Essay

Taste in the age of AI: why generic copy wins nothing, and what 64 industrial websites reveal

Published · Updated 18 min read

Taste in the age of AI is the skill of choosing well when making things has become cheap: knowing which draft is good, which one fits the brand, and which ninety to throw away. The idea is old. What is new is the evidence behind it. Controlled experiments now show that generative AI raises the quality of each person's work while making everyone's work more alike.[1][2] When that happens, the scarce input is no longer production. It is judgment.

This essay brings together three kinds of evidence. The first is experimental: two randomized studies of writers and consultants who worked with and without AI. The second is a language study of what happens to diversity when people write with AI help. The third is our own measurement of how alike industrial companies already sound: a scan of the homepage copy of 64 US mid-cap manufacturing and supply-chain companies. We close with a framework, the Taste Stack, and a weekly method any marketing team can run.

Idea in brief

The problem. AI makes a competent draft free for everyone. In an experiment with 293 writers, access to AI story ideas made stories more novel and more useful, but also more similar to each other.[1]

Why it happens. Everyone draws on the same models, and the models pull toward the middle. Writing with an instruction-tuned model reduced the diversity of essays by different authors.[3] Michael Porter saw the same pattern with management tools in 1996: "The more benchmarking companies do, the more they look alike."[5]

The gap (our data). Industrial copy was generic before AI arrived. Of 64 homepages we analysed, 80% used at least one of eleven stock terms, 44% used the word "solution" or "solutions", and only 14% contained a single number.

The answer. Treat taste as a trained, written standard. Collect references, explain why they work, write kill criteria before generating anything, edit for specifics, and keep one named person who signs off.

Key findings

The essay in 2 minutes
  1. Finding 18.1%

    AI lifts each writer.

    With access to up to five AI story ideas, stories were rated 8.1% more novel and 9.0% more useful than stories written without AI.[1]

  2. Finding 25.2%

    AI narrows the whole field.

    The same AI-assisted stories were more similar to one another, and to the AI ideas themselves: 5.2% and 5.0% more similar to the AI ideas than stories written without help.[1]

  3. Finding 343%

    The floor rises fastest.

    In a field experiment with 758 consultants, those below the average gained 43% from GPT-4, while those above it gained 17%.[2] When the weakest work catches up, average work stops standing out.

  4. Finding 428 of 64

    Industrial copy already sounds the same.

    In our scan, "solutions" was the single most common content word across 64 homepages, used by 28 companies. "Products" and "equipment" were used by 8 each.

  5. Finding 538%

    Leadership is the default claim.

    24 of 64 companies (38%) described themselves with "leader", "leading" or "leadership" in their title, meta description or headline.

  6. Finding 614%

    Specifics are rare.

    Only 9 of 64 (14%) included any number, and 9 (14%) named a place or country.

Chapter 1 · 1 min

What taste is, and what it is not

Taste is the ability to tell good from almost good, quickly, and to explain the difference. It is not personal preference. You may prefer blue. Taste is knowing whether blue is right for this brand, this buyer and this moment, and being able to say why.

For a business, taste shows up as decisions: which headline ships, which claim gets cut, which product line leaves the catalogue. The Hive's entry on what AI marketing is makes a related point: the tools change the mechanics of marketing, not the job.

Terms used in this essay

TermMeaning
TasteTrained judgment about quality and fit, expressed as choices and reasons.
HomogenizationOutputs from different people becoming more alike, measured here as rising text similarity.
Stock termA word or phrase that many companies use about themselves, such as "solutions", "leading" or "innovative".
Kill criteriaWritten rules a draft must pass, set before any drafts exist.
Specificity markerA detail only one company could state: a number, a place, a named product or customer group.
Cosine similarityA score for how alike two texts are, from 0 (no shared content) to 1 (identical), computed from word or meaning vectors.

Rick Rubin, the record producer, is a useful model. CBS News reported in January 2023 that he "barely plays any instruments and has no technical ability." He told 60 Minutes that he is paid for "the confidence that I have in my taste, and my ability to express what I feel has proven helpful for artists."[7] In a business that runs on AI tools, the tools are the engineers. Someone still has to be the producer.

Chapter 2 · 3 min

The evidence, part 1: AI makes each person better and everyone more alike

The clearest study of this trade-off is by Anil Doshi and Oliver Hauser, published in Science Advances in July 2024.[1] Writers recruited online wrote short stories in one of three conditions: on their own, with the option of one AI-generated story idea, or with the option of up to five. A separate group of 600 readers rated the 293 stories without knowing how they were made.

