Separates the offer from the product, then engineers the risk clause: guarantee type, what each costs to honour, who adjudicates it, and which buyers each one attracts.
We have not measured this one. It is here because most of its content is procedural and checkable, not because it beat a run without it.
What it knows, concretely. First, a guarantee table that treats risk reversal as engineering: five guarantee types, what each one costs to honour once you count delivered cost of goods and the processing fee your payment provider may not return, which buyers each type selects for, and the adjudication rule each one requires. The load-bearing claim is that a guarantee with no adjudication rule is not a guarantee, it is a dispute scheduled for later, and the rule has exactly five fields: who decides, what the buyer submits, the deadline, the remedy, and what happens on disagreement.
Second, a ceiling on value stacking with a named mechanism behind it. Adding low-value items drags the perceived value of a bundle towards its average rather than its sum, an effect described in the consumer research literature as the presenter's paradox, from work by Weaver, Garcia and Schwarz published in 2012. The operational consequence is a cap of roughly three to five components, each of which a buyer would have paid for separately.
Third, a proof ladder ordered by how expensive each rung is to fabricate, which turns out to be the same order buyers weight them in: something they can try themselves, a number they can verify independently, a named customer, an unnamed customer, a testimonial, a bare claim. The rule attached to it is that your single most load-bearing claim has to be supported at the top two rungs, and that a seventh testimonial does not substitute.
Who it is not for. If you sell one thing at one price to people who already know you, this is heavier than the decision. If your problem is that nobody arrives on the page at all, this audits the wrong end of the funnel. And if you do not yet know what your buyer would do instead of buying, the audit cannot score half its dimensions and will send you to do positioning first.
Measuring one skill honestly costs about twenty model sessions: five runs with it, five without, on real material, each output graded alone by a session that is not told the other arm exists, against a rubric written by somebody who never saw the skill. We have not spent that on this one yet, so it ships labelled rather than ships silently.
How it would be measured. Tier A spine with a Tier B layer on top. Both arms audit the same three invented sales pages carrying planted defects: a guarantee with no adjudication rule, a nine-item bonus stack, four objective claims with no substantiation, and a proof section resting entirely on unnamed testimonials. The rubric counts planted defects found and false alarms raised on the one page that is deliberately clean, and the rewritten risk clauses then go to blind pairwise preference.
A fair test is buildable here and it is not cheap. The spine is objective: plant a fixed set of defects in an invented sales page, a guarantee with no adjudicator, a nine-item stack, objective claims with no evidence, a proof section made entirely of unnamed quotes, and count how many each arm finds. That part needs no taste at all.
The part that resists scoring is the rewrite. Whether a rewritten risk clause is better than the original is a preference judgement, so it needs blind pairwise comparison with a grader who does not know which arm produced which.
The honest doubt sits with the control. A strong model asked to review a sales page will already say the guarantee is vague. The question is whether it produces the adjudication fields, names the adverse selection each guarantee type attracts, and refuses the seventh bonus. That needs a page where the obvious advice is wrong.
The rule that decides pass or fail was written down before any run was executed and it does not move afterwards. It is in the method note on the hub, along with the full results table including every skill that was tested and cut.
Stated plainly, because a skill that claims everything is useful for nothing.
name and description in the file's frontmatter, so you can also invoke it by name.A conversion copywriter with real customer research beats this comfortably on everything downstream of the offer, and the good ones start by fixing the offer anyway. If you can hire one, hire one.
The regulator's own guidance is free and primary. If your only question is whether a particular claim is permitted, read the source for your market. Any summary, this one included, is a lossy copy of a document you can open yourself.
Your own refund inbox is the cheapest research available and almost nobody reads it in bulk. Twenty refund requests, sorted by reason, will tell you more about which promise is misfiring than any framework can infer from the page.
And the model with no skill at all is a reasonable choice when the offer is already sound and you want a second reading of the page. It is the wrong choice when the structure is broken, because a fluent page with a broken risk clause is a more expensive mistake than a clumsy page with a sound one.
