Ask ChatGPT about your business twice, once with web search on and once with it switched off. Plenty of owners get two different companies back, and one of the two answers is usually empty.
That gap is the most useful 30 seconds anyone can spend on AI visibility, because it separates the two ways a business gets recommended by ChatGPT and every assistant like it. One is retrieval, where the model searches the web while it writes and cites what it finds. The other is training, where the model already knows the business before it looks anything up.
Both end with a customer reading a recommendation. The work behind them has almost nothing in common, and it pays off on schedules that are months apart.
How to check your AI visibility in two questions
With search switched off, you are reading the model’s memory. It formed during a training run that finished before the model shipped, and it holds whatever the public web said about you back then. A national brand comes back described in detail, with categories, products, and reputation. Most local businesses come back as a polite blank.
With search switched on, you are reading today’s web through that assistant’s own retrieval. The model runs a few queries, opens a handful of pages, and builds an answer out of what those pages say. Your website may or may not be among them.
Search on shows what the web says about you today. Search off shows what the model believes about you before it reads anything.
Neither answer is the whole picture, and the difference between them is the diagnostic. A business that appears only with search on is renting its position. One that appears with search off has something closer to a deed.
Vector one, how to show up in AI search right now
This is the vector most marketing teams already work on, sometimes without naming it. When someone asks ChatGPT, Google AI, Perplexity, Claude, or Grok for the best option in a category, the assistant runs a search, opens a few pages and cites them. Getting into that answer means being on the pages that get opened.
Those pages are rarely your homepage. They are review platforms, industry directories, curated best-of articles, comparison pages, marketplace listings, and the occasional deep page of your own that answers the exact question a customer typed. The free AI visibility checker reads a site the way an assistant does and reports which of them your business already reaches.
Retrieval moves fast. A listing added to a platform assistants already cite can change what they say inside a month, which makes this the lane where effort turns into visible movement soonest.
It also splits by platform in ways that surprise people. We tracked dozens of real queries about pizza in New York and found ChatGPT and Gemini picking completely different winners. Prince Street Pizza, viral for years, appeared in 100% of ChatGPT answers and 25% of Gemini answers. Lucali sat at 88% on one platform and 12% on the other. The city and the question were the same, and the source mixes underneath were not.
Retrieval is also where most invisibility comes from, and the reasons are unglamorous. Listings are thin, one company name shows up in three spellings, and the facts on different profiles contradict each other. Our breakdown of why ChatGPT doesn’t recommend your business walks through the patterns that keep showing up in the data.
Vector two, how to get into AI training data for the next model
Every model has a knowledge cutoff, printed in its own documentation, and every model ships months after that date. Between the two sits the pipeline that decides what a model knows without searching. Text gets published, crawlers collect it, a training run consumes the collection, and months later customers meet a model that answers about some businesses instantly and confidently.
A business enters that pipeline at the first step or not at all. Whatever gets published this quarter is written into a model that reaches customers a few quarters from now, and nobody can buy a place in a training run that has already finished.
The asymmetry is easy to see for yourself. With search off, ask an assistant what it knows about a national chain in your category, then ask the same about your own company. The chain gets a paragraph about positioning, price range and reputation. Most local businesses get a sentence explaining that the model has no information. Nothing about quality separates those two answers. One name appeared in enough public text to survive compression into the weights, and the other did not.
Live search rewards one good page in the right place. Training data rewards the same true sentence written in a hundred places by other people.
Volume and variety are the levers here, and the formats that count are mostly the ones retrieval walks past. Video counts, as long as it comes with a transcript, because a recorded walkthrough or a founder explaining how the service works turns spoken words into text that crawlers read. Podcasts and interviews leave behind an episode page and show notes, both carrying your name beside your category on somebody else’s domain. Community threads on Reddit, Quora, industry forums and local groups are public conversation where real people name the business in context, in specific language you did not write yourself.
Social posts and comments are small on their own and add up to a large share of public text. For anything technical, the documentation, guides and course material that teach people how to use a product are read far more often by machines than by humans.
None of these move next Tuesday’s citation count. What they add is copies of the same association, which is what a training run compresses into weights. A model learns that a name and a category belong together, and the strength of that link comes from how many independent places said so. It does not memorize the website itself.
