You get your business recommended by ChatGPT by making it easy to verify: one identity that agrees with itself everywhere, your actual expertise published in plain language a machine can lift, and independent sources that confirm what you say about yourself. There is no submission form, no listing fee and no setting to switch on. ChatGPT does carry advertising now, but an ad slot is not the answer, and nobody sells a place on the shortlist an assistant produces when a buyer asks who to hire. That answer gets built out of evidence the assistant can check, so your job is putting checkable evidence where it will be found.
Most owners arrive at this question holding a chore list. Post more often. Add schema. Get into more directories. Publish an llms.txt file. Some of that has real published evidence behind it. Some of it is folklore that spread fast because it was easy to sell, and you have no obvious way to tell the two apart from the outside.
Here is the sorting rule that helps. An assistant is not deciding whether you are good at the work. It is deciding whether it can defend saying your name to someone who will act on the answer. Those are different tests. Quality gets you into the category. Evidence gets you into the answer.
This guide covers what an assistant checks before it names anyone, which of the popular advice survives contact with published research, where to spend your first ten hours, and whether each AI assistant needs its own separate project.
What ChatGPT Is Actually Doing When It Names a Business
When someone asks an assistant who to hire, it is not sorting a list of pages by score. It is assembling an answer it can stand behind, from whatever it can find, confirm and summarize about the businesses in that category. A recommendation is a claim the machine is making on your behalf, and it will only make claims it can support.
That reframes the whole problem. Your reputation, as your market experiences it, lives in referral conversations, repeat clients and the moment somebody says "call her, she handled exactly this." As Probably Genius puts it: AI can't feel your reputation in a room. It can only work with what it can verify. The room is where you earned it. The record is where the machine reads it.
Nobody outside OpenAI can describe the internals, so treat what follows as a working model drawn from what these systems actually cite rather than a leaked mechanism. The checking runs along three plain lines. Are you real and consistent, meaning the name, people, credentials, locations and profiles agree everywhere it looks. Does anyone other than you say so, in sources you do not own. And does your own material carry knowledge it could not have gotten from the nine other firms in your category. The step by step version of that decision is worth reading once, because watching it happen removes most of the mystery.
Research on which sources get pulled into AI answers points the same way. In a controlled experiment reported in May 2026, researchers ran 252,000 trials across six large language models, testing 18 content factors two documents at a time. Topical relevance and position were the biggest drivers of being cited first, with explicit price information and a recent timestamp helping consistently. Formatting-only changes had little effect. That test measured which source a model reaches for when handed candidates, not who it recommends on the open web, so read it as a pattern rather than a rule. The pattern is blunt enough to be useful: substance decided, decoration did not.
Which is why the honest read on your competitor's appearance is unflattering to nobody. You weren't rejected. You were omitted. The machine never weighed your twenty years against theirs, because it never had your twenty years in a form it could weigh.
The Advice That Holds Up, and the Advice That Doesn't
Start with the piece of advice most often sold as urgent, because Google has now answered it in writing. In its guidance on optimizing for generative AI features, updated July 10, 2026, Google states: "You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities)." It adds that llms.txt files and similar files will neither help nor harm visibility there.
Read that carefully, because it is a statement by Google about Google. It covers AI Overviews and AI Mode, not ChatGPT, and no engine speaks for another. What it does tell you is that the special-file industry grew faster than the evidence for it, and that the same guidance names the thing that does count: content built from what you actually know, with in-depth experience in it.
The same page draws a hard line on volume. Producing many near-identical pages for query variations, primarily to influence rankings or generative AI responses, violates Google's scaled content abuse policy. Publishing more of the same is not a strategy, and at scale it is a penalty waiting to be applied.
Structure still earns its keep, just at the margin rather than the center. The peer-reviewed GEO study presented at KDD 2024 tested content changes against generative engines and found visibility could rise by as much as 40 percent, with the gains varying by domain. Citations, quotations and concrete figures made sources easier to lift. That is real, and it is a multiplier on substance, never a substitute for it.
Then there is the question people ask quietly, which is whether any of this can simply be bought. Ads did arrive: OpenAI announced in January 2026 that it would start showing ads to US users on the free and Go tiers, placed at the bottom of answers when a relevant sponsored product exists, clearly labeled, and separated from the response. OpenAI's own commitment is that ads will not influence the answers ChatGPT gives. Take that at face value and the conclusion is the same either way. You can buy a slot near the answer. You cannot buy the answer.
