Somewhere in Jakarta, a procurement manager types a question into an AI assistant: which cold-chain firms can reliably move frozen seafood from Surabaya to Singapore? The answer arrives in seconds. Three companies, a line on each, a couple of links. Yours is not among them.
Nobody tells you. There is no lost-deal email, no ranking report showing you on page two. The shortlist was drawn up by a machine, read by a human and acted on before you knew the question had been asked.
AI branding means two things at once: using artificial intelligence to build and run a brand, and managing how AI systems such as ChatGPT, Perplexity and Google’s AI Overviews describe, recommend and cite that brand.
Most commentary fixates on the first meaning, because it makes better headlines. The second is where the money is quietly moving.
The shop window moved, and nobody sent a memo
The shift happened fast. Perplexity launched its search engine on 7 December 2022, writing answers and citing the sources behind them.
Google rolled out AI Overviews to everyone in the United States on 14 May 2024 and, before the month was out, had to rein the feature in after it suggested glue as a pizza ingredient and rocks as a snack. It did not retreat. By October 2024 the summaries had spread to more than 100 countries.
OpenAI launched ChatGPT search on 31 October 2024, opened it to every logged-in user that December and dropped the sign-in requirement in February 2025.
What changed is the unit of competition. A results page gave you ten links and a fair chance of a click if you ranked fifth. An answer gives you a paragraph. Three names, perhaps four.
Fifth place is not a worse position. It is no position at all.
Publishers felt it first, because their business was the click. In August 2025 Japan’s Yomiuri Shimbun sued Perplexity, alleging it had free-ridden on 120,000 of its articles. Brands are next, and most have not noticed.
How do AI engines decide what to say about a brand?
Marketers ask this with growing exasperation. The same prompt gives different answers on different days. Analytics cannot see the referral. A brand recommended last month has quietly vanished.
There is no published ranking algorithm to reverse-engineer, and anyone selling you one is guessing. The broad mechanics, though, are not mysterious.
A language model is, among other things, an enormous compression of what has been written about everything. When it searches the web, it reads a handful of pages at the moment of the question and compresses those too. Either way, it repeats consensus.
So the brand that gets described accurately is the one described consistently, in many independent places, in plain language.
What the machine looks for
Entity consistency. The machine needs to be confident that the “Tan & Sons Logistics” on your website, the “Tan and Sons Pte Ltd” in a directory and the “TSL” in a trade article are one company. Conflicting founding dates, three versions of your service list and an About page that says “solutions” eleven times all make that harder. Ambiguity reads as uncertainty, and uncertain models hedge or leave you out.
Independent mentions. Your own website is testimony from an interested party. Models weigh what others say: trade press, review platforms, industry associations, forum threads where buyers compare suppliers. A few genuine discussions in those places can count for more than months of your own blog posts.
The reverse holds too. Mentions decay, and a brand that stops being talked about stops being recommended.
Distinctive, checkable facts. Machines quote specifics. “Founded in 1998, 42 refrigerated trucks, halal-certified cold chain” is quotable. “A trusted partner delivering excellence” is not.
Pages built on clear claims, named sources and concrete numbers get lifted into answers. Keyword stuffing, the old reflex, does nothing for you here and can make matters worse.
Pages that answer. Answer engines lift passages, not whole pages. A real question as a heading, a direct answer beneath it, a fact that survives being quoted on its own: that is the shape they reuse. A wordless brand film or a scanned PDF brochure gives them nothing to work with.
The prize is to become the category
Heinz showed the end state before most marketers knew there was a problem. Its agency, Rethink, asked DALL-E 2, an image generator not yet open to the public, to draw ketchup. The pictures kept coming back looking like Heinz. The campaign that followed, which invited fans to send in their own ketchup prompts, won a Clio Gold.
Asked to draw ketchup, the machine kept drawing Heinz.
Its real lesson was strategic. When a brand has owned a category for long enough, the machine’s idea of the category is the brand.
Ask an assistant for ketchup and you get Heinz; ask it for a cold-chain firm in Surabaya and, today, you get whoever has been written about most clearly. That is the goal. Become the default.
You answer for what the machine says
The second fight is about accuracy, and it cuts both ways.
Other people’s machines get you wrong. In December 2024 Apple Intelligence generated a notification summary, appearing under the BBC News app, that claimed the BBC had reported something it never had. The broadcaster complained, more errors followed, and Apple soon switched off summaries for news apps in a beta update. A brand built on accuracy had been misquoted by software it did not control, on a device it did not make.
Your own machines speak for you. In February 2024 a British Columbia tribunal ruled against Air Canada after its website chatbot gave a bereaved customer wrong information about claiming a bereavement fare refund. The airline argued that the chatbot was a separate legal entity responsible for its own actions. The tribunal disagreed and ordered the airline to pay the fare difference, around CA$800, plus interest and fees.
