AI Visibility
How to become the answer AI gives
Buyers used to search. Now they ask. When someone wants the best option in your category, they open ChatGPT, Claude, Perplexity or Gemini and read the answer, often before they ever reach your site. The brands named in that answer win the consideration. The rest are invisible. This book is the playbook for becoming the brand AI cites, by joining three things most teams still run apart: a clear brand, earned media that credible sources carry, and honest measurement of your place in the answer. It is written for founders and teams building in Asia first, where the opening is widest. Read it straight through, or jump to the chapter you need. Then start the loop.
Chapter 1
A finance manager in Singapore needs payroll software. Two years ago she opened Google, typed "best payroll software for SMEs Singapore", and got ten blue links. She opened five tabs, compared, and clicked through to three vendor sites. Every one of those vendors had a chance to win her.
Today she opens ChatGPT and asks the same thing in plain language. She gets one answer. It names three tools, explains why each suits a small Singapore team, mentions CPF filing, and recommends one to start with. Ask it in English, Bahasa, or Traditional Chinese and the shape of the answer is the same. She does not see a results page. She does not open five tabs. She acts on what she is told.
That is the whole shift in one scene. Discovery used to be a list you scanned. Now it is an answer you receive.
For twenty years, getting found meant ranking on a page of results. The buyer saw ten options and chose one. That page is thinning out fast. More than half of Google searches already end without a click to any website, because the answer now sits at the top. Google puts an AI Overview above the links for a large and growing share of queries, and over a billion people see them. Around 800 million people use ChatGPT every week, and many ask it the questions they used to type into a search box. Perplexity and Gemini work the same way: one synthesized reply, a few named sources, no list to work down.
Gartner has predicted that traditional search volume will drop 25% by 2026 as buyers move to AI assistants. You can watch it in your own analytics. Branded traffic holds, while the broad discovery queries that used to bring strangers to your door are quietly drying up.
When the answer names three brands, those three are the consideration set. There is no fourth link for a curious buyer to find, because there is no page of links to scroll. If you are named, you are in the room. If you are left out, you are invisible, and you do not even get the chance to earn the click, because the click no longer exists.
This is harsher than search ever was. On a Google results page, position six still pulled some traffic, and a sharp title could steal a click from a competitor. In an AI answer there is no position six. The model picks a short list, often marks one as the default recommendation, and stops. Second place stays in the game. Eighth place never appears.
For a founder, the math changes. You are no longer fighting for a slot on a page that shows ten. You are fighting to be one of the two or three names the model trusts enough to say out loud.
SEO optimizes a web page so it ranks on a list. You target keywords, earn backlinks, tune titles and page speed, and win a position. The unit of the game is your page, and the prize is a click.
GEO, generative engine optimization, optimizes so the model names and cites you inside its answer. There is no list to climb. The unit is your brand as the model understands it, and the prize is the mention. You win GEO on four fronts:
SEO and GEO overlap, and a strong SEO foundation still helps, since some engines retrieve live pages before they answer. The difference is the target. SEO wins a rank. GEO wins the sentence that names you.
The last change this large was the rise of Google itself. When search consolidated into one dominant engine in the mid 2000s, an entire industry grew up around ranking on it. Every brand learned SEO, because that was where discovery happened. The interface then held steady for two decades: type a query, scan a list, pick a link.
Now the interface itself is changing. The list is collapsing into an answer. The skills that won the list do not automatically win the answer. The thing you were optimizing for, a ranked page of options, is being replaced by a single reply that names a few brands and stops.
An investor feels this too. Ask Perplexity who the serious players are in Southeast Asian embedded finance, and you get a named shortlist with citations. The companies on that list look real and fundable. The ones missing from it have to explain why they were left out, if they get the meeting at all.
The window is open because most of your competitors have not noticed. Their teams are still filing reports on keywords and rankings while buyers move to the answer. The brands that start shaping how AI describes them now will become the defaults their category inherits, and defaults are sticky. Once a model settles on the three names it trusts in your category, dislodging one of them is slow, expensive work.
Your first move this week: spend one hour. Open ChatGPT, Perplexity, and Google, and ask the exact questions your buyers ask, in the languages they use, for your market: the Singapore payroll question, the Hong Kong logistics question, whatever yours is. For each answer, write down three things: whether you appear, which competitors get named, and which sources the AI cited. That one page is your GEO baseline and your target list, and everything in this book builds from it.
