Quick Answer
Ask what the proposal contains that your SEO retainer does not already cover. Then ask for the baseline before the plan. A credible engagement starts by measuring whether assistants name any business in your category. It then fixes business-information accuracy across the web, and improves content clarity and third-party credibility. According to SOCi's 2026 Local Visibility Index, profile accuracy averaged 68% on ChatGPT and Perplexity against 100% on Gemini. That suggests inaccurate data is a larger constraint than missing content. If a proposal leads with new file formats and schema packages rather than a baseline, ask why.
Start With The Sizing Question Nobody Asks
Before any proposal, establish whether this channel can currently return anything for you.
According to a 2026 Visibility Labs analysis of 94 ecommerce brands, reported by Search Engine Land, ChatGPT referral traffic produced 1.48% of the revenue non-branded organic produced. That was over the same 12 months, and at a conversion rate 31% higher. Good conversion, very small base.
That figure is from ecommerce, and your category may differ in either direction. The point is the shape of the question. A channel can convert well and still be too small to fund separately.
So the first question for any prospective partner is not what they would do. It is what they found when they looked. A vendor who has run category prompts against your market before quoting is doing the work. A vendor whose proposal would read identically for any client in any sector has not looked.
What The Evidence Says The Work Actually Is
This is where most AI search proposals and most published research disagree.
Per the SOCi 2026 index, assistants that assemble business details from across the open web averaged 68% profile accuracy, while Gemini reached 100% by reading Google Maps directly. Roughly a third of what those assistants believed about the businesses tested was wrong.
Wrong information is a different problem from missing content, and it has a different fix. Publishing more pages does not correct an old phone number sitting on six directories. That correction is slow, manual, unrewarding work across third-party listings, and it produces nothing impressive to show in a monthly report.
Which is precisely why it is underrepresented in proposals.
Meanwhile, according to Semrush's 2025 research, business and service websites account for around half of the sources ChatGPT cites. Your own site is already the most likely thing to be cited. The question is usually whether it states your facts clearly and consistently, not whether it exists.
Seven Questions To Ask
Use these in the first call. The answers separate a considered engagement from a renamed one quickly.
1. What did you find when you tested my category? Look for specific prompts, dates and named competitors. A good answer includes the case where no business was recommended at all.
2. Which of these deliverables is not already in a standard SEO retainer? Entity clarity, structured data, site quality and credible mentions are long-standing SEO work. There is nothing wrong with that overlap. There is something wrong with paying twice for it.
3. How will you fix inaccurate business information, and where? You want a list of specific directories, aggregators and profiles, not a line item called "entity optimisation".
4. What will you measure, and from what baseline? Ask for the baseline to be recorded before work starts, including the exact prompts used. Without it, any later claim of improvement is unfalsifiable.
5. What happens if assistants do not recommend businesses in my category? The honest answer is to monitor quarterly and spend the budget elsewhere. A vendor who cannot say that has a fee to protect.
6. Who writes the content, and what makes it citable? Look for answers about clarity, accuracy and source quality rather than volume.
7. What are you not promising? This is the most useful question in the list. A partner who can articulate the limits of their control is describing a real service.
The Promises To Treat With Caution
Some claims in this category are worth pausing on, not because every vendor making them is acting in bad faith, but because the underlying mechanism does not support them.
Avoid guaranteed citations in a named assistant. Nobody can control the output of a model they do not operate, and outputs vary between sessions and users.
Be wary of a guaranteed position or ranking in AI answers. There is no stable ranked list to hold a position in, so nobody can deliver one.
A fixed timeline to visibility. Assistants refresh their information on schedules the vendor does not set.
Visibility as the outcome, with no conversion measurement attached. According to Exploding Topics research cited by Semrush in 2025, fewer than 19% of users routinely click through to the sources cited in AI Overviews. Being mentioned and being visited are different things, and a proposal that measures only the first is measuring the easier one.
A single file or markup package as the solution. Adding a file to your server does not resolve conflicting information about your business elsewhere on the web.
None of this means AI search work is not worth buying. It means the deliverable should be described in terms the vendor can actually control: audits run, listings corrected, pages rewritten, prompts tracked, baselines recorded.
A Simple Way To Compare Two Proposals
| Signal | Considered engagement | Repackaged retainer |
|---|---|---|
| Opens with | A baseline from your category | A list of deliverables |
| Names | Specific prompts and competitors | Generic industry trends |
| Biggest line item | Data accuracy and content quality | Schema and new file formats |
| Measurement | Prompts tracked from a recorded date | Visibility score with no baseline |
| Willing to say | "This may not be worth funding yet" | Always worth funding |
| Overlap with SEO | Stated plainly and priced once | Not mentioned |
The decision rule: if the proposal cannot show you a before state, you cannot verify an after state. Ask for the baseline in writing before you sign anything.
The tradeoff worth naming honestly. The most useful AI search work overlaps heavily with good SEO. That makes it harder to sell as a separate line, and easier to resell as one. A business with weak technical SEO, thin service pages and inconsistent business information will usually get more from fixing those first.
Your five-minute action today: ask an assistant a question a customer would ask in your category, without naming your brand, and note who it recommends. That is your baseline, and it costs nothing.
Tips For Buying This Well
- Ask for the exact prompts a vendor tested and the dates. Undated prompt results cannot be compared later.
- Run two or three prompts yourself before the first call, so you can tell whether their findings match reality.
- Request that the baseline is recorded in a document you keep, not inside the vendor's dashboard.
- Check whether the proposal prices work you are already paying for in an existing retainer.
- Treat a vendor who will name the conditions under which you should not hire them as a better sign than one who will not.
- Ask how business-information corrections will be verified after submission, since directories often revert.
What To Keep An Eye On
Watch whether the engagement reports actions taken or only scores moved, because a rising proprietary visibility score with no underlying work is not evidence. Watch whether corrected listings stay corrected after a quarter. Watch whether any reported visibility gain is accompanied by a change in enquiries, since visibility without conversion is an incomplete result. And watch for the overlap question being avoided rather than answered.
FAQs
Is AI search optimisation just SEO with a new name?
Substantially, but not entirely. The overlap is large enough that an honest partner will say so. The genuinely distinct parts are prompt-level measurement and cross-web information accuracy, which traditional SEO has historically treated as a minor concern.
Can anyone guarantee my brand appears in ChatGPT?
No. The output is controlled by a model the vendor does not operate and varies between sessions. Treat a guarantee here as a reason to ask more questions.
How much should this cost?
There is no defensible number to give you without knowing site size, market, competition, how many listings carry your details, and how much of the work your current retainer covers. Any figure quoted before those are known is a guess.
When is it too early to invest?
When assistants do not recommend businesses in your category at all. That is common in narrow local and B2B niches, and the correct response is a quarterly recheck rather than a retainer.
Where This Leaves You
The uncomfortable summary is that a good AI search engagement looks a lot like good SEO with two additions. Measurement at the prompt level, and serious attention to whether your business information is accurate everywhere it appears. That is worth paying for. It is also worth checking you are not already paying for most of it.
Run one prompt yourself before your next vendor call. If you would rather start from a proper baseline, that is what our AI search optimization work begins with. Our guide to choosing an SEO agency and spotting fake guarantees covers the wider buying question, SEO, AEO and GEO explained covers the terminology, and SEO packages for SMEs covers what belongs in a retainer. For a worked example of visibility rebuilt from the fundamentals up, see our financial services SEO case study.