How do you get your business recommended by ChatGPT and AI search?

You get recommended by ChatGPT and other AI assistants by making your organisation easy to identify, easy to verify and easy to quote. Assistants answer a question about a business category by retrieving current pages from a search index, reading what they find, combining it with what the model absorbed during training, and then writing a short synthesis. Being named in that synthesis depends on four things: a clear and consistent entity identity, corroboration of your claims across independent sources, pages that answer the question directly and completely, and technical accessibility to the crawlers that feed those systems.

This guide is for owners, marketing managers and association executives who have noticed that buyers now ask an assistant before they ask a search engine, and who want a method rather than a trick. It covers how the answers are actually assembled, the four factors that decide whether you are named, what to do about each one, the structured data and technical work that supports it, how to measure results honestly, and what does not work. It is written for organisations in Jamaica and the wider Caribbean, where the corroboration sources differ from those a North American guide would assume.

How an AI assistant actually answers a question about a business category

When someone asks an assistant “who does member portals for professional bodies in Jamaica”, three things happen in quick succession. The assistant decides it needs live information, because names, addresses and service lines change, so it issues search queries of its own, usually rewritten into two or three narrower ones.

It then retrieves a set of candidate pages from a search index and reads them. This is the step people miss. The assistant is not choosing a winner from a ranked list; it is reading a handful of documents and extracting facts from them. A page that ranks fourth but states its answer clearly can contribute more to the response than a page that ranks first and buries the answer under three hundred words of preamble.

Finally it combines what it read with what the underlying model learned during training, and writes a synthesis. Training knowledge is where long-standing reputation lives, and it is built almost entirely out of what other people publish about you. Retrieval is where recency lives. You are therefore being assessed twice, on two timescales: retrieval can be influenced within weeks, training knowledge over years.

Why being recommendable is different from ranking first

Classic search optimisation asks a ranking question: for this query, which page should appear at position one? Assistant optimisation asks a naming question: when the system writes a paragraph about this category, does your organisation appear in it, and is the description accurate?

Those are not the same problem. A search engine can rank a page it does not understand, because links and behaviour signals do much of the work. An assistant has to write a sentence about you, which means it needs a sentence about you to be available. If your website only says you deliver innovative solutions for forward-thinking clients, there is nothing to write. If it says plainly that you are a digital transformation and managed technology company founded in 2010, headquartered in Kingston, serving MSMEs, associations and public-sector organisations across Jamaica, the Caribbean and North America, there is a sentence to write and it will be reused.

The second difference is consistency. Being recommended once is not the objective; being named accurately across repeated askings, by several assistants, is. That comes from agreement between sources, not from one strong page.

The four factors that decide whether you are named

Everything useful in this field reduces to four factors. The table sets out what to do for each and how to check that you have actually done it.

Factor	What to do	How to check it
Entity clarity	Write one canonical description of who you are and what you do, and use it identically on your website, profiles, directory listings and structured data. Keep legal name, trading name, address, phone number and founding date consistent everywhere.	Search your organisation name and read the first twenty results. Do the descriptions agree? Do the addresses match character for character? Ask an assistant what your organisation is and check the answer.
Corroboration	Build independent sources that confirm your claims: accurate listings, reviews, professional body and chamber records, partner directories, press, talks, published research and named client case studies.	List every source outside your control that names you. If the list is short, or all of it is self-submitted, that is your gap.
Quotable content	Answer the question in the first paragraph, in self-contained passages. Use question-shaped headings, tables, explicit definitions, named authors with credentials and visible dates.	Copy one paragraph into a blank document. Does it still make sense on its own? If it needs the four paragraphs above it, an assistant cannot use it.
Technical accessibility	Serve content in the HTML the server returns. Keep robots directives clean and deliberate, publish a sitemap, decide your AI crawler policy and keep responses fast.	View the page source, not the rendered page, and search it for your key sentences. Read your robots file line by line and confirm every rule is one you meant to write.

