Mostly yes. One round of evidence work improves your standing with every major AI assistant, because the assistants ask for the same things before they name a business: a specific page saying what you do and who for, visible signs you are operating now, sources other than you that agree, details that match wherever they are checked, and proof somebody can point at. Build that once and it counts in all five places. What does not carry across is access. Each company runs its own crawlers under its own permissions, and each pulls facts about you as a place through supply lines you do not own. Those get checked per engine rather than rebuilt per engine. It is a short job, and it is nobody's second retainer.
The question usually arrives with a proposal attached. Somebody has quoted you a ChatGPT package, and there is a Gemini one underneath it, and a note about Perplexity being the fast-growing one. You are already paying an agency you like. Nobody in the conversation can tell you whether these are five jobs or one job billed five times, and being unsure about that is not a gap in your knowledge. It is a gap in what anyone has bothered to explain.
So here is the sorting rule, and it holds up under the receipts. Anything that concerns what is true and checkable about your business is built once and serves every assistant. Anything that concerns whether a particular engine can reach that truth is specific to that engine, usually a settings question, and worth ten minutes of somebody competent rather than a second retainer.
This piece covers what five engines said when we asked them directly, the two places where per-engine work is genuinely required, what "once" is honestly worth over time, and the questions that sort a real proposal from a repackaged one.
What Five Engines Asked For
On August 11, 2026 we put ten buyer questions to five models with live web access through OpenRouter: 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. Fifty answers. Then we asked each one to account for the names it had given.
One caution before the findings, because it matters. A model describing its own reasoning is giving you its stated criteria, not a log of what it did. These systems reconstruct after the fact and sometimes get it wrong. Read what follows as five machines telling you what they think they want, which is still the most direct answer anyone has.
They disagreed about where they had looked. Gemini put its answer at nearly all live search. Claude Sonnet 4.6 put its own at roughly 80 to 90 percent training data and called that backwards for a market moving this fast. That split, and what it does to a business that has changed in the last two years, is the subject of our guide to where ChatGPT gets information about your business.
On what would get a firm named, they barely disagreed at all. Four of the five asked for a specific service page that states who it is for rather than a general page mentioning the category. Four described visible evidence of currently operating: recent posts, dated work, service pages touched this year. All five asked for corroboration from sources the business does not control. Three named consistent entity details, so the firm is not confused with a similarly named one. And all five wanted artifacts rather than adjectives: a described method, a case study, something with edges. Even the outlier fits. Claude did not ask for a service page by name. It asked for a named methodology and case studies with verifiable outcomes, which is the same demand wearing a different coat.
They were just as consistent about what got firms dropped. Gemini named vagueness and what it called unverifiable legitimacy, meaning sites too thin to tell whether the business was real and current. GPT-5.2 said plainly 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 absent from all five lists. Not one of them mentioned posting frequency, follower counts or how long you have been in business. What they asked for was checkability, and checkability is the same substance in every direction. That is why one build travels: you are not tuning a page for a platform, you are becoming a business a machine can confirm. The three layers underneath that, and how each one fails on its own, are laid out in what AI needs before it will recommend a business.
The Part That Does Not Travel
Now the exception. It is a real one. Evidence travels. Access does not.
The companies that browse the web publish their own agents and their own robots.txt tokens. Permissions are granted per token, not per industry. Allowing one does not allow another, and the settings are not equivalent even inside a single company.
OpenAI documents this in the plainest terms. Its crawler overview states that each setting is independent of the others, and that a site can allow OAI-SearchBot to appear in search results while disallowing GPTBot to keep its content out of model training. It also states that sites opted out of OAI-SearchBot "will not be shown in ChatGPT search answers," though it notes those sites can still appear as navigational links. One line in one file, and an engine stops quoting you in the place buyers ask their questions. A separate agent, ChatGPT-User, handles fetches a person triggered inside a chat, and because those are user-initiated, OpenAI says robots.txt rules may not apply to them.
Anthropic runs three. ClaudeBot collects content that may contribute to training. Claude-User fetches pages in response to a user's question. Claude-SearchBot indexes for search quality. Anthropic's own documentation says that disabling the second reduces your visibility for user-directed web search, and that disabling the third reduces your visibility and accuracy in search results. Perplexity is built the same way: PerplexityBot is the one that surfaces and links your site, while Perplexity-User handles a fetch a person asked for and, in Perplexity's words, generally ignores robots.txt because a user requested it.
Google supplies the sharpest example of why "one switch" is a fiction. Google-Extended is not a separate crawler at all. It is a robots.txt token controlling whether your content trains future Gemini models and grounds Gemini Apps answers. Google states directly that it "does not impact a site's inclusion in Google Search nor is it used as a ranking signal in Google Search." Read that twice. A block set two years ago by a cautious developer can hold you out of one assistant while every ranking report you receive looks completely healthy.
There is a second trap underneath the first, and it catches more firms than the robots file does. A permission in robots.txt is not the same as a request getting through. Perplexity publishes configuration instructions for Cloudflare and AWS firewalls precisely because bot protection sits in front of many sites and blocks the very crawlers the robots file invited. Three things worth confirming this quarter:
- Which named agents your robots.txt currently allows, one company at a time, rather than assuming a single AI rule covers them.
- Whether your firewall or bot protection is allowing those same agents through, including on subdomains, since a robots file is read per host and Anthropic's own instructions ask you to repeat the rule on each one.
- Whether anything on the checking list is behind a login, a script that never runs for a crawler, or a page that requires interaction to reveal its text.
Not every assistant documents this equally well. The four companies above publish named agents and controls. As of August 2026 we could not find equivalent first-party crawler documentation from xAI for Grok, which is a reason to treat that engine as an unknown and to re-check rather than to assume it behaves like the others.
