When AI gets your brand wrong in a market you do not read

✦Ten days of a wrong answer

During the 2024 United States election, Grok, the chatbot built into X, told users that Kamala Harris had missed the filing deadline and could not appear on the ballot in nine states. The states it named were Alabama, Michigan, Minnesota, Indiana, New Mexico, Ohio, Texas, Pennsylvania and Washington. The deadline had not passed. She could be listed.

The answer was corrected on 31 July. Before that it had reached millions of people on X and beyond.

Nobody signed off on that sentence. No writer wrote it, no editor approved it, no press officer was asked to comment. It was generated, repeated, and believed.

✦Five secretaries of state wrote a letter

On 5 August, five American secretaries of state sent an open letter to Elon Musk. Steve Simon of Minnesota led it, joined by his counterparts in Pennsylvania, Washington, Michigan and New Mexico. They asked for one thing: fix the assistant so that voters get accurate information.

Worth keeping the order straight. The correction came on 31 July. The letter came on 5 August. The letter did not cause the fix.

What matters is the gap in between. While a wrong answer is live, people repeat it. A correction arrives later and travels less far.

✦Introducing yourself to the AI before the customer

The same reporting had a second story in it. One candidate found that the small business policy he had campaigned on did not appear in AI answers at all. So he edited the Q&A section on his own website and updated his Wikipedia entry. The change showed up in AI answers sooner than he expected.

You now introduce yourself to the AI before you introduce yourself to the voter.

That was Jaeseung, and it is the line the whole episode hangs on. Something now speaks about you before you get to speak.

For a company entering Korea this is not a metaphor. Ask a Korean chatbot about your brand and the answer is assembled from whatever Korean-language material happens to exist. For a company that has not launched yet, that is usually a distributor listing, a five-year-old press release, a forum thread, or a Korean brand with a similar name. Almost none of it is yours.

✦What the campaign playbook actually says

A political consulting column, Meridian Star, put out five recommendations for candidates. The episode covered four of them.

• Publish authoritative content yourself: record, positions, results, on a site you control
• Build off-site content: interviews, press, anything published by someone else
• Find wrong answers quickly and ask for corrections
• Archive speeches and policy documents so they stay findable

The second one is earned media in SEO terms. Not the channels you own, the ones other people publish on.

Content published through outside parties rather than by you. Which really just means go out and do things.

Read as a market entry list, it is unremarkable. What changed is the scope. A few channels used to be enough. Now anything that can be cited counts.

✦We searched our own names on air

Rather than theorise, we typed our own names into ChatGPT and Google AI Overviews during the recording.

Mine appears on the first page, but not at the top. Above me sit a Seoul National University professor with the same name and a stage actor with the same name. When ChatGPT described the professor, the sources it cited directly were the university's official site and an author profile at Minumsa, a Korean publisher. One institutional domain and one publisher page. That was the whole recipe.

Jaeseung fared better. ChatGPT introduced him correctly as a digital marketing specialist at Ascent Korea, citing his LinkedIn profile and activity going back to 2021.

Then a second Jaeseung Pi turned up: a director at TBWA Korea. The evidence was a single PDF from a 2016 industry award, and inside it his name appeared once, in a list of judges. One line, ten years old, still describing a person today.

One line in an article, my title and my name, and because of that one thing the internet and the AI already know a director named Jaeseung Pi exists at TBWA.

Turn that around and you have the problem in one sentence. If you have left nothing behind, there is nothing for a model to describe you with. Being modest in public is not a neutral choice here.

✦Why a podcast appearance shows up later

I put this show on Apple, Amazon and several other podcast platforms. There is a reason beyond distribution. When someone searches a guest's name, I want more than one place to answer.

It works. This channel appeared in Jaeseung's own search results, next to his LinkedIn posts and the videos.

Three appearances here and it shows. Your content gets cited, and now there is a Jaeseung Pi who is recognisably not the director.

That is the off-site item from the campaign list, doing its job. Your own site is necessary and not sufficient. You need to exist on other people's.

✦The long tail, explained with chicken stew

Names lead straight into brands.

Imagine a new Korean meal kit brand selling dakbokkeumtang, a braised chicken stew. The domain is a week old and nobody has heard of it. There is no version of the next twelve months where it ranks first for the word dakbokkeumtang.

You do not target the word itself. You start from the long tail and widen out.

The entry points are phrases like how to make dakbokkeumtang taste good, or where to buy the ingredients. Low volume, obvious intent. If you sell meal kits, revenue comes from those, not from the head term.

Foreign brands entering Korea usually get this backwards. The budget goes to the category word, which is owned by domestic incumbents with a decade of Korean-language content behind them. The winnable queries are the specific ones, and they are also the ones AI answers tend to quote.

Name search behaves identically. I will not beat a professor and an actor on my name alone. Add two letters and it is a different question, because I am the one who does that work. We went into the long tail in more detail in the EP 04 note, and into the order of SEO and GEO in the EP 09 note.

✦Naming a brand for a market you have not entered

The most practical stretch of the conversation was about names.

A small shop selling apples cannot call itself Apple. It could register the name, but it would never surface for it. Pick a name that is already contested and you have handed away your own search results on day one.

The same applies to people. If you share a name with an idol or a well known actor, no amount of personal effort matches what an agency publishes about them. In that case you do not compete. You attach your field to your name and make it a different query.

For companies entering Korea this is a launch decision, not a marketing one. Your Korean brand transliteration is a search term. Check it before it goes on the packaging, because changing it afterwards is expensive.

The last thing we tried ran in the opposite direction. We asked for ten recommended SEO specialists in Korea. Agencies came back before individuals. Then we opened the citations to see which channels the answer was built from: LinkedIn, company domains, personal domains. If you want to be in that answer, the list of what to fill in is right there.

✦What to do first

This episode comes down to two moves. Find out how AI currently describes you in Korean. Then fill in the sources that answer was built from.

The candidate in the story did exactly that. He fixed a Q&A page and a Wikipedia entry. No new technology, no special markup. He found the places he was allowed to edit and edited them.

My own name experiment is written up in the EP 13 note. The next note picks up with cases where a wrong AI answer cost real money, and with what Google has shipped in response.

How I run a project is on the service page, and the rest of the episodes are in the episode list.

✦Common questions