ChatGPT described the company we used to be

You ask ChatGPT what your company does, and it describes the business you used to be. It lists products you retired, and when you ask about the product you have just spent weeks writing pages for, it says it does not have enough public evidence to identify it. A buyer who asks the same question gets the same answer.

That happened to us the day after we moved the site. The cost is a buyer forming a wrong first impression before anyone has spoken to them. We did not put a number on it, and nothing here claims one. What we did get was a clean comparison, which is useful for anyone who wants an AI assistant to describe their business correctly.

What was actually going on

A day after the move, the server logs showed heavy visits from every major AI crawler. But a crawler arriving tells you nothing about whether an assistant can say anything about you. So we asked two assistants the same questions, on the same day, with the same crawling permissions.

Perplexity described the platform almost word for word from the new homepage. Asked a buyer’s question that named no brands, which companies offer advertising-management software for billboard operators in India, it listed about ten vendors, and ours was among them, with accurate detail: inventory across billboards, unipoles, gantries and digital screens, availability, proposals, geotagged proof of display, invoicing. Every one of those details sits on a product page. None of it was a claim about being leading or trusted.

ChatGPT, with web search on, described a development, staff augmentation and digital marketing agency and listed products from the old version of the business. Its sources were mostly app store listings and pre-move marketing copy.

Two things were going on. One was stale information: several old addresses were still being served, because our redirects returned “not found” instead of pointing to the new pages. That was our plumbing fault, and we fixed it. The other was that Perplexity had found passages on the new pages that it could lift and credit to us. The answer was mostly a shortened version of sentences we had written. Where a page said only that a product was comprehensive and scalable, no assistant produced anything, because there is nothing in that sentence to carry into an answer.

What we changed

The fix for the old addresses was mechanical: they now redirect properly.

For the writing, we changed what a page must contain for an assistant to lift a passage from it. A passage that gets quoted tends to do four things. It answers a question somebody would actually type. It contains something checkable, such as a mechanism, a number or a named way of going wrong. It stands alone, because it will be pulled out without the paragraph before it. And it is true of you in particular. A sentence equally true of every competitor gives nobody a reason to credit you.

Sections saying what something does not do got picked up disproportionately. Every product page now has one, and every industry page has a passage on when the software is the wrong answer. That is not a trick. A limitation is specific and easy to check, and it is also what a buyer most wants to know.

We also stopped reading the crawler numbers as good news. Crawlers that collect material for training future models were loudest in the first days after the move, several hundred visits each from ClaudeBot and GPTBot. OAI-SearchBot, the one that fetches pages when someone asks a question, had made ten. What the training crawlers take may or may not enter a future model, on a timeline nobody controls, and no model already trained can learn that your site changed. The ones that fetch at question time are the route by which a new site appears in an answer this month, and they need plain pages that load without scripts, are fresh and are specific.

What it did not fix

We cannot make ChatGPT change its answer. Its description of us followed what its search layer already held, and that lags behind a site move. The redirect fix should let it correct itself, but we have not claimed a result.

One test with two assistants, on one day, is an observation and not a measurement. What it supports is narrow: specific text gets quoted and vague text does not.

The pattern, for anyone who wants AI assistants to describe them correctly

Do not infer it from visitor statistics. Once a month, ask three or four assistants the question a buyer would ask in your field, without naming yourself. See whether you appear and whether what they say is right. It takes ten minutes, and it is the only direct reading of the thing you care about.

If you do not appear, look first at your own pages. Are there sentences a competitor could have written? Is there anything on the page that could be checked? And do the old addresses of any page you moved still answer, or do they say “not found”?

Where this ends up

What a page must contain to be quoted, and how to tell the two kinds of crawler apart, is written up in what generative engine optimisation is, treated as an engineering subject: what the page serves, to whom and how fast.