Figure 1. AI ideas made stories better rated

One AI ideaUp to five AI ideas
5.4%
8.1%
Novelty
3.7%
9%
Usefulness

Change in rating (%), 0 to 10

Change in reader ratings compared with stories written without AI, in percent. Source: Doshi and Hauser, Science Advances, 2024.

Access to AI made stories better on both creativity measures, and more access helped more. Most writers took the offer: 88.4% of those who could request an AI idea did so at least once.[1]

The gains were not spread evenly. The study measured each writer's baseline creativity with a standard test, the Divergent Association Task. Writers who scored lower on it gained the most.

Figure 2. Less creative writers gained the most

All writersLess creative writers
8.1%
10.7%
Novelty
9%
11.5%
Usefulness

Change in rating (%), 0 to 14

Change in reader ratings with access to up to five AI ideas, compared with no AI, in percent. Source: Doshi and Hauser, Science Advances, 2024.

The authors write that the five-idea condition "effectively equalizes the creativity scores across less and more creative writers."[1] That is good news for any one writer. It is a problem for a market, because the second half of the study measured what happened to the stories as a group.

Using text embeddings, the researchers scored how similar each story was to all the others in its condition. AI access made stories more similar to the average story, by an amount equal to 10.7% (one idea) and 8.9% (five ideas) of the full similarity range among human-only stories. Stories written with AI were also 5.2% and 5.0% more similar to the AI ideas themselves.[1] The authors' summary is blunt: "With generative AI, writers are individually better off, but collectively a narrower scope of novel content is produced."[1]

The same pattern in professional work

A field experiment with Boston Consulting Group found the same two effects among 758 strategy consultants, about 7% of the firm's individual contributors.[2] On 18 realistic tasks inside the frontier of what GPT-4 could do, consultants using it produced work rated more than 40% higher in quality.

Figure 3. The weakest performers gained the most from AI

Below-average performers (bottom half)
43%
Above-average performers (top half)
17%

Performance gain (%), 0 to 50

Improvement in task performance compared with each group's own baseline, in percent. Source: Dell'Acqua et al., Harvard Business School Working Paper 24-013, 2023.

The researchers also compared the ideas consultants generated for a new shoe aimed at an underserved market. They report "a marked reduction in the variability of these ideas compared to those not using AI," and suggest that GPT-4 "might lead to more homogenized outputs."[2] In a simulated condition where ChatGPT answered the task alone, ideas were the most similar of all.[2]

Why the models pull toward the middle

Vishakh Padmakumar and He He tested this at the level of language. In a controlled experiment, people wrote argumentative essays alone, with a base model (GPT-3) or with an instruction-tuned model (InstructGPT). Writing with InstructGPT, but not with the base model, produced a statistically significant fall in diversity: essays by different authors became more similar, and lexical and content diversity dropped. The authors trace the effect mainly to the text the model contributed, while the text people wrote themselves stayed as varied as before.[3]

Limits of the experimental evidence

Read together, though, they point one way. AI raises the floor and narrows the range. The individual draft gets better while the field gets flatter. In a flatter field, the company that can tell the difference, and acts on it, gets noticed.

If you want an outside read on whether your own copy sounds like you or like the category average, the Brave New audit looks at exactly that.

Chapter 3 · 1 min

The evidence, part 2: search engines judge the output, not the tool

A common fear is that AI-written content will be punished by Google. Google's own guidance says otherwise. Its February 2023 post sits under the heading "Rewarding high-quality content, however it is produced." The same page answers the question "Will AI content rank highly on Search?" with: "Using AI doesn't give content any special gains. It's just content. If it is useful, helpful, original, and satisfies aspects of E-E-A-T, it might do well in Search. If it doesn't, it might not."[4]

That puts the burden on judgment. AI will not sink a page, and it will not save one either. The word "original" in Google's guidance matters here, because the experiments above show that unedited AI help pushes content toward sameness.

Chapter 4 · 3 min

Original research: how alike do industrial companies sound?

The studies above measure writers in experiments. We wanted to see how much sameness already exists in the sector we work in most. So on 3 November 2026 we fetched the homepages of the 91 US mid-cap manufacturing and supply-chain companies in our standing research sample, the same sample used in our earlier AI-readiness scan. For each company we read the three lines a company writes about itself most deliberately: the page title, the meta description and the main headline (H1).

67 of 91 homepages returned a readable page to a research script that named itself. Three of those carried no title, description or headline in their raw HTML, which left 64 companies with copy to analyse. The full method is at the end of this essay.

