Semrush Spotlight: What OpenAI's marketing workflows reveal about the future of marketing

Being able to generate a hundred social assets before breakfast doesn’t mean a marketer should. That was one of the most practical points Leila Woodington, an OpenAI marketing leader, made at Semrush Spotlight in London, and it sat inside a larger argument. She said she was “redesigning work around the realities of the tools that we have today”: rather than fitting another chatbot into the old process, reconsidering both the work and the team structure around it.

She described a lean operation combining people, agency partners, AI-assisted creative iteration and shared data. That is her account of how OpenAI’s marketing works, not an independent audit of its results. But it raises a question I have been testing on a much smaller scale through Editorial Intelligence, a one-person publication and experimental editorial knowledge system: what would it mean to organise editorial work around AI, rather than simply use AI to write faster? The difference in scale is the point of the comparison. It shows which principles might transfer and which depend on money, specialist teams and organisational authority.

Adding AI to the workflow is not the same as changing the workflow

The familiar adoption story starts with a task: writing a blog, producing a campaign brief, preparing social copy or summarising a meeting. Someone discovers that an AI assistant can do part of it faster. The team saves time, perhaps produces more, and calls that transformation.

Sometimes it is a genuine improvement. There is no reason to romanticise work that can be done accurately with less effort. But the sequence of decisions often remains untouched. The brief is still assembled from scattered material, previous research is still hard to find, reviews still repeat old arguments and each campaign still starts with another search through documents and conversations.

If AI only accelerates the production stage, it can amplify those weaknesses. More material moves through a system that has not become any better at knowing what is true, what has already been said or why the next piece needs to exist.

The alternative is to ask a different set of questions. Where does the organisation’s knowledge live? How does a new signal become a decision? Who decides whether something deserves publication? How does a correction travel through related material? What survives when a person leaves, an agency changes or an AI model is replaced? None of these is a question about prompts.

What I built for Editorial Intelligence

Editorial Intelligence began as a place to publish ideas about editorial strategy, research, AI and the changing value of content. As the body of work grew, I encountered a familiar problem: the site remembered the finished articles, but not necessarily the reasoning that connected them.

A new observation might belong in an existing argument. An earlier claim might need qualification. A useful example might apply to several pieces. A normal publishing process encouraged me to write the next article and move on, because reconstructing the relationships was expensive. So I started changing the system around the work.

Today the public website and a separate knowledge base, which I call Editorial Intelligence Core, live in two version-controlled repositories on GitHub, which keeps a dated record of every change. Core holds the methodology, evidence, research and working decisions; my personal working context is kept apart again. None of this is a sophisticated enterprise data platform. It is a deliberate boundary between what I know, what I am still working out and what I choose to publish.

Each article carries structured information alongside the text: whether it is published, when, what it covers and which other pieces it relates to. Both repositories also include written instructions telling AI assistants where the authoritative material lives and what to do when a new idea arrives or an existing piece needs substantial revision.

My practical sequence is no longer simply idea → draft → publish. It is closer to:

New signal → recover relevant context → inspect the existing body of work → decide whether to extend, correct, merge, split or create → exercise editorial judgement → implement and check → publish → retain the history.

Not every stage is automated, and describing the process as an autonomous publishing engine would be misleading. I make the decisions, direct the assistants and check what changes. The system makes the context easier to recover and the implementation easier to manage, and that has changed my work more than fast first drafts ever did.

The first comparison: shared intelligence versus shared files

In a September interview with Tim Healey for The Drum and Little Grey Cells Club, Woodington described AI as “an intelligence layer across everything you do”. At Spotlight she made that concrete, describing a central intelligence layer connecting CRM, email marketing and data dashboards so her team could share customer and campaign context. A marketing organisation working across regions, campaigns, agencies and products needs people to make decisions from sufficiently consistent information. Without that, AI can accelerate the creation of incompatible versions of the same story.

My small-scale equivalent is not a central data warehouse. It is a source-of-truth rule. Core owns the working knowledge and methodology; the website owns the public version. An assistant working on the site is instructed not to treat a finished page as the only source of the underlying thinking.

That separation has a practical consequence. If I ask whether a conference observation should become an article, the right first step is not necessarily to draft it. It is to inspect what I’ve already published and decide whether the observation adds something. If an article already owns the argument, improving that article may be the better editorial decision.

A shared folder of documents would not automatically make this happen. Nor does putting material into GitHub guarantee that a model retrieves the right evidence. The value comes from deciding which source is authoritative, how it should be used and when a human must resolve a conflict.

At enterprise scale, the challenge becomes harder. Sales, product, legal, PR and editorial teams may each have valid but different accounts of the same issue. A central layer cannot simply erase those differences. Someone has to decide who owns the claim, which version is current and what can be used externally, and that governance has to come before another content-generation interface, not after it.

The second comparison: redesigning decisions, not just deliverables

One of the more revealing things in Woodington’s account was the idea of organising teams around available AI capabilities. It implies that job descriptions, handovers and the boundaries between internal teams and agencies are open to reconsideration, which is a much larger proposition than asking every employee to use an assistant.

Her own starting point was more modest than that sounds. She advised the room to begin with one repeated, painful workflow, and described automating her team’s reporting, which she estimated saved roughly six hours a week across the team while a person still reviewed the output for accuracy.