Entity clarity in practice

An entity is a thing a system can identify and hold facts about: your organisation, each named person in it, each service you offer, each location you operate from. Entity clarity means those things are unambiguous.

Start with one canonical description, roughly forty to sixty words, that states what you are, what you do, who you serve and where. Write it once and then use it without paraphrasing. The temptation to reword it for each platform is the single most common cause of a confused entity. Paraphrasing feels like good writing; to a machine comparing sources it looks like disagreement.

Then fix the identity details. One legal name for legal contexts, one trading name used publicly, one address in one format, one primary phone number, one founding year. Where a company has a parent, say so plainly rather than letting two entities blur together. Where you have offices in more than one country, list each with its own address.

Structured data must then match the visible page. If your markup says you are open until 6pm and the page says 5pm, you have created a contradiction and a cautious system resolves it by ignoring both. The rule is simple: never mark up anything that is not visible and true on the page itself.

Finally, audit your profiles. Chamber listings, partner directories, review platforms, professional registers, social profiles and any legacy pages from an old website all count. Every one that disagrees with the others weakens you. Our branding and creative work usually starts here, because a single agreed description of the organisation is a prerequisite for everything downstream.

Why third-party corroboration matters more than anything you publish

Your own website is a claim. Everything else is evidence. Systems that synthesise answers weight independent agreement heavily, because it is their only defence against a well-written page that is not true. This is also the part that cannot be rushed, which is why it is worth starting now. The legitimate routes are:

• Directory and profile accuracy. Not volume. Ten accurate, consistent listings beat sixty inconsistent ones, so correct the old ones before creating new ones.
• Review platforms. Ask satisfied clients, on platforms your buyers actually read, and never offer anything in exchange for a review.
• Professional bodies and chambers. In Jamaica these carry real weight because they are independently verifiable and they are what local journalists and researchers cite.
• Partner and vendor listings. Platform partner directories are independently maintained and difficult to fake, which makes them valuable evidence.
• Press, speaking and teaching. Regional business press, trade publications, conference programmes and event listings are durable, dated, third-party records that name a person and an organisation together.
• Published research. Original survey work or sector analysis gets cited, and citations are corroboration. Most organisations ignore this, which is why we treat research and data services as a growth activity rather than a back-office one.
• Client case studies. Named, approved and specific. One a client will publicly confirm is worth more than ten anonymous ones.

How to write a page an AI system can quote

Assume the reader will only ever see one paragraph of your page, lifted out and pasted into an answer alongside two competitors. Write so that paragraph does its job.

Answer first, in self-contained passages

The opening paragraph must answer the question in the title completely, in two to four sentences. Not an introduction to the answer: the answer. Then make every section stand on its own. Avoid “as we saw above” and pronouns that reach back across headings, and repeat the noun instead. It reads slightly more formally to a human and far more usefully to a machine.

Question headings, tables and definitions

People ask questions, so write headings that match them. “How long does an implementation take?” matches a real query in a way that “Timelines” never will. Tables are the most extractable structure on the web because the relationship between the cells is explicit, so use them for comparisons, ranges and checklists. Define your terms in a single sentence beginning with the term itself, which is the pattern extraction systems are built around.

Named authors, dates and sources

Publish under a real person with stated credentials, not under the company name alone. Show a publication date and, where the subject moves, a last-reviewed date, and cite your sources. An undated, unsigned page is weaker evidence than a dated, signed one, and reasonably so.

The technical layer

None of the above matters if the systems cannot read the page. Four things deserve attention.

Server-rendered HTML. Many retrieval systems parse the HTML the server returned without executing JavaScript. If your content only appears after a script runs in the browser, those systems see an empty shell. Check by viewing source rather than the inspector. This is the most common serious fault we find, and it is invisible to the site owner because their own browser renders everything correctly.

Robots directives. Read your robots file and your page-level meta directives line by line. Blocks added years ago for a staging site, a plugin default, or a firewall rule that challenges unfamiliar user agents will quietly exclude you.

