Girl, You Know It's Generated: AI for Marketing Without Sounding Like a Robot

Law Smith

Founder & President, Tocobaga

Law Smith is the Founder and President of Tocobaga, a Tampa-based ROI-focused marketing agency and SMB advisory. He has strategically advised 1,000+ small and medium businesses, executed 600+ integrated marketing campaigns, and written 30+ business plans across industries ranging from professional services to e-commerce to B2B technology. Tocobaga is a Google Partner and Squarespace Gold Partner headquartered in Ybor City.

On the night of July 21, 1989, two men in bicycle shorts and shoulder-length braids were doing the thing they did better than anyone alive, which was dancing in front of a tape machine. The venue was Lake Compounce, an amusement park in Bristol, Connecticut, on a stop of the Club MTV tour. Rob Pilatus and Fab Morvan, the two faces of Milli Vanilli, had their signature single cued up, the one that had reached number 2 on the Billboard Hot 100 earlier that year. The track started. The crowd screamed. A few bars in, the prerecorded vocal caught on itself and began repeating the same half-line, over and over, while the two men onstage kept miming to a voice that had stopped being a song and become a loop.*

They kept dancing, then ran off the stage. Here is the detail that matters for every business owner currently pasting a ChatGPT draft into a blog post: the crowd did not care. The audience assumed the glitch was part of the show, the story barely made the news, and Milli Vanilli won the Grammy for Best New Artist the following February, with an album that sat at number one on the Billboard 200 for eight weeks and went six-times platinum.* The product worked. The fraud only became a scandal on November 14, 1990, when producer Frank Farian stood up at a press conference in Munich, fired both men and admitted they had not sung a note on the record. Five days later the Recording Academy took the Grammy back, the only time it has ever done so.*

The actual singers were session vocalists named Brad Howell, John Davis and Charles Shaw, plus twin sisters Jodie and Linda Rocco, none of whom Farian put on the album cover and all of whom could sing.* Farian had built a perfectly good product in the studio, attached two attractive, confident, visibly lip-syncing front men to it, and the whole thing held together until the day the tape skipped in public and the day the producer told the truth.

That is the entire problem with AI in marketing, compressed into one 1989 amusement-park set. The technology is not the fraud. A drum machine never got anyone's Grammy revoked. The fraud is a business standing in front of a customer, miming to a voice it did not produce, hoping the tape does not skip. Law Smith, founder of Tocobaga, a lead generation marketing agency and SMB advisory headquartered in Ybor City, Tampa, has watched the Lake Compounce moment arrive in client inboxes for three years now, and the firm's site, SolvingHow.com, gets some version of the same email every week: the AI draft looked great, the AI draft went out, and now a client, a judge, a customer or a Google reviewer has pointed at the loop. Sometimes the tape skips in a blog post that opens with a paragraph about the pace of the modern world. Sometimes it skips in a federal courtroom, in a brief citing six cases that do not exist. The question for a two-partner law firm, a PE-backed HVAC platform with eleven branches, or a Shopify store with one owner and a dog is not whether to use the machine. Most of them already do. The question is who is singing.

Tocobaga is a Google Partner and a Squarespace Gold Partner, and the answer the firm gives across 1,000+ advisory conversations comes with the numbers attached and the Milli Vanilli rule stated up front. Let the machine play the drums. Never let it claim to be the singer.

How many small businesses are actually using AI for marketing?

About six in ten. The U.S. Chamber of Commerce's 2025 Empowering Small Business report, a national survey conducted with Teneo Research, found 58 percent of small businesses using generative AI, up from 40 percent in 2024 and 23 percent in 2023.†

Read that curve again. Adoption more than doubled in two years, and it happened without a sales force, a trade show or a budget line. The tool arrived free, in a browser tab, and the owner who tried it on one email at 11 p.m. was using it for the newsletter by Friday. Surveys that poll marketers rather than owners run higher still, and the exact figure depends on what counts as "using"; a once-a-month proofread and a fully automated blog are both "use". The direction is not in dispute. The majority of competitors on any given street are now drafting with a machine.

For a two-partner law firm, that means the firm down the hall is producing three times the content it did in 2023, and some of it is fine. For an HVAC platform with eleven branches and a private equity sponsor, it means the regional marketing manager quietly stopped writing from scratch a year ago, and nobody at the sponsor has asked what the process is. Neither is wrong to use the tool. The Milli Vanilli record was a hit for a reason. The risk is not the studio. The risk is what gets claimed about who sang.

Pew Research Center set the mood on the other side of the counter in a September 2025 survey of 5,023 U.S. adults: 50 percent said they are more concerned than excited about AI in daily life, up from 37 percent in 2021, and 76 percent said it is extremely or very important to be able to tell whether a picture, a video or a piece of text was made by a person or a machine.‡ Most do not trust their own ear, either: 53 percent said they are not too confident or not at all confident they could tell the difference.‡

That is the market in one sentence. Six in ten sellers use the machine, three in four buyers want to know when it was used, and half the buyers doubt they could tell. The whole game lives in that gap, and the businesses that win it are the ones who never make a customer feel like the Milli Vanilli fans felt the week after the Munich press conference.