A business described three different ways teaches a model three weak associations instead of one strong one, so the description has to stay the same across every format. Mass-produced filler on your own domains works against you too, since deduplication and quality filtering strip that layer out before training starts.
AI search optimization vs AI training data, side by side
| Live retrieval | Next model’s memory | |
|---|---|---|
| What decides it | pages an assistant can find and cite while it answers | how often and how consistently your name appears across public text before the cutoff |
| Where the work lands | directories, review platforms, best-of lists, deep pages of your own | video, podcasts, social, forums, press, community threads |
| How fast it shows | days to weeks | quarters |
| How you check it | ask with search on, watch citations | ask with search off |
| What keeps you out | thin, stale, or contradictory listings | silence everywhere except your own website |
| What it costs to enter late | a listing and an afternoon | a year you cannot buy back |
The vocabulary of classic search covers the left column reasonably well, which is part of why the right column gets ignored. Our note on what GEO changes and what carries over from SEO covers where the old playbook still applies.
How to tell why your business is not showing up in ChatGPT
Run the two-answer test and read the combination.
Empty with search off and named with search on is normal for a local business. Retrieval is carrying you, so protect it and then start the slow work. Everything in the right column takes about a year to surface, which is exactly the head start you have.
Named with search off and missing with search on means the reputation outlived the presence. Assistants know the name and cannot find a current page worth citing, so the answer goes to whoever has one. Check what your site lets crawlers read and which of your pages ever appear as citations.
Empty both ways means starting with retrieval. It is the faster lane, and it produces the evidence that makes the slower work easier to fund.
Named both ways moves the question from presence to share, measured against the rivals appearing in the same answers. Our guide on how to measure AI visibility covers what to track at that stage.
How to rank in ChatGPT now and build visibility for later
Retrieval work runs as a monthly loop. Collect the questions customers ask, look at which pages assistants cite when answering them, and get accurate facts about your business onto those pages. Movement shows up in the same quarter, which makes it easy to justify and easy to keep going.
Training work runs as a publishing habit rather than a campaign. One recorded conversation becomes a transcript, a clip, a post and a thread. Each copy lands on a different platform and carries the same facts about what the business does and who it serves. The marginal cost per format is small, and the multiplication is the point.
The failure mode is predictable. Retrieval work reports numbers this month while the slow vector reports nothing, so the slow vector loses the budget argument every quarter. After a year of that, a competitor’s name is the one models answer with instinctively. A platform that watches both at least keeps the second vector on the same page as the first.
The businesses leading their categories in AI answers next year are running both clocks today. The 30-second test shows which one has been getting ignored.
Frequently Asked Questions
How do I check if my business is on ChatGPT?
Ask the question a customer would ask, without naming the business, and see whether the name appears in the answer. Then ask the same question with web search switched off, which shows whether the model knows the business without looking it up. Both checks take a minute each and they measure different things. A free AI visibility checker runs the same idea across ChatGPT, Google AI, Perplexity, Claude, and Grok at once, since every assistant reads a different mix of sources and answers differently.
Does posting on social media help my business show up in ChatGPT?
Rarely in the answer written today, often in the model trained next year. Live search leans on pages that answer a question directly, such as directories, review platforms, best-of lists, and detailed pages on your own site. Social posts, video transcripts, podcast episodes, and forum threads sit in the public crawl that training runs read later. The two channels pay off on different schedules, which is why social work looks worthless when it is measured only against this month's citations.
How long does it take to show up in AI search?
Weeks through live search, quarters through training data. A new listing on a platform assistants already cite can change an answer within a month. Getting into the weights of the next model means being published, crawled, and included in a training run, then waiting for that model to ship. Every model states a knowledge cutoff earlier than its release date. Plan the first in monthly loops and the second in quarters.
Can a business change what ChatGPT already knows about it?
The weights of a released model stay as they are, and live search is the lever that still works on it. ChatGPT can look a business up while it answers, so presence on the pages it cites carries the near term. Everything published now works on the next model instead of the current one, which is why the slow work starts long before you need the result.
Is AI visibility work worth it for a small local business?
It matters most for businesses customers ask about by category rather than by name. A model that has read your name beside your category in hundreds of independent places offers it without searching, and that position is hard for a competitor to outbid. The cost is a publishing habit rather than a budget line, since the same facts travel across video, audio, social, and community threads at almost no extra cost per format.