ChatGPT gets information about your business from four different places: what it absorbed in training, what it retrieves from the live web when it decides to search, structured listing and place data that comes from providers, and whatever you or the person asking hand it in the conversation. Which of those is doing the work changes from question to question. That is why the same assistant can describe your firm confidently on Monday and vaguely on Tuesday, and why two engines asked the same thing on the same afternoon can disagree about you.
Most owners arrive at this question looking for the switch. You picture a database somewhere with a row for your business, and you want to know how to get into it, or how to correct the row that is already there. It is a reasonable thing to picture. It just is not what is happening.
There is no file on you. Persistent records of your business do exist, in Google's local profiles and similar place directories, and they matter. What does not exist is a stored profile the assistant opens and reads. The picture gets rebuilt every time somebody asks. So if an assistant described your firm thinly last week, that was not a verdict on the firm. It was a description of what was reachable in the seconds it had.
This guide covers the four source types in plain terms, what each of the four big assistants says about where it looks, what five engines told us about their own answers when we asked them on the record, and where your attention is actually worth spending.
There Is No File on Your Business
Start with the four source types, because almost every confusing AI answer about a business traces back to one of them.
Training memory. Everything the model absorbed before a fixed date, held as general knowledge rather than as records it can look up. Anthropic publishes the cutoff for each Claude model, and lists Claude Sonnet 4.6 at August 2025. A model answering from memory alone is telling you what was broadly true then, which is how an assistant ends up naming your old partner, your old address or a service you retired. That failure has its own shape, and you can see it up close in the training data echo.
Live retrieval. When the assistant runs a search and reads results before answering. This is the path most people assume is always on. It is not.
Listing and place data. The structured record of a business as a place: name, category, hours, location, reviews. Google is the most explicit about how this is built. Its Business Profile documentation names four inputs: publicly available crawled web content, licensed data from third parties, contributions from users and from owners who claim their profile, and information based on Google's own interactions with the business. Your own edits are one of those four.
The conversation. Anything supplied in the chat itself: a URL the buyer pastes, their location, an uploaded file, a connected account. This is the source people forget, and it is why a buyer who already knows your name gives the assistant far more to work with than a buyer who does not.
Those four behave differently. Memory is frozen and confident. Retrieval is current and shallow. Listing data is structured and slow to change. The conversation is whatever the buyer happened to bring.
Where Each Assistant Actually Looks
This is one of the few questions where the engines genuinely differ, so it is worth being specific rather than saying "AI" and hoping.
ChatGPT. OpenAI's help documentation for ChatGPT search says the assistant may choose to search the web based on what you ask, or the user can select search manually. When it does search, OpenAI says it turns the request into one or more search queries and may share those disassociated queries with the Bing search engine to return web results, along with structured data sent to Bing or third-party data providers for things like local and timely information. That page documents Enterprise and Edu workspaces, where search can also be switched off entirely by an admin, so treat it as the clearest published account rather than a universal spec. Two practical consequences. Being absent from Bing's index is a different problem from being absent from Google's. And ChatGPT deciding not to search at all is a normal outcome, not a malfunction.
Gemini. Google is the one company documenting both halves: the local business record described above, and an assistant built inside the same company. Google does not publish a map of which profile fields reach a given Gemini answer, so nobody outside Google can claim that link precisely. What Google does publish is unusually direct about what does not help. Its guidance for site owners states there are no additional requirements to appear in AI Overviews or AI Mode, that you do not need to create new machine readable files, AI text files or markup, and that meeting every requirement still does not guarantee Google will crawl, index or serve your content. Google also says plainly that not all Gemini responses include sources.
Claude. Anthropic treats web search as a mode rather than a constant. Its help center says that when web search is on, Claude processes multiple sources and every response includes citations so you can verify them yourself. When it is not searching, the answer is coming from training memory with a published cutoff date.
Perplexity. Perplexity's help center says it searches the internet in real time and that each answer includes numbered citations linking to the original sources. Of the four, it is the one that treats searching as the default rather than a decision to be made per question.
Read those four together and the real difference is not which websites each engine likes. It is whether the engine looks anything up for that particular question, and what it falls back on when it does not. We tested that difference directly in one question, five engines.
What Five Engines Said About Their Own Answers
On August 11, 2026 we ran the same ten buyer questions past five models with live web access enabled: perplexity/sonar, google/gemini-2.5-flash:online, openai/gpt-5.2:online, anthropic/claude-sonnet-4.6:online and x-ai/grok-4.3:online, routed through OpenRouter. Fifty answers. Then we asked each model one follow-up: explain where those names came from.