What are the best ways for a small business to use AI in marketing?

The best uses are the ones where the machine does the volume and a human does the truth. In practice that means ad copy variations, first drafts, review responses, email sequence outlines, call notes turned into proposals, research summaries and repurposing one finished piece into five.

Ad copy is the cleanest case. A Google responsive search ad wants up to 15 headlines and four descriptions, and the honest version of that job is writing 40 candidates and keeping the eight that are true. A machine writes 40 in a minute. The HVAC platform with eleven branches needs the same tune-up offer written for Brandon, Lakeland and Wesley Chapel, and the branch manager in Lakeland is the only person on earth who knows the Lakeland version should mention the 1970s ranch houses with the original ductwork. The machine does the eleven. The human does the one true sentence per city, and that is the sentence that gets the click.

First drafts are where the time goes, because the blank page is the expensive part of writing and the machine does not experience blank pages. A manufacturer's marketing coordinator who owes a 1,200-word explainer on powder coating versus wet paint can have a structurally sound, factually unreliable draft in 90 seconds, then spend the hour that used to go to staring at a cursor on the part that matters: the two jobs last quarter where the customer chose wrong and what it cost them.

Review responses are the use case most owners underrate. A dental practice or a family law firm collecting 30 Google reviews a month can have 30 on-voice draft responses in the time it takes to read them, and the office manager's job becomes adding the one detail per response that proves a human read it: the hygienist's name, the parking situation that got fixed. The one-star reviews get written by a human from the first word. Nobody wants a sympathetic paragraph about a billing dispute from a machine that has never been billed.

Email sequences, proposals and research follow the same rule. A seven-email welcome sequence can be outlined in a minute, and the owner writes the one email that matters, the one with the price. A consultant can drop a 45-minute scoping call transcript into a model and get a proposal skeleton with every number flagged for confirmation. Tocobaga has deployed 1,000,000+ emails, and the rule that holds across all of them is that the structure can come from anywhere, but the subject line and the offer are written by a person who will have to answer for them.

Repurposing pays the most and risks the least. One 3,000-word post becomes a newsletter, five LinkedIn posts and a 30-second video script, and because every fact in the derivative work already lived in the original, the checking is cheap. The machine is excellent at changing the shape of something true. It is dangerous at inventing something new. That distinction, shape versus substance, is the whole job description.

Does AI-generated content actually save time?

Yes, measurably, on the tasks it is good at, and the honest number is smaller than the demo. In a controlled experiment published in Science in 2023, MIT researchers Shakked Noy and Whitney Zhang gave 453 college-educated professionals, marketers and grant writers among them, occupation-specific writing tasks; the group with ChatGPT finished 40 percent faster and independent graders rated their work 18 percent higher.§

Two details from that study matter more than the headline. First, the weaker writers gained the most and the gap between the best and worst participants narrowed, which means the tool is a floor-raiser, not a ceiling-raiser: a business that already writes well gets speed, and a business that writes badly gets competence. Second, the authors warned that real-world gains would be smaller, because in the real world someone has to write the prompt and check the output, and in the experiment nobody was checking anything.§ The 40 percent is the studio time saved. It is not the time saved on the whole album.

The larger study came out of Harvard Business School and Boston Consulting Group the same autumn. Fabrizio Dell'Acqua, Ethan Mollick, Karim Lakhani and colleagues ran 758 BCG consultants through a set of realistic tasks with and without GPT-4. On the tasks the model was good at, consultants with AI completed 12.2 percent more tasks, finished 25.1 percent faster and produced work of significantly higher quality.‖ On a task designed to sit just outside the model's competence, the consultants using AI were 19 percent less likely to produce a correct answer than the consultants working alone.‖

The authors called that boundary the jagged technological frontier, and "jagged" is the word to keep. The tasks on either side of the line look identical from the outside. A product description and a product spec look like the same job. A consultant with an MBA could not tell by looking which side of the line a task sat on, and neither can an owner with a plumbing license. So the time savings are real, and the time savings are also the trap, because the hour the machine gives back has to be partly reinvested in checking, and the business that pockets the whole hour is the one that ends up filing six fake cases.

What happened when a lawyer used ChatGPT for a court filing?

