Dom Carter | Articles

The 3pm blackout is the last thing built for fans

Dom Carter — Wed, 21 Oct 2026 00:00:00 GMT

Every Saturday afternoon in England, between roughly quarter to three and quarter past five, you cannot watch a live football match on television. Not the Premier League, not the Championship, not League Two. The games are happening, tens of thousands of people are in the grounds, and the broadcasters who have paid billions for the rights are not allowed to show them.

That is the 3pm blackout. It dates from the 1960s, when Bob Lord, the Burnley chairman, argued that live television would empty the terraces of the smaller clubs. It is written into UEFA's rules as a protected window that any member association can claim, and England is the one that still does.

I think it is the most interesting rule in British sport, because it is the only part of the fixture schedule that was designed for the person in the ground rather than the person on the sofa. And everything else about how football is sold now points the other way.

What it actually protects

The argument has always been about the lower leagues. If Manchester United against Liverpool is on at 3pm on a Saturday, the theory goes, fewer people turn up at Rochdale. Whether that was ever true is debated. Whether it is true now, when every teenager watches the big game on a dodgy stream anyway, is more doubtful still.

But the rule has quietly done something else. It has kept Saturday at 3pm as the time football happens, and that has held the match-going habit together in a way nothing else has. You know when the game is. You can plan a year of Saturdays around it. The club, the pub, the train, the lift back with your dad: a whole culture is built on one fixed point, and the blackout is the thing that has stopped the broadcasters moving it.

I wrote recently about kick-off times becoming a media product, with Sunday late slots chosen for the sofa audience and the away fan left without a train home. That piece was about the games that have already been moved. This one is about the games that haven't, and why.

Why it won't last

Follow the money. Every rights cycle, the broadcasters pay more and get more games. The direction of travel for 30 years has been towards every match, live, somewhere, for a price. The blackout is the last stretch of the week that isn't for sale, which makes it the most valuable thing the Premier League still has in the cupboard.

The arguments for ending it are already in circulation. Fans abroad can watch 3pm games legally, so why not fans at home. Streaming makes the blackout unenforceable in practice. The lower leagues now have their own streaming deals and might prefer the income. All of these are true, and none of them is the real reason. The real reason is that a 3pm Saturday slot sold to a broadcaster is worth a fortune, and the clubs who vote on it would like the fortune.

When it goes, and I think it goes within two rights cycles, the last argument for scheduling anything around the people who actually attend goes with it. Every kick-off becomes a programming decision, which means every kick-off is decided by whoever is paying for the picture.

The marketing lesson, because there always is one

Most businesses have a 3pm blackout. A rule, usually an old one, that protects the customer who shows up over the customer who scales. A minimum order that keeps the small shop viable. A phone line still answered by a person. A returns policy that assumes the customer is honest.

Those rules get described in the boardroom as inefficiencies, and most of them are. They are also the reason the loyal customers are still loyal, and nobody prices that in until the rule is gone and the customers are too.

Football is about to run the experiment for everybody. Watch the away ends in 2032 and you'll have your answer.

The AI pilot that never lands

Dom Carter — Wed, 21 Oct 2026 00:00:00 GMT

Every agency I talk to has an AI pilot. Some have 4 or 5. They sit in a Notion page somewhere with a name like "Workflow Acceleration Phase 1", a Slack channel with 12 members, and a demo that went well in March.

What almost none of them have is a job that used to take a person and now doesn’t.

That’s the tell. A pilot is supposed to be a test with a decision at the end of it. Keep or kill. But the pilots I see are designed so that neither answer is ever forced. The tool gets used by the 2 people who were already keen. Everyone else carries on as before, and the thing drifts into that permanent, half-alive state where it’s too useful to cancel and too small to matter.

Why it happens

It isn’t laziness and it isn’t fear of the robots. It’s that nobody wants to be the person who says the old process is finished. In an agency, the old process has names attached to it. Saying "we don’t need three rounds of manual QA on reporting any more" is really saying something about the people who do the three rounds, and most managers would rather keep the pilot ambiguous than have that conversation.

So the technology gets blamed for being "not quite there yet", which is a phrase that can be true forever. Or finance arrives first and caps it, which is what happened at Accenture.

What landing looks like

The pilots that turn into something share one boring feature: somebody wrote down, before it started, what would stop being done if it worked. Not "we’ll save time". Which meetings, which documents, which approvals. Then they put a date on it.

That’s uncomfortable, because it means the pilot can fail in public. But it also means it can succeed in public, and that’s the only version that changes how a business runs.

If your AI pilot has been running for more than a quarter and nobody’s job has changed, it isn’t a pilot. It’s a hobby with a budget line, and the people in it already know.

Playbook: Reaching the unreachable

Dom Carter — Wed, 21 Oct 2026 00:00:00 GMT

Two requests shaped the last 18 months of my agency life. Every client wanted us to find efficiencies in the campaigns we ran, everywhere. And every client wanted us to do "something" with AI. This is what came out of the two together, and it is the closest thing I have to a method.

The problem it solves is simple to state. Advertising platforms are very good at targeting the people who already know you. The most popular ones focus almost entirely on the in-market slice of your audience: people who are likely to buy, as long as they see the right ad, at the right time, enough times, over a specific period. That is a real service and it makes money. It is also a ceiling.

Who the unreachable are

Picture three circles, one inside the next.

The inner circle is your customers. You can reach them with your own data and through the platforms, and they are cheap to reach because the platforms know exactly who they are.

Around them is your in-market audience: people who have searched for the category, visited a competitor, or behaved in a way the algorithm reads as intent. You reach them when you share content and when you pay for campaigns. Most of your media budget goes here whether you intend it to or not.

