What to put in a contract for AI search visibility work, clause by clause

A list of questions to ask a vendor is useful before you sign and worthless afterwards. What survives is the document. These are the clauses that decide whether the next twelve months produce evidence or an argument, and every one of them is here because a measurement of ours says it matters rather than because it sounds prudent. The interview stage is covered separately in the 27 question checklist.

Clause one, the question set as a dated annexe with its count

Not a description of the question set. The questions, as an annexe, with a number, frozen before any work starts, and your company name in none of them. Our own audit bank holds 324 written questions of which 318 were asked, and the gap is the point: the 6 that were not asked are the ones naming our own company, and because that batch never ran we hold no prompted figure at all and publish none. A count in the contract makes that kind of gap visible instead of invisible.

Clause two, runs per prompt, written as a number

Three per prompt per engine is the minimum our own evidence supports. On 22 October 2026 we sent 22 buying questions to ChatGPT and Gemini three times each inside six minutes. A brand that appeared at all appeared in all three runs 53.0 per cent of the time on ChatGPT, 35 of 66 brand and prompt pairs, and 26.1 per cent on Gemini, 30 of 115. Of 22 prompts, all three runs named an identical set of companies on 7 of them on ChatGPT and 1 on Gemini. The table is in the run to run measurement.

The specification an engine wrote for this page proposed a contractual minimum of a 250 prompt test set re-measured every 7 days across four engines. That figure is the specification's, not a measured threshold, and we are not putting it in a contract as though it were ours. What our data supports is the runs, not the prompt count.

Clause three, engines and interface named, and reported separately

Name the engines and name how they are read, because the two are different objects. The interface our audit read through accepts four location parameters, gl, hl, location and country, and ignores all four, which we measured, so a market has to be locked in the prompt text. And of 635 readings, 550 ran a live web search while 85 answered from training data. A contract that says five engines without saying how they are read is buying an unknown mixture of those things.

Require one column per engine and refuse a blended score. Our own run shows why: Semrush appeared in 259 of 635 readings and Profound in 154, and the two engines did not agree with each other about the order.

Clause four, two outcome rows, never merged

Cited means a page of yours was material an answer was built from. Named means your company appears in the answer text. They fail for different reasons and a single blended figure hides whichever one is bad. Our own numbers on 22 October 2026 were named 0 of 635 and cited 0 of 635, after being cited 6 of 1,260 two days earlier with 134 pages live against 641. Both numbers are in the 507 page null result and they are never added.

Clause five, the denominator and the floor it sets

Every rate in every report carries its numerator and its denominator. Then one sentence that settles most future arguments: no percentage point change smaller than one reading in the denominator may be reported. With 50 questions, two engines and three runs, one measurement is 300 readings and 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. With a denominator of 18 readings a single answer is more than five points, which is how a report shows a dramatic rise from one extra mention.

Clause six, raw answers retained and exportable by you

The answer text, the sources block and the dates, exportable in a format you can read without the vendor, and still yours if the engagement ends. A stored percentage tells you a number moved. The stored text tells you what moved, which is the only thing you can act on. Export is also the field most often missing on the entry tier of a product rather than on the product: fetched by us on 22 October 2026, tryprofound.com/pricing publishes no data history, no exports and no API on its Trial tier, and all time history with CSV and JSON export and an API on Enterprise, priced as Custom with no figure in any currency on the page.

Clause seven, the counting rule, in writing

Ours is once per reading in which a name appears, never once per occurrence, and the sources block is separated from the answer prose before the prose is read. Write whichever rule you agree on into the document, because the rule moves the totals more than the data does. Two tallies of our own 224 specification file under two different word boundary rules returned method 114 against 96 and dates 152 against 62. Same file, same day, two rules. The pair is published in what makes an assistant cite a vendor website.

Clause eight, the review date, fixed before anybody sees a result

A date in the contract for the re-run, chosen in advance, with the same questions, the same runs and the same scoring. A window moved after the fact to catch a better month is not a measurement. Put the arithmetic for both ends in the document so the follow up is comparable by construction rather than by goodwill.

Clause nine, what is explicitly not promised

List it, so nobody has to infer it. No position in an AI answer, no share of answers that will name you, no date by which an engine will change what it says, and no attribution of a change to the work unless the study carried an untouched comparison set. Anyone offering those in writing is a red flag and the clause is there to make the offer impossible to make quietly.

Clause ten, the conflict of interest statement

Require the vendor to state, in the report, that it sells the remedy it is measuring. Of the 224 page specifications the engines wrote, 71 asked for exactly that disclosure, counted by word boundary. Ours is on every page: we sell AI visibility measurement, every figure here is from our own run, and the figure that costs us most is the zero.

Where a rival beats us, in the contract terms a buyer cares about

Our pricing page renders no SSO, SAML, SOC 2, ISO, DPA or white label, checked by word boundary over 119,203 bytes on 22 October 2026, and a loose search for sso on it matches crossOrigin seven times and subprocessors twice rather than the feature. If your legal team needs the addendum or the certificate, the Enterprise tier at tryprofound.com/pricing publishes SSO and SAML with SOC 2 compliance and up to 9 answer engines against our five, read by us on that date. Our own page also contradicts itself on the Agency plan, which reads businesses covered 1 1 1 1 1 1 10 1 beside a comparison table answering No to more than one business in all eight columns, and both render, so a multi brand clause needs checking against the product rather than against the page.

Questions about contracting for AI visibility work

Is a fixed fee or a retainer right

Access, served facts, markup and the frozen question set are a fixed fee project. Readings are recurring, because three runs per prompt per engine is a repeating multiplication rather than a visit.

How many prompts should the contract specify

Enough that one answer is a small fraction of your denominator. Our data sets the runs rather than the prompt count, so pick the prompt count from the floor you can live with and print the arithmetic.

What happens to the data if we leave

Whatever clause six says. Put the export right and the retention period in the document, because the answer text is the asset.

Should the vendor be allowed to change the question set

Only by written amendment, with both versions kept. A question set edited between measurements makes the pair incomparable and the edit is usually invisible in the report.

Two pages behind these clauses. The purchase decision itself, with our own worst number attached, is in is generative engine optimisation worth paying for, and what our own null result says about it, and the project against retainer choice that clause one depends on is in is AI search optimisation a one off project or an ongoing retainer, decided on measured noise.