He got sanctioned, his firm got sanctioned, and the case became the cautionary tale every bar association now uses. In Mata v. Avianca, a personal injury suit in the Southern District of New York, attorney Steven Schwartz used ChatGPT to research a brief, and the brief cited six judicial opinions that did not exist.¶

The fake cases had real-sounding names. Varghese v. China Southern Airlines. Shaboon v. Egyptair. Estate of Durden v. KLM Royal Dutch Airlines. They came with reporter citations, docket numbers and quoted holdings, and the quoted holdings were exactly what Schwartz's client needed them to say.¶ When opposing counsel could not find the cases, Schwartz asked ChatGPT whether they were real, and the model assured him they were. His partner Peter LoDuca signed the filing without reading a single cited case. When the court asked for copies of the opinions, the lawyers produced them, and the "opinions" were themselves generated by the chatbot.¶

Judge P. Kevin Castel's June 22, 2023 sanctions order fined Schwartz, LoDuca and their firm, Levidow, Levidow and Oberman, $5,000 jointly and severally, and ordered them to mail letters, with the order attached, to their client and to every real judge the chatbot had falsely named as an author of a fake opinion.¶ Imagine the tone of those letters. Castel also wrote the sentence every firm should tape to the monitor: there is "nothing inherently improper about using a reliable artificial intelligence tool for assistance."¶ The sanction was not for using the machine. The sanction was for doubling down after the tape skipped. Castel found the lawyers had been told the cases looked fake and kept standing by them instead of checking, which is the legal-system version of continuing to dance after the vocal has started looping.

A two-partner firm should notice the structure of the failure, because it is a marketing failure as much as a legal one. Nobody had decided in advance which parts of the work the machine was allowed to invent and which parts had to be verified by a human with a Westlaw login, so the humans treated looking true as the same thing as being true.

The same thing happens on a law firm's website every day with lower stakes and no judge. A practice-area page drafted by a model will confidently state a filing deadline, a damages cap or a statute number, and on the hardest kind of question it will usually be wrong. Stanford RegLab researchers tested more than 200,000 legal queries against three leading models in a study published in January 2024 and found hallucination rates between 69 percent and 88 percent on specific, verifiable questions about federal court cases, with the models often overconfident about answers that were wrong.** Those were the models of two years ago, and the current ones are better. Better is not a standard a state bar accepts. A firm's marketing is held to the same truth standard as its briefs, by a different and larger jury.

Will Google penalize AI-generated content?

No, not for being AI-generated, and yes, if it is generated at scale to game rankings. Google's February 2023 guidance says Search rewards high-quality content however it is produced, and that using automation, including AI, with the primary purpose of manipulating rankings violates its spam policies.††

The 2023 post refuses to do the thing owners expect. It does not ban AI content and it does not require an AI label. What it does is restate the standard Google has used for a decade, now abbreviated E-E-A-T: experience, expertise, authoritativeness and trustworthiness, with the first E, experience, added in December 2022, a few weeks after ChatGPT launched.†† That timing was not an accident. First-hand experience of having actually done the thing is the one ingredient a language model cannot supply, and it is precisely the ingredient Google decided to weight more heavily the moment the supply of competent generic text became infinite.

Then came March 2024. Google announced a core update and three new spam policies, and the one that matters here is scaled content abuse, defined as generating many pages for the primary purpose of manipulating rankings rather than helping people, no matter whether automation, humans or a combination produced them.‡‡ Google said the combined changes were expected to cut low-quality, unoriginal content in search results by about 40 percent.‡‡ The word "combination" does the work in that definition. It closes the loophole where a business generates 300 city pages with a model and has an intern change a sentence in each. Google is not asking how the pages were made. Google is asking why. A page that exists because a real customer asked a real question has a why. A page that exists because a spreadsheet had 300 rows does not.

For a professional services firm, the practical reading is almost relaxing. Twelve pieces a year, each answering a question a real client asked and containing something only that firm could know, have nothing to fear from the policy, whether the first draft came from a partner or a prompt. A firm that bought a tool promising 500 "SEO-optimized" articles by Friday has already built the thing the policy was written to find. The tape will skip. It just skips in a ranking report instead of at an amusement park.

Can customers tell when marketing is written by AI?

Mostly no, and that is the uncomfortable part. When the source is hidden, people often prefer the machine's copy; when the source is revealed, they trust the business that used it less.

Start with the blind taste test. MIT Sloan researchers Yunhao Zhang and Renee Gosline showed people marketing and persuasive copy produced by professionals, by GPT-3, and by combinations of the two, without saying which was which. The participants preferred the AI-written content.§§ When the researchers then disclosed the source, ratings for content with human involvement went up, a pattern the authors named human favoritism, while ratings for the AI content did not fall.§§ In the authors' summary, consumers really do not mind machine-produced content; what they value is the knowledge that a person was involved.

That finding looks like permission until it is placed next to the second study. Oliver Schilke and Martin Reimann at the University of Arizona ran 13 experiments, published in Organizational Behavior and Human Decision Processes in 2025, in which people who disclosed using AI for a piece of work were trusted less than people who did not disclose, across every task and evaluator group studied, and across every softer wording of the disclosure the researchers tried.‖‖ The penalty shrank, but did not vanish, among evaluators who already liked technology. And there was one outcome worse than disclosing: getting exposed by someone else, for instance by a third party running an AI detector, which the authors describe as the worst case for trust.‖‖

Put the three findings in a row and the strategy writes itself. People cannot reliably spot machine copy. People like machine copy fine when it is good. People punish a business for admitting it used the machine, and punish it harder for being caught. So a blanket "this content was created with AI" footer on every page is a self-inflicted wound with no upside, and a hidden content farm is a time bomb with a detector attached. The way out is to make the disclosure question moot by ensuring that everything with a byline, a price or a claim in it was produced, or at minimum verified and rewritten, by the person whose name is on it.