Then there is everyone else. The unreachable are potential customers whose online behaviour sits outside what the algorithms and platforms consider worth targeting. They haven't searched, they haven't visited, they haven't bought from a rival this month. They are, by a distance, the largest circle, and most platforms will simply not show your ads to them.

The platforms are not being lazy. They want to give you enough results to keep you spending, and not so much that you dominate. If you pooled all your money into a top-of-funnel campaign you would see nothing you could call a direct result, so they steer you away from that build. From where they sit it is rational. From where you sit it means your growth is capped at the size of the people who were already coming.

Why nobody reaches them

Three reasons, and they compound.

The road from first contact to a tracked sale is long. For an unreachable customer it can take a year, sometimes more, for one full path to complete. Most marketing forecasts are 12 months long. There is no room in a 12-month forecast for a customer who takes 14 months to arrive, so nobody plans for them.

Other channels steal the credit anyway. Even after that long path finishes, the sale will almost always be claimed by a Meta retargeting ad or a brand search click, because those are the last things the person touched. The awareness view that started it 11 months earlier is invisible to a last-click report.

And everyone has to justify a purpose per channel. Marketers, agencies and the platforms themselves all need to show what each pound did. If a channel can't demonstrate a sale, its share of the budget is harder to stake. So the channels that reach new people, which by definition show sales slowly, lose the budget argument every quarter to the channels that harvest existing demand.

When you find a way through this you start to see revenue that was never in the plan. But you have to be brave, commit to it, and trust the data, which is tough, because the data takes a long time to arrive.

The deal you have to make first

Nobody gets to spend more. That is the constraint every client set, and I think it is the right one. To reach new people you need budget for it, and you need it without touching the revenue targets you've already promised.

So the method starts by finding the money inside the existing account. Prevent wastage, discover efficiencies, use technology where it helps. Only then do you move the money up the funnel, and only then do you measure what it does.

Four steps. The first two find the budget. The second two spend and measure it.

Step 1: consolidate the lower-funnel spend

Strip back hard on any spend that contributes to a sale that was going to happen anyway.

The main offender is brand search. If you search your own brand name and your organic listing is already first, the paid ad above it is a lone ranger: it sits there collecting clicks from people who typed your name and would have clicked the free result underneath. Some brand search is worth keeping, particularly where competitors bid on your name. Most of it is not. There is technology that stops you spending on branded keywords when your organic result already holds the top position, and the first thing we did on every account was switch it on.

The second offender is over-indexed retargeting. Retargeting is cheap and it looks brilliant on a dashboard, because it talks to people who already visited you. Past a certain frequency it is paying to follow people who were coming back regardless. Cap it.

Everything freed here goes into a holding pot. Don't spend it yet.

Step 2: block the invalid clicks

Some of the traffic you pay for is not people. Bots, scrapers, click farms, and competitors clicking your ads to burn your budget. The industry's estimate is that invalid clicks are 15 to 25% of what you spend on Google and the other platforms. I'd treat that as a vendor's number. Here is ours.

On one account, over the 30 days from 19 August to 17 September 2025, the blocking tool classified the sources of our paid traffic like this.

Source type	Sources	Clicks	Share of clicks
Legitimate	358,518	420,389	82.47%
Suspicious	5,114	14,481	2.84%
Invalid	50,525	74,849	14.68%

The cost of the traffic from invalid sources in that window was £37,410.07. One account, one month.

The fix is unglamorous. Blacklist the IP addresses flagged as fraudulent or spam, and the ones the tool's models recognise as likely competitors, so the platforms stop serving them your ads and stop charging you for the click. Conversion rate goes up because the denominator stops being padded with robots. Return on ad spend goes up for the same reason.

Now stop for a moment, because this is where a lot of brands do stop. After steps 1 and 2 you have improved return on ad spend and conversion rate without spending a pound more on marketing. It is a good result and you could take it to the board tomorrow.

It does not reach a single unreachable person. It just funds the attempt.

Step 3: map how intent builds

Draw the funnel as a triangle, point down. At the wide top sit the channels that reach the most people at the lowest cost per person with the least relevance: connected TV, video, YouTube, prospecting native and display, Meta awareness and lead-gen. Towards the point sit the channels with the highest intent and the best return: Performance Max, Meta prospecting, brand search, Meta and Performance Max retargeting, retargeting display.

Now mark where the money goes. On most accounts I have looked at it clusters around the point: Demand Gen, Performance Max, Meta prospecting, brand search, retargeting. The top of the triangle is empty or token. That is the shape the platforms recommend, for the reasons in the section above, and it is the shape that guarantees you only ever talk to the two inner circles.

The reason brands can't move the money up is that they still live and die on last click, or on cookies, and both have stopped working.

Cookies are inaccurate on a large share of the web now. Roughly 60% of people opt out of being tracked across browsers, and of those who opt in on Chrome, the consent resets after 7 days, so a path that takes months breaks anyway. Apple's privacy restrictions block unique IDs and restrict IP-based tracking, so iPhones, which are a large share of a premium audience, go dark. And the platforms' own reporting over-counts in their favour while most third-party attribution tools under-count, so the two numbers never agree and nobody trusts either.

What we did instead was track across 4 sources on a single first-party pixel, which is the small piece of code on your own site that records each visit and gives the visitor an ID.

Server side, from the commerce platform (Shopify in our case).
Browser side, from the site itself.
The ad platforms, Meta and Google.
UTM parameters on every link we controlled, including email.

