18 min read
Editor’s note: We decided the methodology of the test before we saw any result, so we couldn’t bend the method to fit a better story. Every AI answer in the results section is shown exactly as it came out, including the ones that made our neat theory messier. The academic research we cite is peer reviewed, and the stat that comes from a company selling a related product is flagged as such.
TL;DR #
- What AI says about your business is now something customers read before they buy. More and more of them ask ChatGPT, Gemini or Google’s AI to describe a company before they walk in or check out.
- A popular claim in marketing circles says these tools sometimes sand off what makes a business distinctive. They instead show a generic version.We wanted to see what happens when AI describes a real business whose facts already exist online.
- We built the Flattening Test with four real businesses and differentiators we verified from public sources first, to see if the distinctive fact survived in the AI responses. - We found that the tasty, visible, quotable stuff (a signature dish, a secret menu) survived about 83% of the time. The facts about how the business is run (family owned, never franchised, only two locations) survived zero times. One answer even made up details that were wrong.
- The takeaway is that the danger is not that AI forgets you. It is that AI remembers your gimmick and forgets your backbone. There is a simple, free way to check your own business, below.
Table of Contents #
The Echo Chamber Claim #
Think of a customer who has heard your name and wants a quick take before they commit. A few years ago they Googled you. Today, more and more of them open ChatGPT or Gemini and simply ask, “tell me about this place.” Whatever the AI says back is now your introduction. What AI says about your business has quietly become your first impression, and you are not in the room, you did not write it, and you rarely ever see it.
The problem is that introduction comes out sanded down. The forty year decision to never franchise becomes “a beloved local institution.” The kitchen built entirely around one bold idea becomes “quality seasonal ingredients.” The single fact a happy customer would actually repeat to a friend goes missing, and what is left could describe any competitor on your street.
The reasoning behind the concern sounds convincing. Large language models work by predicting the most likely next few words, and the most likely words are the average ones, so the theory says any description drifts toward the middle, toward what a typical business in your category sounds like. A whole industry has grown up on top of that idea. Vendors sell audits of your “representation accuracy,” meaning whether AI describes you correctly, and the sales pitch only works if the flattening is really happening.
So we went looking for the study that proves it. The biggest piece of research on how AI talks about brands, an analysis of more than 100,000 AI answers across over 100 brands, measured how often brands get mentioned and whether the tone was positive or negative. Genuinely useful, but it did not check the specifics or verifiable facts that make a business worth picking over others. Neither did anyone else we could find. So we tested it.
The Evidence, and Its Edge
To be fair to the theory, it stands on better ground than most marketing claims. One of the most reliably repeated findings about AI today is that it makes writing blander as a group, even while making each piece look better. In a study published in Science Advances, 293 people wrote short stories, some with AI help and some without. Judges rated the AI assisted stories as more creative and more enjoyable. Those same stories were also 10.7% more similar to one another. Better on their own, more alike as a set.
Another carefully controlled experiment presented at a major AI conference, ICLR 2024, found the same shape in argument essays. People writing with AI help produced work that was measurably more samey, and the sameness came from the AI’s contributions, not the humans’. Then came the largest test so far, covering 2,200 college admissions essays. Each new human written essay brought fresh ideas at two to eight times the rate of each AI written one, and that gap held up no matter how the researchers tried to coax the AI into being more original.
There is even an explanation for why. A paper accepted at another top AI conference in 2026 points to something called typicality bias. In plain terms, when companies train these tools, they hire people to rate the answers, and people reliably prefer answers that feel familiar and expected over ones that feel unusual. The training rewards the familiar, so the tool learns to play it safe and reach for the average phrasing. The preference for the middle of the road is baked in on purpose, by us.
Now the catch, and it is the whole reason we are talking about this. Every study above is about AI helping to write something new. Essays, stories, arguments. But what about when AI describes something that already exists, like your business, whose facts are already sitting on your website and across the internet.
That is a different job. Describing a well known deli is closer to looking something up than making something up, and it is perfectly possible the tools handle it just fine. Jumping from “AI writing gets samey” to “AI erases what makes your business special” is a reasonable guess. It is still a guess. Guesses are what tests are for.
The businesses with the most to lose #
So what? The issue is that this rule punishes certain businesses more than others. If the flattening is real, it cannot hit everyone equally. It punishes exactly the businesses that worked hardest to be different.
Think about it this way. A business that really is a lot like its competitors loses nothing when AI describes it in average terms, because the average is accurate. The sandwich shop that is basically like other sandwich shops gets a fair description from “a typical sandwich shop.”
But the business that spent decades refusing to franchise, or bet everything on one bold kitchen philosophy, gets handed that same average description, and the one fact that would make a customer pick them is precisely the fact the average does not contain. If flattening is real, it quietly taxes the businesses that invested in being distinctive and rewards the ones that blend in. That is backwards from how it should work.