The useful middle is a standard, not a label: a one-paragraph content standard on the About page saying, truthfully, that every piece carries the name of the person responsible for it, that every figure is checked against a source, that software assists drafting the way a word processor assists typing, and that nothing with a price, a legal claim or a client's name in it goes out without a named human's sign-off. That paragraph answers the question the customer actually has, which is not "did a machine touch this" but "is anyone home." Where a state bar rule, a regulator or a platform requires disclosure, disclose fully and first; Google's own guidance treats disclosure as appropriate where a reader would reasonably wonder, and treats the author byline and the first-hand experience behind it as the signals that matter.††

Milli Vanilli's problem was not that Farian used studio singers. Every pop act of 1989 used studio singers. The problem was that the men on the cover claimed to be the voice. A business that lets the machine draft and lets a human own the claim never has to answer the disclosure question, because the honest answer is "a person wrote this, with help," which is how every book, brief and ad has been produced since the invention of the editor.

Why does AI writing sound generic?

Because it is a weighted average of everything ever written on the subject, and the average of a million blog posts is beige. A model predicts the most probable next word, and the most probable next word is, by definition, the one everyone else already used.

The tells are now so consistent that scientists can count them. Dmitry Kobak of the University of Tubingen and colleagues analyzed about 15 million biomedical abstracts published between 2010 and 2024 and tracked words whose frequency jumped abruptly after late 2022. The spike words included "showcasing," "underscores" and "potential," along with a certain three-syllable verb for digging into a topic that now turns up in seemingly every AI draft, and the excess vocabulary implied that as many as 13.5 percent of 2024 abstracts had been processed with a language model.¶¶ Those are scientists, writing about their own experiments, for other scientists, and as many as one in eight of them let the machine's vocabulary through unedited. The odds that a roofing company's blog is doing better are not good.

The vocabulary is only the loudest tell. The others are structural: the paragraph that opens by restating the question, the rule of three applied to lists that naturally have two items or five, the closing paragraph that begins with "ultimately" and says nothing, perfect grammar with no opinion and enthusiasm with no stakes. A reader may not be able to name a single one of these on a quiz, and still feels every one of them as a faint sense that nobody is home.

That feeling is the Voight-Kampff test from Blade Runner, running in the back of every customer's head. In the 1982 film, Rick Deckard identifies replicants, artificial people built to be more convincing than the real thing, by asking a series of emotional questions and watching the involuntary response, and a normal replicant gives itself away in 20 or 30 questions.*** The machine's answers are not wrong. They are too smooth. They lack the one thing a person cannot fake, which is having been somewhere. Generic copy fails the Voight-Kampff test in two or three questions, because it has never been anywhere. The specific Tuesday, the client who chose wrong, the price that went up, the branch in Lakeland with the ductwork: those are the involuntary responses. A machine can be told about them. It cannot have them.

How does a business keep its brand voice when using AI?

By writing the voice down, and by never letting a model guess at it. A voice guide of one page, fed into every prompt and enforced in every edit, is the difference between a draft that sounds like the business and a draft that sounds like the internet.

The guide does not need to be a brand book. The one Tocobaga builds with clients, refined across 500+ brands and rebrands, fits on a single page and answers six questions. Who is talking: a named person, with a title, not "the team." Who is listening: a 44-year-old operations manager at a mid-size manufacturer, not "decision makers." What the business believes that competitors do not: three sentences, stated as opinions, because a voice without an opinion is a weather report. What the business would never say: a list of ten banned words and phrases, starting with the vocabulary in the Tubingen study and adding the industry's own cliches. How the sentences move: short declaratives most of the time, one long sentence when the idea needs it. Three sample paragraphs, written by the owner on a good day, so the model has something to imitate besides the internet.

Then the guide goes into the prompt, every time, in full. A model told "write a blog post about estate planning" produces the average of every estate planning post ever indexed. A model given the one-page guide, two real client questions from last month, the partner's actual opinion about revocable trusts and a banned-word list produces a draft that is at least in the right key. Not finished. In the right key. The remaining distance is closed by the edit, which is where the voice actually lives, because the owner adds the thing that happened on the specific Tuesday, and the model cannot.

A PE sponsor with eleven brands under one platform should treat the voice guide as a portfolio asset. Each brand gets a page, the pages live in a folder the sponsor owns rather than in a freelancer's chat history, and when a brand is sold, the page goes with it. A regional 3PL with four salespeople who each "do their own LinkedIn" has four voices and no brand, and the fix is not to silence the four. It is to give them one page that explains what the company believes about freight, so that four human voices argue for the same thing.

What should a business never let AI write?

