The Search Measurement Contract Every Search Engagement Needs

A search measurement contract is a short written schedule that marketing, finance and the agency sign before any work starts. It fixes what success means, how each number is calculated, which system holds the truth, and what happens when the numbers move or fail to move.

This article walks through the twelve clauses that schedule needs, with weak and stronger wording for each one. The schedule itself is an operational template from Search Agency. It is not legal advice, and anyone who wants to use its wording inside a contract should have a lawyer review it first.

Our SEO measurement framework already argues that success should be agreed before the work begins. This piece supplies the instrument for that agreement, along with the edge cases that tend to break it.

Why search reports become arguments

In our judgment, many reporting disputes begin when two people use the same word for different numbers. Search Agency's own published pages show how easily this happens.

Our article on AI share of voice defines share of voice as your brand's mentions divided by all brand mentions across a set of answers. Our longer piece on AI search measurement describes AI share of voice as the percentage of answers that cite your brand. Both are reasonable measures, and they will produce different figures from the same data.

The same thing happens with recommendation rate. Our measurement page divides recommended mentions by all mentions, while the dictionary we now use for research divides recommendations by all valid answers. A brand that is rarely named but always praised looks strong on the first and weak on the second.

The following scenario is hypothetical. A marketing team reports that AI visibility rose from one quarter to the next. The finance partner sees flat revenue and asks what the rise was worth. The agency explains that three prompts were swapped in the second month because they no longer matched the product range.

Each party is correct by its own definition. Marketing counted mentions, finance counted money, and the agency counted against a prompt set that had quietly changed. Nobody lied, and the meeting still ends with less trust than it started with.

A measurement contract prevents that meeting by settling the definitions, the sources and the change rules while everyone still agrees on the goal.

Separate outcomes and indicators

The contract should name one business outcome and then list the indicators that help explain it. Indicators tell you where to look, and the outcome tells you whether the work paid off.

Citations, visits and revenue each answer a different question for a different reader. Mixing them on one line invites the wrong person to judge the wrong number.

Qualified demand deserves its own definition in the contract. We define it as a commercial outcome recorded in the CRM, which means a visit, a citation or a form submission does not count by default.

The qualification clause then says exactly what turns an enquiry into qualified demand. For a lead business that might be a budget confirmed on a first call. For a hotel it might be a confirmed booking above a minimum length of stay. The point is that sales and finance write the rule, and the agency reports against it.

Write the definitions

Every metric in the contract needs six parts written down. These are the event being counted, the denominator, the deduplication rule, the scope, the unit, and the system that holds the official figure.

Take owned-domain citation rate as the AI example. In our dictionary it is the number of valid AI responses that cite an approved brand-owned domain, divided by all valid responses.

Each word in that sentence needs a rule behind it. A valid response is an answer that was captured successfully for the intended prompt on the intended product surface. A technical failure is excluded and logged, because a failed capture says nothing about visibility.

An answer that loads normally and cites nothing still counts in the denominator, and it scores zero. That answer does not enter the denominator of the citation-conditional rate, which only looks at answers that cite at least one source.

The approved domain list matters too. If a brand runs a main site, a booking engine on a separate domain and a regional blog, the contract states which of those count as owned.

Surfaces are reported separately before anything is combined. ChatGPT, Gemini, Google AI Overviews and Google AI Mode retrieve sources differently, and a blended average hides which surface moved.

The collection method is named as well, because an official model API, a third-party collection service and the consumer app can return different answers to the same prompt.

Now take qualified leads as the commercial example. The event is the CRM status change that sales already uses, recorded once per company per period so that a buyer who submits three forms counts once.

The scope says which channels feed the count and which business units are included. The source of truth is the CRM, and the person who owns the CRM fields owns the number.

Two rules apply to every metric in the schedule. A zero denominator is reported as not applicable and is never shown as zero. A refusal or a no-answer response from an AI surface is reported as its own outcome and is never folded into a zero-visibility result.

Agree on attribution limits

The contract should separate four kinds of evidence and forbid adding them together. Each kind carries a different level of confidence, and finance will rightly discount a report that blends them.

Search Agency's measurement service already labels AI-attributed traffic as correlational and states that modeled branded-search lift is never presented as causal proof. Our internal reporting guidance goes one step further. The observed number is reported as fact, the modeled number is reported as a range, and the two are never blended.

