What is actually being sold
An AI search optimization service sells you presence inside the answers assistants give. The label varies by supplier: AI search optimization, GEO, LLM SEO, AI visibility, answer engine optimization. The name tells you nothing about what is in the engagement.
Real engagements contain some combination of four things, and they differ enormously in cost, speed and how much anyone controls them. Measurement. Technical access. Content. Presence in places you do not own. A proposal that does not separate those four is describing a category rather than a service.
This matters more than usual here because the vocabulary is new enough that a familiar service can be sold as a novel one without anybody quite lying. Both things are on the market, sometimes from the same supplier.
Ask which of the four you are buying and how the hours split. A supplier who cannot answer that has not scoped the work.
The fast half and the slow half
Everything in this work sits on one side of a line, and most disappointment comes from a proposal that blurs it.
The fast half
Confirming the AI crawlers can reach your site, and making your existing pages answer their own questions in the first paragraph rather than the fourth. Both respond within weeks. Both are genuinely cheap. Neither is a programme, and a supplier presenting the crawler check as a phase is padding.
This half is where most sites have easy ground, because commercial writing habitually circles a subject before committing to a statement, and an assistant composing a short answer needs a sentence it can lift.
The slow half
Building genuine depth on a subject you can own, and being referenced in the places that already get cited for your category. Both take a quarter or more, neither is automatable, and the second one involves relationships rather than deliverables.
This half decides whether you get retrieved at all, which is why it matters most and why it is the part most often quietly missing from a proposal. It does not scale, it is hard to bill by the unit, and it looks like nothing is happening for months. See AI search visibility for why retrieval is the binding constraint.
What no supplier can promise
Three claims worth treating as disqualifying, because they describe things nobody controls.
A guaranteed appearance. There is no position to occupy. An assistant names three or four sources and writes a sentence, and the same question asked twice can name different businesses for reasons unconnected to your site.
One timeline across all engines. Perplexity retrieves live and can reflect a change within days. ChatGPT tends to move over ten to fourteen weeks. AI Overviews move when Google rankings move. A single number covering all three has not been checked.
Revenue attribution. The path from a recommendation to a purchase runs through a separate branded search or a direct visit, carrying nothing that identifies its origin. Our page on AI brand mentions covers why the trail disappears. A supplier showing you revenue from AI mentions is modelling and should say so.
What can honestly be promised: the work, a measurement that shows whether it moved anything, and a straight account of which levers are slow.
What month one should look like
The shape of the first month tells you most of what you need to know about a supplier.
A crawler access check, in the first week. Ten minutes of work, binary outcome, and it blocks everything downstream. Confirm GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended can all read the site, which comes down to robots.txt, whether a content delivery network is challenging them, and whether pages render without JavaScript.
A question set written with you, not generated from your website. A generated set reflects your marketing language rather than your buyer's uncertainty, and the whole engagement then gets judged against the wrong thing. You know how your customers speak. A supplier does not.
A baseline before anything changes. This is the step most often skipped, because the numbers will look bad and nobody enjoys opening with that. Skipping it means never being able to show later what moved.
Not a strategy document. If the first month produces a deck and the first measurement arrives in month three, there will be no way to attribute anything that happens afterwards.
How to tell whether it is working
The reporting is where an engagement is either honest or decorative, and it is worth agreeing its shape before signing.
The same question set every month, unchanged. If the questions move, the history is worthless. Any change should be flagged as a new measurement rather than folded into a trend.
Split by assistant, not blended. A single figure hides the case where you are gaining in one and absent from another, which is usually the actionable part.
Every business named, not only yours. That set is your real competitive position, and it is frequently not who you benchmark against commercially.
Actual sentences, quoted. Being named dismissively counts as present in any tally. Reading the wording is the only way to catch a category deciding you are the cheap option.
An honest note on what did not move. In a field this slow, a report with no negative findings has been curated, and several clean months in a row is a warning rather than a result.
What drives the price
Pricing here is unsettled and the headline figure rarely reflects what actually scales.
Measurement scales with questions times assistants times frequency. Twenty questions across four assistants weekly is a few hundred answers a month. Two hundred questions daily is tens of thousands. Those are different products at the same nominal retainer, and competitor tracking multiplies it again.
Content is the largest line in most engagements. Ask for published work rather than a description of a process, and read it as a buyer in your category would.
The slow half is priced by time and cannot be shortcut. Anyone offering to accelerate being referenced is usually offering to buy placements, and bought placements in low-quality venues do not get cited.
The cost people miss is the backlog the measurement creates. A service that reports monthly on which questions you are absent from is generating writing work. Budget for it, or the retainer buys a recurring reminder of a problem nobody is addressing.