Anything that has to be true and that the business will be held to: prices, legal and regulatory claims, guarantees, statistics, citations, testimonials and anything with a customer's name in it. The machine drafts around those parts. A human writes them, from a source, and signs them.

The list sounds obvious until an owner audits last month's output against it. The Shopify store's product page says "ships in two days" because the model inferred it from a competitor. The HVAC platform's landing page promises "licensed and insured in all 50 states" because that is the most probable phrase to follow "licensed and insured," and the platform operates in three. The law firm's blog post states the statute of limitations for a claim, and the number is right for the state next door. None of these were lies in the Munich-press-conference sense. All of them were lip-syncs: a confident performance of a fact nobody in the building had checked.

Prices and guarantees are the obvious category because a wrong one is a contract. Legal and regulatory language is next, and the Avianca lawyers are the reason. Statistics and citations are the quiet killers, because a model asked for a supporting number will produce one, with a source, and the source will be real about as often as the Varghese case was. The rule Tocobaga uses across 600+ campaigns is that every number in a client's marketing traces to a document a human has opened, and every citation is clicked before it is published. Not most. Every.

Testimonials and anything in a customer's voice belong to the customer. A model asked to "clean up" a review will smooth out the grammar and, with it, the proof that a person wrote it, which is the only reason the review was worth anything.

The useful mental model is the one Wolfgang von Kempelen built in 1770. The Mechanical Turk was a chess-playing automaton, a turbaned figure at a cabinet, presented to Empress Maria Theresa in Vienna and exhibited for 84 years as a machine. It beat Benjamin Franklin in Paris and Napoleon at Schonbrunn Palace in 1809, and it kept its secret until a fire in Philadelphia destroyed it on July 5, 1854.††† The secret was that a human chess master sat inside the cabinet, following the game through magnets under the board and moving the arm with levers.††† Everyone marveled at the machine. The machine was never playing. For a business using AI in marketing, the Turk is the inverted lesson: put the human inside the cabinet on purpose, on every move that has to be right, and be honest that the cabinet is a cabinet. The automaton gets the applause. The person inside takes responsibility for the chess.

What is a simple workflow for using AI in marketing?

Five steps, in order, with a human at the start and the end: brief, draft, fact-check, humanize, approve. The machine owns the middle. People own the ends.

Step one is the brief, written by a human, and it is where most of the quality is decided. A usable brief names the audience, the one question the piece answers, the one opinion the business holds, three facts the business knows first hand (a client situation, a number from the firm's own data, a thing that went wrong last quarter), the sources that may be cited, and the voice guide. A brief with no first-hand fact in it produces a draft with no reason to exist.

Step two is the draft, and it is the machine's job. The brief goes in whole, and the model produces a structure, a first pass at every section, and a flagged list of every factual claim it made. Asking for the flagged list is the one prompt habit that matters most, because it turns the fact-check from a search into a checklist.

Step three is the fact-check, and it is a human's job, and it is not optional. Every flagged claim gets a source a human has opened. Every citation is clicked. Anything that cannot be verified is cut, not softened; softening a fake fact with the word "approximately" is how the Varghese case got a page number. The draft at the end of step three is shorter than the draft at the end of step two, and that is the sign it was done.

Step four is humanizing, which is a bad word for a real job. The one person who knows the subject reads the draft aloud, strips the tells, and adds what the model could not know: the specific client, the specific Tuesday, the price, the opinion, the thing the competitor gets wrong. If a competitor could have published the same piece, it is not done.

Step five is approval by someone whose name is on it. The partner, the owner, the brand manager at the platform. Not the intern who ran the prompt. The signature is the whole point, because the signature is what Rob Pilatus and Fab Morvan put on an album they did not sing, and it is what Judge Castel sanctioned. The person who signs is the person who will answer for it, so the person who signs must be the person who read it.

Written as a checklist an owner can screenshot:

• Brief: audience, question, opinion, three first-hand facts, allowed sources, voice guide.

• Draft: full brief in the prompt; ask for a flagged list of every factual claim.

• Fact-check: every claim to a source a human opened; cut what cannot be verified.

• Humanize: read aloud, strip the tells, add the first-hand specifics only the business has.

• Approve: signed by the person whose name is on it, who has read every word.

Milli Vanilli had steps one and two. Farian wrote a brief and the studio produced a draft whose title single sold seven million copies worldwide.* The record had no step five. Nobody whose face was on the cover had done the work, and the business plan depended on the tape never skipping.

How much time does the AI marketing workflow actually save?

Roughly half, once the fact-check and the edit are counted, and that half is worth real money. The honest arithmetic for a small firm goes like this.

Take a two-partner estate planning firm that wants to publish four pieces a month: two blog posts answering real client questions, one newsletter and one set of eight LinkedIn posts. Done entirely by a partner, at the pace most lawyers write, that is about 14 hours a month: five hours per post, two for the newsletter and two for the social set. At a $350 billing rate, the firm is spending $4,900 of partner time a month on marketing content, or $58,800 a year, which is why the firm's blog has three posts on it, the newest from 2023.