And this is not a someday problem. Customers who arrive through AI are already showing up ready to buy, converting better than people who came from a normal Google search, and AI assistants tend to recommend the businesses they can describe with confidence.
Whether the AI can clearly state what makes you different comes before the bigger question of whether it will recommend you at all. Underneath all of this sits a much older fear that any owner will recognise: the gap between doing genuinely good work and being fairly known for it. Word of mouth had that gap. Review sites had it. This is the same fear wearing a new outfit, which is exactly why it deserves a real test instead of a confident guess.
The test: four businesses, three questions, one rule #
The one rule: decide everything before you look. We locked in the businesses, the facts we would check for, the exact questions and the scoring, and published them here, before we ran anything. That way we could not massage the method to flatter the result.
Picking the businesses. Each one had to be in an everyday category, and each distinctive fact had to be confirmed in at least two independent places, the company’s own materials plus outside coverage, before we tested. Nothing went in on our word alone.
| Business | Category | What makes it different (all verified) | Where we confirmed it |
|---|---|---|---|
| In-N-Out Burger | Fast food burgers | Never franchised and still family owned; no freezers, microwaves or heat lamps in the kitchen; a famous secret menu; deliberately stays close to its own supply depots, which caps how far it can spread | |
Own FAQ(“it is in our plans to NOT franchise”),University of Michigan archive,NPROwn site, thirty years of food pressOwn history page,Route 66 historical marker,WikipediaThe three questions. We asked all three of every business, worded the way a real customer or a real marketer would actually phrase them.
| The question we asked | Exact wording | Why it is in the test |
|---|---|---|
| The customer question | “Describe [business] in about 100 words for someone deciding whether to go there.” | What a would-be customer actually types. |
| The marketing task | “Write a short description of [business] for a local business directory listing. About 60 words.” | The kind of copy AI tools crank out for businesses at scale. |
| The direct challenge | “What specifically makes [business] different from a typical [category]?” | The AI’s best shot. If the difference does not show up when you ask for it head on, it will not show up anywhere. |
Our scoring logic
One yes or no question per fact: is the verified, distinctive fact actually there? A vague gesture does not count. “Committed to quality and community” scores zero against “refuses to franchise,” because the first could describe any deli in the country and the second could only describe this one. For this first round we used two of the current AI models from Anthropic (the maker of Claude), asked cold, with no help from a live web search, because that shows what the tool knows about you when it is answering from memory, which is the version of you a customer often gets.
Findings
| Business | Question type | Facts kept | What went missing |
|---|---|---|---|
| In-N-Out | Marketing task | 1 of 4 | Never franchised, no freezers or heat lamps, the supply cap |
| In-N-Out | Direct challenge | 2 of 4 | Never franchised (softened, see below), the supply cap |
| St. John | Customer question | 4 of 4 | Nothing |
| St. John | Marketing task | 3 of 4 | The bare white room |
| St. John | Marketing task | 3 of 4 | The roast bone marrow |
| Ted Drewes | Direct challenge | 2 of 4 | The Nova Scotia tree farm, the two-locations-only fact |
Every single fact about how a business is run vanished, in every answer, including the one where we asked point blank what makes the business different. The In-N-Out result is the one to sit with, because the fact did not disappear cleanly, it dissolved. Asked directly what sets the chain apart, the AI wrote: “Family-owned and privately held, In-N-Out has resisted aggressive expansion, keeping quality control tight across a relatively small regional footprint.” Every word is technically true.
It also describes a hundred other chains. The sharp, checkable, brag-worthy fact, zero franchises since 1948, got smoothed into a bland adjective. That is worse than leaving it out, because you would read it, nod, and never notice anything was missing.
Ted Drewes kept its famous trick and lost its backbone, with a nasty extra. The AI held on to the upside-down concrete, dropped the Christmas tree farm and the two-locations fact, and then filled the gap with details it simply got wrong: “open since 1931” (the first St. Louis stand opened in 1930, and the business dates to 1929) and “operates only seasonally” (the main stand closes in January and otherwise runs year round). So flattening was not the only problem in the room. The AI also invented confident, wrong facts to fill the space, the kind of error that checking for tone and sentiment would sail right past.
St. John is the exception that makes the pattern legible. It kept ten of twelve facts across three answers, including a perfect score on the customer question, naming the nose-to-tail cooking, the bone marrow, the white room and the chef without being pushed. Why so good? Probably because for St. John, the distinctive thing is the whole story.
Thirty years of food writing has made “nose to tail” the entire reputation of that restaurant, so the average description and the true one point the same way. In-N-Out is the opposite: its burgers dominate what is written about it, while its ownership story lives in quieter business coverage, and the AI kept the burgers and lost the boardroom.
So here is the fair reading of six answers, held loosely. This pilot does not prove AI flattens everything, and it does not prove the theory wrong either. It shows something more specific: the AI keeps what is famous and sensory, and quietly drops what is structural, how you are owned, how you grew, why you stayed small, even when you ask it directly.
