AI didn’t make marketing mediocre. Marketing was already mediocre.
AI simply made mediocre marketing almost free.
That distinction decides what you do next. If AI caused the problem, the fix is a policy about AI. If AI removed the cost of a problem the industry already had, then the fix is a much harder conversation about what your marketing was ever actually good at.
Start with the number that gives the game away. Eighty seven per cent of B2B marketers say AI has made them more productive. Fewer than two in five say their content is performing better. Both figures come from the Content Marketing Institute and MarketingProfs, surveying 1,015 B2B marketers in mid 2025. The same study found 58 per cent believe their content quality has improved.
So quality is up, output is up, productivity is up, and results are flat.
That is not a paradox. It is what happens when everyone gets better at the same time. Improvement that is universally available stops being improvement and becomes the new floor. AI made competence abundant. And once competence is widely available, it stops commanding an advantage.
Every marketing advantage dies the same way
Think about what has genuinely delivered advantage in marketing over the last thirty years, and when each one stopped.
Having a website. Understanding how search worked. Buying media efficiently. Owning a marketing automation stack. Publishing consistently. Running paid social before your category noticed it existed.
Every one of those was a real edge. Every one is now hygiene. And each died the same death, which had nothing to do with the capability getting worse. It got easy.
Advantage is not a property of a capability. It is a property of that capability’s scarcity.
The industry keeps missing this, because it keeps mistaking the capability for the advantage and then acting surprised when the returns evaporate. Media buying threw off outsized returns while it required specialist knowledge and relationships. Programmatic made competent buying available to anyone with a budget, and the advantage moved on. Search threw off outsized returns while most companies did not understand it, and then everyone hired for it. Marketing automation was a moat until the moat came in a subscription.
Nothing that can be installed is a moat.
Notice too that the half-life keeps shortening. Websites took years to travel from advantage to table stakes. Marketing automation moved faster. Content marketing faster again. Generative AI has compressed the cycle dramatically.
By the time the Content Marketing Institute fielded that survey in mid 2025, 89 per cent of B2B marketers were already using AI to generate or optimise copy. A capability sitting at 89 per cent adoption is not an advantage. It is a baseline. If your team is still writing the business case for AI adoption, the thing you are building the case for stopped differentiating anybody some time ago.
This is the honest answer to “how do we get an edge from AI”, and nobody wants to hear it. You don’t. Not from adoption. Adoption is now the cost of staying in the game, and the edge has already moved to whatever AI left scarce.
Hold that question. The rest of this follows from it.
The mediocrity came first
Before blaming the tools, look at what marketing was producing when the tools were expensive.
System1, working with LinkedIn’s B2B Institute, tested 1,600 B2B ads, mostly television, with around six million people globally over four years. Seventy five per cent scored one star or less on System1’s emotional response measure, the level that contributes effectively nothing to long term market share growth. Not one of the 1,600 reached five stars.
That was 2021. That was work with budget behind it, agencies attached, and senior sign off. It still landed as noise.
System1’s later research with eatbigfish and Peter Field put a figure on the flatness. In the UK market, 60 per cent of responses to B2B television advertising registered as emotionally neutral, against 52 per cent for consumer advertising. B2B is the duller category by measurement, not by reputation. And at the same level of excess share of voice, non dull advertising delivered around 1.3 points of annualised share gain while extremely dull advertising delivered around 0.1. An order of magnitude difference in growth, decided by whether the work registered at all.
Peter Field’s work for the IPA traced the same decline across nearly 600 case studies. Creatively awarded campaigns were roughly twelve times as efficient as non awarded campaigns in the twelve years to 2008. By 2018 that advantage had fallen below four times. Over the same period, the share of awarded campaigns that were short term rose from under 5 per cent to almost 40 per cent.
None of that required a language model. The industry manufactured its own mediocrity through short termism, measurement pressure and the steady substitution of activation for brand, which is the argument the previous essay in this series makes in full.