One caution before the findings. A model describing its own retrieval is giving you its stated criteria, not a log. These systems reconstruct their reasoning after the fact and sometimes get it wrong. Read the answers as self-description, useful for what they reveal about criteria, not as forensics.
The answers themselves already varied in form. Three of the five wrote source links directly into their answers, 157 distinct domains across the run. Perplexity used numbered footnotes instead of visible links. Claude named firms with no source links at all.
The follow-ups were sharper. Gemini said that for a question this new, "nearly 100% of the recommendations came from live web search results." Grok put its split at roughly 70 to 80 percent live results, with the rest from prior knowledge. GPT-5.2 said something more useful than either: "I did not run live web searches before recommending those firms," and estimated its first answer at effectively all prior familiarity and no live search, then browsed afterward to check itself. Claude Sonnet 4.6, whose published knowledge cutoff is August 2025, put its own original answer at roughly 80 to 90 percent training data and called that backwards for a market moving this fast. Perplexity said it could not reconstruct the session and flagged several names as not supported by the results in front of it.
So on one afternoon, with the same ten questions, some engines were reading the current web and some were, by their own account, quoting a year-old memory of it. If your firm has changed in the last two years, that gap is not academic. It is the difference between being described as you are and being described as you were.
The Part Where All Five Agreed
The engines disagreed about where they had looked. They agreed almost completely about what would get a business named.
Four or five of the five, unprompted, asked for versions of the same short list, item by item: corroboration and proof artifacts from all five, a specific service page and evidence of current operation from four, consistent entity details from three. A specific service page that states who it is for and what it actually does, rather than a general page that mentions the category. Visible evidence that the business is operating now, such as recent posts, current service pages or dated work. Independent corroboration: mentions, profiles and reviews on sources the business does not control. Consistent entity details so the firm is not confused with a similarly named one. And proof artifacts, meaning described methods or case studies rather than adjectives.
They were equally clear about what got firms dropped. Gemini named vagueness and what it called unverifiable legitimacy: sites too thin to tell whether the business was real and current. GPT-5.2 said flatly that "we guarantee you'll be #1 in ChatGPT" positioning tends to reduce credibility. Claude, reviewing its own answer, downgraded names it had taken from a listicle where the publisher had ranked itself first.
Notice what is missing from all five lists. Not one of them mentioned posting volume, follower counts or how long you have been in business. What they asked for was checkability, which is the whole reason a strong firm can be skipped while a thinner one gets named.
Our own working frame says the same thing in three parts: AI recommends businesses it can verify, that others vouch for, and that know something it doesn't. Those map neatly onto the engines' own stated criteria, which is reassuring, because we did not write theirs. The verification half of it starts with being one coherent, findable thing, which is the plain-English argument in entity identity for non-technical owners, written for people who glaze over at the word schema.
Where to Put Your Attention First
Knowing where the engines look only helps if it changes what you do on Monday. Four moves, in this order.
- Answer the buyer's question on your own pages, specifically. Who this is for, what it costs to get wrong, what you actually do about it. A page that mentions your category is not the same as a page that answers a question.
- Fix the record of you as a place. Owner edits are one of Google's four inputs, so the crawled pages and third-party data have to agree with your profile. Conflicting addresses, old phone numbers and half-claimed listings are exactly the kind of disagreement that keeps a record from confirming itself.
- Build the corroboration you do not own. Independent mentions, credible profiles and reviews are what every engine asked for and the one thing your website cannot supply about itself.
- Check across engines, on a schedule. We test the five engines your buyers actually ask, ChatGPT, Gemini, Perplexity, Claude and Grok, with the questions they actually ask. One screenshot on one model is an anecdote. Five engines, repeated monthly, is a measurement.
One boundary worth stating plainly, because plenty of people in this market will not state it. Nobody outside the companies that build these models controls what they say, and anyone promising you a permanent place in an answer is selling something they cannot deliver. What you control is whether the evidence exists, agrees with itself and can be found by a machine that only has a few seconds to look. That part is entirely buildable, and it is the part almost nobody has finished.
If you want a plain read of how the engines currently describe you, our free Recommendation Check runs the questions your buyers ask and shows you what comes back. It is a diagnostic, not a pitch in disguise, and the report is yours either way.
Want to Learn More?
Probably Genius builds and operates the recommendation layer around expert-led businesses, so the record an assistant checks actually reflects the work you do. That is the job inside The Answer: your expertise on the record, structured for machines to read, and corroborated beyond your own site. The published methodologies show exactly how it runs, including the 109-Point Diagnostic that sits behind the free check.
You are probably a genius at what you do. Our job is making sure the machines can tell.