A whole government in one field

Since the end of September, the United States has been answering questions put to its government in a single field. What a government has built there, every brand will face: when all that is left at the front is a field, the answer is the brand experience.

Someone in Ohio, 67 years old, retires in January. The husband is already on Medicare, while their own coverage still runs through work. What needs doing now, and when? Until now, that meant clicking through the websites of several agencies. On america.gov you type the question into a field and get an answer with links to the sources. The example comes from the site itself.

Behind the field sit roughly 29,000 government websites, as was said at the launch. In front of it, almost nothing is left. The sitemap of america.gov lists seven addresses; the answers only come into being at the moment of the question. Joe Gebbia, Chief Design Officer of the United States, recalled the entrance halls of great public buildings at the launch. Before anyone there said a word, marble and murals told visitors that they deserved dignity and respect. In the field there is no hall. There, the answer has to show the respect.

Websites can already become such fields today. In 2025 Microsoft presented an open tool that turns a website into an answer field, developed by the creator of Schema.org. Chrome is testing a standard through which sites offer agents their functions. ChatGPT’s browser already uses it, and at the end of September Shopify opened its merchants’ checkout to agents through it. It is not widespread yet. Among the most visited domains, hardly any offers agents anything of its own.

You could object that brand management has always given answers, with every page and every ad. True. Only now a machine builds the answer, anew for every question, from everything a brand has ever published.

Half the job

Brands are currently preparing their content for machines. That is right, and it solves half the job. Afterwards the machine finds everything a brand has ever said, including what no longer applies.

Those who advise brands on AI search mostly recommend preparing content for machines, with clear formats and structured data. Behind this lies a plausible assumption: whoever prepares their content cleanly gets correct answers. For america.gov, Gebbia puts it like this: the site is trained on “one official source of truth”, the government. By his own figures, this one source consists of 800 million pages. It does not agree with itself. On launch day the White House fact sheet promised that Medicare enrollment would be possible on america.gov “later this year”. The site itself announces it for 2027.

A website may contradict itself. On the page, every statement carries its context: a press release from 2013 is recognisably one from 2013, and the small print is small. In the answer, that falls away. There, one voice speaks for everything the house has ever said, the old included. Brands are, to use a term from computer science, stateful. They carry their history with them, and in the answer field it becomes audible.

How common this is I checked in early October in a small sample, at seven brands from Germany’s DAX index. One question per brand that customers often ask, and every answer to it on the brand’s own pages. Six could be checked. At two of them, an outdated value is online next to the valid one. According to its terms of May 2026, Zalando grants a voluntary 30-day right of return. The corporate page on its own history speaks of an “up to 100-day right of return”, without noting that 30 days apply today. On top come press releases from 2013 onwards and two old versions of the terms as PDFs on Zalando’s own server. At a major bank, two pre-contractual information sheets on the current account sit side by side. One gives 12.80 percent for unarranged overdrafts, the other 13.05, and which one applies is revealed only by the date in the small print. Of the rest, three answer consistently, and one gives no time frame at all.

Not every difference is an error. A press release from 2013 may say 100 days; it is just that the answer no longer says anything about 2013. One account model may carry different interest rates than the next. But a few things have to hold in every context. In Brand as an MCP Server I called that selection brand management.

Pages can be collected. An answer has to be decided.

Right data, wrong answer

Whether an answer is right does not depend on the data alone. Even a system that draws only on verified sources answers wrongly as long as nobody has committed to which statement applies.

The physician Céline Gounder put 37 health questions to america.gov for CBS News. For chest pain and a numb left arm, its first words were “Call 9-1-1 now”, as they should be. Where CDC pages contradicted each other, the machine said so and quoted both versions. Asked how it chooses, it described a rule: first the agency that runs the programme, then the most recent official source. That sounds reasonable, and it went wrong twice, in opposite directions. On a vaccine recommendation for newborns it took the newest page three times, although a federal court had paused the change and the earlier recommendation applied again. On insulin it described a pilot programme from 2021 instead of the 2022 law that replaced it. The rule explains the first error, not the second. All the sources were official.