The attribution clause also states what the contract does not promise. It should say plainly that no party will assign a rupiah value to a single AI mention, because no method can support that figure.

If the business wants a causal answer, the contract should budget for a test. Our article on measuring search influence describes the observed and self-reported layers in more detail.

Freeze the baseline responsibly

A baseline is the set of figures the work will be judged against, recorded before any work begins. The contract should freeze it in writing and still allow honest changes through a version log.

The baseline clause names the date range, the exclusions, any known tracking faults and the seasonal pattern the business expects. A retailer that starts measuring in a festive peak needs that peak written into the baseline, or every later month will look like a decline.

The prompt panel for AI measurement gets the same treatment. The clause lists how prompts were selected, the language of each one, the surfaces, the location setting, the number of repetitions and the panel version number.

Repetitions matter because AI answers vary from run to run. Our published method runs each prompt several times, and repeated runs of one prompt are treated as repeated measurements of one question and never as separate buyers.

Change control is the clause that is easiest to leave vague. Products launch, prompts stop matching what buyers ask, and tracking breaks. Freezing everything forever is as unhelpful as changing things without notice.

The change clause lists the events that trigger a new version. Typical triggers include a product launch or withdrawal, a tracking fault, a new AI surface, a change to the CRM qualification rule, or a merger of business units.

Each change gets a log entry with the date, the reason, the approver and the effect on past figures. The entry also states whether earlier periods will be restated on the new basis or reported in two series side by side.

Restatement is possible when the raw data still exists, which is a reason to keep raw answers and exports for the full term. When it is not possible, the report shows a visible break in the line, so nobody reads a method change as a performance change.

Assign delivery dependencies

A contract that judges the agency on outcomes must also record which work depends on the client. Otherwise an unshipped fix shows up in the report as agency underperformance.

Search Agency's methodology lists three dependencies on the client side. These are a release path for technical fixes, regular access to a subject-matter expert, and someone at the monthly readout who can make a decision.

The same page says engagements stall at the release path more than anywhere else. It also says that when numbers fail to move, the usual cause is fixes written and never deployed, or content signed off and never published.

The responsibilities clause should therefore track every work item through four states. Those states are recommended, approved, deployed and verified live, each with a date and an owner.

Outcome measurement for an item starts from the verified live date. A schema fix recommended in month one and deployed in month five has had one month to work, and the report should say so.

The clause also names who owns analytics. Search Agency's methodology page states that repairing broken GA4 and tag setups is outside its scope. A contract with us therefore names the party that will fix tracking if the baseline review finds a fault.

Agree on decisions

Measurement only earns its cost if it changes a decision. The contract should list the decisions in advance, who makes each one, and the evidence that triggers it.

Five outcomes cover most situations. The parties can continue, investigate, re-scope, pause or stop.

• Continue when indicators move in the agreed direction and delivery is on schedule.
• Investigate when a figure moves sharply in either direction, before anyone celebrates or panics.
• Re-scope when the evidence shows the original question was the wrong one, such as a category where buyers rarely use AI tools.
• Pause when a client-side dependency blocks delivery for longer than the agreed period.
• Stop when the agreed review shows no movement in qualified demand after the agreed maturity window, with delivery complete.

The thresholds belong to the parties, and this article will not supply them. A growth target written to make the clause look measurable does more harm than an honest statement that the first quarter sets the baseline.

What the contract can fix is the form of each rule. It names the metric, the comparison window, how missing data is treated, the size of change that counts as movement, and who decides.

Time horizon belongs in the same clause. Leads that arrive this month may take a full sales cycle to close, so a cohort is judged only once it has had time to mature.

Use the measurement schedule

The schedule below condenses the twelve clauses into one page that can sit beside an engagement agreement. Each row shows the kind of wording that causes disputes and the kind that prevents them.

The downloadable version adds a metric-definition sheet, a change log template and a decision-rights table. The metric sheet uses the same dictionary as this article, so the terms in a report match the terms in the agreement.

We would normally include a completed example from a real engagement here. We do not yet have one we are permitted to publish, so the schedule ships blank with notes on how to fill each row.

It works best when the finance partner helps fill it in. Data rights is an easy clause to skip, and it matters most at handover, when the client needs the raw answers and exports to keep the trend line going.

If you are about to commission search or AI search work, align your measurement plan before commissioning the campaign. Search Agency's measurement team can work through the twelve clauses with your marketing, finance and analytics leads before any reporting starts.