Each visitor gets a hashed signal ID from the pixel. It persists for up to 3 years. When a visitor eventually makes a purchase, the system resolves the customer record back to that signal ID, and every earlier interaction, with the ads, the emails and the website, is now attached to a real sale. Not just for customers, for every single website visitor.

That gives you a view like this. Someone clicks a Facebook ad. Ninety days pass. They click a Google retargeting ad. They buy. A last-click model gives the Google ad 100% of the credit and the Facebook ad nothing, which is how your Facebook prospecting campaign ends up looking like a cost centre.

The attribution model built on the joined-up data reconciles your actual sales with what the platforms reported and relinks the hidden conversions to the campaigns that started them. On a typical journey of ad view, social click, email click, search click, sale, the model we used weighted it roughly 30, 25, 20, 15 and 10. Those weights are not gospel. The point is that the view at the top of the funnel gets a share, and that share is calculated from your own sales data rather than assumed.

The consequence is the thing that matters. Awareness becomes a trackable, measurable campaign type. Connected TV, video and YouTube, the channels at the top of the triangle, now have a return on ad spend figure, measured over the real length of the journey. You can make conversion-led decisions about top-of-funnel allocation, which until now was an act of faith.

Step 4: split the strategy

With the budget found and the measurement in place, divide the account in two and judge each half on its own job.

Budget 1 is new customer acquisition. It is built on the models of who the new customers turned out to be, which is to say the unreachable, once you can finally see them. Expect a lower return on ad spend and a higher cost per acquisition. In exchange, every customer is brand new and 100% incremental. The creative is stronger and more brand-level, introductory rather than direct sales, because you are talking to someone who has never heard of you. Over time the measurement gives you something you never had before: proof that these customers would not have arrived otherwise.

Budget 2 is retention and in-market. The expected sales. Higher return, lower cost per acquisition, some new customers and some returning, and a blended return on ad spend of around 5x on the accounts we ran. This is where the platforms want all your money, and now it gets the share it deserves and no more.

The split is what stops the argument. When new-customer activity is judged against in-market activity on the same return figure, it loses every time and gets cut. When each has its own target, the question becomes whether the business wants to grow, and that is a question for the board, not the ad platform.

What it cost us

The method took time, and the hardest part was not technical. It was holding the line through the months where the top-of-funnel spend showed nothing, because the journeys hadn't completed, while a retargeting campaign next to it posted a 9x return. Every monthly review was a chance to lose nerve.

That is why step 3 has to come before step 4. If you move the money up the funnel without the measurement, you are asking the client to trust you. If you build the measurement first, you are asking them to trust the data, and the data turns up.

The platforms are not going to do this for you. They are paid on the inner circles. The outer one is yours to find, and the budget to find it is already in your account, being spent on people who were coming anyway.

Playbook: Paid media as a creative test lab

Dom Carter — Tue, 20 Oct 2026 00:00:00 GMT

Luxury is about consistency. Culture is about change. The brief we were handed was to build a brand fluent in both, and the two audiences involved wanted opposite things.

The brand is a Swiss performance cycling label with a long heritage and a customer base that is loyal to the point of being territorial. Those customers did not want change. The brand also needed a younger, more progressive customer it had never really spoken to, and that customer wouldn't come without change. Retain one, attract the other, lose credibility with neither. Any CMO who has tried to do this knows it is elegant to describe and brutal to execute.

This is how we did it, and what I'd tell anyone with the same problem.

The mindset shift before the work

The existing audience didn't want change. The new audience wouldn't come without it. There is no creative brief that resolves that, because the answer isn't a campaign, it's a process.

So we replaced assumption with experimentation. Test everything, be analytical about it, learn fast, and scale what resonates. Written down that sounds like every agency deck ever produced. The difference is what we tested on, which was paid media, and what we measured, which was not sales.

The framework we worked to had 5 parts, and I'd keep all of them.

Identify the segments and what they value. The new audience was defined by values and life, not demographics.
Define creative hypotheses. Reels, stills, carousels, each carrying a specific guess about what the new audience would respond to.
Experiment through feedback loops, using paid mechanics to put each hypothesis in front of the right people quickly.
Amplify what works, not what's loud. Repeat.
Leave what doesn't behind. Don't repeat.

Number 5 is the one organisations find hardest. A piece of content somebody senior loved, that tested badly, still has to go.

Rebuilding the content engine

The old model was a top-down content calendar. Decide in January what the brand will say in June, produce it, publish it, hope. One brand voice, applied to everyone.

We stopped asking "what should we say?" and started asking "what are they reacting to?"

In practice that meant weekly creative drops rather than quarterly campaigns. Structured feedback from the metrics that show real reaction, which are saves, direct messages and comments, rather than likes. And fast iteration: what worked this week shapes what we make next week, not next quarter. The brand's own marketing lead rebuilt the internal process to support that cadence, which is the part agencies tend to skip over when they present the results.

Paid media as the test lab

Here is the part that I think is genuinely useful beyond cycling.

We used paid campaigns to test message and aesthetic. The metric wasn't return on ad spend. It was resonance.

The mechanism is simple once you see it. The platforms charge less to show people content they respond to, because responsive content keeps users on the platform. Run 6 pieces of creative with the same targeting and the same budget and they come back with 6 different costs per thousand impressions. The cheapest one is the one the audience is telling you it prefers. It is the fastest honest feedback a brand can buy, and it costs a fraction of a campaign.

The combinations with the cheapest costs per thousand became the ones we used to drive new awareness at scale. Paid stopped being a performance channel and became the insight function. The sales still happened, but they were the result of the test, not its purpose.

