AI Marketing Best Practices for a Local Business: A Pre-Post Checklist

Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI Marketing Best Practices for a Local Business: A Pre-Post Checklist.
Coding Liquids tutorial cover featuring Sagnik Bhattacharya for AI Marketing Best Practices for a Local Business: A Pre-Post Checklist.

Before a local business publishes AI-written marketing, check twelve things: facts, claims, prices and dates, images, reviews and testimonials, consent for people and homes shown, disclosure, competitors, industry rules, links and contact details, originality, and who approved it. Log the result.

AI is fluent and confident, which is the problem: it writes a plausible price, an invented guarantee and a warm testimonial with equal ease. The list below was built by running each check on one realistic post, so you can see what each one catches.

Follow me on Instagram@sagnikteaches

The post we will check

An illustrative post from a small interior design studio. The owner asked an AI assistant for an Instagram caption and a matching website paragraph about a recent living room refresh, and attached a photo of the finished room and one line from a client message.

Connect on LinkedInSagnik Bhattacharya
Our award-winning team transformed this tired 1970s living room in just two weeks, for a total of $4,000. Our clients say it's the best design in the area and they'd never go anywhere else. Book your free consultation today, limited slots this month! #interiordesign #livingroom #beforeandafter

It reads well. It is also close to unpublishable. Each of the twelve checks below finds something in it, and the pattern of what goes wrong is the same in most trades.

Subscribe on YouTube@sagnikteaches

Group one: what the post says is true

Check 1. Every fact against a source you hold

Take each number, date, name and place in the post and tick it against a record: the invoice, the project folder, the price list. In the sample, "two weeks" and "$4,000" go against the project file. The invoice says four weeks from start to handover, and $4,000 was the design fee, not the total with furnishings. The post would mislead anyone budgeting from it.

The rule: the AI may not supply a figure. You supply the figure, from a record, in the prompt. Where the AI offers one you did not give it, delete it.

Check 2. Claims you could not prove in writing

Superlatives and guarantees are claims. "Award-winning" needs a named award you can point to. "Best design in the area" cannot be proved at all, and "never go anywhere else" is a quote nobody gave. Swap each for something checkable: "shortlisted for [named award] in [year]", "refreshed in four weeks", "our client's words: 'we use the room every day'".

Regulated trades need more care. A mortgage broker's post that says "lowest rates" or "approved in 24 hours", or a financial adviser's that implies returns, can breach advertising and conduct rules even when the idea is harmless. Use your normal compliance wording and route any post that mentions rates, returns or approval through whoever signs those off.

Check 3. Prices, offers and dates still valid on the day

"Limited slots this month" is fine only if slots are limited this month. Check the diary. An offer with an end date needs the end date in the post, and a reminder in your calendar to remove it. Posts live for years; offers do not.

Check 4. Comparisons and competitors

Comparing yourself to named competitors needs evidence you could show. Comparing to "the average studio" invites a challenge. If the AI wrote either, cut it unless you hold the proof.

Group two: the images, the people and the reviews

Check 5. Is the image what the caption says it is?

If the caption says "this living room", the picture must be that living room, as it was photographed, not an AI render or a polished stock image. AI-generated or heavily edited images are fine when labelled as such, but a "before and after" that is partly generated is a false claim of results. Where you generate illustrations, say so: "AI-generated concept, not a finished project".

For stock and AI images, check the licence or the tool's terms for commercial use. If you sell to customers in the EU, transparency rules on synthetic content are tightening, and realistic AI images of real-looking people are the area to watch. Look at the platform's labelling tools and check with an adviser what applies to your business.

Check 6. People, homes and children in the frame

A photo of a client's living room shows their home, sometimes their family photos, a street number or a window view. Get written permission to use the images, say how long you may use them, and look for mail, certificates and children in the shot before posting. A permission slip kept with the project file takes two minutes and answers any later query.

Check 7. Reviews and testimonials are real, whole and not bought

  • Real: never generate a review or a quote with AI. If the AI added "they'd never go anywhere else", that is an invented testimonial.
  • Whole: quote a customer's words exactly. You may trim with an ellipsis; you may not change the meaning. Ask permission to use their first name.
  • Not incentivised: Google's policies, as checked in September 2026, bar offering rewards for reviews, selectively asking only happy customers (review gating), and staff quotas for reviews. Keep review requests neutral.

The sample's only real source was one line from a client message. It said, "We love the room and use it every day." That is the quote to use, with the client's permission, in place of the invented line. For AI-drafted replies to reviews, should you let AI auto-post replies to Google reviews? covers the extra risk of letting them post unseen.