Three kinds of supplier wearing one label
Proposals in this space look alike and the businesses behind them are very different. Three groups, worth telling apart before comparing prices.
Search agencies who added the service. The largest group. Genuinely competent at the underlying work, because most of it is the work they already did, and honest ones will say so. The risk is a proposal that is entirely on-site because that is what they have always sold.
Specialists who started here. Newer, usually stronger on measurement and on the parts that are genuinely different, sometimes thinner on the fundamentals underneath. Ask what they would do if the measurement showed the problem was ordinary search authority, because that is the most likely finding.
Tool vendors selling services around their product. The measurement is usually excellent, since it is their software. The question is who does the work the measurement points at, and whether that is in the price or a separate conversation.
None of the three is wrong. What matters is matching the supplier to which half of the work your measurement says you actually need, which is another reason the baseline comes before the retainer rather than after it.
What to keep in house
The honest split, which cuts against a supplier's interest to explain.
Keep the fast half. The crawler check is an afternoon. The baseline is an afternoon. Rewriting existing pages to lead with the claim is a week and is mostly editorial judgement rather than technical skill. A small team can get a long way on this alone, and should, before paying anyone.
Buy the writing if the measurement says pages are missing. That is a straightforward commission and easy to judge on delivery.
Buy the slow half if you can find someone genuinely good at it. Building depth and earning references is sustained work with relationships attached, and it is the hardest thing to do alone and the hardest to buy well.
Do not sign a retainer before a baseline exists. Without one there is no way to separate work that helped from work that merely happened, which suits the supplier and nobody else.
Measure first, do the fast half yourself, then decide what is worth buying. Any other order is buying before knowing.
How to brief one so the work is judgeable
Most disappointing engagements here were badly briefed rather than badly executed, and the fix is unglamorous.
Give them the questions, or write them together. The single most useful thing a client can do, and it costs an afternoon. You know how your buyers speak and a supplier does not. A set generated from your website measures your marketing language rather than your market.
Name what success looks like, in advance and in writing. Not improved visibility. Something checkable: named in at least a third of the shortlist questions on one named assistant by a named date. It may turn out to be the wrong target. A wrong target you can test still beats a vague one you cannot.
Say which half you are buying. If it is the fast work, expect it finished in weeks rather than retained. If it is the slow work, accept that the first quarter looks like nothing is happening, because that is what building depth looks like from outside.
Ask for the baseline before any other deliverable. Everything else is unjudgeable without it, and a supplier who resists is telling you how they expect to be assessed.
A brief containing those four makes a good supplier better and makes a weak one visible inside a month, which is the point of writing one.
The question set is the real contract
Almost every dispute in this category traces back to the same omission. The engagement was agreed in terms of activity, so many pages, so many hours, and nobody wrote down the questions the work would be judged against. Six months later both sides are arguing about whether it worked, and there is no artefact that could settle it.
The question set is that artefact. It is the list of things a buyer might ask an assistant where you would want your name to come up. Twenty to fifty of them, in the words buyers actually use rather than the words your category uses. Agreed at the start, kept fixed, and run on a schedule from before the work begins.
Fixed is the part people get wrong. A set that gets edited each month cannot show a trend, because every edit resets the baseline. Add to it deliberately and keep the originals running unchanged alongside anything new, so that the old series stays readable.
There is a second reason to insist on it, which is that it forces an early conversation about whether your buyers reach assistants at all. Writing fifty questions a real customer would ask is harder than it sounds, and a business that cannot fill the list is usually discovering something important about its own market rather than about the supplier. It also protects you in a way that is easy to miss. A supplier who agrees a question set has agreed to be measured on questions you chose, which makes flattering reporting much harder to produce. A supplier who resists one is telling you something worth hearing. The mechanics of running it are covered in measuring something that leaves no trail, and it is worth reading before the first invoice rather than after the fourth.
What to walk away from
Five, roughly in order of how much they should worry you.
A guarantee of appearing. Nobody controls this, and offering to is either a misunderstanding or a misrepresentation.
Reporting that leads with traffic. Most recommendations produce no click, so a supplier measuring referral sessions will report failure during genuine progress and success during a fluke.
Refusing to show the question set. The set determines everything the engagement is judged on. Not being allowed to see it is not a defensible commercial position.
A volume promise as the headline. Any number of pages per month is a promise about output rather than outcome, and this is the category where publishing width instead of depth fails hardest.
A proposal entirely about your own website. It has quietly omitted the half that usually decides whether anything gets retrieved at all. Our guide to what moves visibility separates the two.