Two things to say about this. One, it only works if you are disciplined about holding targeting and budget constant, otherwise you're measuring the media buy, not the creative. Two, it changes the conversation with finance, because a campaign that "lost money" might have found the message that made the next 3 campaigns work, and you need a way to say that before the review rather than after.

Creators as participants, not adverts

The new audience was not going to arrive through the brand's own voice, however well tested. It needed voices that already belonged to that audience.

We introduced real creators, not influencers, and the distinction matters. UK-based female cyclists, chosen for alignment with the audience we wanted rather than for reach. Not millions of followers. People whose life on a bike looked like the life of the customer we were trying to reach.

They weren't promoting. They were participating. The storytelling was co-created rather than handed down in a templated brief, which meant it sometimes went places the brand wouldn't have chosen and that was the point.

One creator run produced more than 500,000 organic impressions, more than 2,500 new engaged followers, a revenue uplift and a spike in site activity. The comment sentiment was curiosity and admiration with conversion signals in it. One comment said: "This is the first time a cycling brand felt like me."

That line is the whole case. Emotional resonance showed up as commercial impact, in that order.

The weekly rhythm, if you want to copy it

Monday, look at last week's numbers: cost per thousand by creative, saves, comments, shares, and the direct messages the social team logged. Decide which 2 or 3 hypotheses survived.

Tuesday to Thursday, make the next drop. Three to six pieces, each a variation on something that survived or a fresh guess at something the segment might value. Keep production light, because most of it will be discarded and expensive content is hard to kill.

Friday, the drop goes into paid with targeting and budget held constant across every piece, so the only variable is the creative. Small spend. Enough impressions to read a cost per thousand with confidence, not enough to matter if the whole drop fails.

Then the following Monday you do it again. After about 8 weeks the pattern is hard to miss, and the pieces that keep winning become the awareness campaign, with real budget behind them, to the new audience only.

The core audience sees the proven work later, or not at all, depending on whether it fits them. That is the whole protection mechanism and it costs nothing but discipline.

Protecting the core

None of this would have been worth doing if the loyal customers had noticed and left.

Heritage wasn't sacrificed. It was strengthened by being recontextualised. The legacy creative was evolved rather than erased. The performance-first tone was kept for the core audience, who kept seeing the brand they loved. Creative divergence was applied only where it was strategically needed, which meant the new-audience work ran in paid to the new audience, and the core saw almost none of it until it had been proven.

We didn't break the brand. The marketing lead would tell you that the brand's own people were the harder audience to carry than the customers, and I'd agree.

What changed

Before and after, in one view.

From a top-down content calendar, to control-group testing through paid ads.
From a single brand voice, to multi-dimensional storytelling with creators carrying part of it.
From assumption, to insight.

The outputs got better. The process change is what made them better, and it's the process that transfers to any brand with the same tension.

What I'd tell every CMO in the room

Test the emotions, not just the formats. Feedback is your new influence, and the audience will tell you what they want if you build the loop to hear it. Creators aren't shortcuts, they are strategic voices, and choosing them for reach is how you end up with expensive content nobody believes. If you aren't iterating, you're stagnating, however good the brand book is.

And the one I'd put on the wall: data doesn't kill creativity. It sharpens it. Every creative director who has watched a piece they didn't rate come back with the cheapest cost per thousand on the account has learned something they couldn't have learned in a room.

The brand didn't abandon its past. It future-proofed its relevance. That isn't rebranding. It's reinvention, done in weekly increments, with the audience voting on every one.

Playbook: Where AI earns its place in ecommerce content

Dom Carter — Mon, 19 Oct 2026 00:00:00 GMT

Open the men's ready-to-wear section of a luxury fashion site and count the products. When we did this for one house in 2025 there were 588 in a single category. Then open a few product pages and look at what each one actually carries.

The studded leather jacket near the top of the range had everything. Product shots from several angles, a model wearing it, campaign imagery, video. The down jacket a little further down had product shots and a model. The Chelsea boots and the holdall had product shots on a white background and nothing else. No model, no motion, no context.

That pattern is not unique to the house we looked at. It is what every large catalogue looks like, and it is the shape of the opportunity.

The three tiers

Every big ecommerce catalogue sorts itself into 3 tiers of content, whether or not anyone decided it.

Top tier. The hero products and the new collection. Product, model, video and campaign imagery. These pages get the creative director's attention, the shoot budget and the homepage. They are a small share of the products and a large share of the brand.

Secondary. Product and model stills shot in the studio, sometimes with campaign or lifestyle imagery alongside. Good pages. Not special ones.

Tertiary and accessories. Product-focused imagery only. Flat shots, often excellent, of a product on white. No model interaction, no video, no context for what it looks like on a person or in a life.

The pattern makes sense from a production point of view. A shoot costs what it costs, so it goes on the products that justify it. The problem is that revenue doesn't follow the same curve. Individually, a tertiary product sells less than a hero. In aggregate, the long tail is a large part of what the site turns over, and every one of those pages is asking a customer to spend real money on a product they can see less of than they would in a shop.

Where AI earns its place

Here is where I'd point generative tools, and where I wouldn't.

Not at the top tier. The campaign imagery is the brand. A luxury house's hero content is the result of a creative team's judgement, and the fastest way to make a luxury brand look cheap is to let a model generate the bit that was supposed to be art.

At the tertiary tier, the choice is different. Nobody is going to shoot the holdall on a model. The alternative to AI there is not a photographer, it is nothing. So the question becomes whether a model-on-body image generated from the existing product shot, or 4 seconds of motion that shows how the bag hangs, or a lifestyle context around a product that only ever existed on white, moves more people from the page to the basket than the flat shot does on its own.