--- name: offer-audit description: Audits an offer as a construct separate from the product: scope, price and terms, the remedy when it does not work, time to a result, what the buyer has to do, and who carries the risk. Includes a guarantee engineering table covering cost to honour, adverse selection and the five-field adjudication rule, the damaging admission as a credibility mechanism, the ceiling on value stacking, the substantiation duty attached to objective claims, a proof ladder ordered by how hard each rung is to fabricate, and a scored checklist with a rule about what to fix first. This skill should be used when writing or reviewing a sales page, a pricing page's offer block, a proposal, a guarantee or refund policy, or any bundle of bonuses. --- # Offer construction audit ## The claim this skill is built on The product is what you built. The offer is what the buyer is being asked to agree to. They are different objects, and almost all weak offers are strong products with the risk left entirely on the buyer. You can see this by writing the two lists side by side. The product list is capability. The offer list is six things, and only the first of them is about capability at all: | Component | The buyer's question | Where it usually goes wrong | | --- | --- | --- | | What is included | What exactly do I get | Scope stated as capability, with the exclusions omitted | | What it costs | What do I pay, when, in what currency, for how long | Price present, terms and renewal absent | | What happens if it does not work | What is my remedy | A guarantee sentence with no rule behind it | | How long it takes | When do I have a result | No timeframe, or one with no condition attached | | What I have to do | What work is mine | Silence, which the buyer reads as none, and then discovers otherwise | | What I am risking | Money, time, data, or my own credibility internally | Only the money is addressed, and it is usually the smallest of the four | The last row is the one people miss. In business purchases the money is frequently not the largest risk. The buyer is risking the political cost of having recommended you, the switching cost of moving data in and back out, and the time of the people who will have to learn it. An offer that reverses only the financial risk has addressed the cheapest quarter of the problem. ## The risk clause, engineered Risk reversal is usually written as a slogan. It should be written the way you would write a warranty, because that is what it is. | Guarantee type | What it costs to honour | Who it attracts | Adjudication needed | | --- | --- | --- | --- | | Unconditional money back, any time | Full refund, plus delivered cost of goods you cannot recover, plus the processing fee your provider may not return, plus the support time | Everyone, including the small group who extract the full value and then ask for the money back | None, which is both its point and its cost | | Time-bounded money back, for example 30 or 60 days | The same, capped by the window and by how much value is deliverable inside it | Cautious buyers, if the window is longer than your time to first result | A date check, and a rule for the request that arrives one day late | | Conditional on the buyer doing something | Lower, because the condition filters | Buyers who intend to do the work, which is the population you want | Real: you must be able to verify the condition from artefacts you already hold | | Performance based, a named outcome or you do not get paid | The entire delivery cost, with no revenue | Buyers with the worst baseline and the least measurable data, which is the opposite of what you want | Heavy: a named metric, a named source of truth, a baseline period, and a named person who reads the number | | Remedy in kind, credit, extension or extra service | Cheapest of all, mostly marginal cost | Nobody in particular, because it reverses very little | Light, but it fails the buyer who wants out rather than more | Two rules come out of this table. **Rule one: the window must be longer than your time to first result.** A fourteen-day guarantee on a product whose first meaningful output arrives in week four is not a guarantee, it is a deadline that forces a defensive refund from anyone who has not got round to it yet. Lengthening the window removes the deadline. It usually reduces refund requests rather than increasing them, because the pressure to decide before the clock runs out disappears. **Rule two: a guarantee with no adjudication rule is a dispute waiting to happen.** Write the rule in five fields and publish it next to the claim. 1. Who decides. A named role, not "the team". 2. What the buyer submits. The specific artefact: an export, a screenshot, a completed worksheet, a date. 3. The deadline, on both sides. Theirs to claim, yours to answer, in days. 4. The remedy. The amount, the form, and how it arrives. 5. What happens if you disagree. Because you will, roughly once in every hundred. If the rule does not fit in four sentences it will not be applied consistently, and inconsistent application is worse than no guarantee, because it becomes a payments problem. A buyer who cannot claim a remedy that was advertised does not go away quietly. They raise a card dispute, which costs you the transaction, a dispute fee, and a contribution to a ratio the card networks monitor with published thresholds under about one percent. Ask your payment provider for the current numbers before you assume this is theoretical. ## The damaging admission A claim that concedes something real buys credibility for the claim standing next to it. The mechanism is straightforward: the concession costs you something, so the reader stops reading you as a seller for a moment, and that suspension carries over to the adjacent sentence. The structure is fixed. Concede a specific, true, checkable limitation that the buyer would otherwise discover on their own. Place the load-bearing claim immediately next to it. Do not soften the admission. **Example one, a data import.