Finding 1: the same eleven words do much of the talking

We counted how many companies used each of eleven stock terms anywhere in those three lines. The company's own name was removed first, so "Advanced Drainage Systems" does not count as saying "advanced".

Figure 4. "Solutions" is the most common thing industrial companies say about themselves

Solution(s)
43.8%
Global / world
39.1%
Leader / leading
37.5%
Innovation / innovative
26.6%
Advanced / engineered
15.6%
Future / tomorrow
14.1%
Critical / mission-critical
9.4%
Best / excellence
9.4%
Partner
7.8%
Sustainable
6.2%
Trusted / reliable
6.2%

Share of companies (%), 0 to 50

Share of 64 industrial companies using each term in their homepage title, meta description or H1, in percent. Source: Brave New Media scan, 3 November 2026.

"Solutions" was the single most common content word in the whole sample, used by 28 companies. For comparison, "products" and "equipment" were used by 8 companies each. In a sample of firms that make bearings, railcars, decking, filtration systems and drone aircraft, the most frequent self-description names no product at all.

This is how it reads in practice. One company describes itself as a "Leading commercial and industrial HVAC solutions manufacturer" that "provides innovative, energy-efficient products." Another is "the global leader in stored energy solutions, delivering innovative batteries, chargers, and power systems." A third "delivers innovative manufacturing, construction, and packaging solutions backed by expertise and scale." All three are real, and all three could swap most of their words without anyone noticing.

Finding 2: four in five companies use at least one stock term

Figure 5. Most companies stack several stock terms

20.3%
0
21.9%
1
14.1%
2
23.4%
3
20.3%
4 or more

Share of companies (%), 0 to 30; x: number of stock terms used

Number of the eleven stock terms each company used in its homepage title, meta description and H1, as a share of 64 companies, in percent. Source: Brave New Media scan, 3 November 2026.

51 of 64 companies (80%) used at least one stock term, 37 (58%) used two or more, and 28 (44%) used three or more. The average was 2.2 stock terms in three short lines of text.

Finding 3: specifics are the exception

Specific details are the opposite of stock terms. A number, a founding year, a named region, a voltage or a count of locations is something only one company can say.

Figure 6. Generic claims outnumber specific ones three to one

Uses "solution(s)"
43.8%
Claims to be a leader or leading
37.5%
Contains any number (year, count or spec)
14.1%
Names a place or country
14.1%

Share of companies (%), 0 to 50

Share of 64 industrial companies whose homepage title, meta description or H1 contains each marker, in percent. Source: Brave New Media scan, 3 November 2026.

The few specific lines stand out at once. One distributor writes that it has "over 670 locations". A structural-hardware maker signs off with "Trusted by builders since 1956." A power-components maker's headline is "Optimized for 48V power delivery." A shipping line lists Hawaii, Alaska, Guam, Japan and China. Each of these tells a buyer something they could check, and something a rival could not copy without it being false.

We also scored how similar each company's copy was to every other company's, using a standard word-weighting method (TF-IDF cosine similarity) across all 2,016 pairs. Scores were low overall, which is expected for texts this short. The most similar pairs were not always direct rivals: the closest pair joined a defense supplier and a space-and-defense supplier, and the third closest joined a battery maker and a power-component maker. Pairs in the same sub-industry were only about one and a half times as similar as pairs from different industries (average 0.044 against 0.028). Read cautiously, that suggests much of the shared vocabulary is not industry language. It is generic corporate language.

What the scan does and does not show

This copy was written before AI tools became common, or without them, or with them. We cannot tell which, and it does not matter for the finding. The point is the starting position. When AI drafts are added to a category that already describes itself in the same eleven words, the experimental evidence says the result drifts further toward the middle.[1][3] Porter described the same mechanism for operations: "The more benchmarking companies do, the more they look alike."[5]

Chapter 5 · 1 min

A framework: the Taste Stack

The research suggests that taste is not one act but a stack of habits, each resting on the one below. A sign-off does little if there is no written standard to sign off against, and a standard is weak if nobody can say why its references work.

LayerQuestion it answersEvidenceWhat to check
5. Sign-offWho is accountable for what ships?Readers rated AI-assisted stories higher, but the stories converged; someone has to notice the convergence.[1]One named approver for every piece under the brand
4. Edit for specificsDoes each piece say something only we could say?Only 14% of industrial homepages in our scan contained a number.Numbers, places, named products and customer groups in every key line
3. Kill criteriaWhat must every draft pass?Writing with an instruction-tuned model reduced diversity across authors.[3]Five to seven written rules, set before any drafts exist
2. ReasonsWhy does each reference work?The weakest performers gained most from AI, so polish alone no longer separates good from average.[2]One sentence per reference: "This works because..."
1. ReferencesWhat does good look like for us?80% of industrial homepages in our scan used at least one stock term.A library of work you admire, from inside and outside your industry

Most teams start at the top, with a manager approving drafts by feel. The research suggests starting at the bottom, because judgment needs material to work from.