Crawler policy. There are two distinct decisions here and they are frequently conflated. One is whether you allow crawlers that fetch pages to answer a user’s question right now, which is how you get cited and linked. The other is whether you allow crawlers that collect content for model training, which is how you might appear in future training knowledge but which you may object to on commercial or licensing grounds. Most organisations that want to be recommended should allow the first. The second is a genuine business judgement, and a publisher with valuable proprietary content may reasonably decline it. Make the decision deliberately and write it down, because the specific crawler names, user agent strings and behaviours in this area change frequently and should be reverified against current documentation before you edit any file.

Sitemaps, summary files and speed. Keep an accurate XML sitemap. Consider publishing an llms.txt file: a plain-text summary at your site root describing what your organisation is, what your main pages cover and where the authoritative version of each fact lives. Adoption is not universal and it replaces nothing, but it is cheap and it forces you to write the canonical description you needed anyway. Keep server responses fast, because retrieval systems work to short timeouts and a slow page is simply skipped. Our managed cloud hosting and website development work treats these as baseline requirements rather than optimisations.

The structured data that matters most

Structured data does not make you recommendable on its own; it removes ambiguity, which is worth a great deal. Concentrate on a few types and maintain them properly rather than marking up everything badly.

• Organization on your home page or about page: legal name, alternate name, founding date, logo, contact points, and the list of profiles that represent you elsewhere.
• LocalBusiness for each physical office, with its own address and phone number.
• Person for your named authors and senior people, with role, credentials and employer.
• Service for each distinct service line, with the area served.
• FAQPage on pages that genuinely contain questions and answers visible to the reader.
• Article on every guide, with author, publication date and modified date.

Two rules govern all of it. Everything in the markup must be visible and true on the page. And identifiers must stay stable: once you have established the canonical identifier for your organisation, do not change it because you redesigned the site.

How to measure it honestly

Assistant answers vary between runs, accounts and regions, so a single screenshot proves nothing. Measure it as a sampling exercise. Build a fixed prompt set of twenty to forty questions a real buyer would ask, in their words, covering your categories, your locations and your named services. Run the same set against several assistants on a fixed schedule, monthly is enough for most organisations, and record four separate things for each run: whether you were mentioned at all, whether you were recommended rather than merely listed, whether you were cited with a link to your own pages, and whether the description was accurate. The fourth is the one people forget and often the most valuable, because a confidently wrong description of your services is worse than absence and it tells you exactly which source needs correcting.

Alongside that, watch your analytics for referral traffic from assistant domains. Volumes are usually small relative to search, but intent is high and the trend matters more than the absolute number. Treat both measures as directional; anyone offering a precise share-of-voice figure for AI answers is selling certainty that does not exist.

What does not work

• Keyword stuffing. Repeating a phrase does not make a passage more quotable. It makes it less coherent and harder to extract a clean sentence from.
• Mass-produced thin content. Sixty shallow pages covering every phrase variation dilute the signal about what you actually do and compete with your good pages.
• Fake or incentivised reviews. Corroboration works precisely because it is independent. Manufactured reviews are detectable and the reputational cost of being caught is real.
• Claiming credentials you do not hold. Certifications, partnerships and awards are verifiable against registers, and a claim that fails verification undermines every other claim on the page.
• Any scheme promising to submit you to AI systems. There is no submission process and no paid inclusion. Anyone selling a direct line into an assistant’s recommendations is selling you the work described in this guide at a premium, or selling you nothing at all.

A ninety-day sequence

If you are starting from nothing, work in this order. The sequence matters, because corroboration built on an inconsistent entity only multiplies the inconsistency.

