AI ad copy fact check: prices, dates and exclusions

An AI ad copy fact check compares already written promotions with one approved offer, including the price, timing and exclusions. In fifteen minutes, you can create a fact card, inspect three fictional variants and record which claims must be held instead of asking a model to approve the campaign.

Suppose an independent stationery shop has a newsletter draft, a social caption and a short advertisement ready for review. The wording differs appropriately, but so do the promises. One version broadens the eligible products; another adds delivery. A polished tone makes those changes easy to miss.

This is a release check for an existing offer, not a guide to choosing prices, generating advertisements or obtaining legal clearance. Every business, price, date and advertisement in the worked exercise is fictional. Content updated October 15, 2026.

What belongs on an approved fact card?

A fact card is a compact record of the offer the responsible person has approved. Build it from that approval, not by summarising an advertisement you are supposed to check. Otherwise the original mistake can quietly become your reference. Give the card a version and keep that copy unchanged while reviewing.

The card does not establish a delivery charge. Missing information is not a zero charge. If the business wants another fulfilment option, the offer owner must approve that separately. For the different task of setting a price, use the AI pricing workflow. Here the price is an input, not a recommendation.

Add the approver, card version and review date. When the offer changes, retain the earlier version rather than replacing it silently. You need to know which promise a previous review actually checked.

Keep the three existing variants intact

Label the drafts A, B and C before making edits. These deliberately flawed examples are exercises, not copy-ready advertisements for your own business. Their differences make the expected check visible.

1. A: Every desk accessory is USD24. Offer ends November 6!
2. B: Desk set S1 is USD24 per set, November 3, 2026, 09:00 through November 5, 2026, 17:00 UTC. Shop collection only. Gift wrapping excluded.
3. C: Desk set S1 for USD24 with free delivery everywhere, while stocks last.

A expands eligibility beyond S1 and changes the end date. It also omits the start and precise ending time. B preserves the five fields on the supplied card. C introduces a delivery promise unsupported by the card and inconsistent with collection only. Its stock-based ending does not preserve the fixed period in F3.

A short format does not make missing conditions correct. If the intended placement cannot communicate the offer accurately, mark that version as needing work rather than simplifying the underlying business promise. The email writing templates cover the separate drafting task. This exercise starts after those drafts exist.

Ask the assistant to locate differences

You can use a text assistant to propose a comparison. No particular model, paid feature or file-upload capability is required by this method. Practise with the fictional card instead of an unreleased customer campaign. For the broader approach, the practical AI guide starts with a small result you can inspect yourself.

Open gemini.google.com in your browser. The fifteen-minute exercise assumes you are already signed in; registration or sign-in may take additional time. Paste the instruction below into the message field. Replace [V1] with the complete offer card above and [A, B, C] with all three complete drafts. Send the message, then look for exact wording, field references and issue statuses in the response.

Use only approved offer card V1 below. F1–F5 are fixed facts. Compare existing advertisements A, B and C separately. For each issue, quote the exact wording, identify the relevant field, state what the card says and assign a status: matches, contradicts, missing condition or unsupported addition. List a question for the offer owner when needed. Do not calculate a new price, generate new advertisements, broaden eligibility or approve publication. Card: [V1]. Drafts: [A, B, C].

An answer saying everything looks fine is not a completed comparison. Ask for exact expressions and field references. Google's explanation of Gemini responses notes that convincing answers may still be inaccurate. Treat the assistant's comparison as a proposed issue list, then inspect the card yourself.

Pay particular attention to small words with large effects: all, only, included, free and until. A sentence about a set must not silently become a price per individual item. If the model bundles several claims together, split the row so that a correct price does not hide a wrong deadline.

Record the differences, not just a verdict

Use a short matrix with the variant, exact expression, reference field, discrepancy, owner and next action. Another person should be able to understand why a sentence was held without replaying the chat. Here are three expected findings from the exercise; they do not exhaust every issue in A and C.

Unsupported does not mean the service could never exist. It means this approval does not support the promise. Keep the claim out of the release until the responsible person resolves it. Do not ask the model to fill the gap from assumptions about how shops usually operate.

The general document citation audit explains checking whether source passages support claims. Here the distinct job is comparing channel variants against a numbered offer card and tracking their release locations. You do not need PDF page references or an assistant-generated summary to perform it.

Invalidate old checks when the offer changes

Suppose the owner changes the end time and approves card V2. Draft B does not remain approved merely because it matched V1. Keep its earlier result and mark it as requiring review against V2. A change to the reference must reach every affected variant, including scheduled copies.

Create a placement register containing the variant label, channel, draft location, scheduled time, card version and responsible person. A draft location can be a document link or the identifier used in your publishing tool. Its purpose is to distinguish the file someone corrected from the older copy still waiting elsewhere.

Open the actual scheduled draft after the correction is transferred. Do not assume an edited document automatically updates a campaign tool. Where you cannot inspect the relevant copy, leave that placement unchecked. This step is a human handoff practice, not a claimed integration between applications.

Google Ads' unavailable-offers guidance says advertisements must not promise unavailable offers, and its example addresses updating or removing an expired promotion. That supports returning to released copy when the offer changes. This exercise does not certify Google Ads approval or compliance with every advertising requirement.

Finish the first pass with explicit actions

Use three minutes to freeze the card and label the drafts, five minutes for the proposed comparison, five for your own field-by-field review and two to assign corrections. This is a suggested exercise schedule, not a performance benchmark. If time runs out, record unchecked fields rather than treating them as matches.

For this sample, make sure the review catches A's eligibility and date changes plus C's delivery and timing claims. B's textual match does not prove that every placement is correct. An image may contain old wording, a caption may be truncated, or a scheduled copy may still use a previous version.

The acceptance checklist in the Firefly campaign workflow addresses the surrounding visual asset. Use that separately when inspecting the final advertisement. A correct background and a correct fact comparison solve different parts of the release.

In his essay about perfection versus effectiveness, Anar describes how pursuing perfection delayed his own projects. The application here is checking three concrete variants before building a larger campaign system. This is not a claim that he tested these advertisements or achieved a measured sales result using this method.

Keep a release record someone can follow

Copy the complete V1 card, the three drafts and the discrepancy rows you checked. On Windows, search for Notepad in Start, open a new document and paste them. Press Ctrl+S, choose Desktop and save as ad-check-V1.txt. Close the file, then double-click it on your desktop to reopen it. Confirm the card version, disputed wording, owner and next action are still present.

Retain the card version, checked wording, findings, corrections and final disposition together. Before release, open the exact version that will be published. A filename containing final is not proof that it is the reviewed text. Name the person who will revisit it if the offer changes.

The useful output today is not another advertisement. It is an approved-input card, a populated discrepancy matrix and explicit actions for the held variants. Keep the distinction between a proposed correction, a verified correction and a released placement visible in the record.