Chapter 2
Ask ChatGPT, Claude, Perplexity or Gemini who makes the best payroll software for a Singapore startup, and you get a confident, tidy answer with three or four names. It reads like the model simply knows. It does not. Behind that paragraph sits a supply chain, and once you can see it, you can feed it.
A language model is not a database of facts. It is a pattern engine that has read an enormous amount of text and learned what tends to follow what. When you ask a question, it reaches for the names, claims and phrasings that showed up most often, and most consistently, across the sources it has seen. If twenty credible pages call a company the leading player in a category, the model treats that as the safe, expected answer. If only your own homepage says it, the model has nothing to lean on.
This is the first thing to internalise. The model is trying to give a defensible answer, not a creative one. It wants to say what it can stand behind. So the real question changes. Stop asking what you say about yourself. Start asking what the sources the model trusts say about you, and whether there are enough of them.
It follows that a model prefers claims it can lift without risk. "Founded in 2019, the company serves 4,000 SMEs across Singapore, Malaysia and Indonesia" is liftable: concrete, attributable, easy to cross-check against other pages. "The leading end-to-end solution for modern businesses" is unliftable: vague, unverifiable, and identical to a thousand competitors. Specificity also matches you to the question. A reporter who calls you "a Singapore fintech focused on cross-border payroll" has handed the model the exact phrase that connects you to "payroll software for a Singapore startup." Own the specific words buyers use.
Models know things in two ways, and the difference shapes your whole strategy.
The first is training memory. During training the model digested a snapshot of the web, books, forums and licensed archives up to a cutoff date. That knowledge is baked in. It is broad, it is instant to recall, and it is stale. It also arrives with no citations: the model cannot tell you where it learned a fact, it simply absorbed it.
The second is live retrieval. When a tool browses, ChatGPT with search, Perplexity, Gemini's grounding, Google's AI Overviews, it runs a real search in the moment, pulls a handful of current pages, and writes the answer from those. This is where the visible citations come from. It is fresh, it is specific, and it is winnable this quarter, because you are competing for a slot in a live search result, not waiting to be absorbed into the next model.
Most high-stakes commercial questions now trigger retrieval. "Best", "vs", "pricing", "alternatives to", anything a buyer actually types, sends the model to the live web. So you fight on two fronts: earn enough consistent mention over time to live in training memory, and rank in the fresh sources that retrieval reaches for today.
Not all pages count the same. Retrieval systems, and the training process before them, lean hardest on sources with a track record of being right and being independent. In rough order of weight:
Notice the pattern. The weight rises with independence. The further a claim sits from your marketing department, the more a model trusts it.
Here is the single lens that ties it together. Before a model names you, it is effectively asking: who, other than this company, says so, and do they agree? Two forces decide the answer.
Consensus across independent sources. One article is an anecdote. The same claim, in five unrelated outlets, from different reporters, becomes a fact the model will state plainly. Models are built to distrust the lone voice and trust the chorus. This is why one big hit matters less than coverage that stacks and repeats.
Freshness. In live retrieval, recent beats old. A 2026 roundup outranks a 2023 one. A quiet year makes you look dormant, and the model drops you for a competitor who kept showing up.
Put plainly: AI does not reward the brand that talks the most about itself. It rewards the brand that the most credible, independent, recent sources talk about in agreement. That is the whole game, and it is why brand, PR and citation measurement are now one job instead of three.
Your takeaway: list the ten questions a buyer would ask an AI before choosing you. For each one, open ChatGPT or Perplexity, ask it, and write down the sources it actually cites. That citation list, not your homepage, is your real competitive map. Every chapter after this is about getting your name, in specific and verifiable terms, into those sources, from enough independent voices that the model has no honest way to leave you out.
Chapter 3
A model cites what it can repeat without risk. A journalist quotes what they can place in one sentence. Both need the same thing from you: a clear, specific claim they can lift and stand behind. If your brand is vague, you hand them nothing to lift, and you get skipped.
This is the step most teams skip. They rush to pitch before they are clear, then wonder why the coverage is thin and why AI answers name a competitor. Brand comes first. Positioning, messaging, and voice are the foundation every pitch, page, and profile sits on. Make them specific and everything downstream becomes liftable. Leave them fuzzy and no amount of outreach fixes it.