** Before: "The fastest way to move your data across." After: "The importer handles the four formats listed below in one pass. If your data sits in a custom schema it will fail, and you will be using the manual mapper, which takes about two hours for fifty thousand rows. Everything in those four formats has never needed a human." **Example two, team size.** Before: "Works for teams of any size." After: "Under about five people this is more process than you need and a shared document is genuinely better. Between five and fifty it replaces the four tools listed here, and that band is what every design decision was made for." Three constraints. The admission must be true, or you have made your credibility problem legally interesting rather than solved it. It must be specific, because a vague concession reads as modesty and buys nothing. And it must not be a disguised boast: "we are too thorough" is the tell that the technique has been understood as a tone rather than a trade. ## Value stacking and its ceiling Adding components to an offer raises perceived value up to a point and then lowers it. The mechanism is an averaging effect: people evaluate a bundle closer to the average of its parts than to their sum, so appending a low-value item to a strong set drags the whole set down. In the consumer research literature this is the presenter's paradox, described by Weaver, Garcia and Schwarz in work published in 2012. The operational consequences: - Cap the stack at roughly three to five components. Eleven reads as padding and invites the reader to price the weakest one. - Every component must be something a buyer would plausibly have paid for on its own. If it would not survive as a product, it is not a bonus, it is a feature, and it belongs in the scope list. - Every component needs an external price reference, or it has no value to add. "A template pack" prices at nothing. "The onboarding session we sell separately at 400" prices at 400 and is checkable. - Remove before you add. The fastest way to raise the perceived value of a six-item stack is usually to delete the two weakest items. ## Specificity, and the duty that comes with it A claim with a number, a timeframe and a named condition is checkable, and checkable is why it is believable. The same claim without them carries no information and the reader knows it. | Noise | Checkable | | --- | --- | | Save time on reporting | Cuts the monthly close from six working days to two, for teams already on the accounting systems listed below | | Trusted by leading companies | Used by eleven insurers, four of whom will take a reference call | | Fast setup | Live in one working day if your data is in one of four formats, about a week otherwise | | Great support | Replies within four working hours, from the engineer who wrote the feature, published response times for the last quarter | The duty attached is not optional in most markets. Objective claims that a reader would take as factual generally require the advertiser to hold evidence for them before publication. The UK advertising code states this directly at rule 3.7, which requires documentary evidence for objective claims capable of substantiation. In the United States, the Federal Trade Commission's long-standing advertising substantiation policy requires a reasonable basis for objective claims before they are made. Testimonials and reviews carry their own rules, and they have tightened: the FTC finalised a rule on consumer reviews and testimonials in 2024 covering fake reviews and undisclosed connections. Check the current text for your market before you rely on any of this, and treat this paragraph as a prompt to look rather than as the answer. The audit consequence is simple. Every sentence with a number in it gets a source, or it comes off the page. Reducing the volume of an unsupported claim does not fix it; replacing it with one you can evidence does. ## The proof ladder Ranked by how hard each rung is to fabricate, which is also, and not coincidentally, the order buyers weight them. 1. **Something they can try themselves.** A trial on their own data, a sandbox, a public calculator, source code. Cost to fake: you have to build it. 2. **A number they can verify independently.** A status page with history, a third-party benchmark, an audited figure, a public filing. Cost to fake: fraud, with a paper trail. 3. **A named customer.** Strongest when they will take a reference call, because faking it requires a real person to lie on a scheduled call. Weaker as a logo on a wall, which requires a logo file and, usually, permission you should actually have. 4. **An unnamed customer.** "A mid-sized logistics operator." Cost to fake: a paragraph. 5. **A testimonial.** Cost to fake: a paragraph and a photograph, which is precisely why regulators wrote rules about this rung specifically. 6. **A claim you make about yourself.** Cost to fake: nothing. The rule: your single most load-bearing claim, the one the purchase turns on, must be supported at rung one or two. Everything below rung three is decoration on top of that, useful for warmth and useless as evidence. If the entire proof section lives at rungs four to six, the offer is unproven no matter how much of it there is, and the fix is not more of it. ## The audit procedure Score an existing page on eight dimensions, zero, one or two each, for a maximum of sixteen. 1. The buyer's alternative is acknowledged somewhere on the page. 2. Scope is stated, including at least one explicit exclusion. 