites. Assistants look for agreement across independent sources. One site claiming excellence is marketing; five unrelated sources describing the same company the same way is evidence.

What counts, roughly in order of effort: client case studies published on the client's own domain, a directory or marketplace profile with real detail, a conference talk or podcast with a written page attached, a genuinely useful answer in a community where your customers already are, original data or research nobody else has published, and coverage in a publication your industry actually reads.

Consistency is what makes it compound. Describe yourself with the same words on every surface. If your site says 'design and engineering studio for digital products' and your directory profile says 'IT solutions provider', you've split your own identity across two entities, and the assistant is now less sure who you are, not more.

How do you measure any of this?

Imperfectly, and that's fine. There's no equivalent of a rank tracker with a reliable position number, because two people asking the same assistant the same question can get different answers. What you can do is build a small, honest measurement habit:

1. 01Write down the fifteen questions a good-fit customer would actually ask an assistant before hiring someone like you. Not keywords — questions, in their words.
2. 02Ask them across ChatGPT, Perplexity, Google's AI mode and Claude. Record whether you're named, whether you're cited, and which page got picked.
3. 03Repeat monthly. The trend matters; any single answer is noise.
4. 04Watch referral traffic from assistant domains in your analytics. It'll be small. Judge it by what those visitors do, not by volume.
5. 05Track the qualitative signal too: how often new enquiries already know what you do. That's the effect showing up before the numbers do.

On timing: don't expect same-week results. Perplexity tends to pick up new pages within days. ChatGPT's search index, which leans on Bing, usually takes one to three weeks. Google's AI Overviews follow ordinary Google indexing, so four to eight weeks is a fair expectation for a new page. Publish, then leave it alone long enough to be judged.

What to ignore

• Anything sold as a guaranteed AI ranking. There's no placement to buy and no dial to turn.
• Keyword stuffing, reborn as 'prompt stuffing'. Assistants read for meaning; padding a page with question phrasings makes it worse to read and no likelier to be cited.
• Publishing volume for its own sake. Thin AI-written filler is the easiest thing in the world to produce right now, which is exactly why it carries no weight. Twelve pages that answer a real question well outperform two hundred that don't.
• Hidden text aimed at crawlers. It's detectable, it violates every platform's guidelines, and it puts your ordinary search visibility at risk to chase a citation.

Where to start on Monday

In order, because the order matters: check your robots file for AI crawler blocks. Confirm your key pages render their content without JavaScript. Take your five most important pages and rewrite the opening of each so the first paragraph answers the page's question outright. Add organisation and article structured data if it's missing. Then pick one off-site surface — a client case study, a directory profile, a talk — and do it properly.

None of that is exotic, and all of it also improves ordinary search performance and human readability, which is the reassuring part: the work doesn't become worthless if the platforms change again next year. Clear, specific, well-structured, verifiable content has been the right answer for two decades. AI search just made it legible in a new place.

We do this work as part of digital marketing engagements, and we run the same checklist on our own site. If you want a frank read on where your business currently stands in AI answers, the discovery call ends with three written recommendations you could act on tomorrow.