That answers the question of whether brand management is now becoming a data business. The data was there, maintained by the responsible agencies. Others built and ran the field, with language models from outside. What was missing was the decision as to which statement applies when two official ones contradict each other. Data decides nothing. This decision is brand work, and it only becomes data once someone writes it down so that a machine can apply it.

Who decides when nobody has decided

The machine decides the individual case, and that will stay so. What is open is by whose rule.

On 1 October I asked america.gov when I could enrol in Medicare there. The answer: not today, only in 2027, plus a link to its own preview. That the White House had promised “later this year” did not come up. Only when I asked about the contradiction did the machine lay out both sources side by side, with their dates. “Both pages are official, and they do not give the same timing.” So it had long since decided, quietly and by its own rule. It only told me because I already knew.

America.gov also shows what happens when a boundary only emerges in operation. The promise was that you could ask “any question”. While the launch was still under way, answers changed. Afterwards the field declined questions as “political questions”, a day later with different wording, while it went on answering on inflation. If a boundary consists of a single word, the machine interprets it from question to question. The UK did it the other way round. GOV.UK Chat ran for seven weeks without an announcement, more than 7,800 people asked over 15,000 questions, and the boundary was fixed beforehand: “it does not attempt to provide advice”.

In companies, the rule on which statement applies sits in the instructions given to the chatbot. Whoever writes them decides how the brand answers. For our trade, that is the next task: we already shape the tone, and the precedence rule belongs with it.

The one answer that counts

Without a commitment, a brand loses in the answer field either way. If it commits, it gives the only answer a customer can rely on.

If the small print wins, the brand sounds like its terms and conditions. If the campaign wins, the brand has to keep its promise. America.gov writes into its terms of use that the government is not liable for losses resulting from reliance on an answer. For companies it is different. In May 2026 the Higher Regional Court of Hamm ruled that a company is liable for the misleading information given by its own chatbot. That only correct data had been fed in did not help. What decided it was that the company could steer the answers through instructions and filters. The case concerned competition law and a chatbot the company ran itself. The judgment is final because nobody appealed. Germany’s Federal Court of Justice has therefore never examined the question.

For customers, this liability is the value. Many machines answer questions about a brand today, for recommendations mostly from sources the brand does not control. They are asked even where a wrong answer gets expensive. At the end of August, the trade publication Nikkei Cross Trend surveyed around 1,000 people in Japan who had used AI when shopping. For travel and financial products, 50 to 60 percent of them had asked an AI for advice; for everyday goods and groceries, around 30 percent. Whether that holds outside Japan is open. In its own field, the brand decides what the answers are drawn from, and stands behind them.

The objections

The first: many companies have had a single source of truth for a long time.

America.gov has one too; it is called the government. A single source does not replace the rule on which of its statements applies.

The second: a government is not a brand. Right, it has a monopoly and can exclude liability. For brands, that makes the task harder. Their customers can ask elsewhere at any time, and their own answer is binding.

Where to start

Long before the field arrives, the work begins with an inventory of your own contradictions.

You take the twenty questions customers ask most often, from customer service and from the search on your own website. For each, you lay all your own sources side by side, from the product page via the terms and old PDFs to the chatbot’s answer, and mark where they contradict each other and where that is intended.

For each contradiction you then commit to what applies, including the exceptions that beat the newest version, such as a court ruling or a recall. Every commitment gets a date and a sender, as I proposed for brand rules in Open Brand Repo. It is stored so that a machine can act on it: since when it applies, what it replaces and what it overrides. Structured data so far mostly marks up what something is, a return period for instance. That it replaces an older one is not stated there. With Machine Readable Brands, a brand exists as a specification that can be applied, and that includes which of its statements applies.

One commitment is missing at america.gov: that the answer says so of its own accord, every time, when its own sources contradict each other. In the CBS test it did; with me, only when I asked. Testing uses the same twenty questions, before launch and after every change of model. The UK took seven weeks for that.

Read next: Brand as an MCP Server: How a Brand Answers When an Agent Asks. →

Terms used in this text: See the glossary →