3. Price and terms are visible without a form: amount, period, renewal, what happens at the end. 4. The risk clause exists and carries all five adjudication fields. 5. Time to a result is stated with the condition it depends on. 6. The buyer's own work is stated honestly. 7. Claims are specific, and every number has a source. 8. The load-bearing claim is supported at proof rung one or two. **What to fix first.** In this order, regardless of which score is lowest: - Anything creating legal exposure. An unsupported objective claim, an undisclosed connection in a testimonial, a guarantee you cannot honour. These are fixed today, by deletion if necessary. - Then the lowest scorer among dimensions four, seven and eight, because those three cost nothing but writing and carry the most weight. - Then scope and terms, which are cheap and prevent refunds rather than causing sales. - Then everything else. **Thresholds.** Below eight, the offer is the problem and copy edits are wasted effort. Between eight and twelve, fix the two lowest dimensions and ship. Above twelve, stop auditing the offer: your next finding is in traffic quality or in the audience, not on this page. **When you cannot tell.** If nobody can state what the buyer would do instead of buying, dimensions one, five and eight cannot be scored, and any score you produce is invented. Stop the audit and do the positioning work first, because the guarantee, the timeframe and the proof are all statements about a comparison, and you do not yet know what the comparison is. ## Worked example An invented case: a compliance training subscription sold to mid-sized manufacturers. **Scores.** Alternative acknowledged: 0, the page never mentions the in-house spreadsheet and annual half-day session that every prospect currently runs. Scope: 1, modules listed, no exclusions. Price and terms: 1, price shown, auto-renewal buried in the terms. Risk clause: 0, "satisfaction guaranteed" with no window, no adjudicator, no evidence requirement. Time to result: 0. Buyer's work: 0, and the product needs an administrator to upload a staff list, which nobody has said. Specificity: 1, two of six numbers sourced. Proof: 0, nine testimonials, first names and initials only. Total: three of sixteen. **Legal exposure first.** "Reduces incidents by up to 60 percent" is an objective claim with no evidence behind it. It comes off the page today. **Then the risk clause.** Rewritten with all five fields: a 60-day window because first certification lands at about week three, the remedy is a full refund minus nothing, the buyer submits their completion export, the head of customer operations decides within five working days, and a disagreement escalates to a named person with a final decision in ten. **Then proof.** Two of the nine testimonials belong to customers willing to be named and to take calls, moving the load-bearing claim from rung five to rung three, and a self-serve sandbox with sample staff data moves it to rung one. The other seven quotes stay, doing decoration. **Then the stack.** Six bonuses become three: the policy template pack, the annual regulator update briefing, and the administrator setup call, each with an external price. The two removed items were a checklist and a newsletter, neither of which anyone would have bought. **Verdict: hold the page.** The unsupported claim is a same-day deletion. The offer is not ready to test until the risk clause and proof rungs are rebuilt, and doing copy passes on the current page would improve the wording of an offer that has nothing behind it. ## Failure modes **A guarantee with no adjudicator.** Recognisable because the first awkward claim is decided by whoever happens to answer, and the second is decided differently. The visible symptom is card disputes rather than refund requests. **A window shorter than the time to first result.** Symptom: a cluster of refunds in the last three days of the window, from accounts with almost no usage. **Stacking past the ceiling.** Symptom: a bundle of eleven items where a buyer in a sales call asks what two of them actually are. **Reversing only the financial risk.** Symptom: buyers who say the price is fine and still do not proceed, because the risk they were carrying was the switching cost or their own credibility. **Numbers with no source.** Symptom: nobody in the company can say where the figure came from, and it has been copied onto three other pages in the meantime. **Proof entirely at the bottom rungs.** Symptom: a long testimonial wall and a competitor winning deals with one public benchmark. **The fake damaging admission.** Symptom: the concession is a boast in disguise, and the sentence immediately after it gets no lift at all. **Auditing the offer when the traffic is wrong.** Symptom: the page scores well, the audit finds little, and conversion is still poor, because the visitors were never in the segment. ## What this skill does not do - It does not write the sales page. It produces a score, an ordered fix list and a rebuilt risk clause. Turning that into a page is a separate job and a copywriter is better at it. - It does not substantiate claims, hold your evidence, or tell you whether your specific claim is lawful in your market. It flags which sentences carry the duty. - It cannot price the guarantee for you, because that needs your refund rate, delivery cost and margin, none of which it can see. - It does not test anything. Every judgement here is a prior, and a live audience overrules it. - It will not construct a risk reversal for a product that fails often, and it treats that as a product finding rather than a wording problem.
These skills all ask your assistant to check things against your actual codebase, your actual schema, your actual design system. Locul keeps that context current on its own, from the files you already have, on your machine. Mac and Windows, free to start.