What generative AI changed is the price. Producing forgettable work used to cost a budget, a team and a quarter. Now it costs a prompt.
AI lowered the cost of being forgettable, and volume responded accordingly.
Average is now infinitely scalable
There is a mechanism behind the convergence, and it is better documented than most of the commentary suggests.
Anil Doshi at UCL and Oliver Hauser at the University of Exeter randomised 293 writers across three conditions: no AI, one AI generated idea, and five. Six hundred independent evaluators assessed the results, published in Science Advances in July 2024.
Individual work got better. Novelty rose 5.4 per cent with one AI idea and 8.1 per cent with five. Usefulness rose 3.7 and 9.0 per cent.
Collective similarity between the stories rose too, by 10.7 per cent of the measured range with a single AI idea.
Every writer improved. The body of work became more alike. For one marketer that reads as a win, because your draft beats the one you would have written alone. For a category it reads as convergence, because every competitor’s draft improved in the same direction at the same time.
There is a real complication worth knowing. A University of Michigan study with more than 800 participants across 40 countries found the opposite effect, with high AI exposure increasing collective idea diversity. The designs differ in one respect. Doshi and Hauser gave people AI output to work from. Michigan used AI to surface ideas participants would not otherwise have reached.
Substitution converges. Expansion diverges. Which one you get is a decision about process, not a property of the technology, and most teams have never made that decision consciously.
The volume side is now measurable. Graphite, sampling 43,000 URLs from Common Crawl, estimates that roughly half of newly published web articles are primarily AI generated. Its detector has published false-positive and false-negative rates, so treat the exact percentage cautiously. The important finding is not whether AI has technically crossed 50 per cent. It is that human and primarily AI-generated articles are now being produced at roughly comparable volumes.
What happens to that content is the more interesting finding. Across 31,493 keywords, of the articles ranking on the first two pages of Google results, 86 per cent were human written. Only 7 per cent of top position results were AI generated, half the baseline rate. Citations inside AI answer engines split roughly 82 per cent human to 18 per cent AI.
Production share and visibility share have come apart.
Meanwhile the surface being competed for is shrinking. Pew Research Center tracked 68,879 Google searches from a panel of 900 US adults. When an AI summary appeared, 8 per cent of visits produced a click on a traditional result, against 15 per cent when no summary appeared. One per cent clicked a link inside the summary itself. Ahrefs, comparing 150,000 keywords with AI Overviews against 150,000 without, reported a 58 per cent average drop in click through rate for top ranking pages, up from 34.5 per cent eight months earlier. SparkToro, using Similarweb clickstream data, put the share of US Google searches ending without any click at 68 per cent for early 2026.
More competent content, produced faster, competing for a materially smaller surface.
And the audience is not neutral about it. The Nuremberg Institute for Market Decisions ran controlled experiments alongside a 3,000 person survey across the US, UK and Germany. Identical advertising rated lower on emotional impact when labelled AI generated than when labelled human made, and only a quarter of respondents thought they could recognise AI content in the first place.
Which makes disclosure the wrong thing to be anxious about. Undifferentiated work fails on its own terms, long before anyone forms a view about how it was made.
Differentiation was never the mechanism
Here the industry’s own framing gets in the way, and it is worth clearing out.
Almost every article about B2B sameness quotes a statistic along the lines of “86 per cent of B2B buyers see no real difference between suppliers”, usually credited to CEB and Google. We went looking for the primary source. It does not exist. The figure circulates through agency blogs and conference decks citing each other, and it has been doing so for a decade.
Here is what is documented. Gartner surveyed more than 1,100 customers in late 2020 and found 64 per cent could not distinguish one B2B brand’s digital experience from a competitor’s. Gartner also names perceptions of difference between supplier offerings as one of three drivers of buyer confidence, and its earlier work found that buyers overwhelmed by high quality information were 153 per cent more likely to settle for a smaller, less ambitious purchase than they originally planned.