Run the same calendar through the five-step workflow. The brief is partner work, 15 minutes per piece, because the partner holds the opinion and the first-hand facts: one hour. The draft is paralegal work, ten minutes per piece to run the prompt and clean up the output. The fact-check is paralegal work too, 45 minutes per blog post, because statutes and deadlines have to be opened and read, and 15 minutes each for the newsletter and the social set, which repurpose the posts: two hours. The humanizing edit is the partner's job, 45 minutes per blog post and 15 minutes for each of the other two pieces, plus a final sign-off read of a few minutes each: two and a half hours. Add 50 minutes of paralegal formatting and scheduling, and the total comes to about seven hours a month, three and a half of them partner time and three and a half paralegal time at $75 an hour.

The new monthly cost is about $1,225 in partner time plus about $265 in paralegal time, or $1,490, against $4,900. The firm saves roughly $3,400 a month, about $41,000 a year, and actually publishes, which is the real gain, because a blog with 48 posts answering real questions produces leads and a blog with three does not. The Noy and Zhang 40 percent and the BCG 25 percent are the lab numbers.§‖ The firm's number lands in the same neighborhood once checking is paid for, and the checking is what keeps the firm out of Judge Castel's mailbox.

Take the same math to the eleven-branch HVAC platform, where eleven localized landing pages, 22 seasonal email campaigns a year and a steady flow of review responses used to need a full-time coordinator plus an agency retainer. With the workflow, one coordinator and one accountable brand manager handle it, with each branch manager contributing the one true local sentence per page. The savings there are a headcount and a retainer, and a PE sponsor reading a quality-of-earnings report knows exactly how that flows to the multiple.

How can a business tell if its marketing sounds like a robot?

Run the Voight-Kampff test on it before a customer does. Read the piece aloud, hunt for the eight tells, and ask the one question that matters: could a competitor have published this word for word?

The eight tells are quick to check. One: the opening restates the question or the title. Two: a sentence begins with "In today's" or a phrase like it. Three: two or more "whether it's X, Y or Z" triplets. Four: any of the spike words the Tubingen team counted, "showcasing," "underscores," "potential," plus the usual business suspects.¶¶ Five: a closing paragraph that begins with "ultimately" or "overall" and adds nothing. Six: no first-person detail, no named person, no date, no price, no place. Seven: no opinion a competitor would disagree with. Eight: perfect grammar with zero friction, which real writers never produce because real writers have a point of view and a deadline.

Zero or one tell and the piece is probably fine. Three or more and it fails the test, and it fails the way the Blade Runner replicants failed: not by being wrong, but by having no involuntary responses.*** The fix is never to run it through a second tool that promises to "humanize" it. The fix is to send it back to the one person in the building who has actually done the thing the piece is about and ask for two paragraphs only that person could write.

The same audit works on a competitor's site, and a Tampa law firm or a Florida manufacturer that reads its three nearest competitors' most recent posts aloud will usually find that two of the three fail, which means the bar for sounding like a person who has been somewhere is, for the moment, on the floor. That is an opening. It will not stay open.

Tocobaga runs this audit at the start of most client relationships, and after 300+ websites, the pattern is reliable: the pages that convert are the ones with the owner's own sentences in them, and the pages that do not are the ones where every sentence could have come from anywhere. The firm's fee for the work is flat and does not scale with a client's ad spend, the contract does not auto-renew, and every deliverable, account and asset belongs to the client from the first day, with no agency name at the bottom of the site. Tocobaga is also not for everyone, and when a business is better served by a different kind of shop, the firm says so and names the shop. For an owner who wants to know which side of the test the current site sits on, a 20-minute fit call at (813) 534-1261 or hello@solvinghow.com, or through SolvingHow.com, is usually enough to find out.

What is the one rule for using AI in marketing without losing trust?

Let the machine play the instruments and never let it claim to be the singer. Every piece a business publishes needs a human who did the work, checked the facts and signed it, and the machine's job is to make that human faster, not absent.

The economics make the rule easier to follow than it sounds. The first draft of a 1,200-word article now costs pennies and 90 seconds, so any agency still charging the 2022 price for the 2022 product is charging for a studio singer and sending two dancers to the gig. What an owner should pay for now is the part the machine cannot do and the part the machine makes more dangerous: the brief, the fact-check, the voice, the first-hand specifics and the signature. A PE-backed platform evaluating an agency across eleven brands should ask who owns the voice guides and the prompts, whether every statistic traces to a source, and whose name is on the work. If the honest answer to the last question is "the agency's," the platform is Farian, the agency is two dancers, and the tape will eventually skip in front of a buyer running diligence of their own.

Return to Bristol, Connecticut, for the last time. The tape skipped on July 21, 1989, and the crowd forgave it instantly, because a crowd at an amusement park is not running a Voight-Kampff test on a pop act. What the crowd did not forgive, 16 months later, was being told the truth by a producer in Munich.* The refund program that followed was open to an estimated 10 million buyers, and by one account about 50 people bothered to ask for their money back, which tells the whole story about what the fraud actually cost.* The buyers did not want the money. They wanted the album they thought they had bought, and they could not have it, because that album never existed. The music was real. The claim was fake. The claim was the product.

