A business profile you stop updating is worse than none

OpenAI opened Businesses in ChatGPT in beta on 10 November. A business can now claim a record inside the Business Hub, correct what it says, attach a product feed with prices and availability, and read which conversation topics it appears in.

That sounds like a listing. It isn't. It changes the status of your business facts inside the assistant. Before, those facts were inferred from whatever the crawler could parse off your site. Now they can be supplied by you, which means the assistant no longer has to guess. And that is precisely where the trouble starts, because a profile you claim and abandon is worse than one you never claimed.

What the claim actually changes

The mechanical part is small. You claim the business, you edit the fields, and OpenAI states plainly that ChatGPT may use that information to answer customer questions. Retailers can push a feed carrying pricing and availability. The Hub also shows how often the business is mentioned and which discovery topics it shows up in. Business Agents, the part that answers customers on your behalf, start with eligible ecommerce and online service businesses in the U.S.

The interesting sentence in OpenAI's own copy is the conditional one: may use this information. The profile is not a directory entry that sits passively in a database. It is an input to answers. Fields you fill become candidate facts for the model to state.

So the claim is not a marketing checkbox. It is an assertion, made on your behalf, in the exact place buyers now ask their first question.

Inferred facts get hedged. Asserted facts don't.

Here is the shift that costs people money, and it is easy to miss because both states look like "the assistant knows about us."

An assistant working from crawled pages is doing inference. It reads your homepage, your service pages, maybe an old press mention, and it stitches together a description. Because the input is incomplete and contradictory, the output carries qualifiers: a company typically does this, you may want to contact them, hours vary. The hedging is not politeness. It is the model signalling that it is reconstructing rather than reciting.

A claimed profile removes that signal. The facts now arrive as declarations with a first-party source attached. The assistant answers from them with less uncertainty, and the uncertainty it drops is the useful part.

Which means the risk profile inverts. An unclaimed business risks being described vaguely. A claimed business risks being described wrongly, confidently, and at scale, using details its own team stopped maintaining eighteen months ago. Vague is a missed call. Confidently wrong is a customer who drove across town to a closed door.

This is also why profiles compound in a way pages don't. Your website is re-crawled and re-read, so an outdated sentence gets softened by fresher pages around it. A profile field is a single authoritative value with nothing to contradict it. When it drifts, it drifts alone.

We watch this pattern in client work constantly, and the pattern is not technical. It is ownership. The claim gets made by whoever is excited about AI search that quarter, and the maintenance falls to nobody. Nobody is a bad owner of a field that answers "are you open right now."

The same logic governs how any of this gets cited, which is why citation formatting matters as much as the profile itself: see how ChatGPT, Perplexity and Gemini format the sources they cite.

What I measured: hours are covered, prices are not

To find out where the gap actually sits, I fetched twelve pages on 11 November 2026. The set was the business-operated results ranking in the top twenty for a local commercial query (emergency plumbing, Austin, Texas), one page per business, directories excluded. For each page I parsed the JSON-LD and checked which business facts were declared in machine-readable form.

This is a probe, not a census: twelve sites, one category, one metro, one day. It tells you which fields businesses bother to declare, not what the internet does in aggregate.

Read the first and last rows together. Every one of the twelve tells a human it is available around the clock. Eight of twelve also declare hours in structured data, and ten declare where they serve. Hours and service area are fields another platform spent a decade training businesses to maintain, and it shows.

Price is the outlier. Nine of the twelve show a dollar figure somewhere in the visible copy, mostly a coupon or a "starting at." Three declare a price range in the markup, five declare an offer of some kind, and not one of the twelve attaches a price to it. The field every buyer asks about first is the one nobody has structured.

Now line that up against the product: the new feed is built around pricing and availability. The field with the weakest existing support is the field the platform just promoted. That is not a coincidence to celebrate; it is the maintenance debt, arriving early.

Every field needs an owner or an empty state

The practical rule I'd apply before touching the Hub: no field gets filled unless a named person owns its refresh interval. Fields decay at wildly different rates, and treating them as one block is how profiles rot.

• Hours and holiday exceptions. Decays daily. Owner: whoever opens the door. A single missed holiday change turns a working profile into a wrong answer on the highest-intent question there is.
• Price and availability. Decays fastest of all if it comes from a feed. Owner: whoever owns the catalog. A price that has drifted is worse than a price that was never published, because the buyer arrives already committed to a number.
• Category. Rarely changes, huge blast radius. Owner: whoever owns positioning. The category decides which comparison set you appear in, so a wrong one puts you in a conversation you cannot win.
• Service area. Rarely changes, fails silently. Owner: ops. Excluding a city you actually serve means you are absent from every query in it, and nothing alerts you.
• Attributes such as parking, accessibility or payment methods. Cheap to verify once, rarely wrong. Fill them; they cost little and settle small doubts.

The empty state is legitimate. An unfilled field is a question the assistant has to answer from your site, which is usually accurate and always hedgeable. A filled field that nobody maintains is an assertion you are no longer standing behind.

The Hub's numbers are a topic map, not a funnel

The insights view reports mentions and discovery topics. Both are relative figures measured inside ChatGPT. Neither reconciles with your analytics, and trying to make them will waste a quarter.

That is not a flaw to complain about; it is the same measurement boundary that shows up everywhere in AI search, where referral traffic is collapsed into direct or never attributed at all. If you are trying to separate assistant traffic from human traffic, that problem is already its own discipline: agent traffic and customer traffic look identical in a log.

What the Hub's numbers are genuinely good for is the topic list. The discovery topics tell you the questions your content has to answer, in the buyer's own phrasing. Read them as a brief, not a scoreboard. A mention count that grows while the topic list stays generic usually means the assistant is naming you as an example, not recommending you.

Who should wait

• Businesses outside the U.S. should expect the profile and the insights long before agents. Claim and correct the facts; don't hold the work until agents open in your market.
• Businesses with volatile prices or stock should either automate the feed or leave those fields empty. A manual feed with a monthly cadence is a liability with a delay on it.
• Anyone without an owner for the maintenance should claim anyway, but fill only the fields that change annually. Partial and true beats complete and stale.

Worth knowing: this is not the only surface where a business now has a profile to maintain. Google has been pushing the same idea from the other direction, which is why publisher and business profiles keep multiplying. The count of places holding a copy of your facts is going up, and none of them share a source of truth with the others.

Record a baseline before you change anything

You cannot measure a change you didn't freeze first. Before claiming anything, ask four questions and keep the answers verbatim with the date:

1. What does this company do?
2. Who are its main competitors?
3. Does it offer one specific service you actually sell?
4. Where is it based, and when is it open?

Ask the identical wording again after the profile propagates, and again after two weeks. Publish or file both versions side by side, including every question where nothing changed. The unchanged answers are the finding, not the failure: they tell you which facts the assistant was already confident about, and which ones it invented from an old page.

Fixed wording matters more than sample size. Change one word between runs and you are measuring the prompt, not the profile.

The point of claiming was never to be present. It was to be correct on the day someone asks. Those are different jobs, and only one of them ends when you press the button.

If you want the fields mapped and owned before you claim, our SEO and GEO work starts with exactly that inventory, and you can tell us what your assistant currently says about you.