Sit with that one. The information was good. Volume alone was enough to shrink the deal.
Now the deeper point, which cuts against how most B2B teams think about this entirely.
In a 2007 paper in the Australasian Marketing Journal, Jenni Romaniuk, Byron Sharp and Andrew Ehrenberg showed a low level of perceived differentiation between competing brands across many categories and two countries. Buyers largely did not see meaningful differences. They bought anyway. The authors’ conclusion was to stop putting perceived differentiation at the centre of brand strategy and put distinctiveness there instead: unique associations that make a brand easily identifiable.
If distinctiveness rather than differentiation is the engine, then “everything sounds the same” describes a memory failure, not a positioning one. The question is identifiability. Does anything about your work make it retrievable later?
Retrievable when, though, is the part that gets skipped. John Dawes’s work at Ehrenberg-Bass with LinkedIn’s B2B Institute is the useful frame here, shorthanded as the 95-5 rule: up to 95 per cent of business buyers are not in market at any given moment. Dawes presents that as a heuristic rather than a measured constant, so use it as a way of thinking and not as a number in a board deck.
The measured version comes from 6sense’s B2B Buyer Experience Report for 2025, drawing on nearly 4,000 responses. Ninety five per cent of winning vendors were already on the buyer’s day one shortlist. Around four in five deals went to the vendor preferred before any seller contact. Buyers reached first contact roughly 61 per cent of the way through their process, earlier than the 69 per cent recorded the year before. Companion research put 85 per cent of buyers as having prior experience with the vendor they selected.
That is a statement about winners rather than about your odds, so do not turn it into a conversion rate. What it establishes is enough: by the time you reach a competitive evaluation, most of the outcome is already sitting in the buyer’s memory of you.
Which is where Jenni Romaniuk’s category entry points come in. Buyers retrieve brands by situation: the problem, the moment, the trigger. Attach yourself to a handful of the situations that matter and you get recalled when one occurs. Publish competent, unmemorable content attached to nothing in particular and you do not.
So the real cost of sameness lands well before the comparison. You were never in the set to be compared.
What abundance leaves scarce
Back to the question the second section left open. Every capability that becomes abundant stops conferring advantage. So what has survived every one of these cycles without becoming abundant?
Two things, and they are not the same thing, though the industry uses them interchangeably.
Judgement determines what should be done. Taste determines how it should be expressed.
Judgement is the commercial call: which market, which problem, which trade off, what to stop doing. The expressive call comes later and answers a different question, about what makes the work land, what to cut, and what deserves to exist at all.
Neither has ever been available in a subscription. You cannot install taste, and no prompt can make the final judgement for you.
The evidence for the judgement premium is strong and getting stronger. Fabrizio Dell’Acqua, Ethan Mollick, Karim Lakhani and colleagues ran an experiment with 758 consultants at Boston Consulting Group, published in Organization Science in 2025. On tasks inside AI’s capability frontier, consultants with AI access completed 12.2 per cent more tasks, 25.1 per cent faster, at higher quality. On tasks outside that frontier, consultants with AI access were 19 percentage points less likely to produce a correct solution than those working without it.
Same tool, same people. The variable was whether the task sat inside the model’s competence or outside it, and nobody hands you a map of where that line falls. Finding it is a judgement call, and the study prices what getting it wrong costs.
AI does not reduce the need for judgement. It raises the price of not having any, because a confident wrong answer produced in four seconds travels a great deal further than a slow one.
The labour market is already repricing this. Stanford’s Digital Economy Lab, using ADP payroll data covering millions of US workers, found employment among 22 to 25 year olds in AI exposed occupations now sits 19 per cent below where it would be had it kept pace with less exposed peers. Experienced workers show no comparable gap. When this series covered that research in the second essay, the figure was 16 per cent. Four months later it is 19, and the researchers note the divergence has widened steadily since they first documented it.