An unplaceable company is unquotable. When a reporter cannot say in one line what you do, for whom, and why it matters, you do not make the story. When a model cannot find a consistent, checkable claim across your site, your LinkedIn, and the coverage about you, it will not risk putting your name in an answer. It names the brand it can describe without guessing.
Vague reads as risk. "Leading provider of innovative solutions" gives a model nothing to verify and a journalist nothing to print. There are ten thousand leading providers. The model cannot rank you, so it drops you. The reporter has no angle, so they move on.
Specificity gives both a handle. "Payroll for 400 SMEs across Singapore and Malaysia, PDPA compliant" can be checked, repeated, and ranked. A model can match it to a query like "payroll software for small firms in Malaysia." A reporter can place you in a story about SME compliance. Same company, two fates, decided by how clearly you describe yourself.
A liftable claim has four traits: it is specific, it is verifiable, it carries a number or a named scope, and it says the same thing everywhere. Watch what that does to ordinary marketing copy.
Each "after" can be lifted into an answer or a news sentence without the writer taking on risk. The number anchors it. The scope, "across the Klang Valley" or "in Singapore and Malaysia", tells a model which queries you belong in. The named standard, PDPA or SOC 2 or ISO 27001, is checkable against public records. Five rules make a claim citable:
A model builds its picture of you from many sources at once: your homepage, your About page, a founder's LinkedIn, a directory listing, an old interview. When those sources disagree, the model sees noise and lowers its confidence in you. When they agree, it sees a pattern it can repeat.
A message house is one page that fixes what you say, so every surface says the same checkable things. It holds five parts:
Then make every surface match it. The boilerplate on your site is the boilerplate in your pitch is the boilerplate in your LinkedIn. When the same specific claim shows up in five places and in the coverage you earn, a model reads the convergence and starts giving you as the answer. One source of truth also kills quiet contradictions. "40 staff" on the site and "a small team" in an interview is a shrug to a human and a confidence hit to a model.
Do this before any outreach. It is about a week of work, and it decides whether the outreach lands.
Before you send a single pitch, pass the one-line test and publish one message house that every page, profile, and proof point repeats with the same numbers. A brand a model can describe in one checkable sentence is a brand it will cite. Clarity is not a branding nicety here. It is the precondition for being the answer.
Chapter 4
A clear brand gives the model something to say about you. Earned coverage gives it a source to cite. This chapter is the second half of that job: getting your story into the places large language models read, trust, and repeat when a buyer asks them who to work with.
Here is the mechanic that makes PR matter again. When someone asks an AI assistant for the best payroll software in Singapore, or a fintech lawyer in Hong Kong, the model answers from what it has read. It leans on sources it has seen often, from outlets it has learned to trust. One article in the right trade title can be pulled into thousands of answers. A release on a wire that no one reads will be pulled into none. So you stop chasing coverage for its own sake and start chasing the specific coverage a model will cite.
Most founders start PR by buying a media list: a spreadsheet of 500 email addresses sorted by region, blasted the same release. The response rate is close to zero, and the few pickups land on low-trust syndication pages that models skip.
Work the other way. Find the ten to twenty outlets that actually cover your category in your market, that publish often, and that models already cite. In Southeast Asia that short list usually includes the regional startup and tech desks, Tech in Asia, e27, KrAsia, et Vulcan Post; the business press, The Business Times and The Straits Times in Singapore, the South China Morning Post tech desk and Nikkei Asia across the region; and the deal desks like DealStreetAsia when you have a funding angle.
Then add the trade press for your category, because a model trusts a title that covers only your world. Fintech has Fintech News Singapore and The Asian Banker. Marketing has Marketing-Interactive and Campaign Asia. Retail has Inside Retail Asia. Travel has WiT and Skift. These titles rank for the exact questions buyers ask, they get crawled constantly, and a feature in one carries more weight in an answer than a dozen generic pickups. To confirm your list, ask the model the questions your buyers ask, then read which outlets it names and links. Pitch where it is already looking.
A reporter does not want your news. They want a story their editor will approve and their readers will finish. Give them one. A pitch that lands has five parts and fits in one short email:
Keep it under 150 words. Lead with the angle, not the boilerplate. If the story only works as an announcement about you, it is an ad, and a reporter can tell in the first line. Tie it to something the desk already covers: the Singapore fintech licensing change, the Indonesian logistics squeeze, the Hong Kong family-office inflows. Offer one outlet an exclusive when the story is strong enough to carry a front page. Offer the data widely when it is a trend piece.