Chapter 6 · 2 min

How to build taste in the age of AI: a five-step method

Step 1: Collect references

Keep a folder of work you admire: ads, emails, packaging, landing pages, trade-show stands, even product manuals. Include work from outside your industry. A valve maker can learn from a bank's clarity. If you want a reading list behind the collection, the Hive's branding canon lays out which books to read and in what order.

Step 2: Explain why each one works

A reference without a reason is a mood board. For every item, write one sentence that starts "This works because...". "It's clean" is not a reason. "It makes one promise and backs it with one number" is. Once you can say why something works, you can brief a freelancer, a new hire or an AI tool, and you can check whether you got it.

Step 3: Write kill criteria before you generate

Kill criteria protect you from falling for whatever comes back first. Five to seven rules are enough. For example:

The fourth line matters more than it looks. The Hive's entry on distinctive brand assets explains why being recognisable as yourself often does more commercial work than being different. Our scan shows how few industrial companies pass that test today.

Tip: Turn Figure 4 into your do-not-use list. Search your homepage, your LinkedIn page and your last three proposals for the eleven terms. Every hit is a place where a specific fact could go instead.

Step 4: Edit for specifics

Treat every AI draft as a first draft from a fast, eager junior who has never met your customers. Cut the opening paragraph, which is usually throat-clearing. Delete any sentence that could appear on a rival's site unchanged. Replace "innovative solutions" with the thing you make and the person who buys it.

Read the draft aloud. If you would not say a sentence to a customer across a counter, rewrite it. The Hive's entry on processing fluency explains why plain, easy-to-read copy also tends to feel more true.

Step 5: Keep one named human sign-off

Name one person who approves everything that goes out under the brand. This matters for AI search too. The Hive explains why brand consistency matters for AI search: answer engines build a picture of a company from what it publishes, so off-brand pieces blur the description that gets repeated to buyers.

Which step to do first

StepTaste Stack layerEffortLikely impactWhen
Stock-term search on your homepage and profilesEdit for specificsLowHighThis week
Written kill criteriaKill criteriaLowHighThis week
Name one approverSign-offLowMediumThis week
Reference library with reasonsReferences and reasonsMediumHighOngoing
Rewrite key pages around specificsEdit for specificsMediumHighThis quarter

Chapter 7 · 1 min

A made-up but typical case: the distributor's newsletter

This case is made up, but it's typical. Say a regional electrical-components distributor starts using AI to write its monthly email to contractors. The drafts are polished. They promise "industry-leading quality" and "innovative solutions", the same words our scan found across the sector. Open rates drift down.

The marketing lead applies the method. She collects five emails she enjoyed and notices they share one specific detail per issue. She writes kill criteria: one product, one story, one fact only the distributor knows, such as a lead time it holds on a hard-to-get breaker or a new line now in the warehouse. She deletes every stock term.

The new emails are shorter and sound like the company. When the distributor expands into a new region, that checkable fact does more for a stranger than any polished claim.

Chapter 8 · 1 min

What this means for leaders

For a manufacturing or supply-chain business, the research moves the advantage. When every supplier can produce a clean website and a tidy proposal in an afternoon, polish stops separating anyone. AI pulls output toward the middle, and our scan shows that industrial copy already starts there.

That makes taste a growth skill, not a matter of style. A specific, checkable line about what you make, for whom and where, is what lets a buyer in a new market tell you apart from the twenty other suppliers they found. At Brave New Media we build AI flows: workflows that take a company's message into more channels, more markets and more languages without adding headcount at the same pace. The flow is the easy part. The standard it runs against is the hard part, because a flow without taste only produces more of the average, faster. With a written standard behind it, the same flow can carry one sharp idea to many buyers. The anonymised Brave New case studies show how that plays out in practice.

If you want an AI flow that takes your brand to more markets without losing what makes it yours, start with the Brave New audit.

Chapter 9 · 1 min

Common mistakes

If you want a book that makes the case for odd, human choices over neat logic, read the Hive's notes on Rory Sutherland's Alchemy.