The finding underneath the headline is the one to carry. Declines concentrate where AI substitutes for human tasks. Where it complements, employment is flat or rising, particularly for experienced workers.
Now taste, which the industry treats as a soft word for something with a very hard price tag.
Go back to the System1, eatbigfish and Field research on dullness. In their UK B2B sample, of roughly £103 million in advertising spend, £53 million sat in the extremely dull quartile. More than half of the money went into the worst quarter of the work.
That is not a budget problem. Every one of those campaigns was funded, approved, and produced to a professional standard. Somebody signed it off. What was missing was the discernment to look at competent work and say it is not good enough to be remembered, and the standing to say it before the money went out the door.
Taste is expensive to lack and impossible to buy. Anu Atluru’s definition is the most useful one available: taste is discernment expressed. In a world of scarcity we treasure tools. In a world of abundance we treasure taste.
Marketing is now firmly in the second world, and most of its processes were built for the first.
There is a complication the profession has barely started to face. Judgement and taste are built by doing the work: running the campaign, defending the budget, watching a decision fail and understanding why. David Duncan argued in Harvard Business Review in February 2026 that AI is absorbing the repetitive early career tasks which historically produced professional discernment. Put that next to the Stanford data and the shape is uncomfortable. Fewer entry level roles, and less formative work inside the roles that remain, squeezes the pipeline that produces the one capability the market is now paying a premium for.
The scarce asset is getting scarcer. Which is good news for anyone who already has it, and a serious problem for the profession.
Five decisions worth making now
Diagnosis without a decision is commentary. Here is what changes if you accept the argument.
1. Stop looking for advantage in adoption.
A capability your competitors can buy this quarter is not, by itself, a strategy. Ask a harder question about anything you are about to invest in: how long before this is available to everyone in my category, and what happens to my position then. If the answer is under two years, you are buying hygiene and should budget for it as such rather than pitching it as an edge.
2. Measure retrieval, not output.
Volume metrics reward the one thing AI made free. If your reporting is built on pieces published, keywords ranked and sessions delivered, you have instrumented the part of the system that no longer differentiates you, in an environment where the click surface is contracting.
Add a memory side measure: whether your brand is retrieved for the buying situations that matter, tracked over time. The trade off is real. It moves slowly and is much harder to defend in a monthly review than a traffic chart. Make the case before you need it, not during a budget cut.
3. Decide your category entry points before you brief anything.
Identify the buying situations you want to be retrieved for and keep the list short enough to be memorable. Filter each candidate on three tests: is the association credible for you, is it competitive in the category, and is it common enough among buyers to be worth owning.
Everything you publish attaches to one of them or it does not get made. If your content calendar is currently a list of topics, converting it to a list of situations is the highest leverage change available to you this quarter.
4. Move AI upstream of the draft.
The homogenisation pattern comes from taking AI output as a starting point and polishing it. The diversity pattern comes from using AI to reach options you would not have reached alone. Same tools, different placement in the process.
Use it to widen the option set, pressure test a position, argue against your own case, find the evidence that contradicts you. Then reject the first plausible output as a matter of routine. This is slower per asset, and it only works if someone in the room has the standing to say no. Both of those are the point.
5. Protect the apprenticeship in your own team.
If the Stanford pattern reaches marketing and entry level work keeps thinning, an individual team develops a judgement gap long before the profession does. Hand juniors decisions rather than tasks. Let them own a call, defend it, and be wrong occasionally with support.
The cost is short term output. Weigh it against a team that can operate the tools and cannot tell when the tools are wrong.
What this comes down to
Every marketing advantage of the last thirty years died by becoming abundant. Websites, search, media buying, automation, content. AI is running the same cycle at four times the speed, and adoption crossed into hygiene while most teams were still building the business case.
Competent execution is no longer where sustainable advantage lives. It cannot be, because competent execution is getting cheaper and more widely available by the day.
