Context engineering for marketing

AI can write the email, summarise the call, draft the campaign, and turn one idea into twenty assets.

That's useful. It's also the easy part.

The harder question is whether the work reflects what's true now.

Your AI probably doesn't know your current positioning. Or that a competitor changed pricing last week. Or the message your latest campaign tested, the objection that keeps coming up in sales calls, or the claim on your homepage that no longer matches the product.

Faster output built on an old picture just moves you in the wrong direction faster.

I've spent around twelve years in B2B marketing, and the last stretch of it rolling AI out across an agency and its clients. The pattern is the same everywhere. Teams pour their effort into prompts and skill libraries while the context behind them goes stale.

That's why context engineering matters. Prompts are the small lever. The advantage sits in the freshness and the connectedness of the context behind the work.

What is context engineering for marketing?

Context engineering is the practice of designing and maintaining the information an AI system needs before it can do useful marketing work.

The term comes from AI engineering. Anthropic defines it as curating and maintaining the optimal set of information a model works from, and searches for the term spiked around 1,900% in 2025 as agent builders ran into the same wall marketers are hitting now. Notice the word maintaining. Even the engineers put it in the definition.

A prompt describes the task. Context explains the world that task has to succeed in.

For a B2B marketing team, that world includes

• The business model, sales motion, pricing, margins, and objectives
• The audience, jobs to be done, triggers, objections, and buying process
• The product, use cases, constraints, proof, integrations, and roadmap
• The market, competitors, alternatives, new entrants, and category narrative
• The positioning, messaging, campaigns, brand rules, and existing assets
• The buyer signals appearing in search, reviews, social conversations, and AI answers
• The performance data showing what you did and what happened next

Context engineering gives AI a maintained picture it can reason from.

Without one, the model can still produce a polished answer. It just can't know whether the answer is right for your company.

Why better prompts and skills aren't enough

Prompting matters. A clear task, output format, audience, and constraint will improve a response.

And most teams have moved past one-off prompting by now. They've built skills, packaged instructions that teach the AI how to do a job the way the team does it, shared so nobody reinvents the same prompt every Monday. Better prompts were the 2023 fix. Skills are the current one, and they're worth building.

They still don't solve this problem, though. A skill captures how to do the work. Whether the work reflects your company today depends on the context behind it.

Ask an AI to "write a differentiated homepage for our ICP", with or without a beautifully built homepage skill. If its view of your ICP comes from a deck written eighteen months ago, its answer starts from an old assumption. If it can't see that every competitor now uses the same claim, it will happily reinforce a position the market has already crowded. And if it doesn't know your pricing changed, it will write a promise the product no longer supports.

The output can sound confident and still be wrong. And a skill written six months ago carries the same risk as the deck, because it encodes assumptions the market has since moved past.

Prompts and skills are both instructions, and instructions only tell the AI what to do. Context tells it what's true. Guardrails set what it must and mustn't do. Verification tests whether the answer can be trusted. Almost all the effort I see goes into the instructions, and the bigger opportunity sits in the other three.

I've said this on LinkedIn before. If your outputs feel generic, don't tweak the prompt. Upgrade the context.

AI has made stale context dangerous

Marketing context has always gone stale, because marketing is the most dynamic function in the business. It sits closest to the market, and markets don't stand still. What's true today might not be true next month. A blue ocean this quarter is a crowded one next quarter, and a position that feels differentiated right now blends in as competitors shift and catch up.

AI turned that from a nuisance into a hazard. Working with AI is like hiring a super enthusiastic, high-potential junior employee. Enormous output, zero tenure. Give them context and they're the best hire you've ever made. Give them nothing and they'll run off and enthusiastically produce slop with no basis in reality, at a pace no human junior could match.

Before generative AI, a team might ship a campaign each month and update a handful of core assets each quarter. Misalignment was possible, but the surface area moved slowly.

Now a team can create ads, emails, pages, sales collateral, social posts, and campaign variants every day. Each asset is another place where an outdated claim, a stale audience assumption, a crowded position, or a price that changed last week can spread.

AI compounds it in three ways.

1. More volume

Every asset gets produced from the same underlying context, so get that wrong and the mistake ships fifty times instead of once.

2. More velocity

Work goes from idea to published in hours now, which leaves almost no window to notice that reality changed along the way.

3. More surface area

Marketing, sales, product, agencies, freelancers, and AI systems are all producing customer-facing material, and each one might be working from a different version of the truth.

There's a name starting to stick for this gap, and I've been writing about it for a while. Marketing drift. The distance between what's true about your company and market, and what your marketing communicates.

You'll recognise the symptom. The homepage says one thing, the ads say another, sales tells a third story, and the product changed two months ago. Nobody decided any of that. It happens because nothing in the operating model stops it.

Every era of AI raises the value of context

The first era of AI products was chat. You typed, it answered, and everything it knew about your business had to fit in the message you gave it. Stale context cost you a mediocre reply.

The second era, the one we're in now, is agents. Software that plans and completes multi-step work on your behalf. The stakes rise here, because nobody reviews every step. An agent working from an eighteen-month-old ICP comes back with a finished piece of work built on a wrong assumption.