That is a conversion question, and conversion questions can be tested.

The specific things we proposed, because they already existed as content types on the top-tier pages and so had a known effect on the customer, were these.

Model-on-body imagery for products that only have flat shots, generated from the product shot itself so the garment is right.
Short motion from stills for products with no video, enough to show drape, texture and how the product moves.
Lifestyle context for accessories, placing the product in the kind of setting the campaign imagery uses for the hero range.

In each case the point is to raise the content on the pages that currently receive less production attention towards the standard of the pages that receive more, without an additional shoot or additional physical production.

How to run it

Treat it as a conversion programme with a brand gate, in this order.

Audit. Pull the product list with the content each page carries: image count, model shot yes or no, video yes or no, campaign imagery yes or no. Add revenue and sessions per product. You now have a map of where the money sits relative to the content, and the gap between the two is the list of pages worth working on first.

Set the bar. The brand's, not the tool's. Decide who signs off AI-assisted imagery and make it the same person who signs off a campaign image, against the same standard. If a generated image fails on fabric, fit or finish, it doesn't ship. You will learn quickly which product categories the tools handle and which they don't, and in 2025 the honest answer was that knitwear and anything with a complex drape were still hard, while structured bags, shoes and outerwear were workable.

Test by tier. Pick a tertiary set, add the new content to half the pages, leave the other half alone, and measure conversion and return rate over a sensible period. Return rate matters because content that misrepresents a product converts once and comes back. Then do the same on a secondary set, where the uplift will be smaller because the pages were already decent.

Scale what passed. Roll the content types that beat control across the tier they beat it in, and keep the tiers that didn't move as they were.

The three things that go wrong

Disclosure. If a customer finds out the image of the jacket on a person was generated, after they bought it, you have a trust problem on top of a return. Say it, plainly, where the image sits. Most customers care far less than the brand team fears, provided they weren't misled.

Rights. A generated model is not a model with a contract, but the product shot it was built from has a photographer, and a lifestyle context built from reference imagery has a source. Keep the chain clean, because the saving on the shoot is not worth a dispute with a photographer.

Tone creep. The cheapest tier is the biggest tier, and once it carries generated content the house risks having most of its product pages set by a tool rather than a creative director. The gate exists to stop that. The hero range sets the standard the generated work has to reach, not the other way round.

Why this is the AI argument I actually believe

Most of what gets pitched to brands as AI creative is a way to make the hero content cheaper. I don't think that works for luxury, and I'm not sure it works for anyone whose brand is the point. The version that works is less glamorous: find the hundreds of pages production will never reach, bring them up to the standard the customer already expects from the rest of the site, and measure whether it sold more.

It is a conversion rate project with a creative director holding the door. That is a smaller claim than "AI will transform your content" and it has the advantage of being true.

ChatGPT ads will take your customer list

Dom Carter — Sun, 18 Oct 2026 00:00:00 GMT

OpenAI started showing ads in ChatGPT in Europe in late August. Free users and the cheaper Go tier see them, paying subscribers don't. Earlier this month the company said it is testing a visual format too, which appears while the tool is generating an image for you. Ads passed $1bn in annualised revenue in under 200 days, according to The Next Web, on a product OpenAI says 1.2 billion people use every week.

That was the story. The more useful one arrived this week, when 404 Media ran two campaigns on the platform to see how targeting actually works.

It works the way it works everywhere else. You upload a list.

Custom audiences, new room

An advertiser can give OpenAI a file of email addresses, phone numbers or the advertising identifier on an Android phone. OpenAI checks which of those belong to a ChatGPT account, and the matches become your audience. You need at least 25,000 entries to build one, the company says it deletes the file within 24 hours, and it doesn't tell you which individuals matched.

If you have ever run a campaign on Meta, that will sound familiar, because it is Custom Audiences with the serial numbers filed off. The targeting machinery is the same one I wrote about when an ad slot delivered spyware, just with a different buyer at the keyboard. The list you built from your CRM, the one you already match against Facebook and Google, now has a third place to go. There is nothing novel in the mechanics and I would be wary of anyone telling you there is.

What is new is where the ad lands. On Meta it arrives in a feed, between a cousin's holiday and a video of a dog. On ChatGPT it arrives inside a conversation the person believes is private. They have asked a machine for help with a cover letter, a symptom, a break-up, a redundancy consultation. OpenAI says it checks whether a conversation is "emotionally vulnerable" before an ad appears and that ads never influence the answers. I take that at face value. It is still a different room, and most people in it have not noticed the door is open.

The opt-out that doesn't

Users can switch off personalised ads in their settings. 404 Media read the small print: that setting stops ads being chosen using your past chats and preferences. OpenAI's own wording is that ads "may still be based on your current chat." So the thing the person is typing right now remains a targeting signal, opt-out or not, and a list match on their email address sits outside the setting entirely.

I am not writing this to be outraged about it. Every platform has an opt-out that is narrower than the button suggests. I am writing it because somebody at your agency, or at your client, is going to be asked to put budget in, probably this quarter, and the pitch will lead with the 1.2 billion.

The reach is real. So is the fact that most of it is people who have never bought anything from a chatbot and have no idea an ad can find them there by their email address. Early adopters of a new placement usually pay for the learning. Decide whether that is you.

The bit that will matter in January

Here is the part of the European story that the US coverage skipped. In August the European Commission designated ChatGPT a very large online search engine under the Digital Services Act, on 159 million monthly EU users against a threshold of 45 million. From January, that brings a public advertising repository: a year of every ad that ran, with the advertiser, the dates, the targeting parameters and the reach.