A customer who hires a two-partner firm, buys from a Shopify store or signs a service agreement with an HVAC platform is buying the same thing the Milli Vanilli fans were: a claim, made by a person, that the person stands behind. AI changes the cost of producing the sentences. It does not change what the sentences are for. A business that uses the machine for the drumming, the arrangement, the forty headline variations and the first draft of the newsletter, and keeps a person on the vocal, the price, the statute and the signature, gets the speed without the Munich press conference.

The tape is going to skip. It skips for everyone. The only question a business controls is what it was claiming when the loop started.

How can a small business use AI for marketing without sounding like a robot?

Use the machine for volume and a person for truth. Let AI draft ad variations, first drafts, review responses, email outlines and repurposed content, then have a named human fact-check every claim, strip the generic tells, add first-hand specifics only the business knows, and sign the piece. A one-page written voice guide fed into every prompt keeps the draft in the right key. Anything with a price, legal claim, statistic or customer name is written and verified by a person.

Does Google penalize AI-generated content?

Not for being AI-generated. Google's February 2023 guidance says Search rewards high-quality content however it is produced, and judges it on experience, expertise, authoritativeness and trustworthiness. What Google does penalize, under the March 2024 scaled content abuse policy, is generating many pages primarily to manipulate rankings, whether automation, humans or a combination produced them. A business publishing useful, first-hand, verified content is fine regardless of which tool typed the first draft.

What should a business never let AI write?

Prices, guarantees, legal and regulatory claims, statistics, citations, testimonials and anything containing a customer's name. Language models produce confident, plausible text, and on verifiable facts they are frequently wrong; a Stanford study found hallucination rates of 69 to 88 percent on specific legal questions. A human writes those parts from a source and signs them. The machine can draft the structure around them.

Should a business tell customers it uses AI for marketing?

Disclose where a professional rule, a regulator or a platform requires it, and where a reasonable customer would want to know. Research from the University of Arizona across 13 experiments found that disclosing AI use lowers trust, and being exposed by a third party lowers it more. The safer approach is a published content standard: every piece carries a responsible person's name, every figure is checked against a source, and software assists drafting the way a word processor assists typing.

How much time does AI save on marketing content?

In controlled studies, 25 to 40 percent on writing tasks the model handles well, and the authors warn real-world gains are smaller once prompting and fact-checking are included. A practical estimate for a small firm is about half. A two-partner law firm spending 14 partner hours a month on content can bring that to roughly seven mixed hours, saving on the order of $3,400 a month, while actually publishing consistently, which is where the leads come from.

What happened in the ChatGPT fake cases lawsuit?

In Mata v. Avianca, a New York attorney used ChatGPT for legal research and filed a brief citing six judicial opinions that did not exist, complete with fake quotes and citations. When challenged, the lawyers doubled down instead of checking. On June 22, 2023, Judge P. Kevin Castel sanctioned the two attorneys and their firm $5,000 and ordered them to notify their client and every judge falsely named as an author. The judge noted that using a reliable AI tool is not itself improper; failing to verify is.

Sources

*

Wikipedia. "Milli Vanilli." Accessed October 14, 2026. The July 21, 1989 Lake Compounce performance in Bristol, Connecticut; the single's number 2 Billboard Hot 100 peak; eight weeks at number one on the Billboard 200 and six-times platinum certification; Frank Farian's November 14, 1990 Munich press conference; the November 19, 1990 Grammy revocation; the session singers; the refund settlement covering an estimated 10 million buyers with roughly 50 refunds requested. https://en.wikipedia.org/wiki/Milli_Vanilli

†

U.S. Chamber of Commerce Technology Engagement Center, with Teneo Research. "Empowering Small Business: The Impact of Technology on U.S. Small Business." August 18, 2025. National survey; 58 percent of small businesses use generative AI, up from 40 percent in 2024 and 23 percent in 2023. https://www.uschamber.com/technology/empowering-small-business-the-impact-of-technology-on-u-s-small-business

‡

Pew Research Center. "How Americans View AI and Its Impact on People and Society." September 17, 2025. Survey of 5,023 U.S. adults, June 9 to 15, 2025; 50 percent more concerned than excited; 76 percent say it is extremely or very important to be able to tell AI content from human content; 53 percent not confident they can. https://www.pewresearch.org/science/2025/09/17/how-americans-view-ai-and-its-impact-on-people-and-society/

§

Noy, Shakked, and Whitney Zhang. "Experimental evidence on the productivity effects of generative artificial intelligence." Science, 2023; summarized by MIT News, July 14, 2023. Experiment with 453 college-educated professionals; ChatGPT cut task time 40 percent and raised graded quality 18 percent; gains concentrated among weaker writers; authors caution real-world gains would be smaller. https://news.mit.edu/2023/study-finds-chatgpt-boosts-worker-productivity-writing-0714

