The day after launch, ChatGPT described the old us
A day after our new site went live, ChatGPT still described Sazinga as a marketing business. What an assistant reads, and what we fixed.
Somebody asks an AI assistant about a company like yours, and the answer either leaves you out or describes you as something you are not. The buyer never sees your website. They see the assistant’s paragraph, and they act on it.
That happened to us. Our new site went live on 10 September. The day after, we asked ChatGPT, with web search on, what Sazinga does. It described us as development, staff augmentation and digital marketing, and listed products from old app-store listings and the previous website. It said in terms that it did not have enough public evidence to identify Sarva, the platform our products sit under. Not one of the seven products appeared by name.
We are an engineering company. The description came from the old site, which read as a marketing business, and the assistant was faithfully repeating what it had found.
What was actually going on
Answer engines do not hand back ten links for a person to weigh up. They collect passages from pages, assemble an answer, and sometimes say where it came from. What they say about you depends on what they can fetch and what they find there, which makes it an engineering question: what does the site actually serve, to a machine, on the day it asks?
We looked at our own server logs rather than guess. Since the cutover, ClaudeBot had made 503 requests, bingbot 493, GPTBot 419 and Googlebot 46. The site was being read. But three things were wrong.
Old addresses were returning an error instead of an answer. All 140 entries in our map from the old website to the new one had been written with a trailing slash, and the server only matched that exact form. Search engines hold many of these addresses without the slash, so they got a “not found” where we meant a permanent redirect. That teaches a crawler nothing, and it is why stale snippets survived. A search on Bing’s index still returned an old page from the previous site, with old marketing copy.
The two assistants we tested disagreed. Perplexity had the new site: asked what Sarva is, it returned our About paragraph almost word for word, and in a question about software for billboard operators it listed Sazinga accurately among ten vendors. ChatGPT did not. Same site, same day, and the difference was whether the assistant had something specific to lift.
And Google had indexed only five of twelve pages we checked.
What we changed
We rekeyed the redirect map so every old address, with or without the trailing slash and even with a tracking suffix, now returns a permanent redirect to the right new page. We wrote a small tool that reads the map off the server and requests every entry in three forms: 141 entries, 420 requests, all passing. We added an RSS feed, because the old site had one and Bing still remembered it. We resubmitted the sitemap.
The rewrites on the site itself followed the same principle. State, in plain sentences that stand alone, what we are, what the products do, and what they do not do, in the page’s served text and not built in the browser afterwards. A model can only quote a passage that is clearly there.
What it did not fix
We could not make ChatGPT change its answer, and nothing we did guarantees it will. An assistant that has already been trained cannot know a site exists. Training crawlers give no timeline and nothing you can submit. The engines that can pick the site up quickly are the ones that fetch pages at the moment of the question, and whether they cite us cannot be proved from here. The only test is to ask the assistants directly and record what they say.
The pattern, for anyone whose company is described wrongly
Ask the assistants your buyers use the questions your buyers ask, write down the answers, and repeat it. Treat what they say as a readout of what your site, your old listings and other people’s pages say about you. Then check what is actually served: that the important text is in the page itself, that old addresses redirect properly instead of returning an error, and that every page says one or two precise things a stranger could quote on their own.
Do not pay anyone who promises placement in AI answers. It is not a guarantee anyone can make.
Where this ends up
What a site has to serve, redirect and state for a machine to describe it correctly is engineering work on the site itself, and it is part of how we approach custom software development.