Read that again with your customer list in mind. If you upload 40,000 emails to reach lapsed buyers with a win-back offer, the fact that you targeted a matched list, and roughly how many people it reached, becomes a public record anyone can search. Your competitors will search it. So will the journalist writing about which brands advertise next to vulnerable conversations.

That is not an argument against doing it. Plenty of advertisers will be fine with the record and some will be proud of it. It is an argument for deciding in advance what you are happy to have written down, because for the first time the targeting choice will be published alongside the creative.

Most of us have spent 15 years making those choices in a dashboard nobody outside the agency could see. That was never the same as nobody caring.

The checkout is leaving your website

Dom Carter — Sat, 17 Oct 2026 00:00:00 GMT

TikTok announced two things on 5 October that belong together. The first is a shopping assistant, a chat agent that answers questions about sizing, shipping and stock and remembers what you told it. The second is one-click checkout from the For You feed, so you can buy from a brand without leaving the video you were watching. TechCrunch's report says both are being built with Salesforce, Shopify, Shoplazza and Stripe.

Ten days later Google launched Universal Cart in the US, a basket that lives in search and the Gemini app, collects products from any retailer, tells you when the price drops, and checks out through a protocol Google wrote for AI agents to talk to shops.

Neither of these is a feature. Together they are the end of a 20-year arrangement.

The arrangement

Since roughly the launch of Google Shopping, the deal between platforms and retailers has gone like this. The platform owns discovery, the retailer owns the checkout. You pay for the click, the customer lands on your site, and everything from that moment belongs to you: the email address, the basket, the returns, the second purchase.

That second purchase is the whole business. Nobody in ecommerce makes money on a cold acquisition. They make it on the third and fourth order from someone they already know, bought with an email that cost nothing. The customer relationship is the asset and the website is where it was kept.

It is also where the margin was kept. A checkout you own is a checkout you don't pay a platform a percentage of, and that percentage is what every intermediary in this story is really after.

TikTok and Google are now offering to keep it for you.

What the retailers are doing about it

The Wall Street Journal looked at how merchants are responding, and PYMNTS summarised it. Three different answers from three different sizes of company.

Amazon has blocked Meta's agent and OpenAI's from shopping on its platform. It owns discovery, checkout, delivery and the data, and it has no reason to hand any link in that chain to someone else. Target lets agents in from an approved list. Walmart has gone the other way and partnered with OpenAI and Google to bring customers in through them.

BrightEdge, a search optimisation firm quoted in the piece, reckons agent purchases are growing 43% a month and have roughly quadrupled since April. The same firm says only about 3% of the ecommerce businesses it works with let an agent complete a purchase. So the demand is arriving and almost nobody has the door open. That gap is where the next 2 years of ecommerce strategy sits.

The question nobody is asking in the pitch

When an agent buys on someone's behalf, the retailer gets an order and not much else. The person behind it is represented by software that may tell you nothing about who they are. You can't retarget them, you can't email them, you may not even know whether they are new or returning. A Federal Reserve governor, Christopher Waller, put it in payments terms: the question shifts from whether the buyer is authorised to pay to whether the agent is authorised to pay for them. Chargebacks and returns follow from that.

Agencies are going to be asked to "optimise for agents", and some already are. Before anyone does, it is worth re-reading the brief nobody talked to a customer about, because an agent is the one customer you definitely never met. The honest version of that advice is uncomfortable. Yes, structure your product data so a machine can read it. Yes, let the approved agents in, because Walmart is right that the customers will arrive through them whether you like it or not. But go in knowing what you are trading. Every order that completes inside TikTok or Universal Cart is a customer you never meet.

For a brand with a strong direct relationship and a decent email list, that is a real cost, and it belongs in the forecast next to the extra revenue. For a brand that lives on paid acquisition anyway, it might be no worse than what Meta already takes.

Either way, decide it on purpose. The platforms have.

The $29 cap on the tool that was working

Dom Carter — Fri, 16 Oct 2026 00:00:00 GMT

Accenture, the consultancy that has spent 3 years telling clients to put AI at the centre of everything, has told its own staff they can spend $29 a month on Claude. 404 Media reported it this week. The cap took effect on 1 October, usage is now charged by consumption, and anyone who goes over has to find a project or team willing to pick up the rest of the bill. Some employees had been running up $500 to $1,000 a month. One said their access was cut off entirely.

In sterling that is about £22, which is roughly what a decent Manchester lunch costs for two.

It is not just Accenture. Earlier this month The Information reported that Meta and Microsoft are steering staff away from Claude too. Meta's Claude Code users have fallen from around 60,000 to 30,000, partly through layoffs, after the company spent more than $105m on the tool in a single 28-day stretch. Microsoft has cut the amount an individual can consume from as much as $100,000 a month to about $10,000 and trimmed its projected Anthropic spend by more than a third.

Before anyone enjoys the irony too much, it is worth saying what is actually going on, because it is about to happen in every agency in the country.

The leaderboard

The detail I can't get past is from Meta. Staff built an internal leaderboard, reportedly named Claudeonomics, to compete on how much AI they were using. The report calls it "tokenmaxxing". People were racing each other to consume the most.

That is what happens when you tell a workforce that the future belongs to whoever adopts fastest and then don't tell them what adoption is for. Usage becomes the metric because it is the only one that exists. Finance sees a bill with no output attached to it, and finance does what finance does.

I have sat in that meeting. It is not an unreasonable meeting. Someone is looking at a number that doubled in a quarter and asking what it bought, and the honest answer is that nobody set it up so that question could be answered.