Group three: permission, disclosure and rules

Check 8. Personal data and consent

If the post names a customer, shows their face, or draws on their message, you hold their personal data and need their agreement for this use. Marketing emails and texts need consent you can show, plus a working way to stop them. Do not paste customers' private messages, enquiry details or contract terms into a public AI tool to get a caption; use a plan where your data is not used for training, or summarise without identifying anyone. For wording, see the compliance checklist for AI-written emails.

Check 9. Disclosure of what is paid, gifted or generated

If a supplier gave you the furniture, or a creator was paid to post, say so clearly, using the platform's own tools, not a hashtag buried at the end. If AI produced an image that looks like a real photograph, label it. There is rarely a rule against using AI, but plenty against letting people believe something false.

Check 10. Industry and platform rules

Each trade has its own. Health claims, financial promotions, legal advice and anything to do with children or alcohol are the common ones. Platforms add their own: for example, Instagram has capped hashtags at five per post since December 2025, so the sample's hashtags are fine but a list of fifteen would not be. When in doubt, read the platform's current policy page, not a third party's summary of it.

Group four: the finishing checks

Check 11. Links, contact details, hours and originality

Click every link. Read the phone number and address aloud against your own door sign. AI sometimes writes a plausible but wrong opening hour, or a booking link to a page that does not exist. For originality, paste a distinctive sentence into a search engine in quotes: assistants can echo phrases from other sites, and a post copied from a competitor looks bad even if nobody sues.

Check 12. Who approved it, and when

The last check produces a record. A one-line log per post takes seconds and answers "who said we could say that?" a year later.

DatePostWritten byChecked byClaims verified againstPermissions on fileResult
14 OctLiving room refresh, Instagram and siteOwner with AI draftStudio managerProject file P-0231, invoiceClient photo and quote consent, 12 OctPublished after edits 1, 2, 5, 7

The post after the twelve checks

This is the sample caption, rewritten with every unsupported claim removed or replaced with something checked:

A 1970s living room, refreshed over four weeks. The design fee for this project was $4,000; furnishings and fitting came to a further $6,200. Our client says: "We love the room and use it every day." Photo is the finished room, shared with permission. We have two consultation slots left in November. #interiordesign #livingroom #beforeandafter

It is shorter on adjectives and longer on things a reader can act on. It also happens to be a better advert, since a reader can now tell whether the budget fits theirs. Honest specifics usually outperform the generic praise the AI first produced, though I cannot give you a statistic for that; judge it from your own enquiries.

Using AI to run the first pass, and what it misses

A second AI pass can screen a draft before a human does. It catches the obvious, and it misses what only you know.

Review this draft social post for a local interior design studio. List, as bullets, any:
1. numbers, dates or prices (I will verify each),
2. superlatives, guarantees or awards,
3. statements about customers or quotes,
4. claims about images,
5. offers with a time limit,
6. anything that sounds like a promise.
Do not rewrite the post. Do not say whether it is acceptable.

DRAFT:
{draft}

Illustrative output for the sample post:

1. Numbers: two weeks; $4,000. 2. Superlatives: award-winning; best design in the area. 3. Customer statements: "clients say... never go anywhere else". 4. Images: none stated. 5. Time-limited: "limited slots this month". 6. Promise: "free consultation".

It found the claims. It did not find that "two weeks" was wrong, that "free consultation" is not an offer you actually make, or that the photo needed permission. Those need the project file and the owner's memory. Treat the AI pass as a way to build the verification list, not as the verification. Giving the AI your real facts up front, as described in how to give AI your business context, reduces the number of inventions to catch.

Three more posts, run through the list quickly

A signage company's before-and-after of a shopfront

The photo shows a customer's shop and its name. Check 6 asks for the owner's permission; check 2 queries "doubled their footfall" (no source, so delete); check 11 catches a phone number from the old brand. Result: publish after three edits.

A financial adviser's seasonal reminder about pension contributions

Checks 1 and 10 dominate. The AI quoted a contribution limit that might be out of date, and the post needs the firm's standard risk wording. Result: the post goes to the compliance reviewer, not straight out, and the limit is removed in favour of a link to the official source.

A mortgage broker's rate-watch email

The AI invented a "typical rate". Check 1 fails: no rate may appear unless it came from a lender's current published figure on the day. Check 3 flags the date, since the email will sit in inboxes. Result: rewrite without any rate, linking to a rates page you maintain.

The twelve checks on one page, with how to verify each

Copy this into a shared document and keep one copy per post. The third column is what makes it a check instead of a wish.

