Before you hire anyone to improve your AI visibility, expect four plain answers: what ships every month, who does each job, what every piece has to clear before it goes out under your name, and what the provider will not promise. A proposal that cannot give you those four is not an engagement yet. It is a posture with a price attached.
Here is the part that makes this purchase feel slippery. You cannot inspect the thing you are buying influence over. Nobody outside the model companies can see how an assistant assembles a recommendation, which means nobody can audit that mechanism for you, however confidently they describe it. What you can audit is the supplier. You are not being asked to judge a technology. You are being asked to judge a vendor, which is a thing you have done many times before.
The difficulty is that every proposal in this category uses the same eight words. Verification. Entity. Answer engines. Schema. Authority. Recommendation readiness. A monitoring subscription, a content calendar you have to feed, and a full operating program can all be described in that vocabulary, and the price ranges overlap. The words do not sort them. The inventory does.
So this is a hiring standard rather than a category explainer: what a legitimate month contains, the questions that separate one proposal from another, the promises that should end a meeting, and who is left holding the risk when a provider gets it wrong. Probably Genius sells this work, so read what follows as the standard we publish and hold ourselves to, not the verdict of a neutral referee. Every test below is one you can run on us.
What Makes This Purchase Different
When you hired an SEO agency, there was a number at the end of it. You could open a rank tracker and see position four become position two. Argue about attribution all you like, the output was checkable by a person who was not selling it to you.
This category arrived without that number, because the output is unstable by design. Ask two assistants the same buyer question and you get different names. Ask the same assistant twice and you often still get different names. That is not a fault in your provider's work, and it is the reason no honest one will quote you a rank. What replaces the rank is an appearance rate measured across repeated runs, which is a real number with a date and a denominator attached, and we walk through how to read one in measuring whether AI visibility work is actually working.
The second difference is that there is no license here. No certification, no governing body, no agreed standard of care. The service category grew up faster than any standard of care did, and nobody can be struck off for practicing this badly. For now, the buyer is the regulator.
What does exist now is a paper trail, and it is more useful than most buyers realize. In its guidance on hiring search help, updated June 5, 2026, Google walks through interviewing a provider, checking references, granting only read access to Search Console during an audit, and finding someone else if they guarantee you first place. The same page now addresses this category by name, asking whether advice on "AEO" or "GEO" services is aligned with Google's official guidance on optimizing for generative AI features. The category has no license. It does have documentation, written by a party that is not trying to sell you the service, and it is yours to read before the first meeting.
What a Legitimate Month Contains
A serious engagement has a shape you can describe to your accountant. It opens with a setup phase that does two jobs: getting your actual expertise out of your head and onto the record, and making your identity consistent everywhere a machine checks it. Then it settles into a monthly rhythm that repeats.
The rhythm should be countable. Published expertise arriving on a fixed cadence, written from what you know rather than from category boilerplate. Upkeep on the entity, so your name, people, credentials and locations keep agreeing with each other as they drift. Evidence built beyond your own website, because self-testimony has a ceiling that no amount of on-site work raises. Measurement on a schedule, using the same buyer questions each time. And a report that shows both halves: what shipped, and what the engines answered.
In our own program that reads as 12 done-for-you thought-leadership articles a month, planned to create approximately 60 clear Knowledge Entries, plus entity and schema maintenance, external validation, and a monthly probe across the five engines your buyers actually use: ChatGPT, Gemini, Perplexity, Claude and Grok, with the answers saved word for word. The full inventory, line by line, sits in what you actually receive from a done-for-you program.
Your side of a done-for-you arrangement should be small and specific: a recorded conversation, and confirming facts only you can confirm. That is the honest trade. If drafting, scheduling or chasing starts migrating onto your desk, the program has quietly changed shape, and the month you were buying back has gone again.
The test is not whether a provider can talk about all of this. It is whether they will write it down as a list. Ask what ships, who makes it, and what it has to clear.
The Questions to Ask Before You Sign
Six questions do most of the sorting. Ask them in one meeting and the shape of what you are being sold becomes obvious, usually within the first two answers.
- What do you measure, and can I see last month's raw answers? Not a summary slide. The actual text an engine returned, dated, so this month can be compared with the next one. A provider who keeps the answers can prove movement. A provider who keeps only the chart is asking you to trust the chart.
- What gets built that is not on my website? Independent citations, listings, third-party references, published expertise that others can point to. This is the part buyers most often discover is missing after they have paid for a year.
- Who writes the material, and what does it have to pass before it publishes? You want a named threshold, not an adjective. If the standard is not written down, there is no standard.
- What access do you need from me, and when? Google's guidance is specific here: at the audit stage, grant read access to Search Console rather than write access. A provider who needs publishing rights before they have shown you anything is asking for trust in the wrong order.
- Which official guidance supports each recommendation? Google now asks buyers to check exactly this. Note the limit, though. Google documents its own generative features, and the assistant makers publish very little of the same kind, so this question tests one engine well and the others barely at all. Where doctrine runs out, ask for measurement instead.