So the cap is not stupid. The cap is the predictable result of launching a tool with no definition of what good looks like. The mistake was made 6 months before the cap, when nobody wrote down which jobs the tool was supposed to change. It is the same failure as the AI pilot that never lands, seen from the finance side.

What a $29 cap actually does

Think about who it lands on. The person spending $800 a month on Claude is, in most organisations, one of two things. Either they have found something that genuinely works and are running it hard, or they are mucking about. A flat cap can't tell the difference. It stops both, and it stops the first one harder because they had further to fall.

The people spending $29 are not the ones you need to study. The ones at $800 are, and the right response to that bill is to go and sit with them for an afternoon and find out what they are doing. If it is real, you have found a process worth rolling out and the $800 was the cheapest research you will ever buy. If it is not, you have a conversation to have, and it is a conversation about judgement, not a line in a budget.

Instead the person at $800 now has to go and ask a project lead to fund them. Some will. Most will quietly go back to doing it the old way, because asking is embarrassing and the old way still works, and nobody will ever know what was lost.

The agency version

We are at the stage where most agencies have handed out licences and are starting to get the bill. The bill is lumpy and it does not match headcount, because a handful of people use the tools properly and everyone else opens them twice a week. That lumpiness looks like a problem to a finance director and like a signal to anyone running the adoption.

If you cap it before you have looked at it, you pay twice. Once for the licences nobody uses, and once for the person who was getting somewhere and went back to the slow way because the fast one needed a form.

The ad slot that delivered spyware

Dom Carter — Thu, 15 Oct 2026 00:00:00 GMT

Last week Citizen Lab, the research group at the University of Toronto that tracks surveillance tools, published a report on something it calls AdFection. It is spyware. What makes it different from the usual kind is how it gets onto your phone. You don't click a link or open a file. An ad loads on a page you are reading, and that is enough.

Engadget's write-up has the numbers. The company behind it, an Israeli firm called Discover-Tech, presented itself as an advertising business. According to the report it displayed 442 million ads across more than 120,000 websites, including the BBC and the New York Times, and inside apps like Duolingo and Grindr. The targets were journalists, staff at NATO and the European Commission, and people at Nvidia and Arm. Google has since blocked the firm from its platforms after a joint investigation with Citizen Lab and Confiant.

If you work in media buying, the detail that should stop you is not the spyware. It is the mechanism.

It used the same auction you use

Every time a page with programmatic ads loads, an auction runs in a few milliseconds. The page offers up a slot and some information about the person looking at it, buyers bid, the winner's ad appears. That is real-time bidding, and it is how most display advertising is bought now, including the display we run for clients.

Discover-Tech didn't hack that system. It used it as designed. The report says the firm used the bidding process to profile and de-anonymise specific people, then won the auction to show them an ad, and the ad carried the code. The same machinery that lets a brand reach "people in Manchester who looked at running shoes" let a surveillance company reach "this particular person at the European Commission".

That is the bit I keep coming back to. The targeting worked exactly as advertised. It was just pointed at someone.

What brand safety was looking at

Most brand safety spend in this industry goes on checking the page. Is the article about terrorism, is the video a hoax, is the site a made-for-advertising farm. Those are reasonable questions, and the tools answer them well enough.

None of them look at the ad next to yours. None of them look at who else won an auction on the same page, or what their creative actually contained. Your pre-roll on a BBC article was safe by every measure we use, and the display unit below it was delivering spyware to the person reading. The page was fine. The pipe was not.

I don't think the brand safety vendors are to blame for that. They sold what the market asked for, which was protection against embarrassing adjacency. The market never asked whether the exchange itself could be used as a weapon, because nobody had shown that it could. Now somebody has.

The question to ask your buyer

The honest answer to "where does my money go" in programmatic is still, for most advertisers, a shrug. There is the agency, then the demand platform, then one or several exchanges, then the supply platforms, then the publisher. Each one takes a cut and each one is a place where something you didn't buy can sit next to the thing you did.

The fix is not to stop buying programmatically. The same auction is about to get a new buyer in ChatGPT, which I've written about in ChatGPT ads will take your customer list, and the supply chain question applies there too. It is one of the few channels that can reach a specific audience at a sensible price, and the alternative is paying rate card for a homepage takeover. The fix is to shorten the chain. Fewer exchanges. Direct deals with publishers you actually want to be on. Supply path reporting that names every intermediary, which the good platforms will give you if you ask and the bad ones will explain is very complicated.

Citizen Lab's senior researcher, Bill Marczak, told the Financial Times the firm had been using the advertising system this way for years, turning trusted sites into a delivery platform. Years. This was not a clever new trick that got caught quickly. It ran, on the pages your brand was paying to appear on, for a long time.

Your logo was on that pipe. The money you spent to keep it safe was looking at the article.

The brief nobody talked to a customer about

Dom Carter — Wed, 14 Oct 2026 00:00:00 GMT

You can usually tell within 2 pages whether a brief was written by someone who has talked to a customer. The ones that haven’t use a particular language. The audience is "time-poor professionals". They "value authenticity". They are "digitally native" and "seeking brands that share their values".

None of those sentences are false. They’re just not about anyone.

I’ve sat in enough pitch rooms to know how this happens. A brand team is asked for a brief by Friday. The last customer research is 3 years old and was done for a different product. So the audience section gets written from memory, which means it gets written from the assumptions of the people in the room, and the people in the room are not the audience. They’re marketers in their thirties describing someone they imagine.

The agency then builds something for that imaginary person. The imaginary person loves it. The real ones don’t notice.

The 3 questions that find it

I don’t ask clients for research any more. I ask three things.