Buyers do not only ask open questions. They ask the model to compare. Best HR platform in Singapore. Xero versus QuickBooks for a Malaysian SME. Top three PR agencies in Jakarta. The model answers those from comparison pages, roundups, and listicles, the same pages that rank in search.
These are the highest-leverage earned media in the AI era, because each one maps directly onto a buyer's decision prompt. Get into them on purpose:
One well-ranked roundup that names you can feed every comparison answer in your category for a year.
A story in English reaches investors and the regional desks. It does not always reach the buyer in Surabaya or Da Nang. Models answer in the language of the question and pull from sources in that language. When a buyer in Jakarta asks in Bahasa Indonesia, the model reaches for Indonesian sources first. So place the story in-market, in-language:
Do not run the English release through machine translation and call it done. Brief a local writer, or work with the outlet in the language, so the quotes read like a person and the terms match how buyers actually search. In-language coverage is how the story travels from the regional headline to the end market where the sale happens.
Here is the full loop. You pitch a real story to a trade desk that models trust. The reporter files it with your named spokesperson and your data point. That article gets crawled, linked from a roundup, and picked up in a local-language title. Months later, when a buyer asks an assistant who to trust in your category, the model has seen your name in a credible source, in the right market, in the right language, and it says you.
Takeaway: build a list of ten to twenty outlets that cover your exact category and market, confirm which ones your target model already cites by asking it your buyers' questions, then pitch each one a single real story with a named spokesperson and one piece of data they cannot get elsewhere. One clip in the right place beats a hundred on a wire.
Chapter 5
Your buyer does not start at your website. They start at a blank prompt box. They type a question, read the answer, and often act on it before they ever see your name. If your category is being researched inside ChatGPT, Gemini, Claude, or Perplexity, the prompt is the first screen of your funnel. You do not control what the model says back. You can control how well you understand the question, and that is where the work begins.
Most teams skip this step. They decide "we need to show up in AI" and start writing content aimed at phrases they invented in a meeting. The models do not answer invented phrases. They answer the questions real people actually type. So before you write a word or pitch a single journalist, you build a list of the exact prompts your buyers use. Everything else in this book hangs off that list.
You are hunting for the exact words your buyers use when they talk among themselves. Four sources give them to you fast and for free.
Ask your sales and support team. They hear the real questions every day. Sit with two or three of them for an hour. Ask what a prospect raises on the first call, and what they ask right before they decide. Write it down word for word. "Is [competitor] better for a small team?" is a prompt. "Does this work with Xero?" is a prompt. Your reps are already carrying a prompt list in their heads.
Mine your won and lost deals. Pull your last 20 closed deals, won and lost. For each one, find the single question the buyer asked right before the decision. That question is the highest-value prompt you own, because someone with a budget reached for it at the moment of truth. Lost deals matter more here. They show you the question where a competitor got named instead of you.
Ask the models directly. Open ChatGPT, Gemini, Claude, and Perplexity. Type the questions you just collected. Watch who gets named and who gets cited. "Best PR agency in Singapore for fintech" returns a list. Read whose names appear, which articles the model links to, and which outlets it trusts. Run the same prompt three times. The answers shift, and the shifts tell you how stable each brand's position is. Do this across all four tools, because they pull from different sources and answer differently.
Mine the communities where your buyers talk. The models learn from public conversation, so the places your buyers gather are the places the models read. In Asia that means Reddit threads, LinkedIn posts and their comments, local Telegram and WhatsApp groups, and trade forums specific to your industry. Search those for your category and read the real phrasing. A founder in a Jakarta startup WhatsApp group asking "ada rekomendasi agency PR?" is handing you a prompt in the exact words, and the exact language, your buyer uses.
Not every prompt deserves your effort. A buyer early in research types broad, open questions. A buyer close to spending types narrow, commercial ones. You prioritize the second kind. Three signals tell them apart.
Score each prompt on these three and the list sorts itself. You will usually find that 10 to 15 prompts carry most of the commercial weight. That short set is where you aim your brand work, your PR, and your measurement.
In Asia, one prompt is really several. The models answer each version separately, so you track them separately.