‖

Dell'Acqua, Fabrizio, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, Francois Candelon, and Karim R. Lakhani. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality." Harvard Business School Working Paper 24-013, September 2023. Field experiment with 758 BCG consultants; 12.2 percent more tasks, 25.1 percent faster and higher quality inside the frontier; 19 percent less likely to be correct on a task outside it. https://www.hbs.edu/faculty/Pages/item.aspx?num=64700

¶

United States District Court, Southern District of New York. Mata v. Avianca, Inc., No. 22-cv-1461 (PKC), Opinion and Order on Sanctions, June 22, 2023 (Castel, J.). Six nonexistent opinions cited; $5,000 sanction imposed jointly and severally on Steven Schwartz, Peter LoDuca and Levidow, Levidow and Oberman; letters ordered to the client and to each judge falsely named as an author; the court's statement that using a reliable AI tool for assistance is not inherently improper. https://www.nhd.uscourts.gov/sites/default/files/pdf/Mata-v-Avianca-sanctions-order.PDF

**

Dahl, Matthew, Varun Magesh, Mirac Suzgun, and Daniel E. Ho. "Hallucinating Law: Legal Mistakes with Large Language Models are Pervasive." Stanford Institute for Human-Centered Artificial Intelligence, January 11, 2024. More than 200,000 legal queries tested against GPT 3.5, PaLM 2 and Llama 2; hallucination rates of 69 to 88 percent on specific, verifiable questions; models overconfident about wrong answers. https://hai.stanford.edu/news/hallucinating-law-legal-mistakes-large-language-models-are-pervasive

††

Sullivan, Danny, and Chris Nelson. "Google Search's guidance about AI-generated content." Google Search Central Blog, February 8, 2023. Search rewards high-quality content however it is produced; automation used primarily to manipulate rankings violates spam policies; E-E-A-T (experience added December 2022); disclosure appropriate where readers would reasonably wonder. https://developers.google.com/search/blog/2023/02/google-search-and-ai-content

‡‡

Tucker, Elizabeth. "What web creators should know about our March 2024 core update and new spam policies." Google Search Central Blog, March 5, 2024. Scaled content abuse defined as generating many pages primarily to manipulate rankings, whether automation, humans or a combination are involved; combined changes expected to reduce low-quality, unoriginal content in results by about 40 percent. https://developers.google.com/search/blog/2024/03/core-update-spam-policies

§§

Zhang, Yunhao, and Renee Richardson Gosline. "Human Favoritism, Not AI Aversion." Summarized in MIT Sloan Ideas Made to Matter, "Study gauges how people perceive AI-created content," October 3, 2023. Participants preferred AI-generated marketing copy when the source was hidden; disclosure boosted human-made content without penalizing AI content. https://mitsloan.mit.edu/ideas-made-to-matter/study-gauges-how-people-perceive-ai-created-content

‖‖

Schilke, Oliver, and Martin Reimann. "The transparency dilemma: How AI disclosure erodes trust." Organizational Behavior and Human Decision Processes 188 (2025), Elsevier; summarized by University of Arizona News, May 22, 2025. Thirteen experiments; disclosing AI use lowered trust across tasks and disclosure framings; third-party exposure lowered it further; penalty attenuated but not eliminated among technology-favorable evaluators. https://news.arizona.edu/node/31718

¶¶

Kobak, Dmitry, Rita Gonzalez-Marquez, Emoke-Agnes Horvat, and Jan Lause. Study of LLM-assisted writing in biomedical publications measured through excess vocabulary. Science Advances, July 2025 (DOI 10.1126/sciadv.adt3813), as summarized by Scimex, Australian Science Media Centre, July 4, 2025. Analysis of about 15.1 million PubMed abstracts, 2010 to 2024; abrupt rise in words such as showcasing, underscores and potential; as much as 13.5 percent of 2024 abstracts could have involved language-model processing. https://www.scimex.org/newsfeed/ai-chatbots-might-be-helping-to-write-more-than-1-in-10-biomedical-research-papers

***

Wikipedia. "Replicant." Accessed October 14, 2026. Plot facts from Blade Runner (1982): the Voight-Kampff test provokes emotional responses to distinguish replicants; a typical replicant is identified within 20 to 30 questions, Rachael required more than 100; Nexus-6 four-year lifespan; based on Philip K. Dick's Do Androids Dream of Electric Sheep? https://en.wikipedia.org/wiki/Replicant

†††

Wikipedia. "Mechanical Turk." Accessed October 14, 2026. Built by Wolfgang von Kempelen and first displayed in 1770 to impress Empress Maria Theresa; exhibited for 84 years as an automaton; a hidden chess master moved the pieces by levers and magnets; defeated Benjamin Franklin and Napoleon (Schonbrunn, 1809); destroyed by fire in Philadelphia on July 5, 1854. https://en.wikipedia.org/wiki/Mechanical_Turk