Open your homepage and your closest competitor's in two tabs, cover both logos, and see whether anyone could tell which is which.
Gather four things before you open a chat window: how customers describe the problem in their own words, why your last ten deals were won or lost, what your product specifically does that a competitor’s doesn’t, and your competitors’ current homepage copy. Feed those in, ask for options rather than an answer, then run the swap test: if a competitor could put their name on the line unchanged, it isn’t a value proposition. Specificity is the part that carries commercial weight, and there’s field evidence it moves customer satisfaction and spend.
Try this before you read any further. Open your homepage in one tab and your closest competitor’s in another, then cover both logos with your hand. Read the first line of each out loud.
Most of the time you can’t tell them apart, and neither can anyone else.
Gartner put a number on this in a survey of more than 1,100 B2B customers: 64% said they cannot tell the difference between one B2B brand’s digital experience and another’s[1]. That’s the customers saying it, not the marketers, which is the part that should sting slightly.
What I find more useful than the number is the reframe it forces. The question stops being “is our value proposition any good” and becomes “could a buyer tell our page from theirs.” Those get different answers, and only one of them is checkable by someone other than you.
The reason this matters more now is that everybody’s first draft comes from the same handful of models. If you and four competitors all ask for a value proposition for a project management tool for mid-size agencies, you’ll all get variations on the same sentence, and you’ll all improve it slightly, and you’ll all end up in the same place.
There’s a specific mechanism here, and it’s been measured, which surprised me when I went looking.
When 293 writers were given AI-generated ideas to work from, their output got better and more alike at the same time. Judged more creative, better written and more enjoyable, especially for the less confident writers, and also measurably more similar to each other than the human-only work was. The authors call it an increase in individual creativity “at the risk of losing collective novelty,” a social dilemma where everyone is individually better off and collectively narrower[2].
That was short fiction rather than marketing copy, and nobody has run the equivalent on value propositions, so take it as a mechanism worth knowing rather than proof about your homepage.
The mechanism is anchoring, and it’s the actionable bit. Writers’ output moved toward whichever AI idea they saw first. You read draft one, it’s decent, and every subsequent version becomes a variation on it. You never go back and ask whether the framing was right, because framing feels settled once something readable exists.
Feed it your customers’ words, not your own adjectives.
Which points at the fix. A model asked to invent a value proposition has nothing to work from but the average of everything ever written about your category, so it returns the average. Give it material nobody else has, and it stops averaging and starts editing.
This is the unglamorous part and it takes about an hour. It’s also the entire difference between output you’ll use and output you’ll quietly abandon.
| Input | Where to get it | What it looked like for one B2B software team |
|---|---|---|
| 1. The problem in customers’ words | Sales call notes, support tickets, review sites, onboarding forms | “I can’t tell what my team is actually working on without asking three people.” Nobody said “lack of visibility.” |
| 2. Why the last ten deals were won or lost | Your CRM, or twenty minutes with two salespeople | Won on the approvals workflow four times. Lost on price twice, on integrations twice. |
| 3. What your product specifically does that theirs doesn’t | Product, not marketing. Ask for features, not benefits | Approval chains can branch on spend threshold. Two named competitors can’t do this. |
| 4. Competitors’ current homepage copy | Copy and paste the actual text of four of them | Three of four led on “streamline your workflow.” All four used “seamless.” |
The four inputs to gather before prompting, with a worked example from a B2B software team. Input 1 does the most work; input 4 is what lets the model avoid rather than reproduce the category’s clichés.
Input one is where people cheat, and I understand why. It’s much faster to write down what you think customers say than to go and read what they actually said. But the phrase “I can’t tell what my team is working on without asking three people” is worth more than any sentence you’ll invent, because it’s the thing a real person typed at eleven o’clock at night.
Input four is the one people skip entirely. Pasting in four competitors’ actual copy lets you instruct the model to avoid it, which is far more effective than asking it to be original. Original is a direction; avoid these fifteen specific phrases is an instruction.
If you’re short on time, here’s the order I’d cut things in:
If you don’t have a clear picture of who you’re writing for yet, do that first. Our guides to building an ideal customer profile with AI and analysing customer feedback at volume both feed straight into input one.