Language. Your buyer in Jakarta asks in Bahasa Indonesia. Your buyer in Taipei or Hong Kong asks in Traditional Chinese, which the model treats as distinct from the Simplified Chinese a Shenzhen buyer uses. The English answer and the Bahasa answer are built from different sources and name different brands. If you only check English, you are blind to most of your own market. Collect each priority prompt in every language your buyers actually use.
Place. Asian buyers attach a location to almost every commercial question. "Best X in Singapore." "Alternatives to Y in Vietnam." "PR agency for F&B in Bangkok." The model answers each with a local shortlist, and your brand either sits in the Singapore answer or it does not. A strong position in one market does not carry to the next. You win them one market at a time.
So a single category question becomes a grid: the prompt, times the languages, times the markets you sell in. One English prompt can become eight or ten real prompts once you cross it with Bahasa Indonesia, Traditional Chinese, and three countries. Build the grid. It shows you exactly where you are absent.
Put all of this in one simple sheet. One row per prompt. Columns: the prompt in the buyer's words, the language, the market, the source you found it in, its type (research or decision), whether a competitor is already named, and how often you hear it. Sort by intent, then competitor presence, then frequency. Cut the list to the top 15.
That sheet is the control panel for your whole GEO program. Every chapter after this one acts on it. Your brand positioning answers these prompts. Your PR earns the citations the models trust when they respond. Your measurement tracks your name climbing inside these exact answers over time.
Do this now: open a sheet. Pull your last 20 won and lost deals and write down the one question each buyer asked before deciding. Add the questions your sales and support team hear most. Run every one through ChatGPT, Gemini, Claude, and Perplexity, in each language and market you sell to, and note who gets named. You will end the week with a ranked list of the 10 to 15 prompts that decide your deals. That list is where everything else starts.
Chapter 6
You have done the hard part. You know the claim your brand owns, you have earned coverage in sources buyers trust, and you understand how AI systems decide who to name. One question remains: how do you know it is working, and how do you make it work harder each quarter? This final chapter is the scoreboard, and the loop that keeps the score climbing.
First, retire the old metric. For twenty years the goal was the click: rank on Google, win the visit, convert the visit. When a founder in Singapore or a marketing lead in Hong Kong asks an AI assistant "who is the best option for a company like mine," there is often no click to win. The assistant reads, decides, and names two or three brands. Yours is named, or it is not. Traffic numbers tell you almost nothing about that moment. You have to measure the answer itself.
Four numbers tell you where you stand. Track them and you can stop guessing.
Read these as trend lines, not a single reading. An AI audit done once is a photograph of moving water. The answer shifts for reasons that have nothing to do with you: models retrain and reset what they know, your best coverage ages and loses weight, a competitor lands a strong placement and jumps the ranking, a reporter who cited you moves on. The brand named this quarter can be gone next quarter.
So build an instrument, not a snapshot. Fix a list of decision prompts, twenty to forty to start. Run them on a schedule, weekly if you can, monthly at worst, across the assistants your buyers use: ChatGPT, Claude, Gemini, Perplexity. Log appearance, position, competitors named, and sources cited every time. A one-week dip means little. A four-week slide in share of voice is your cue to pitch again.
Here is where brand, PR, and measurement stop being three projects and become one engine. The loop runs in four moves, and each turn sharpens the next:
Run these apart and each one underperforms. Brand sits in a slide deck. PR chases logos that feel good and move nothing. Measurement becomes a report no one opens. Run them as a loop and a single cited placement teaches you where to aim the next three. The score compounds: every win lowers the cost of the next one.
You will not be the answer in week one. Honest progress looks like this:
The exact figures vary by category and how contested it is. The shape holds: baseline, then citation, then named for the prompts that drive revenue.
You do not need a platform to begin. You need an afternoon.
Do that and you are ahead of nearly every competitor, most of whom are still counting clicks that no longer come.
That is the whole playbook in one motion. A brand clear enough to be repeated. Earned coverage in the sources buyers and machines both trust. Measurement that shows which coverage became a citation and feeds the next campaign. Brand, earned media, and measurement are one loop, and the loop compounds: every cited placement makes the next one easier to win. The companies that will own the answer across Asia are not the loudest or the best funded. They are the ones who start the loop now and keep it turning. Your buyer is already asking AI who to trust. Begin this week, and make the answer your name.
Cites X runs the whole loop: build the brand, win the coverage, and become the answer AI gives.
Demander l'accès → More Field Guides →