The First Draft Is the Risk

Marketing teams are no longer short of copy, ad variants or channel suggestions. The risk is that platform-generated drafts, AI search summaries and automated social tools make brands sound interchangeable while leaving the business responsible for the outcome.

This week’s marketing stories were not really about better generation. They were about who gets to shape the version customers actually see. Google Ads terms, ChatGPT Ads, agent-readable websites, AI search visibility and social media agents all point to the same uncomfortable shift: production is moving outward, but responsibility is not.

The draft moved upstream

Google’s updated Ads terms make the new deal fairly plain. Platforms can automate more of the campaign, including formats, generated components, campaign assets and destinations, but the advertiser remains responsible for what appears in front of the customer.¹ That is not a minor operational update for media buyers. It changes where the first draft lives, and it changes who has to catch the mistake before it reaches the market.

OpenAI is moving in a similar direction with ChatGPT Ads. Search Engine Land reported that OpenAI has a feature that can generate ads for advertisers to review, edit and approve.² MediaPost’s wider point is that marketing is shifting from ads placed around browsing to commercial messages inside answer-seeking experiences.³ That makes the draft more powerful because it appears closer to the moment of decision. It gives generated copy the feel of helpful context, not a separate asset waiting to be ignored.

The old ad workflow had clearer boundaries. A team wrote the ad, selected the offer, picked the landing page, reviewed the campaign and then handed it to the platform for distribution. There was still plenty of automation, but the brand could usually point to a human decision behind the words. Now the system can shape more of the message before the marketer has even formed a strong opinion about what the message should be.

That is the awkward part for anyone running paid media. If the platform generates ten versions, someone still has to know which one sounds wrong. If the system proposes a destination, someone still has to know whether the destination fits the promise. If the same environment can answer the customer’s question, sell the ad slot and draft the creative, the brand needs a sharper internal standard than “review before publish”.

More output, muddier learning

Google’s test of AI-generated summaries under Search ads shows why this is not only a brand voice problem. Search Engine Land reported that some advertisers saw generated summaries beneath ad descriptions, with the summaries created independently and carrying a warning that they can make mistakes.⁴ Paid search has always been valuable because it teaches marketers quickly. You write something, the market reacts, and the data tells you what to change next.

An extra AI layer makes that loop harder to read. Was the click driven by the ad description, the generated summary, the brand name, the offer, the position or the user’s prior knowledge? Did the summary clarify the promise, or did it nudge the customer towards an expectation the landing page could not meet? Performance marketing does not become useless in this setup, but it does become noisier.

That matters because the industry has spent the last two years treating content volume as if it were a productivity metric. More variants can be useful when the hypothesis is clear. More copy can be useful when the team knows what it is testing. But if no one can explain what the machine changed, why it changed it, and what the customer actually responded to, the team has not learned much.

This is where AI content marketing can quietly become worse than old-fashioned inefficient marketing. At least the older process forced teams to make choices slowly enough to notice them. The new process can fill a campaign, calendar or landing page with polished material that teaches the team almost nothing. The first draft is not dangerous because it is synthetic. It is dangerous because it can look finished before the thinking is finished.

The web has another reader

The same problem is showing up in search. Search Engine Land reported that ChatGPT accounted for 92.4% of AI referral traffic in one study of 6.77 million LLM-driven sessions, with monthly LLM sessions in that dataset rising to 644,478 in May 2026.⁵ Other Search Engine Land coverage this week pushed marketers towards prompt-level visibility and a rethink of SEO priorities for AI search.⁶⁷ The message is blunt: visibility is no longer only a page ranking against a keyword. The old habit of measuring visibility at page level now misses too much of the buyer’s path.

Digiday reported that publishers including Time and The Economist are experimenting with versions of their sites that are easier for AI agents to read.⁸ Search Engine Land’s GraphRAG analysis points in the same direction, with AI search becoming more concerned with entities, relationships and credibility than isolated text matches.⁹ For marketers, that means the brand is not represented by a homepage alone. It is assembled from product pages, reviews, third-party articles, customer language, videos, profiles, comparison content and whatever else a model considers credible. A site that is clear to a person but ambiguous to a machine can still be misread when a recommendation is formed.

This breaks the fantasy of total control. A polished page may say one thing, while the wider web says another. A brand may describe itself as premium, local, sustainable or expert, while customers, creators and comparison pages use very different words. AI systems are not sentimental about the approved positioning deck. They will often prefer the pattern they can verify.

The practical implication is not to publish more bland pages for machines. That would be a neat way to make the problem worse. The better move is to make the business clearer in the places where people and systems already look: product information, FAQs, opening hours, menus, service descriptions, customer examples, reviews, policies and plain explanations of what the business actually does. The goal is not to write for robots. The goal is to stop forcing both humans and machines to guess.

The better use is practical

The DRESSX try-on study is a useful counterweight to the gloomier version of the AI marketing conversation. Marketing Tech News reported that DRESSX studied AI try-on use across luxury fashion ecommerce platforms, covering 1.2 million shoppers across 216 countries and 83 languages.¹⁰ The interesting point is not that the technology looks clever. It is that it attacks a real moment of hesitation: will this item look right on me?

That is a better model for AI in marketing than endless asset production. A shopper does not need a brand to generate a fictional campaign at higher volume. They need help making a decision with less uncertainty. In ecommerce, that might mean fit, sizing, colour, styling or return confidence. In Instagram content strategy, it might mean AI captions for Instagram business posts that are grounded in the actual product, not generic hype about “new arrivals”.

The Verge’s interview with Digitas CEO Amy Lanzi lands in the same place from a different angle. Her argument was not that AI has no role in advertising, but that it cannot save advertising from weak ideas or bad taste.¹¹ That should be printed above every generative workflow. AI can speed production, analyse signals and reduce dead ends, but it cannot rescue a brand that has nothing clear to say.

The strongest use cases are the ones that remove friction without removing judgement. Try-on helps a shopper answer a real question. A good planning tool helps a business keep a calendar moving without making every caption sound imported from a template. A good ad workflow helps a team test faster without forgetting what it is testing. The better machine output gets, the more visible weak human direction becomes.

The approval layer matters

The uncomfortable conclusion is that the marketer’s job is not being reduced to pressing publish. It is becoming more editorial, more systems-aware and more accountable for decisions made across tools the marketer may not fully control. That is not a small shift. It asks marketers to understand creative, data, workflow, policy, AI search, ad formats and customer trust at the same time.

This is why “final approval” is too thin a safeguard. Approval only works when the reviewer knows what they are looking for. Does the claim need evidence? Does the caption sound like the business? Does the offer create the wrong expectation? Does the generated summary change what the customer thinks they are clicking into? Does the answer engine have enough credible information to describe the brand accurately?

The better teams will build clearer approval layers before the volume arrives. They will define what must never be promised, which phrases do not sound like them, which claims require proof, which customer questions deserve a direct answer and which channels should not be automated. They will treat AI as a drafting and organising layer, not as a substitute for taste, memory or responsibility. That is how to stay on brand with AI content without turning every review process into a bottleneck.

The first draft has become cheap, fast and increasingly external to the brand. That can be useful, but it should make marketers more careful rather than less. The brand still owns what customers see, even when the platform helps create it. The advantage will belong to teams that can use the machine’s speed without letting the machine’s default voice become their own.

Sources