Almost every CMS vendor now describes its platform using the same vocabulary.

AI-powered.

AI-first.

Agentic.

Intelligent.

Autonomous.

Contextual.

The terminology has changed so quickly that it is becoming difficult to distinguish a genuine architectural shift from a product team adding a chatbot to the editorial interface.

Adobe now describes Experience Manager Sites as an agentic CMS. Sanity positions itself as a Content Operating System for the AI era. Sitecore promotes an AI-powered platform with pre-built agents, while Contentful increasingly discusses agentic architecture, AI governance and AI-driven content operations. Optimizely positions Opal as an agent and orchestration platform across the digital-experience lifecycle.

Every major vendor has an AI story.

They are not all telling the same story.

AI-assisted is not the same as AI-native

The first wave of CMS AI features was relatively simple.

A text box was added to the editing interface.

The editor could request:

• a headline;
• a summary;
• a translation;
• an SEO description;
• or a different tone of voice.

These features were useful, but they did not fundamentally change the CMS.

The AI was effectively an external writing assistant embedded in the product.

It did not necessarily understand the content model, the site structure, the workflow, the audience, the experimentation history or the organisation’s governance.

An AI-native CMS should go further.

AI should be capable of operating on the same structured content, permissions and workflows as the rest of the platform.

It should not merely generate text.

It should understand what the text represents, where it belongs, how it may be used and what must happen next.

The eight tests of an AI-native CMS

Rather than comparing lists of AI features, buyers should evaluate the architecture beneath them.

1. Does the AI understand structured content?

A generic model sees words.

A useful CMS agent should also understand:

• content types;
• field definitions;
• validations;
• references;
• taxonomies;
• localisation;
• variants;
• and publication state.

Ask a vendor to demonstrate what happens when an agent creates content.

Does it return an unstructured block of prose for an editor to copy and paste?

Or can it create a valid content item that respects the schema?

Can it distinguish between a product, campaign, article, location and regulated statement?

Can it identify required fields and explain why content is invalid?

Structured content is the foundation of reusable digital experience. It is also essential for making information consistently machine-readable to downstream agents and AI-driven discovery systems.

2. Is AI embedded in the workflow?

A CMS becomes more interesting when AI can participate in the content lifecycle.

That includes:

• planning;
• briefing;
• creation;
• enrichment;
• review;
• translation;
• approval;
• publishing;
• experimentation;
• analysis;
• and optimisation.

A disconnected generation tool may improve copywriting speed.

An embedded agent can help move work through the system.

For example, it might identify missing metadata, create a draft, request the correct review, flag a legal risk and prepare regional variants.

The critical question is whether AI operates inside the workflow or merely beside it.

3. Does it respect permissions and governance?

Many AI demonstrations are performed using an administrator account.

That proves very little.

A genuine enterprise implementation must behave differently for different users.

A regional editor should not be able to instruct an agent to modify global content merely because the natural-language request sounds reasonable.

A contractor should not gain access to sensitive campaign plans through a summary tool.

A content agent should not publish regulated claims without approval.

The CMS must apply its permission model to AI operations just as rigorously as it applies it to human operations.

Governance should include:

• role-based access;
• approval requirements;
• brand rules;
• legal constraints;
• audit logs;
• model and agent versioning;
• and traceable outputs.

Optimizely, for example, has introduced agent versioning, execution logs and expanded roles within Opal, illustrating the direction in which enterprise agent governance is moving.

4. Can the AI take action?

There is a significant difference between recommending an action and performing one.

An assistant may suggest that an outdated article should be updated.

An agent can:

1. retrieve the article;
2. identify the obsolete sections;
3. draft replacements;
4. update the content item;
5. preserve the previous version;
6. initiate review;
7. and notify the owner.

This is where the word “agentic” should be tested carefully.

Ask what tools the agent can actually use.

Ask which operations are read-only.

Ask which actions are reversible.

Ask whether human approval can be inserted before publication.

Ask how failures are handled.

If the agent can only generate text, it may be AI-powered, but it is not meaningfully agentic.

5. Can it work across the wider digital-experience stack?

Content does not exist in isolation.

Its value is connected to:

• audiences;
• customer data;
• commerce;
• analytics;
• experimentation;
• search;
• personalisation;
• and campaign operations.

A stronger AI architecture can reason across these areas.

For example:

Identify pages with declining conversion, review their current audience targeting, recommend content changes and create an experiment.

That workflow requires more than CMS access.

It requires connected operational context.

Adobe is positioning Agent Orchestrator as a coordination layer across Adobe and third-party agents. Sitecore promotes agents spanning content and marketing workflows, while Optimizely is developing Opal across CMS, experimentation and orchestration scenarios.

