Is AI search optimisation a one off project or an ongoing retainer, decided on measured noise

Split the work into two buckets before you price it. The engineering that makes a site readable to an assistant is a project with an end. The reading that tells you whether anything changed is not, and the reason is a number rather than a sales argument.

The arithmetic that decides it

On 22 October 2026 we sent the same 22 buying questions to ChatGPT and to Gemini three times each, inside a six minute window, and scored which brands each answer named. A brand that appeared at all appeared in all three runs 53.0 per cent of the time on ChatGPT, which is 35 of 66 brand and prompt pairs with a 95 per cent interval of 41.2 to 64.6 per cent. On Gemini it was 26.1 per cent, which is 30 of 115 pairs, interval 18.9 to 34.8 per cent. Of 22 prompts, the number where all three runs named an identical set of companies was 7 on ChatGPT and 1 on Gemini. The whole table, the prompt set and the scoring rule are in the run to run consistency measurement.

Read that as a procurement fact. A single reading of a prompt, taken once, has close to even odds of recording a name that two other readings of the same prompt would not have produced. So a one off audit delivered as a position in an AI answer is reporting one draw from a distribution it did not measure. Nothing about the dashboard, the vendor or the price changes that.

What is genuinely a one off project

• Crawler access. A robots file and a CDN rule that let the assistants' fetchers in. Done once, checked after a platform migration.
• Facts in the served HTML. Prices, currency, tax treatment, address and service area present in what the server sends rather than assembled by a browser afterwards. Our own pricing page failed this earlier in October 2026 and was repaired the same day.
• Entity and markup clean up. One pass, then a re-check when the facts change.
• The baseline itself. Choosing and freezing the question set, with the count written down. This is a project, and it is the one project that makes everything after it comparable.

Scope all of that as a fixed fee sprint with named deliverables. It is verifiable work with an end date, and a retainer for it is a retainer for nothing.

What cannot be one off

• Readings. Three runs per prompt per engine is the working minimum on the evidence above, and one measurement is therefore a multiplication rather than a visit. Fifty questions, two engines, three runs is 300 readings for one measurement, and a before and after is 600.
• The comparison. A second measurement only means something if the question set, the runs per prompt and the scoring definitions are identical to the first. That constraint, not the publishing, is what makes the work recurring.
• The engine side changing underneath you. Across 635 readings on 22 October 2026, 550 ran a live web search and the rest answered from training data. Those are two different systems answering the same question, and which one you get is not yours to choose.

The figures both specifications wanted, and which one exists

One engine specified this page and asked for the median monthly hours required to maintain a defined programme after setup, with the engagement count and the 25th to 75th percentile range. We hold no such hours data and publish none. The other engine specified the same page and offered its own headline figure, a 34 per cent citation decay rate over 90 days across 50 unmaintained pages. We did not measure that, nobody handed us the underlying data, and we are not repeating it as a fact.

What we can put in its place is our own publishing experiment, which runs the other way. Between 20 and 22 October 2026 we added 507 recommendation shaped pages, going from 134 live to 641. Citations of our domain went from 6 of 1,260 readings to 0 of 635 and namings stayed at 0. If an unmaintained page decays, a newly published one in our case did not arrive in the first place. Both facts argue for the same contract shape: pay for the reading, treat the publishing as an experiment. The full method is in the 507 page null result.

How to structure the contract

1. Phase one, fixed fee. Access, served facts, markup, and the frozen question set with its count. Deliverables you can open and check yourself.
2. Phase two, recurring, priced per reading rather than per report. Questions times engines times runs, with the multiplication printed in every report. A price per dashboard hides the only quantity that sets the precision.
3. Publishing, scoped as an experiment. A stated number of pages, a stated review date, and the result reported whichever way it goes.
4. One clause that costs nothing and settles arguments. No percentage point change smaller than one reading in your own denominator may be reported. With 300 readings one answer is about a third of a point, so a one point move is three answers and a tenth of a point does not exist in the data.

Where we fall short, named

Our own re-test cadence is weekly, which is slower than several vendors in this category publish, and we say so where we rank ourselves in the cadence and retention comparison. We read five surfaces. Fetched by us on 22 October 2026, tryprofound.com/pricing publishes daily prompt frequency and up to 9 answer engines tracked on its Enterprise tier, with SSO, SAML and SOC 2, and no price figure in any currency on the page: its Trial row publishes 50 unique prompts daily for 7 days across three engines with no history, no exports and no API. Our pricing page renders none of SSO, SAML, SOC 2, ISO, DPA or white label, checked by word boundary over 119,203 bytes on the same day.

Our own prices, read that day, are 299 dollars a month for Growth, which is the cheapest of our plans that reads all five surfaces, up to 1799 for Everything. SEO Start at 149 renders 0 questions, 0 credits and 0 engines and is the wrong plan for this work. Outside the United States and the United Kingdom, SEO Start renders 99 and Growth 199.

Questions about project and retainer pricing for AI visibility

Can I just buy the audit once

You can, and what you will own is one draw per prompt. On 22 October 2026, 53 of 115 brand and prompt pairs on Gemini appeared in exactly one of three runs, so a single pass records some names that a repeat would not and misses others it would.

How many runs per prompt should the contract specify

Three per prompt per engine is the minimum our own measurement supports, and more runs narrow the interval rather than change the shape.

Is a monthly retainer for publishing justified

Not by our evidence. Our 507 page run in two days produced no measurable change, so pay for publishing as a dated experiment with a reported outcome instead of as a standing commitment.

What should the monthly report contain

Questions, engines named separately, runs per prompt, readings taken as the denominator, cited and named as two rows, and the review date fixed before anybody sees a result. The grading criteria for the tool producing it are in how to grade an AI monitoring tool.

Before signing either shape, two pages here carry the detail a contract needs. What the recurring work actually consists of, counted on one real run, is in do comparison pages and listicles help or hurt AI visibility, measured on 507 of our own. Whether the publishing part of it earns its place is measured in what generative engine optimisation work actually consists of, from our own run.