What is left is the judgement to decide what deserves to exist, and the taste to make it worth remembering. Neither can be bought, installed, subscribed to or prompted. Both are built slowly, by people doing the work, which is why the supply is tightening at the exact moment the price is going up.
That is an uncomfortable place for the profession. It is an extraordinary opportunity for anyone willing to build the thing that does not scale.
Evidence informs. Judgement decides.
This is the fourth essay in a series on what marketing becomes when execution stops being the constraint. Subscribe to get the next one, which goes deep on the system that actually produces growth and where AI sits inside it.
If you want to build this properly rather than read about it, FP Collectiv’s courses take the same evidence led approach across B2B marketing fundamentals and marketing with AI.
Sources
Content Marketing Institute and MarketingProfs, B2B Content and Marketing Trends: Insights for 2026, published October 2025. Survey of 1,015 B2B marketers conducted June to August 2025. Findings used: 89 per cent use AI to generate or optimise written copy; 87 per cent report improved productivity; 58 per cent improved content quality; 39 per cent improved content performance. The productivity and performance figures are separate items with separate bases and are not presented in the essay as a single computed gap.
System1 and LinkedIn’s B2B Institute, research on B2B creative effectiveness, reported in Marketing Week, January 2021. 1,600 B2B ads tested with approximately six million people globally over four years. Findings used: 75 per cent scored one star or less; none reached five stars. Sample skews heavily to television advertising.
System1, eatbigfish and Peter Field, The Extraordinary Cost of Dull. Findings used: 60 per cent neutrality for UK B2B television advertising against 52 per cent for consumer; annualised share gain of approximately 1.3 points for non dull advertising against approximately 0.1 points for extremely dull at equivalent excess share of voice, reported as approximations and described in the essay as an order of magnitude rather than a precise multiple; of roughly £103 million in UK B2B advertising spend analysed, approximately £53 million sat in the extremely dull quartile.
Peter Field, The Crisis in Creative Effectiveness, IPA, 2019. Analysis of nearly 600 IPA case studies. Findings used: efficiency advantage of creatively awarded campaigns fell from approximately twelve times to below four times between 2008 and 2018; share of awarded campaigns that were short term rose from under 5 per cent to almost 40 per cent.
Anil R. Doshi and Oliver P. Hauser, “Generative AI enhances individual creativity but reduces the collective diversity of novel content”, Science Advances, July 2024. 293 writers randomised across three conditions, 600 independent evaluators. Findings used: novelty up 5.4 per cent with one AI idea and 8.1 per cent with five; usefulness up 3.7 and 9.0 per cent; collective similarity up 10.7 per cent of the measured range with one AI idea.
Joshua Ashkinaze, Julia Mendelsohn, Li Qiwei, Ceren Budak and Eric Gilbert (University of Michigan), How AI Ideas Affect the Creativity, Diversity, and Evolution of Human Ideas, ACM Collective Intelligence Conference, 2025. More than 800 participants across 40 countries. Finding used: high AI exposure increased collective idea diversity without affecting individual creativity.
Graphite, AI content volume research, May 2026. Method: 43,000 URLs randomly sampled from Common Crawl; detector validated at 4.2 per cent false positive and 0.6 per cent false negative rates. Observed share of new articles primarily AI generated was 49.9 per cent in Q1 2026 and 50.9 per cent in Q4 2025. Applying the published error rates puts the true share nearer 48 per cent, which is why the essay declines to treat the 50 per cent mark as a milestone and reports comparable volumes instead.
Graphite, AI content in search and answer engines, October 2025. Method: 31,493 keywords, first two pages of results. Findings used: 86 per cent of ranking articles human written; 7 per cent of top position results AI generated; approximately 82 per cent of AI answer engine citations human written. Correlation only; AI generated content may skew toward lower authority domains for unrelated reasons.
Ahrefs, AI Overviews click through research, May 2026, across 300,000 keywords. Finding used: 58 per cent average click through rate reduction for top ranking pages where an AI Overview appears, up from 34.5 per cent eight months earlier. Vendor research.
