All ten sections with bracketed fields: objective, scope, method, client inputs, deliverables, exclusions, acceptance criteria, timeline and change control, fee, and the responsible-use statement, plus signature blocks.
Markdown templateWhat the buyer is actually purchasing
Before any section, get the frame right, because every language pitfall in audit proposals comes from misstating what is for sale. The client is buying three things: a baseline another practitioner could reproduce, findings that trace to saved evidence, and a roadmap with owners. They are not buying a visibility outcome, because nobody outside the platforms controls one. A proposal that implies otherwise wins the deal and loses the relationship at the first quarterly review, and the honest version is genuinely easier to sell to a skeptical buyer, as covered in how to sell AI visibility audits.
One gate before you write anything: if you cannot fill in the scope fields below from real information, do not guess and do not pad the fee for uncertainty. Sell a short paid discovery phase first. A proposal written before discovery is a fee attached to assumptions.
Objective and scope: where disputes are prevented
The objective section needs exactly one sentence pattern:
[Agency] will assess how [audience] can encounter and understand [client brand]
while deciding [business decision], across [declared platforms], in [geography]
and [language], producing a reproducible baseline, material gap findings, and a
prioritized action plan for the next [period].
Say “assess how the brand can encounter and understand,” never “improve how the brand ranks.” The first is observable work; the second is a promise about platform behavior. Pitfall language to strike from drafts: “dominate AI search,” “become the recommended answer,” “optimize for every model.” Each one is an objective claim about outcomes you cannot substantiate, which is the FTC’s reasonable-basis problem wearing a sales hat.
Scope needs six declared fields, and each one exists because its absence has a specific failure mode:
- Brand and offers. One legal or trading name and named offers. Without it, findings sprawl into service lines nobody asked about.
- Platforms and surfaces, including account state and location context. “AI search” is not a platform; “ChatGPT search, signed out, US context” is.
- Prompt panel: count, strata, run count, and date window, frozen before observation. This imports the whole method of the buyer-decision prompt set into the contract.
- Priority pages, listed or defined by a selection rule.
- Competitors, named or selected by a rule written before observation, always measured on the same panel and denominator.
- Technical review, itemized: public response, robots and crawler access, indexability and snippet controls, canonicals, rendered text, structured-versus-visible facts.
Method, inputs, and deliverables
The method section is short and load-bearing. Commit in writing to freezing the panel before observation, preserving exact prompts and full answer evidence, classifying mentions and citations separately, grading statements against a client-approved fact ledger, and labeling findings observed, verified, inferred, or unknown. Then add the sentence that protects every future engagement: like-for-like comparisons use the same declared method version, and a material method change starts a new baseline. This matches the open measurement protocol, and referencing a published method in the proposal is itself a trust signal.
Client inputs deserve their own numbered list with a deadline: one authorized owner, written approval of everything in scope, the approved fact ledger with a source per claim, agreed tool access, and response times for factual questions. The pitfall is politeness. If the input section has no consequence attached, the fact ledger arrives six weeks late and the delivery date is still yours. Tie delays to a day-for-day extension and to change control.
Deliverables should be countable artifacts: the scope and method record, the dated fact ledger, the technical eligibility record for a stated number of pages, the frozen panel and observation log, the mention, citation, representation, and source baseline with competitor comparison, findings traced to evidence, the client report, the handoff session, and a 90-day roadmap with owners and acceptance checks. Each maps to a section of the 47-point audit checklist, which functions as your internal quality gate; keep the internal scoring out of the proposal unless the buyer benefits from seeing it.
Exclusions: the section that sells
List, as items not included unless added in writing: guarantees of any ranking, mention, citation, recommendation, sentiment, traffic, lead, or revenue outcome; a census of all prompts, users, models, or answers; access to or claims about platform-internal systems; implementation work; continuous monitoring past the observation window; additional languages, regions, brands, or platforms; and legal or regulatory advice.
This section also pre-answers the most common sales objection, “how do we know it worked?”, with something better than a promise: acceptance criteria.
Six of the 47 checks live in the proposal, not the auditOne named business decision, recorded owner consent, fixed geography and language, declared platforms, bounded offers, and written exclusions. All six are settled in this document, and the checker grades them alongside the other 41.
Score scope and consentAcceptance, timeline, and fee
Acceptance criteria convert “done” from a feeling into a checklist: every scope field filled, the frozen panel preserved, every material finding traced to evidence, the four evidence layers reported separately, every roadmap item owning an owner and an observable completion check, limits and unknowns visible, deliverables opening in the agreed formats. A client who co-signs these criteria cannot later demand acceptance conditioned on a visibility outcome, because the document already defines acceptance.
Timeline language has one pitfall: quoting a start date before inputs exist. Anchor the window to a start condition (“five business days after section 4 inputs are complete”) rather than a calendar date you do not control. Put change control next to it: added prompts, runs, platforms, pages, competitors, languages, or deliverables change the fee and window, presented in writing before the added work begins. Every scope addition multiplies observations, not just effort; one added platform on a 30-prompt, three-run panel is 90 new observations to capture and classify.
Price the declared scope with the scope-first pricing model rather than working backward from a competitor’s number whose scope you cannot see. A fixed fee is safe exactly when the scope section above is complete; that is the commercial reason the scope rigor pays for itself.
The responsible-use statement
End with the paragraph that makes the whole document coherent:
Generative answers vary by platform, model, location, account state, wording,
and time. This audit reports a controlled, declared sample and available
first-party evidence; it cannot guarantee how any platform will behave.
Recommendations are designed to improve public clarity, technical eligibility,
evidence quality, and buyer usefulness even where sampled answers do not change.
The last sentence is the quiet test of the engagement’s worth. If your roadmap would be worthless when sampled answers stay unchanged, the audit was theater, and this statement will feel dangerous to include. If the roadmap fixes real representation conflicts, real eligibility failures, and real evidence gaps, the statement costs nothing and reads as confidence. Write the proposal so that sentence is true, deliver against it, and hand the results over in the client report template.