Where the Facts Come From
The second per-engine difference is supply. It is the one owners least expect. You do not publish into an assistant. You publish into supply lines that some assistants read and others do not.
Google is the most explicit about this. Its Business Profile documentation says profile information is compiled from four sources: publicly available crawled web content, licensed data from third parties, contributions from users and from owners who claim the profile, and information based on Google's own interactions with the business. Your edits are one input of four. The same page notes that profiles are shown on Search and Maps and also on third-party sites and apps through interfaces such as Google Maps Platform, which means the record travels outward as well as inward.
Licensing runs between companies that do not answer to you either. Yelp tells business owners, in its own help center, that it "provides Apple with access to some of our business page information." It adds that it does not control how, whether or when Apple integrates it. Read that as a shape rather than a fact about any one assistant. Both examples run the same way: business descriptions get licensed between companies, so a wrong detail can arrive somewhere you never visited, from a source you never chose.
Google's generative surfaces add one more condition, and this is the part your SEO agency genuinely owns. Google states that to be eligible for its generative AI features a page must be indexed and eligible to be shown in Google Search with a snippet, and that meeting every requirement still does not guarantee crawling, indexing or serving. On Google's side, ordinary search hygiene is the entry gate. It is also not the finish line, which is the distinction we map in the coexistence guide for firms that already have an SEO agency and, at the definitional level, in whether AI visibility is the same as SEO.
Google's own guidance also happens to endorse the least technical part of this work. Its advice for generative AI features asks for non-commodity content, which it describes as unique expert or experienced takes that go beyond common knowledge, and warns that producing many near-identical pages to influence AI responses violates its scaled content abuse policy. The company with the most documentation is pointing at judgment rather than volume.
What Counts as Doing It Once
Put the two halves together and the sorting rule becomes usable. Ask of any proposed task: is this a fact about my business, or a route to it? Facts are portable. Routes are local.
One qualification before the list, because the honest version matters more than the tidy one. Portable evidence is not uniform lift. The engines still weight live retrieval and training memory differently, so the same record can land on one assistant this month and not on another. Portability means the work counts everywhere. It does not mean it lands everywhere at once.
Facts get built once, and they count everywhere. Your identity stated the same way in every place a machine checks. Service pages that answer a buyer's real question in the first sentence instead of describing your category. Corroboration on surfaces you do not own, which is the one thing your website cannot supply about itself. Judgment on the record, in your own name, that the other firms in your category have not published. None of that is engine-specific, and none of it expires the week a model updates.
Routes get checked per engine, on a schedule, and they are cheap. Crawler permissions. Firewall rules. Indexing eligibility. The place record and the profiles feeding it. Then the reading itself, because the engines answer differently enough that one of them is a sample rather than a verdict, which is the argument in one question, five engines.
Two honest qualifications on the word "once." The first is upkeep. In our own monthly runs, the sources behind an answer can change substantially month to month, so a record built and abandoned goes quietly stale while nothing on your side appears to break. The second is stature. In one large vendor study (Ranqo, 100,000 or more AI answers across 100 or more brands, March to May 2026), household names appeared in 73 percent of relevant unbranded AI answers. Established mid-market brands appeared in 44 percent, and niche or small brands in 11 percent. That is one vendor's sample and one method, and it is not a verdict on your firm. Read it as a map. The questions where household names already sit are the expensive ground. The specific, situational questions they never answer are where a smaller expert firm can be the best available name, and choosing that ground deliberately is most of the strategy.
So "once" is accurate about the build and misleading about the operation. The evidence base is built once. Keeping it true, extending it, and reading what the engines do with it is a rhythm, and the honest version of that rhythm is described in how to measure whether AI visibility work is actually working.
What to Buy, and What to Send Back
If a proposal on your desk has a line item per assistant, you now have the questions that settle it. None of them makes you technical. None of them fires anybody.
- Ask what is actually different between the ChatGPT work and the Gemini work. Access and place data are legitimate answers. "Different algorithms" is not one, because no outside vendor can inspect how these assistants assemble an answer.
- Ask what gets built off your own website, and who builds it. Corroboration is the layer every engine asked for and the layer most retainers never scoped.
- Ask how results get reported. An appearance rate across a fixed set of buyer questions and several engines, with a date attached, is a measurement. A screenshot is an anecdote with good lighting.
- Ask who checks the crawler permissions and the firewall, and how often. If nobody in the room owns that, the evidence work can be excellent and still not arrive.
The version we sell is deliberately one system rather than five. Your expertise mapped and published under your name, the entity details repaired and kept true, corroboration built beyond your own pages, and the same buyer questions put to five engines every month so you can see what moved. The full inventory of that is written out in what you actually receive from a done-for-you program, including the part that stays yours.
One boundary, stated plainly, because the market is not always careful with it. Nobody outside these companies controls what a model says, and any promise of a permanent place in an answer is a promise nobody can keep. What is buildable is readiness: evidence that exists, agrees with itself, and can be found by a machine that has a few seconds to look. Readiness is the part almost nobody finishes. As Probably Genius puts it: AI can't feel your reputation in a room. It can only work with what it can verify.
If you want a reading before you buy anything, our free Recommendation Check puts the questions your buyers ask to five engines and scores what comes back across 109 checkpoints. It is a diagnostic rather than 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 that one clean record does the work in every place a buyer might ask. That is the job inside The Answer: your expertise on the record, the entity kept coherent, corroboration built beyond your own site, and five engines read every month. The published methodologies show how it runs, including the 109-Point Diagnostic behind the free check.
You're probably a genius at what you do. We make sure AI gets the memo.