The third era is starting to take shape. Persistent AI coworkers. OpenAI's Tara Seshan, who leads product for Codex and ChatGPT's work mode, described it on Lenny's Podcast as working "with a persistent co-worker who is able to get things done with you". As the models improve and our trust in them grows, the job shifts from rowing to steering. We delegate bigger loops and check in less often, trusting these systems to think and act for us for longer.

Each step raises the value of the context behind the system. Seshan made the point in hiring terms. An isolated agent is like "a colleague who you hire, who you lock into a room, never give them access to Google Docs and Slack and the company database". They could be brilliant and they'd still be useless to you.

Think about what happens when a person actually joins your marketing team. Their first weeks are pure context acquisition. The docs, the tools, who the customers are, what the company sells, what's been tried, what's off limits. After that, a human keeps absorbing context ambiently, in meetings and Slack threads. An AI coworker gets none of that for free. The picture it holds of your company is exactly the picture someone maintains for it.

We give these systems a longer leash as trust grows. A coworker that keeps citing last year's ICP never earns a longer one, and shouldn't.

So as these systems become more capable of acting on our behalf, the constraint moves. Intelligence is becoming abundant. A current, connected, sourced, and correctable picture of your company and market is the part that doesn't come with the model.

The seven layers of useful marketing context

A warning about scope before the list. Dumping every company document into a knowledge base gets you a bigger pile, and the pile was never the problem. The useful version is smaller, organised around the decisions marketing actually makes.

These seven layers are where I'd start.

1. Business context

How the company makes money, and what the business needs marketing to change.

Useful inputs include

• Business model and pricing
• Sales motion and buying cycle
• Revenue goals and margin constraints
• Strategic priorities
• Geographic or segment focus
• Operational and legal guardrails

Without business context, AI can recommend activity that looks good in a channel and does little for the business.

2. Audience context

Move beyond the static persona slide.

Useful audience context includes

• Jobs to be done
• Trigger events
• Problems and desired outcomes
• Objections and perceived risks
• Buying committee roles
• Language used in calls, reviews, communities, search, and support

This layer should change as buyer behaviour changes. A maintained audience view is evidence rather than an archetype frozen in a deck.

3. Product context

Your AI needs to know what the product does today rather than what the launch document said last year.

Include

• Products and use cases
• Features and limitations
• Pricing and packaging
• Integrations
• Proof and customer outcomes
• Roadmap changes that affect market-facing claims

Product context prevents overclaiming and keeps campaigns aligned with what customers can actually buy.

4. Market context

The changing external picture around the company.

It includes

• Competitor pricing, positioning, products, and campaigns
• New entrants and substitute solutions
• Category narratives and emerging vocabulary
• Buyer sentiment and reviews
• Search demand and AI visibility
• Market news that changes urgency, risk, or opportunity

This is where most static knowledge bases fail. Storing information isn't enough. The system needs to notice when it changes.

5. Brand and messaging context

The rules for how the company communicates.

Include

• Positioning and differentiation
• Messaging hierarchy
• Voice and tone
• Approved and prohibited claims
• Visual and verbal brand rules
• Existing campaigns and core assets

This layer does more than police style. It helps the system spot when different surfaces have fallen out of sync.

6. Objective and channel context

A useful recommendation depends on where and why it will be used.

Capture

• The objective
• Success metrics
• Budget and timeline
• Channel constraints
• Funnel stage
• Required formats
• Review and approval rules

The same market signal might lead to a homepage change, a sales enablement update, an ad test, or no action at all. Objective and channel context determine which.

7. Performance and feedback context

Close the loop between what you did and what happened next.

Useful signals include

• Traffic and rankings
• Conversion performance
• Campaign and ad results
• Pipeline and revenue
• Win/loss themes
• Sales objections
• Support conversations
• Qualitative customer feedback

Without this layer, the system can recommend activity but can't learn whether the activity worked.

From static knowledge base to living market context

Most teams already have most of this information. It sits in Notion, Drive, CRM records, analytics dashboards, sales calls, Slack threads, research files, campaign briefs, and somebody's spreadsheet named FINAL-v3.

I've built that spreadsheet more than once. Each version died the week I stopped having time for it.

The problem goes beyond storage. The hard part is maintenance, and I can put numbers on that. I've run more than 30 customer discovery interviews for Moso since the spring. Of the interviews I've run over the last few months, 60% raised stale or hard-to-maintain context without me prompting it. A third had built their own version, in Claude Code, Notion, n8n, or Airtable, and hit the same wall. And three people, independently, reached for the same phrase. Keeping the context current is "a full-time job". One marketing consultant gave me the full version. "You need another full-time job just building all of that stuff in AI and maintaining it."

A useful context layer needs four properties.

It stays current

The system knows when a source changed, when it was last checked, and which claims may have expired.

It stays connected

A competitor pricing change connects to your pricing, positioning, buyer objections, live campaigns, and performance, instead of sitting as an isolated alert.

It stays sourced

Every material claim points to where it came from. If sources disagree, the contradiction stays visible.