The sum was trivial. The principle was not. Your chatbot is your brand talking. If it invents a policy, you have a policy.
Should brands use AI to make their advertising?
Now the first meaning, the one that trends.
In November 2024 Coca-Cola released three AI-generated holiday films reworking its 1995 “Holidays Are Coming” advertisement, produced with the studios Secret Level, Silverside AI and Wild Card. Online, the reaction was brutal: distorted proportions, delivery-truck wheels that did not turn, and a chorus of film and TV workers who felt their craft had been cheapened for a soft-drink ad. Yet in emotional-response testing with ordinary viewers in the US and UK, the same films earned a top score.
Coca-Cola’s Holidays Are Coming, the long-running film the brand has since reworked with AI.
Coca-Cola appears to have drawn its own conclusion. It made another AI holiday film in 2025. The reviews were no kinder.
Duolingo learned a harsher version of the lesson. In April 2025 it declared itself “AI-first” and said it would gradually stop using contractors for work AI could do. The backlash came from its own users, many of whom cancelled subscriptions and ended their learning streaks in protest, attacking the very habit the brand had built.
In May the company removed every video from its TikTok and Instagram accounts. By August its chief executive was telling investors the effect on the financial metrics was not material.
Read the two cases together and the conclusion is uncomfortable for both camps. Audiences rarely punish the tool. They punish what the tool reveals about your priorities.
Duolingo’s announcement read as a company caring less about quality and the people who built it, which stung because so much of the brand’s charm depended on feeling human. Coca-Cola’s harshest critics were people who already disliked the trade-off. Most viewers simply saw a Christmas advert.
The AI work nobody complains about
There is another way to use the technology, and it rarely trends because nobody is angry. During the pandemic, Cadbury Celebrations in India and Ogilvy used machine learning to create ads featuring Shah Rukh Khan that also promoted thousands of struggling local businesses.
Cadbury used AI to lend Shah Rukh Khan to thousands of neighbourhood shops.
One of the country’s biggest film stars appeared to endorse neighbourhood shops that could never have afforded him. The work won a Titanium Lion at Cannes. AI was used to give something away, so the brand looked generous rather than cheap.
In Singapore, DBS has taken the quieter route at scale. Its virtual assistants, with generative AI features across Singapore, Hong Kong and Taiwan, now serve more than ten million users, and seven in ten of its employees use generative AI tools at work. Klarna made a louder version of the same move in 2024, claiming that its OpenAI-powered assistant had handled two-thirds of customer chats in its first month, the work of 700 full-time agents.
Service AI faces a simple test. Did the problem get solved? When it does, nobody writes an op-ed.
So the question for any AI creative decision is whether it makes the brand more itself or merely cheaper. Customers can usually tell which.
A working plan for being described well
Picture a family-run kueh maker in Joo Chiat that supplies cafés and hotels. Here is AI branding at that scale, without buying a single new tool.
- Write your fact sheet, then make everything match it. One canonical name, founding year, address, what you make, who you serve and what sets you apart. Then reconcile your website, Google Business Profile, LinkedIn page, marketplace listings and directory entries against it.
- Put the distinctive facts in plain sentences. “Made by hand every morning since 1987 to a grandmother’s pandan recipe” will be repeated. Mood photography with no words will not.
- Earn mentions where buyers talk. Food writers, trade associations, local press, review platforms, forum threads. These cannot be faked for long, so do not try.
- Publish the answers buyers need. Minimum orders, halal status, shelf life, delivery areas, each with a plain answer on your own site.
- Audit the machines monthly. Put a buyer’s questions to three or four AI tools. Note if you appear, how you are described and which sources are cited. Track the pattern over months, because single answers wobble.
None of this is exotic. It is sound branding made legible to a new kind of reader, one that never visits your shop, reads everything, trusts the majority and forgets anyone who falls silent.
Quick answers
What is AI branding?
AI branding covers two linked disciplines: using artificial intelligence to create and manage brand work, and shaping how AI tools such as ChatGPT, Perplexity and Google’s AI Overviews describe and recommend a brand.
Is generative engine optimisation different from SEO?
It builds on SEO rather than replacing it, because AI answers often draw on pages that search engines already index. The difference is the goal: being named or quoted inside an answer, not ranking as one link among ten.
Does AI-generated advertising damage a brand?
Not automatically. Backlash tends to follow when AI use signals that a brand values cost over the qualities its audience loves, while AI that improves service or gives customers something extra is rarely criticised.
When the machine describes you, make sure the true story is the easiest one for it to find; Firefish’s digital branding team can help you write it.