The most valuable AI story may therefore belong not to the best text generator but to the platform with the richest connected context.

6. Is the architecture open?

Organisations should be cautious of AI architectures that only function with one model, one interface or one vendor-controlled ecosystem.

Questions to ask include:

• Can we choose or change the underlying model?
• Can agents use external tools?
• Can external agents access governed CMS capabilities?
• Is MCP supported?
• Are APIs available for agent operations?
• Can we export execution logs?
• Can we integrate our own knowledge sources?
• Can we build custom tools and agents?
• Can we apply our own security controls?

An AI-native CMS should not become a new form of lock-in.

The model market will continue to change rapidly. Organisations should be able to evolve their model and orchestration strategy without replacing their content foundation.

7. Can the AI observe outcomes?

Content generation is not optimisation.

A platform cannot claim to optimise content unless it can observe what happened after publication.

This may include:

• engagement;
• conversion;
• experiment performance;
• search visibility;
• generative-engine discoverability;
• audience behaviour;
• and commercial outcomes.

The strongest systems will create feedback loops.

An agent produces or modifies content.

The platform measures the result.

The agent compares performance against the objective.

The system recommends or performs the next action.

Optimizely has a natural advantage in this area because experimentation is part of its broader platform. Its 2026 product updates, for example, include contextual multi-armed bandits using Opal to reallocate traffic based on observed response data.

Whether every organisation should allow continuous autonomous optimisation is another question.

But the underlying loop—create, test, learn, improve—is more meaningful than generation alone.

8. Can the vendor explain the operating model?

One of the simplest ways to test an AI product is to ask the vendor to explain exactly how it works.

What context is sent to the model?

Where is that context stored?

Which model processes it?

Is customer data used for model training?

How are credentials managed?

How long are prompts and outputs retained?

What happens when a tool fails?

How are hallucinations detected?

Which actions require approval?

How are model changes tested?

How can an administrator investigate a problematic result?

An “AI-native” platform should have credible answers.

Vague statements about responsible AI are not an operating model.

A practical CMS evaluation scenario

Instead of asking vendors to demonstrate generic content generation, give each one the same enterprise scenario:

A product has been withdrawn in three markets. Find every affected experience, prepare appropriate replacement content, preserve required legal language, update structured records, route the changes to regional reviewers and show how success will be measured.

This scenario tests:

• content discovery;
• semantic understanding;
• structured content;
• dependencies;
• permissions;
• localisation;
• workflow;
• auditing;
• and analytics.

It is much harder to fake than generating a campaign headline.

Where the major vendors are differentiating

The market is beginning to separate into several AI strategies.

AI as a creative and experience ecosystem

Adobe is connecting content, assets, customer data and agent orchestration across the Experience Cloud. Its strength is the breadth of that ecosystem.

AI as programmable content infrastructure

Sanity emphasises structured content, schema-aware automation and agent access to its content backend. Its positioning is particularly strong for developer-led and highly composable environments.

AI as composable content operations

Contentful is building around structured content, hosted MCP access, AI actions and agentic workflows within a composable platform.

AI as unified marketing operations

Sitecore is positioning agents across content, personalisation and marketing workflows, including pre-built agents and an Agentic Studio.

AI as experimentation-aware orchestration

Optimizely’s strongest opportunity is connecting content creation with experimentation, personalisation and measurable optimisation through Opal.

None of these approaches is inherently superior for every organisation.

They reflect different architectural heritages and commercial strategies.

The Optimizely and Opal angle

Optimizely should resist competing solely on the number of available agents.

Agent counts are easy for competitors to match.

The stronger differentiator is the relationship between content and evidence.

Opal can potentially operate across:

• CMS content;
• campaign planning;
• experimentation;
• analytics;
• personalisation;
• and optimisation.

That allows Optimizely to make a more credible argument than “AI helps you create more content.”

The better proposition is:

AI helps you decide what content to create, governs how it is produced, tests whether it worked and improves the next decision.

That is much closer to an AI-native digital-experience platform.

Final thoughts

The CMS market is entering an awkward phase in which every product sounds almost identical.

Every platform has agents.

Every platform generates content.

Every platform promises productivity.

The differences become visible only when buyers look below the interface.

A genuinely AI-native CMS should understand structured content, participate in workflows, respect governance, use tools, observe outcomes and operate across the wider digital-experience ecosystem.

The winning platforms will not be the ones that add AI to the most screens.

They will be the ones that make their underlying content and operational context useful to both humans and machines.