Pew Research Center, “Google users are less likely to click on links when an AI summary appears in the results”, July 2025. 68,879 searches tracked across a panel of 900 US adults, March 2025. Findings used: 8 per cent click rate with an AI summary present against 15 per cent without; 1 per cent click rate on links inside the summary. General consumer search rather than B2B specific.
SparkToro with Similarweb data, June 2026. Finding used: 68 per cent of US Google searches ended without a click, January to April 2026. General consumer search rather than B2B specific.
Nuremberg Institute for Market Decisions, Transparency without trust, Fabian Buder and Matthias Unfried, 2024. 3,000 respondents across the US, UK and Germany plus two controlled experiments. Findings used: identical advertising rated lower on emotional impact when labelled AI generated; 25 per cent believed they could recognise AI generated content, which is a confidence measure rather than a test of ability.
Gartner, press release May 2021, survey of more than 1,100 B2B customers conducted late 2020. Finding used: 64 per cent cannot distinguish one B2B brand’s digital experience from a competitor’s, with perceptions of difference between supplier offerings named as one of three drivers of buyer confidence.
Gartner, press release July 2019, survey of more than 1,000 B2B customers. Finding used: buyers experiencing information overload are 153 per cent more likely to settle for a smaller, less disruptive purchase than originally planned.
Jenni Romaniuk, Byron Sharp and Andrew Ehrenberg, “Evidence concerning the importance of perceived brand differentiation”, Australasian Marketing Journal, 2007. Finding used: low perceived differentiation across competing brands with purchase continuing regardless; distinctiveness proposed in place of differentiation. Not conducted in B2B categories, which are higher consideration than those studied.
Jenni Romaniuk, Ehrenberg-Bass Institute with LinkedIn’s B2B Institute, category entry points research. Concept used: buyers retrieve brands by buying situation, and candidate entry points are assessed on credibility, competitiveness and commonality.
6sense, B2B Buyer Experience Report for 2025, nearly 4,000 responses, and In It (Multiple Times) To Win It, November 2025. Findings used: 95 per cent of winning vendors were on the day one shortlist; around four in five deals went to the pre contact favourite; buyers reached first contact at 61 per cent of the journey, down from 69 per cent; 85 per cent of buyers had prior experience with the vendor selected. These describe winners rather than conversion odds.
John Dawes, Ehrenberg-Bass Institute with LinkedIn’s B2B Institute, The 95-5 Rule, 2021. Presented by the author as a heuristic rather than a measured constant.
Fabrizio Dell’Acqua, Edward McFowland III, Ethan Mollick, Hila Lifshitz-Assaf, Katherine C. Kellogg, Saran Rajendran, Lisa Krayer, François Candelon and Karim R. Lakhani, “Navigating the Jagged Technological Frontier”, Harvard Business School working paper 2023, published in Organization Science, 2025. 758 BCG consultants, AI access randomised, tasks classified as inside or outside AI’s capability frontier. Findings used: inside the frontier, 12.2 per cent more tasks completed, 25.1 per cent faster; outside it, 19 percentage points less likely to produce a correct solution.
Erik Brynjolfsson, Bharat Chandar and Ruyu Chen (Stanford Digital Economy Lab), Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence, August 2026 version. ADP payroll data covering millions of US workers. Findings used: employment of 22 to 25 year olds in AI exposed occupations 19 per cent below the counterfactual, revised up from 16 per cent in the November 2025 version; no comparable gap for experienced workers; declines concentrated where AI substitutes rather than complements. Covers AI exposed occupations across the economy rather than marketing specifically.
David S. Duncan, “How Do Workers Develop Good Judgment in the AI Era?”, Harvard Business Review, February 2026. Argument rather than original research.
Anu Atluru, “Taste is Eating Silicon Valley”, Working Theorys, September 2024. Argument rather than research. Definition used: taste is discernment expressed.