1. Week one: baseline. Run your prompt set against three assistants and record what they say about you today, including anything inaccurate. It is the only before-and-after you will get.
2. Week two: write the canonical description. One paragraph, agreed by whoever has authority to agree it, plus the fixed set of identity facts.
3. Weeks three and four: fix the entity. Correct every profile, listing and legacy page so they carry that description and those facts, and make the structured data match.
4. Week five: technical audit. Confirm server-rendered HTML, clean robots directives, a working sitemap and fast responses, then decide and document your crawler policy.
5. Weeks six to ten: publish. Write the four to eight pages that answer the questions your buyers actually ask, following the quotable-content rules. Depth over quantity.
6. Weeks eight to twelve: corroboration. Approach clients for reviews and named case studies, update professional and chamber records, pitch one piece of press or one talk, and publish one piece of original research.
7. Week thirteen: re-measure. Run the same prompt set, compare against the baseline on all four measures, and set the next quarter’s priorities from the gaps.

What this means for your organisation

The work that makes you recommendable by an AI assistant is, almost entirely, the work that makes you legible to a human buyer doing careful research: say clearly who you are, be consistent everywhere, let other people confirm it, answer real questions properly and keep the site fast and readable. There is no shortcut, and the absence of one is good news for organisations that do real work.

Start with the baseline measurement and the canonical description, because they cost almost nothing and they tell you how large the problem is. If assistants already describe you accurately, your job is corroboration and content. If they describe you wrongly or not at all, entity clarity comes first and everything else after.

One caution to carry throughout. This field moves quickly. Crawler names, user agent strings, the behaviour of individual assistants, the weight given to particular structured data types and the adoption of conventions such as llms.txt all change on a timescale of months. Treat the principles in this guide as durable and the specifics as something to reverify before you act on them. Our digital ecosystem assessment covers the entity, content and technical layers together, and digital marketing takes the corroboration and publishing programme forward, but the method above is the whole of it and you can run most of it yourself.

Frequently asked questions

How do you get your business recommended by ChatGPT?

You get recommended by making your organisation easy to identify, verify and quote. Publish one canonical description of what you do and use it identically everywhere, build independent sources that corroborate it such as reviews, chamber records, partner listings and press, write pages that answer real questions directly in self-contained passages, and make sure your content is in the server-rendered HTML that crawlers can read. There is no submission process and no paid inclusion.

Is AI search optimisation different from SEO?

It overlaps but asks a different question. SEO asks which page should rank first for a query. AI search asks whether your organisation gets named when the system writes a paragraph about your category, and whether the description is accurate. Ranking helps because retrieval draws on a search index, but a clearly written page that ranks fourth can contribute more to an answer than a vague page that ranks first.

Should I block AI crawlers on my website?

Treat it as two separate decisions. Crawlers that fetch pages to answer a user's question in real time are how you get cited and linked, so most organisations that want to be recommended should allow them. Crawlers that collect content for model training are a commercial judgement, and a publisher with valuable proprietary content may reasonably decline. Verify current crawler names and behaviours before editing any file, because they change often.

How do you measure whether AI assistants are recommending you?

Build a fixed set of twenty to forty questions a real buyer would ask, run the same set against several assistants on a monthly schedule, and record four things for each run: whether you were mentioned, whether you were recommended rather than just listed, whether you were cited with a link, and whether the description was accurate. Add referral traffic from assistant domains in your analytics. Treat both as directional, not precise.

What is llms.txt and do I need one?

An llms.txt file is a plain-text summary placed at the root of your website describing what your organisation is, what your main pages cover and where the authoritative version of each fact lives. Adoption is not universal and it does not replace structured data, sitemaps or good content. It is cheap to maintain and writing one forces you to agree the canonical description you need anyway, so it is usually worth doing.

Can you pay to be included in AI assistant recommendations?

No. There is no submission process and no paid inclusion programme for appearing in an assistant's recommendations. Anyone selling a direct line into AI systems is either selling the ordinary work of entity clarity, corroboration, content and technical accessibility at a premium, or selling nothing at all. The legitimate route is to be accurately described, independently corroborated and technically readable.