Most advice about AI visibility is strategic. This one is not. This is the plumbing — the layer that determines whether a model can parse your business accurately before any question of content quality arises.
Start with access, because it is binary
Before anything else, confirm the crawlers can reach you. The ones that matter most are GPTBot, PerplexityBot, ClaudeBot, Google-Extended and CCBot.
Check two places, not one. Your robots.txt is visible and easy to audit. Your CDN or web application firewall is neither, and bot-protection rules there routinely block AI crawlers independently of anything in that file. It is a common issue, and it is often not deliberate — it is inherited configuration nobody revisited.
Blocking these crawlers is a legitimate choice for some businesses. Discovering eighteen months later that you made it by accident is not.
JSON-LD: state facts unambiguously
Schema markup is how you assert facts in a form that needs no interpretation. Prose can be misread; structured data cannot.
The types that carry most of the weight:
Organization — your legal name, trading name, address, contact details, licence identifiers, parent company. This is the anchor everything else attaches to.
LocalBusiness or its sector-specific subtypes — for anything with a physical presence and service area.
Service and Product — what you actually sell, described as discrete entities rather than as paragraphs on a page.
Person — named professionals with credentials, qualifications and roles. In regulated sectors this can be an important signal to publish.
FAQPage — question and answer pairs, which map closely onto the question-and-answer shape of an assistant's response.
The test that matters: could a model answer a factual question about your business from your markup alone, without reading a word of your prose?
Entity consistency: the unglamorous one
Models cross-reference. Your business name, address, licence details and category should read identically everywhere something might check — your site, your Google Business Profile, the relevant regulator's register, LinkedIn, industry directories, trade press.
In Dubai this takes real work. Free zone entities, DED licences and trading names routinely produce several legitimate versions of one company's name. Commercially that is normal. Computationally it means a model has multiple candidate entities and no confident way to merge them, and it may describe someone whose identity is unambiguous instead.
Pick a canonical form. Use it everywhere. Reconcile the rest. This is administrative rather than technical.
llms.txt: a convention, not a standard
llms.txt is a proposed file at your domain root that points AI systems to the content you consider authoritative, in a form that is cheap to parse. A companion llms-full.txt carries the expanded content.
Be clear-eyed about its status. It is a community convention that has gained traction, not a ratified standard, and adoption across AI providers is uneven and changing. It costs very little to publish and may help; it is not the foundation of anything.
Anyone selling llms.txt deployment as the core of an AI visibility programme is selling you the cheapest part of the work.
Content structure: writing for retrieval
When a model retrieves from your site, it works with passages rather than whole pages. It needs a chunk of text that stands alone — a specific claim, a number, a direct answer — liftable into a response without the surrounding context.
Practically that means: real headings that describe what follows. Direct answers before elaboration rather than after. Self-contained sections. Specific figures, dates, qualifications and ranges rather than adjectives. And honest hedging where hedging is warranted.
The tension is real and worth naming: this is close to the opposite of how persuasive marketing copy is structured. Resolve it across pages rather than within one — reference content built to be cited, conversion pages built to convince.
The order to do it in
- Crawler access. Binary, fast, and everything downstream is wasted without it.
- Organization and Service schema. The foundation other markup attaches to.
- Entity consistency. Tedious and unglamorous, but foundational.
- Person schema, if you are in a sector where credentials matter.
- Content restructuring on your highest-value pages only.
- llms.txt, last, because it is cheap and speculative.
What this does not do
Structured data makes you parseable. It does not make you authoritative. A model that can read your business perfectly still needs a reason to prefer you, and that comes from corroboration in sources it trusts — which is a slower, harder, more valuable piece of work.
Do the plumbing first because it is cheap and gating. Do not mistake it for the whole job.