The unhelpful advice usually shares one tell: it promises a mechanism inside the model. Nobody outside the model companies has that. OpenAI has never published a ranked list of the signals behind naming one business over another in a recommendation answer, so anyone selling you the ChatGPT formula is describing a guess with confidence. And being cited is not the same as being recommended, which is where a lot of reporting quietly loses the plot.
Where to Put Your First Ten Hours
The work has an order, and doing it out of order is how firms spend a year publishing into a void. Four steps, in this sequence.
- Step one: read the room. Ask the assistants your buyers actually use the questions your buyers actually ask, and write down the answers. Not one screenshot. The same questions, across several engines, recorded so you can compare them next quarter. One question put to five engines tells you more in ten minutes than a month of guessing.
- Step two: make your identity agree with itself. Same business name, same address, same people, same credentials, everywhere a machine can look, connected with structured data so the pieces resolve to one entity. This is unglamorous and it is the foundation. Entity identity is the plain-English version of what that means.
- Step three: put your judgment on the record. Answer the real questions buyers ask you on the phone, in your own words, with your reasoning visible. The generic version of your category's advice is already in the training data. The version with your specific experience in it is the only part worth citing.
- Step four: get corroborated off your own pages. Independent mentions, business listings, reviews and press that tell the same story your site tells. Self-testimony has a ceiling, and third-party agreement is what lifts a claim over it.
Notice what is not on the list: posting frequency, follower counts and clever prompts. None of them change the record, and the record is the thing being checked. If ten hours is all you have this month, spend them on steps one and two, because everything after them compounds and nothing before them does.
One caution on step three. Your expertise probably feels too obvious to write down, which is exactly why your website is usually the thinnest part of your evidence. The judgment you consider unremarkable is the part no competitor can copy and no model already holds.
Do the Different Assistants Need Different Work?
Mostly no, with one practical exception. The evidence standard looks broadly common across the assistants we test: each of them needs a business it can verify, corroborate and describe accurately before it will put a name forward. Build that once and it tends to travel, because you are not optimizing for a platform. You are becoming checkable.
The exception is access. Engines that browse the live web reach your pages through their own named crawlers, and permissions are set per crawler in your robots.txt file. Allowing one does not allow another, and a well-meaning block from two years ago can quietly remove you from an entire engine while your site looks perfect to you. Check what you are allowing before you conclude the evidence is the problem.
Surfaces differ too. Google publishes explicit guidance for its generative features and ties eligibility to normal Search indexing, which makes it the most documented target of the group. Others publish less. Treat each engine as a separate reading rather than a separate strategy, and never generalize one platform's behavior into a claim about "AI," because the answers genuinely diverge by engine, phrasing and week.
That variance is also why a single screenshot proves very little. An assistant can name you on Tuesday and skip you on Thursday for the same question. The useful unit of measurement is a repeated panel of buyer questions across several engines, scored the same way each month, which is what turns an anecdote into a trend you can manage.
What to Expect, and What Nobody Can Promise
No honest provider will guarantee that ChatGPT recommends you. Model answers are probabilistic, the companies change them without notice, and nobody outside those companies controls the output. What you can build is readiness: verifiable identity, expertise on the record, independent corroboration. Then you measure what happens. Readiness is buildable. Outcomes are earned.
Waiting is the expensive option, because the step where a buyer would have found you by reading three websites is thinning out. Pew Research Center tracked 900 US adults and found that when a Google AI summary appeared, users clicked a traditional result in 8 percent of visits, against 15 percent without one, and 26 percent of those visits ended the session rather than 16 percent. That is Google, not ChatGPT, and it is the same direction of travel: fewer comparisons, more answers.
Accuracy matters more here than in any other marketing work you have done. A stretched credential on the public record is not a headline you can rewrite next quarter. It is evidence against you, filed permanently where every machine can read it. That is why every claim we publish traces to a source or it does not ship, and every piece is scored 0 to 100 through the Integrity Gate: nothing publishes under 80.
Done properly this is an operation rather than a project. Your expertise already exists, so the work is getting it out of your head and onto the record, then strengthening that record beyond your own pages month after month. You do the work only you can do. We make sure buyers, Google and AI can find it, verify it and understand why it matters.
Before you buy anything, get a reading. The free Recommendation Check scores your visibility across 109 checkpoints and shows you what five AI engines answer today when buyers ask who to trust in your category, so you start from what is actually missing instead of a chore list.
Want to Learn More?
Probably Genius is a done-for-you operation for expert-led firms: we map what you know, put it on the record in a form machines can verify, and build the outside corroboration that makes it credible. If your work is excellent and the introductions have gone quiet, the problem is probably not the work. It is the record. You're probably a genius at what you do. We make sure AI gets the memo.