#CheckHow to verify itTypical time
1FactsTick each figure against the invoice, project file or price list3 minutes
2ClaimsCan you show proof in writing today? If not, remove or soften2 minutes
3Offers and datesCheck the diary; put an end date and a reminder to take it down1 minute
4ComparisonsKeep only those you hold evidence for1 minute
5ImagesIs it the real thing, an edit or generated? Label accordingly; check licence2 minutes
6People and homesWritten permission on file; scan for addresses, documents, children2 minutes
7ReviewsExact words, permission, no incentive, no generated quotes2 minutes
8Personal dataConsent for the use; nothing private pasted into a public tool2 minutes
9DisclosurePaid, gifted or generated content flagged using platform tools1 minute
10Industry and platform rulesRead the current policy page; route regulated wording to compliance2 to 15 minutes
11Links, details, originalityClick, read aloud, search a distinctive line in quotes3 minutes
12Approval logOne line: who, when, against what1 minute

Added together, that is roughly 20 minutes for a post with an image and a testimonial, and about ten for a plain text post. The time shrinks as you feed the AI better facts, because fewer inventions appear to catch.

When a check fails: fix, soften or drop

A failed check does not always mean a rewrite. Use a simple rule for each failure.

  • Fix when the fact exists and the AI got it wrong. Replace the figure with the right one from your record.
  • Soften when the idea is fair but the wording over-reaches. "Best in the area" becomes "a studio people return to", only if you can show returning clients.
  • Drop when you hold no evidence at all, or the permission is missing. An invented testimonial is always a drop.
  • Escalate when the post touches a regulated claim, a legal question or someone's complaint. The owner should not decide that on a Friday afternoon alone.

If a post needs more than three fixes, the prompt was the problem. Go back and add the facts you left out, rather than patching the text.

Prompting image tools so there is less to check

Most image problems begin with the request. Asking an AI image tool for "a beautiful modern living room" produces a believable photograph of a room that does not exist. If you then post it under "our latest project", you have made a false claim without writing one false word.

  • Use real photographs for any post about your work, results or people.
  • Use generated images for concept art or mood boards only, and caption them as such: "Concept image made with AI, not a finished room".
  • Avoid generating realistic people to stand in as customers or staff. A made-up face with a made-up name reads as a fake testimonial.
  • Keep the prompt and the tool name in your log, so you can say how the image was made.

For Google Merchant Center product images, Google requires AI-generated images to keep the IPTC "TrainedAlgorithmicMedia" metadata, so do not strip it when you resize or edit. That is a useful habit for any image you generate.

Where to keep the approval step so it gets used

The best place is wherever the draft already lives. If you draft in a shared document, use its comment and suggestion tools, and have the checker leave a comment saying "checked 1 to 12" with the initials. If you plan posts on a board, add a column called "Checked" that a post cannot leave until a second person moves it. For a longer discussion of review steps that do not slow the team down, see human review for AI work without slowing down.

Keeping the habit going

A checklist you do not use is decoration. Three things help it stick.

  • Put it where the post is written. Paste the twelve lines at the end of your drafting document, and tick as you go.
  • Split the load. The writer runs checks 1 to 4 and 11; a second person runs 5 to 10 and 12.
  • Review the log monthly. Which checks catch the most? If check 1 fails every week, your AI prompt is missing facts; add them to a standing business-context note.

If your team has several people using AI for marketing, a short written policy helps; an acceptable-use policy for a professional firm gives a model. And if you want the checks built into how your posts are actually drafted and approved, that is something my AI implementation consultation can map with you.

Pre-post checking questions

Who should sign off a post, the owner or whoever wrote it?

Someone other than whoever prompted the AI, where you can manage it. A second pair of eyes catches the claim the first person stopped seeing. In a one-person business, wait an hour or read the post aloud before you approve it.

Do I have to say a post was written by AI?

There is no single rule for text. Disclosure matters most for images or video that look like real photographs of real people or real work, and for fake-looking testimonials, which you should never produce. Check the platform's labelling rules and the law where your customers live, and ask a qualified adviser if unsure.

Can I use AI to write customer reviews for my page?

No. Invented reviews are deceptive and breach the review policies of the major platforms. You can use AI to draft a request asking real customers for an honest review, and to help reply to reviews you have received.

How long should this check take?

About ten minutes for a routine post once the list is familiar. Posts with prices, comparisons, before and after photos or testimonials need longer. A regulated trade such as lending or advice needs your usual compliance step on top.

Further reads

Sources: Google's own review and contributed-content policies, platform advertising and labelling guidance, and your own industry body or adviser for regulated claims.

Want a checking routine your team will actually use?

On a 1:1 call we'll look at how your posts get written and approved today, build a short checklist to fit, and decide which checks a tool can do for you.

Book a 1:1 call with me