Here’s the prompt. It’s long, and that’s deliberate: almost all of its length is your material rather than instructions.
| Paste this, replacing the bracketed sections with your own inputs |
|---|
| You’re helping me write a value proposition. Don’t write one yet. Here’s how our customers describe the problem, in their own words: [paste 10 to 15 real quotes] Here’s why our last ten deals were won or lost: [paste] Here’s what our product does that these named competitors don’t: [paste, features not benefits] Here’s our four main competitors’ current homepage copy: [paste the actual text] Step 1: List every word or claim that appears in more than one competitor’s copy. These are banned from anything you write for me. Step 2: Tell me the three most specific things in the customer quotes that none of the competitor copy addresses. Step 3: Write six value proposition options, each built on a different one of those three angles, two per angle. Label the angle above each pair. Rules for all six: under 15 words. No word from your Step 1 banned list. Each one must name something concrete our product does, not a feeling it produces. If an option would still be true with a competitor’s name on it, don’t give it to me. |
A complete drafting prompt. Steps 1 and 2 are the parts that matter; they force the model to analyse your material before generating anything, which is what stops it defaulting to the category average.
The “don’t write one yet” opener does real work. Left to its own devices, a model will produce a value proposition in its first sentence and then spend the rest of the conversation defending it, and you’ll anchor on it just as reliably as the writers in that study did.
Steps 1 and 2 are also worth reading yourself before you look at step 3. The banned-word list is usually funnier and more useful than anything that follows it. When three of four competitors lead on “streamline your workflow,” you don’t want a better version of that phrase. You want to vacate it entirely.
You’ll get six options and two of them will feel good. Feeling good is not the test, because a well-constructed generic sentence feels good, that’s what makes it dangerous.
Score each candidate out of ten. It takes two minutes and it’s the closest thing I have to an objective standard.
| Check | Score 2 | Score 1 | Score 0 |
|---|---|---|---|
| Competitor swap. Put a named competitor’s name on it. Is it still true? | No, it’s now false | It’s a stretch for them | Yes, entirely true |
| Specificity. Does it name something the product does? | A concrete mechanism | A category of thing | A feeling or an outcome only |
| Customer language. Would a customer use these words? | It quotes them almost directly | Recognisable to them | It’s internal or industry language |
| Reason to believe. Can you prove it on the next screen? | Yes, with a specific proof point | Partially | You’d have to be trusted |
| Salesperson test. Would your best salesperson say it out loud on a call? | They already do | They might | They’d be embarrassed |
Score each candidate 0 to 2 per row. Bands: 8 to 10, ship it and build the page around it. 5 to 7, the angle is right and the wording isn’t; go back to step 3. Below 5, wrong angle, start again from the customer quotes. A zero on the competitor swap row is an automatic fail whatever the total.
If a competitor could put their name on it, it isn’t a value proposition. It’s a category description.
The salesperson row is the one I’d add if you only run one. Marketing copy that a salesperson would never say out loud is copy that has drifted from what actually persuades people, and the gap between those two things is usually the whole problem. Sit in on two sales calls and listen to how your best person describes the product when nobody’s writing it down.
Every option that fails the swap test fails it the same way: it’s abstract. And abstraction is the default output of a model that doesn’t have your details, so this is the edit you’ll make most often.
There’s better evidence for this than “be specific” usually gets. Packard and Berger went through more than 1,000 real customer and employee interactions and found concreteness tracking both satisfaction and spend: across 200 recorded service calls at a US apparel retailer, one standard deviation more concrete language was associated with an 8.9% rise in customer satisfaction, and across 941 email interactions with real spend data attached, roughly 30% more spend over the following 90 days, which they carefully qualify as “for this particular firm”[3].
Their explanation for why is the part I keep coming back to. Customers infer that people who speak concretely have actually been listening.
The difference between “I’ll go look for that” and “I’ll go search for that t-shirt in grey” is what they measured. That’s a small enough gap to be reproducible in your own copy this afternoon.
| What AI gives you first | What it says after you supply the detail |
|---|---|
| Streamline your approval workflows | Approvals that branch automatically at your spend thresholds |
| Get complete visibility into your team’s work | See what everyone shipped this week without asking anyone |
| Powerful reporting that drives better decisions | Your Monday report, written from your data, before you’re at your desk |
| Seamless integration with the tools you love | Two-way sync with Xero, so you stop reconciling twice |
| Enterprise-grade security you can trust | Your data stays in the UK, and we’ll name the data centre |
Five abstract-to-concrete rewrites. The pattern in every row: replace the category verb with the specific mechanism, and replace the promised feeling with the observable result.