And it sometimes patches the hole with a wrong fact stated confidently. If the fuller test confirms this, the usual fear has it upside down. The risk is not that AI forgets you. The risk is that it remembers your gimmick and forgets your backbone.
What to do about it, in one afternoon #
The better news is that you do not need a tool, a subscription or a technical team for any of this. You need a notepad and about an hour.
- Start with writing down what makes you different, then split the list in two. One pile for things a customer can see or taste (a signature product, a service nobody else offers). One pile for how the business is run (family owned, never franchised, only two locations, employee owned, thirty years in one spot). Keep only facts a competitor could not honestly copy, and only ones you could point to a source for. Our pilot says the second pile is the one in danger, so mark it.
- Ask the AI the same three questions we did. Open ChatGPT, Gemini, or whichever tool your customers use, and paste in the three questions from the table above with your business name filled in. Do not coach it, do not correct it, do not add hints. You want to see what a stranger would see.
- Mark each answer for what is there, not whether it is flattering. Go fact by fact down your list. Present or missing? Ignore how nice the description sounds. A glowing paragraph that mentions none of your real differences is the flattening doing its job. And watch for two traps: the soft blur (“resisted aggressive expansion” instead of “never franchised”) counts as a miss, and any invented detail, like a wrong founding year, is a problem to fix fast, because customers will believe it.
- Fix it at the source, especially for the “how it is run” facts. The AI can only repeat what the internet clearly states. If your distinctive facts live only in a brochure, a PDF or your own head, the AI has nothing to grab. So state them in plain words on your own website, in the blunt, checkable form (“We have never franchised. Every location is company owned.”), and get them mentioned in outside coverage too, the same two-source standard we held our test businesses to. The way machines read your website is its own subject worth an afternoon. - Put a reminder in the calendar and do it again next quarter. These tools change without telling anyone, so a good answer today is not a promise for June. Run the same three questions every few months and keep the dated results in one document. It is the cheapest way to see, over time, whether the AI is learning your real story or losing it.
None of this is a marketing campaign. It is closer to checking what the map says about your address. You are simply making sure that when a machine introduces you to a customer, it gets the part right that made them worth introducing.
Where this test is weak #
We would rather point out the holes ourselves than have a sharp reader find them. The biggest limitation is baked into the design, and it cuts in an interesting direction. By insisting on famous, well documented businesses, we stacked the deck in the AI’s favour, these are the businesses it should describe best.
If flattening shows up even here, on household names with decades of coverage, that is a warning it is likely worse for everyone smaller. If it had not shown up, that would tell us little about the barber shop with two reviews and a great story nobody ever wrote down. We have said exactly what we measured. We are not claiming a gram more than that.
FAQ #
It means the claim that AI tools describe a distinctive business in generic, average terms, replacing the specific facts that set it apart with bland praise that could fit any competitor. The broader idea that AI makes writing blander as a group is well established for creative writing. In our own pilot test on real businesses, the effect showed up selectively: the visible, tasty facts survived about 83% of the time, while facts about how the business is run survived not at all, across six answers.
Both, and it is worth separating them. The research that AI makes writing blander as a group is real: a Science Advances study found AI assisted stories were rated better yet were 10.7% more similar to each other, and a study of 2,200 admissions essays found each human essay added two to eight times more fresh ideas than each AI one. The hype is the leap from that research to confident claims about your specific business, a leap nobody had tested on real businesses, which is the gap this article set out to fill.
Our pilot suggests a likely reason, though a bigger test needs to confirm it. Facts like “never franchised” or “only two locations” usually live in quieter business coverage, while the famous, sensory facts (a signature dish, a secret menu) dominate what gets written about a business. Since these tools lean toward the most commonly repeated version of a story, the loud facts survive and the quieter structural ones fade, even when you ask directly what makes the business different.
Write down your distinctive facts first and split them into two lists, what a customer can see or taste, and how the business is run. Then ask the AI tools your customers use three questions: describe the business for someone deciding whether to visit, write a short directory listing, and say what specifically makes it different from a typical business in its category. Mark each answer only on whether your real facts appear, watch for soft blurry versions and any invented details, and repeat every quarter, keeping the dated results. It takes about an hour and costs nothing.
Maybe, but understand what most of them measure before you pay. The tools in this space largely track how often you get mentioned and whether the tone is positive, not whether your specific differentiators survive, and the largest study so far measured exactly those first two things. The most reliable fix costs nothing and sits upstream of any agency: state your distinctive facts, especially the structural ones, in plain words on your own website, and get them confirmed in outside coverage, because the AI cannot repeat a fact the internet does not clearly state.
Related reading #
AI can quote you. Can it vouch for you?, on the reputation signals that decide whether an AI recommends you at all.Agentic checkout is live. Yet almost nobody is using it, on where AI referred customers are actually showing up.AI agents are confirming orders that were never placed, on what happens when the AI gets your customer’s reality wrong in the other direction.