Chapter 10 · 1 min

Methodology

Sample. The 91 companies listed on Wikipedia's S&P MidCap 400 and S&P SmallCap 600 constituent pages under the GICS Industrials sector in eleven manufacturing and freight sub-industries, with official websites from Wikidata (property P856), corrected by hand for five companies. This is the same sample used in our earlier AI-readiness scan.

Fetch. On 3 November 2026 a script requested each homepage once, identifying itself with the user agent "BraveNewResearch/1.0 (+https://brave-new.com; research)". It read raw HTML only, ran no JavaScript and did not retry blocked requests. 67 of 91 homepages returned HTTP 200. The rest returned 403 (17), 202 (3) or another status (4), and are excluded.

Text. For each readable homepage we took the title tag, the meta description and every H1 headline, removed duplicates, and removed the company's own name. Three homepages had none of these in their raw HTML, leaving 64 companies.

Measures. A company counted as using a stock term if any of the eleven patterns in Figure 4 appeared anywhere in its text (case-insensitive, word-start match; for example "innovat" matches "innovative" and "innovation"). "Contains any number" means any digit. "Names a place" means any of a fixed list of countries, regions, US states and cities. Similarity is TF-IDF cosine similarity over all 2,016 company pairs, with common function words removed.

Limits. These are three short lines per company, not whole websites. The term list is ours, and a different list would give different shares. Some companies whose sites build their text with JavaScript are missing. We did not measure whether generic copy costs sales, and we cannot tell which copy was written with AI. The scripts, raw text and results are published with this essay.

Chapter 11 · 1 min

Frequently asked questions

What does taste mean in the age of AI?

Taste in the age of AI is the skill of choosing well when options are unlimited: recognising which draft is good, which fits the brand and buyer, and what to cut. Experiments show AI raises the quality of individual work while making everyone's output more alike, so the human ability to select and refine becomes what sets one business apart.

Does AI make content more generic?

The evidence says it tends to. In a Science Advances study of 293 writers, AI story ideas made stories better rated but more similar to each other. Writing with an instruction-tuned model reduced diversity across authors in an ICLR 2024 study. The effect depends on how people use the tool, and careful editing can counter it.

How do you develop taste as a business owner?

Study a lot of good work and write one sentence on why each piece works. Keep references from inside and outside your industry. Write rules every draft must pass before you generate anything, judge drafts against them, and log why you reject things. Over a few months that habit sharpens judgment more than any course.

Does Google penalize AI-generated content?

No, not for being AI-generated. Google's February 2023 guidance rewards high-quality content "however it is produced" and says using AI "doesn't give content any special gains." Content made mainly to manipulate rankings breaks its spam policies whoever wrote it. Usefulness, originality and trustworthiness are what count.

How can a manufacturer stand out when everyone uses AI marketing tools?

Say things only you can say. In our scan of 64 industrial homepages, 44% used the word "solutions" and only 14% included any number. Replace stock terms with checkable specifics: what you make, for whom, where, and since when. Keep one person accountable for signing off everything published under the brand.

Your next step

This week, copy your homepage title, meta description and main headline into one document. Search them for the eleven terms in Figure 4 and highlight every hit. Then, for each highlight, write the specific fact that could replace it: a number, a place, a product or a customer. That list is your first set of kill criteria.

If you want a team to run the same check across every market and channel you sell in, and an AI flow that keeps the standard as you grow, start with the Brave New audit.

Brave New Media is a marketing agency that builds AI flows to help companies scale their ideas, grow their customer base and expand into new markets, with experience across supply chain and manufacturing.

See also

References

1. Doshi, A. R. and Hauser, O. P. "Generative AI enhances individual creativity but reduces the collective diversity of novel content." Science Advances 10(28), eadn5290, 12 July 2024. science.org/doi/10.1126/sciadv.adn5290

2. Dell'Acqua, F., McFowland, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F. and Lakhani, K. R. "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, 2023. mitsloan.mit.edu

3. Padmakumar, V. and He, H. "Does Writing with Language Models Reduce Content Diversity?" ICLR 2024. arxiv.org/abs/2309.05196

4. Google Search Central. "Google Search's guidance about AI-generated content." 8 February 2023. developers.google.com

5. Porter, M. E. "What Is Strategy?" Harvard Business Review, November to December 1996. hbr.org/1996/11/what-is-strategy

6. Brave New Media. Homepage copy scan of 91 US mid-cap industrial company websites, 3 November 2026. Data published with this essay.

7. Nelson, K. "Rick Rubin's stripped down approach to making music." CBS News, 15 January 2023. cbsnews.com