It stays correctable

You can confirm or correct the system's understanding, and your corrections persist instead of disappearing with the next prompt.

This is what turns a repository into an intelligence layer.

How the decision loop works

Context engineering earns its keep when it changes a decision. The loop looks like this.

1. Observe

Watch the company, competitors, buyers, category, visibility, and performance for meaningful change.

2. Verify

Check the evidence. Keep the source and date attached, along with anything that contradicts it.

3. Connect

Relate the change to your product, positioning, audience, campaigns, and results.

4. Assess materiality

Decide whether the change is urgent, important, worth watching, or irrelevant to the business.

5. Recommend

Propose a response with the reasoning, assumptions, evidence, and confidence visible.

6. Review

Apply strategy, taste, timing, and reputational judgment. This step belongs to a marketer.

7. Act and learn

Ship the work, watch the result, and feed the outcome back into the context.

What comes out the other end is a better-informed play rather than more content.

An example: a competitor changes pricing

Say a competitor moves from predictable monthly pricing to usage-based credits.

A tracking tool can detect the page change. A summary tool can explain the new model. Neither answer is enough.

A system with maintained context asks what a good marketer would ask.

• Does your product offer predictable pricing?
• Do your buyers care about cost certainty?
• Are reviews or social conversations showing frustration with credit models?
• Does your homepage currently make pricing predictability clear?
• Are active ads using a message that now has more market relevance?
• Has this competitor move shown up in conversion or sales conversations?

The recommendation might be to claim predictable pricing more directly, starting with the homepage and following through the ads and the sales narrative.

Or it might be to do nothing, because the affected buyer segment isn't yours.

The market change is the same. Company context changes the play.

AI-enabled and AI-native marketing are not the same

An AI-enabled team uses AI to do familiar tasks faster. It drafts the email, summarises the meeting, creates campaign variants, and researches an account. The workflow stays mostly the same, with a faster tool inside it.

An AI-native team redesigns the operating model around maintained context and continuous decision loops.

Rather than opening a blank chat every time something gets difficult, it builds systems that know the business, watch for change, apply guardrails, show their sources, and improve as the team corrects them.

I've written about this split before. AI-enabled teams accelerate the old job. AI-native teams rethink it.

None of this means automating everything. The aim is to strip out the repeated monitoring and synthesis work that keeps marketers from applying judgment where it matters.

Why fragmented copilots don't create shared intelligence

The previous era of martech produced a point solution for every use case. The AI era is adding a copilot to each one.

That may improve individual tools. It doesn't solve the fragmented picture.

Your advertising assistant knows the ad account. Your CRM assistant knows the pipeline. Your SEO assistant knows rankings. Your social assistant knows engagement. Each can answer a narrow question inside its own boundaries.

Marketing decisions cross those boundaries.

A drop in campaign performance might trace back to a competitor move, a pricing change, a shift in buyer language, a broken landing page claim, or a category narrative that moved. No isolated copilot can see the full chain.

I've made this argument about fragmented martech before. Twenty disconnected AIs are still twenty disconnected views.

The fix is one maintained intelligence layer that the tools and people share. That doesn't require a single tool that executes every task.

Keep the human in the control layer

Context engineering should make AI more useful. It shouldn't make marketing autonomous by default.

Marketing is full of decisions that resist a clean pass/fail test, like whether the idea is distinctive, whether the audience will care, whether the timing is right, whether the message feels credible, and whether the upside is worth the reputational risk.

AI can watch more sources than a person, connect evidence across systems, identify patterns, and recommend a response. A marketer still makes the call.

I've argued that AI should do the work, not just help. Marketing still needs a human in the loop, because taste and judgment can't always be tested against a simple number.

The number of people thinking stays the same. What shrinks is the time spent monitoring tabs, moving information between tools, rebuilding the same context from scratch, and working out which change actually matters.

How to audit your marketing context

Before you add another AI tool, assess the context your team already has.

Ask five questions.

1. Is it complete?

Can the system see the business, audience, product, market, brand, objectives, and performance, or only one slice?

2. Is it current?

Can you tell when each important claim was last verified? What happens when pricing, positioning, or buyer behaviour changes?

3. Is it connected?

Can an external market change be linked to your own campaigns, positioning, buyers, and results?

4. Is it sourced?

Can a marketer inspect the evidence behind a claim? Are contradictions visible?

5. Is it correctable?

Can the team amend the system's understanding? Do those corrections persist?

If the answer is no to several of these, a new generation tool won't solve the underlying problem. It will produce from the same weak foundation faster.

Where the next advantage comes from

AI has made average production cheap. That moves the advantage somewhere else.

The teams that win will hold a clearer picture of reality than their market, and they'll act on material changes sooner. Volume alone will decide less and less.

That takes more than a prompt library or a folder of skills. It takes living company and market context, connected performance feedback, visible sources, durable corrections, and a human control layer.

This is what we're building at Moso.

It starts with your URL, builds a market graph around your company, and surfaces what changed, why it matters, and the plays worth making. Every rendered claim carries a source. Your corrections outrank the machine. Your team decides what ships.

Build your market graph and get your first briefing the same day.