When did someone on this team last speak to a customer who wasn’t a complaint? What surprised you? And what does the customer call the product, in their own words, when they’re not talking to you?

If the third one gets a blank look, the brief has a hole in it, and no amount of creative will fill it. A campaign built on a made-up audience can still look good. It can win awards. What it can’t do is move the number the client actually cares about, because the person it was built to persuade doesn’t exist.

Why this is the expensive one

Media is where the money goes, so media is where the blame lands when nothing happens. And media is about to get stranger, with the checkout leaving your website and agents buying on people's behalf, so the brief had better describe a real person. But by the time the campaign is live, the mistake is months old. It was made in a meeting room, with good intentions, by people who were too busy to go and ask.

Go and ask. It takes an afternoon. The alternative costs the media budget.

Kick-off times are a media product now

Dom Carter — Wed, 07 Oct 2026 00:00:00 GMT

If you want to know who a sport thinks its customer is, don’t read the mission statement. Look at the fixture list.

A Sunday 4.30pm away game in a city 200 miles from your own tells you everything. There’s no train home. The club knows there’s no train home. The broadcaster knows. The fixture went where it went because that slot pulls a bigger audience on a sofa than the 3,000 people who’ll make the trip, and that audience is the one the rights deal was priced on.

I don’t think that’s a scandal. It’s a business decision, and quite a rational one. The money that pays for the squad comes from the screen, not the turnstile. But it’s worth being honest about what it means, because clubs still talk to supporters as though they’re the point, and the schedule says they’re the backdrop.

The backdrop problem

Broadcasters need the ground to be full and loud. An empty away end on television looks like a product nobody wants. So the match-going fan is essential to the picture and incidental to the decision, which is an odd position to put your most loyal customers in.

Most industries would call that a retention risk. Football calls it tradition and assumes it’s permanent.

What marketers should take from it

Every business has a version of this. In football, the 3pm blackout is the one piece of the schedule still arranged the other way round. The customer who shows up, pays and complains, versus the customer who scales. The pressure is always to design for the one who scales. The mistake is forgetting that the one who shows up is the reason the scalable one is watching at all.

Football will find out what happens when the away end gets thin. The rest of us can learn it cheaper by looking at our own schedules and asking who they were really written for.

The robo boss now needs a witness

Dom Carter — Fri, 02 Oct 2026 00:00:00 GMT

On 30 September Gavin Newsom signed a bill called the No Robo Bosses Act. From 1 July 2027, an employer in California will not be able to discipline or sack someone on the output of an automated system alone. If software is the main reason for the decision, a human has to check it first, and not by ticking a box. The reviewer has to corroborate it against evidence that exists outside the system: supervisor evaluations, the person's own work, peer reviews, interviews. If they can't, the output can't be used.

The worker then gets a written notice, in plain language, at the moment they are told, saying a system was involved, that a human checked it, and giving them a named person who can explain it. They can also ask once a year for the last 12 months of their own data that the system used. The penalty is $500 a violation. Fisher Phillips, an employment law firm, has a clear summary for anyone who needs the detail.

Newsom vetoed a broader version last October, objecting to what he called unfocused notification requirements on any business using "even the most innocuous tools". The version he signed is narrower. It covers employees and not gig workers, it drops the rule that you name the AI vendor, and it leaves the phrase "primarily relies" undefined, which the lawyers are already enjoying.

The sponsor, Senator Jerry McNerney, said "no worker should ever be fired or disciplined by a robo boss." Fine. The more interesting question is how many of us already have one.

You don't need AI to have a robo boss

Nobody at an agency fires people on a dashboard. But look at what decides who gets the difficult conversation.

Utilisation reports. Timesheet compliance. The client satisfaction score. The number of tickets closed, the average response time, the revenue per head. These are automated decision systems in everything but the marketing. They are built from inputs nobody fully trusts, they produce a number per person, and that number sorts the team into people who get left alone and people who get a meeting.

The meeting is where the human arrives. By then the decision about who is in the room has already been made, by the report, and the human's job has quietly become delivering it. That is the thing the California law is actually aimed at, and it did not need a language model to exist. It just needed a spreadsheet and a busy manager.

What the law asks for is just management

Strip the legal language away and the requirement is this. Before you act on the number, go and look at the work. Talk to the people who sit next to them. Read what they produced. If the number and the work disagree, the work wins.

That is not a compliance burden. It is the job. It is what a decent head of department did before any of this existed, and it is what gets skipped when there are 40 people in the team and the report is sitting right there with a red cell next to one name.

The honest thing to admit is that AI makes the skipping easier, not harder. It is the flip side of the $29 cap: the same tools that get rationed for cost get trusted for judgement, and neither decision gets looked at. A system that writes the performance summary for you, in confident prose, with a recommendation at the bottom, is more persuasive than a red cell. It reads like someone already did the looking. California's answer is to require that someone actually did, and to put their name on the notice.

What I would do with it

We are not in California and the law does not reach us. I would adopt the standard anyway, for one reason that has nothing to do with being nice.

A decision made on a report alone is a decision you cannot defend when the report turns out to be wrong, and reports are wrong more often than anyone admits. The utilisation figure that didn't count the pitch work. The response time skewed by one client in a different timezone. The satisfaction score filled in by someone having a bad week. When you act on that and it is later picked apart, the fact that you had a number does not help you. The fact that you went and looked does.

So write it down somewhere. Nobody in this business gets a difficult conversation off the back of a report until a person has read the work. Then give that person's name to whoever is in the meeting.

The dashboard will still decide who it flags. The difference is that the human is now doing something before the meeting, not just at it.