- What happens at month twelve if the reading has not moved? Listen for a real answer about diagnosis and course correction. Both extremes are informative: a shrug, or a promise that it definitely will.
If you already have an SEO agency, run these questions past them too before you buy anything. Often the honest reply is that some of it sits outside their scope, which is useful rather than damning, and we mapped that whole conversation in what "we handle the AI stuff" actually covers.
The Promises That Should End the Meeting
Some answers are disqualifying, and not because they are ambitious. They describe control nobody has.
- "We guarantee the top spot," or "we guarantee ChatGPT will recommend you." Google states plainly that no one can guarantee a number one ranking, and tells buyers to find someone else if a provider promises first place. Assistant answers come with no public ledger of positions at all, so a guarantee about them is a bigger claim resting on thinner ground than the ranking promise Google already tells you to walk away from.
- "Our schema will get you into AI answers." Structured data genuinely helps machines understand a page. Google's own structured data guidelines, updated July 10, 2026, say directly that correct markup is not a guarantee of appearing in results. Eligibility is not placement, and a provider who blurs the two will blur other things.
- "We have a relationship with the model companies." Google's hiring page warns specifically about claims of a special relationship or a priority submission. No assistant maker publishes a ranked list of the signals behind naming one business over another, so anyone selling you the formula is selling a guess with good posture.
- "We will generate reviews for you." The FTC's consumer reviews and testimonials rule took effect on October 21, 2024, and the Commission's own guidance says advertising agencies, public relations firms, review brokers and reputation management companies are not immune from liability under it. That one is not a matter of taste. It is a federal rule, and the Commission's guidance says courts can impose civil penalties for knowing violations.
- "You will see results in six weeks." Nobody can date this. Google notes that a recrawl alone can take from a few days to a few weeks and still does not guarantee inclusion, and that is one of the better documented steps in the chain. The rest is slower and less visible.
- "We will publish a hundred pages a month." Volume is not the strategy it appears to be. Google's guidance says mass-producing pages without adding value can violate its scaled content abuse policy, so the aggressive-sounding option is the one carrying the compliance risk.
A guarantee in this category is not confidence. It is a tell. The defensible version of the promise sounds smaller and holds up better: readiness, not causation. A provider can make your business easier to verify, trust and recommend, then measure what happens. Ours is written on the wall in exactly those words, and you should expect any provider you hire to state their boundary as plainly.
Who Carries the Risk If It Goes Wrong
This is the question buyers skip, and it is the one that costs the most. The work does not stay in a campaign account. It goes onto the public record under your name.
Google puts the position bluntly in its hiring guidance: you are ultimately responsible for the actions of any company you hire, which is why it advises knowing exactly how they intend to help. The FTC's rule points the same way from the legal side. A business that puts testimonials on its own website is disseminating them rather than merely hosting them, and if those testimonials are fake or false, the business itself can be liable.
Apply that to a stretched credential, an invented statistic or a case study that describes work you did not do. A bad marketing campaign ends when you stop paying for it. A bad published claim keeps sitting there, indexed, quotable and available to every machine that looks you up, long after the invoice is settled. The record outlives the retainer.
So ask the provenance question directly. Where does each claim in my content come from, and what happens when a fact cannot be sourced? The answer you want describes a rule, not a proofreader, and the difference is the whole subject of how to stop AI content from inventing your expertise. For our own work the rule is a scored gate: every piece is scored 0 to 100 through the Integrity Gate, and nothing publishes under 80. You do not have to adopt our number. You should insist on somebody's.
If You Are Not a Household Name
There is one more thing worth understanding before you set expectations, because it changes what a good plan looks like for a firm your size.
In a large vendor study by Ranqo, covering more than 100,000 AI answers across over 100 brands between March and May 2026, household names appeared in 73 percent of relevant unbranded AI answers, established mid-market brands in 44 percent, and small or niche brands in 11 percent. That is one vendor's dataset rather than a law of nature, and the middle rung matters as much as the ends: the ladder is a gradient, not a wall.
Read it as a map instead of a verdict. The broad category questions are where the biggest names already sit, and a plan that spends your year competing there is spending it badly. The winnable ground is the specific, situational question a buyer actually asks, the one where your twenty years of judgment is genuinely the better answer and the household name has nothing on the record at all. So add a seventh question to your list: which buyer questions do you intend to win for me, and why those?
Before you sign anything, get a reading you did not pay for. The free Recommendation Check is our 109-point AI visibility diagnostic: what five AI engines answer today when your buyers ask who to trust, which of your claims cannot currently be verified, and where the useful opening sits. It takes about an hour to present and carries no obligation. Take the results into any meeting you like, including one with somebody else. A provider worth hiring will be glad you arrived with a baseline.
Want to Learn More?
Probably Genius was built by Jacquie Baker and Christopher Shaw, who spent more than 40 combined years working out what makes an expert the obvious choice in a room, and then translating it for the systems that now assemble the shortlist. This piece is the standard we ask you to hold us to, published before you ask. You're probably a genius at what you do. We make sure AI gets the memo.