For my own publication, the corresponding change was editorial routing. Before AI, I could theoretically have reread my archive whenever a new thought arrived. In practice, the coordination cost meant I often wouldn’t, and it was easier to publish another piece. Now I can interrogate the existing body of work, compare arguments and ask whether a proposed article is genuinely distinct. Sometimes the outcome is a new piece. Sometimes it is a correction, a revised section or a decision to do nothing.

I saw the difference during an earlier round of site revisions. A thought about the value of video turned out to belong in an existing article rather than a new post. Another observation exposed a weakness in how I described the publishing cycle, so I revised that essay instead. Then I found that the site needed a better way to show material revisions without pretending an old article had just been published for the first time, and that became a change to the publishing model itself.

The AI did not supply the editorial reason for any of those changes. It reduced the work required to test and implement them while I was still thinking. This is where I find the phrase AI-native most useful: a step that used to be too cumbersome to do routinely becomes part of the normal process.

The third comparison: what humans are still responsible for

It would be easy to turn the OpenAI session into a story about smaller teams producing more, but that would leave out the difficult part. When the cost of generating material falls, the relative importance of deciding what deserves to exist can rise.

At Spotlight, Woodington described a shift from “trading in deliverables to a world of trading decisions”, an idea she attributed to a speaker she had heard at Cannes. That was the context for the hundred social assets before breakfast. Someone still has to distinguish a useful insight from a plausible sentence, a customer observation from a general claim, and a verified outcome from a marketing aspiration.

That is why I use different AI environments for different parts of the work. ChatGPT is often where I develop and challenge an emerging argument. Tools that can work directly inside the repositories can then inspect the actual files, test whether an idea duplicates something, propose changes and leave a visible record. Neither has editorial authority simply because it can produce convincing text.

The separation is intentional. I don’t want a system that agrees with me more efficiently. I want one that makes it easier to encounter the evidence and discover when my preferred argument does not hold.

There are limits. Models can misread sources, miss relevant files, invent connections and confidently improve prose while weakening its meaning. A version-controlled mistake is still a mistake, and structured metadata can make a publication easier to maintain without making its claims true.

Woodington put one boundary particularly sharply in the September interview: “AI doesn’t have taste.” It was a point about the importance of decisions when producing deliverables becomes easier. I would add that taste alone is not enough. Editorial judgement also needs evidence, accountability and the willingness to correct a persuasive but mistaken claim.

Human judgement therefore moves into the architecture of the process: deciding which evidence counts, when to stop, what must be checked and who is accountable for the published result. In a larger organisation, this also includes permissions, privacy, legal review and conflicting stakeholder interests. Those are not irritating legacy steps that AI-native design can wish away. They are part of the work that needs designing.

What doesn’t transfer

There is a risk in treating the practices of an AI company as a universal template. OpenAI operates in an unusual environment: it develops the technology, attracts exceptional attention and can organise specialist people and partners around it. A typical B2B marketing team cannot reproduce those conditions by buying subscriptions.

My own example has an important limitation too. I can decide where Editorial Intelligence’s knowledge lives because I am the owner, editor and publisher. I do not have to negotiate with ten departments about who may change a claim, and I can make an editorial decision and alter the site in the same session.

That freedom is a useful laboratory, but it is not proof that the model would work in a multi-author company. Shared ownership introduces problems my current system has not solved at scale: permissions, audit requirements, conflicting evidence, approval delays and the question of who maintains the knowledge when the initial enthusiast moves on.

The lesson is not to copy OpenAI’s structure, or my repositories. Woodington’s own advice points the same way: start with one repeated, painful workflow, find where it loses context or makes people repeat work, and test whether changing the structure improves the decision. For one team, that might be an evidence register behind customer stories. For another, it might be a reliable record of product claims and their owners. For a third, it might be a way to check whether new thought leadership repeats an argument already made. The smallest useful change often teaches more than an elaborate AI transformation programme.

What I took away from the conference

I left the session less interested in how quickly an AI-assisted team can make an asset than in how it decides what to make, which information to trust and where human attention belongs. Woodington’s line from the September interview, “People change slowly, even when technology changes fast.”, is usually read as a warning about adoption. I think it also says where the value sits. The machinery changes quickly, but trust in what an organisation says is earned slowly, through claims that hold up and corrections that reach every place they need to.

That is what my Editorial Intelligence experiment keeps teaching me. I haven’t proved that it scales across a large company, or that it improves revenue, conversion or AI citations. What I can show is a working editorial architecture and a different pattern of decisions: I return to arguments, find the evidence behind them, change my mind and update the right part of the publication instead of adding another piece.

If a team’s honest answer to what AI has changed is that it can publish more, the work has become faster. Whether it has become better is a separate question, and the one worth measuring: whether people find trustworthy evidence more reliably, whether fewer unsupported claims reach publication, whether previous research is reused and whether reviews settle decisions rather than reopen them.

---

This article draws on a marketing session at Semrush Spotlight London on 13 October 2026, Tim Healey’s September 2026 interview with Leila Woodington, and my documented Editorial Intelligence publishing experiment. Direct quotations are drawn from both the September interview and my recorded Spotlight session transcript; the source of each is named in the text. Descriptions of OpenAI’s operating model are attributed to Woodington’s accounts, not independent verification of internal performance. For the implementation detail behind my own system, see How to build an AI-native publishing system and I could not have published like this before AI.