Specific is the only kind of different that survives contact with a buyer.
Notice that the right-hand column is riskier. “Your data stays in the UK” is a claim somebody can hold you to, where “enterprise-grade security” commits you to nothing. That’s precisely why it works, and also why abstract copy is so persistent inside organisations: nobody ever got asked to justify a vague sentence in a legal review.
A value proposition isn’t a homepage headline, though it usually becomes one. It’s the thing that has to survive being said in four places without changing shape:
That last one is the real measure, and it’s the only one you don’t control. You’ve got it right when customers repeat it back to you in words close to your own.
Havas’s 2025 Meaningful Brands research puts a fairly bleak frame around what happens if you don’t: people wouldn’t care if 78% of brands disappeared overnight, and the figure was up five points year on year[4]. There’s a verbatim in that report from a respondent that I think is worth more than the percentage. Under a heading about repetition, someone said: “A lot of brands just repeat the same message and products. It gets dull.”
Not offensive. Not wrong. Dull, which is worse, because dull is invisible.
So, this week. Do the swap test on what’s already on your homepage, before you write anything new. Cover the logo, put your closest competitor’s name on the first line, and see if it’s still true. If it is, you don’t have a copywriting problem to solve with a better prompt. You’ve got an hour of reading customer quotes ahead of you, and that’s the part that actually changes the sentence.
Once you’ve got the line, our guide to writing landing page copy with AI covers building the rest of the page around it without losing the specificity you just fought for.
Because a model asked to invent one has nothing to work from except the average of everything ever written about your category, so it returns that average. It isn’t a prompting failure and a cleverer prompt won’t fix it. The fix is supplying material nobody else has: your customers’ actual words from sales notes and support tickets, why your last ten deals were won or lost, and the specific mechanisms your product has that named competitors don’t. Given that, the model stops generating from the average and starts editing your material.
Four things. Ten to fifteen real customer quotes describing the problem in their words rather than yours. The win and loss reasons for your last ten deals. A list from your product team of specific things you do that named competitors don’t, written as features not benefits. And the actual pasted text of four competitors’ homepages, which lets you instruct the model to avoid the category’s shared clichés rather than vaguely asking it to be original.
Run the swap test. Put a named competitor’s name on the line and ask whether it’s still true. If it is, you’ve written a category description, not a value proposition. Score it across five checks: the competitor swap, whether it names a concrete mechanism rather than a feeling, whether a customer would use those words, whether you can prove it on the next screen, and whether your best salesperson would say it out loud on a call. Anything scoring zero on the swap check fails regardless of the total.
There is credible evidence for the mechanism, though not from a study of marketing copy specifically. A preregistered experiment with 293 writers found that people given AI-generated ideas produced individually better work that was measurably more similar to each other’s than human-only work was, and that writers anchored on whichever idea they saw first. That study used short fiction. Treat it as a reason to generate many options and feed in proprietary material, rather than as proof about your homepage.
Specific enough that somebody could hold you to it. ‘Enterprise-grade security you can trust’ commits you to nothing; ‘your data stays in the UK, and we’ll name the data centre’ is checkable, which is exactly why it works. There’s field evidence behind this: research analysing over 1,000 real customer interactions found concrete language associated with an 8.9% rise in customer satisfaction across 200 service calls, and the authors’ explanation is that customers infer someone speaking concretely has actually been listening.
The Gartner figure comes from a Gartner press release reporting a December 2020 survey of more than 1,100 B2B customers, and the date is given in the text rather than presented as current research. A widely circulating statistic claiming 68% of B2B buyers say brands sound the same, attributed to Gartner, was checked against gartner.com and could not be found anywhere on their site; it has been dropped rather than repeated. The Doshi and Hauser study is open access and was read in full via PubMed Central; it studied short fiction rather than marketing copy, which is stated plainly here instead of being extrapolated. Hansen and Wanke’s 2010 finding that concrete language is judged more truthful was considered and dropped, because a preregistered high-powered replication in 2019 failed to reproduce it. Packard and Berger’s field research is used instead. The Havas figure comes from the report PDF rather than the landing page, where it appears only inside a graphic.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.