The DXP Scorecard logo
Independent Platform Evaluation
Scored on implementation experience
Not vendor briefings
Analyst EditorialAI

What CMS Is Winning at AI in 2026?

Every CMS vendor now sells AI. None of them built the model behind it. The race is being won on what surrounds the model, and most buyers cannot yet tell.

DXP Principal Analyst GroupPublished 11 min read

The short answer

No CMS is winning at AI by building AI. Every platform here wraps models from OpenAI, Anthropic or Google, so the race is decided by what surrounds the model: agents with approval gates, an MCP server, a choice of model and an audit trail. Contentstack, SitecoreAI and Contentful lead on that measure. Optimizely SaaS CMS and Adobe Experience Manager follow for enterprises, Sanity for teams building their own agents, and HubSpot Content Hub for marketers without developers.

Key takeaways

  • No CMS vendor trains a frontier language model. SitecoreAI runs on Microsoft's Azure OpenAI service, Optimizely's Opal on Google's Gemini and Anthropic's Claude, HubSpot's model cards name OpenAI, Anthropic and Google feature by feature, and Adobe's only home-grown generative model is Firefly, for images. What you are buying is the wrapper.
  • There are two strategies. Wrap a lab's model into named agents and meter it in credits, or open a door with an MCP server and bring-your-own-model so your own agents and assistants do the work. The platforms at the top of this editorial do both, and that is why they are at the top.
  • Text generation and MCP access are now table stakes. Across the 41 platforms we score, the thin spots are semantic search, machine-learning personalization, AI audit trails and usage dashboards. That is exactly where the surprise bills and the compliance problems appear.
  • Agentic is the most abused word in CMS marketing this year. Several vendors' agent suites are early access, preview or a vendor-run service dressed as a product. Buy what is generally available, with an approval step you can see and a log you can export.
  • Marketers are winning where the agent sits inside the editor with brand rules and an approval gate. Developers are winning where the door is open. Legal is winning only where the vendor names its model provider in writing and bars training on your data.

Our analysts’ take

Walk any CMS vendor's website in October 2026 and the word AI appears before the word content. SitecoreAI renamed the whole product around it. Kontent.ai put it in the company name. Optimizely, Adobe, Contentstack and HubSpot each sell an agent platform, and even the open-source projects ship an AI foundation in core. A marketing leader is told, repeatedly, that the platform will write, translate, audit, personalize and publish on its own. What nobody says on the home page is where the intelligence comes from.

The answer is the same everywhere. It comes from OpenAI, Anthropic or Google, rented by the token and resold inside the product. That is not a scandal; it is the only sensible way to build software in 2026. But it changes the question a buyer should ask. If every vendor has access to the same handful of models, the model is not the differentiator. The differentiator is what the vendor built around it: whether the agents can do real work in your content, whether a human approves before anything publishes, whether you can see what it cost, whether you can swap the model when your legal team or your budget says so, and whether your own AI tools can reach the content through a standard door.

This editorial takes a position on that question. It names the two strategies vendors have chosen, says which platforms are executing each one and which are mostly announcing, and reads the evidence across all 41 platforms we score to answer the question buyers actually have: is anyone winning yet, and is it you? This is where the Principal Analyst Group stands in October 2026.

What to look for

  1. 1

    Whose model, really

    Ask which lab's model runs each AI feature, whether the vendor can swap it without telling you, and whether you can bring your own keys. The answer decides your data agreements, your exit options and who profits from the markup.

  2. 2

    Agents that do work and stop

    A real agent takes a brief, builds or changes many pages, and queues the result as drafts for a human. A chat box that rewrites a paragraph is an assistant. Ask for the multi-step demo and for the screen where you reject the result.

  3. 3

    A door your own tools can use

    An official MCP server lets ChatGPT, Claude, Copilot or the agents you build read, write and publish under your permissions. Check whether it is hosted or local, whether it respects roles, and whether every action lands in the activity log.

  4. 4

    Guardrails legal will accept

    An audit trail that separates human from AI edits, brand and legal rules the AI cannot skip, a contract that bars training on your data, and ideally ISO 42001 certification of the vendor's AI management. Without these, the feature stays switched off.

  5. 5

    A bill you can forecast

    Every platform meters AI in credits, interactions or tokens. Ask for the credit price, the included allowance, the cap and the dashboard, and ask what a campaign launch with translations consumes. The license is rarely the expensive part.

Two ways to put AI in a CMS, and nobody owns the model

Strip away the branding and every CMS vendor has made one of two bets, or both. Neither involves training a model. Adobe's Firefly is the lone exception, and it generates pictures, not pages. Everything that writes, translates, tags or reasons in these products is rented from OpenAI, Anthropic or Google, usually through Microsoft Azure, Amazon Bedrock or Google Vertex, and resold to you in credits.

The wrapper: pick a lab, build agents, meter it

The vendor chooses the model, builds brand kits, agents and approval flows around it, and sells usage back in credits or interactions. SitecoreAI on Azure OpenAI, Optimizely's Opal on Gemini and Claude, HubSpot's Breeze on OpenAI, Anthropic or Google depending on the feature, Webflow with OpenAI and Anthropic, Xperience by Kentico and Sitefinity on Azure OpenAI, Hygraph on Claude alone, Bloomreach's marketing agent on Gemini. The upside is finished features a marketer can use on day one. The downside is that the vendor owns the model choice, the markup and the roadmap, and swaps the model when it suits them.

The door: an MCP server and your own keys

The vendor exposes content, schema, workflow and publishing through a Model Context Protocol server and lets you bring your own model keys, then gets out of the way. Drupal, WordPress 7.0, Umbraco, Directus, Strapi, Magnolia and Jahia all work this way, and so does Sanity for teams building agents. The upside is control: your contracts, your data residency, your choice of Claude, ChatGPT or Copilot. The downside is that the vendor has shipped plumbing, and someone on your side has to build the house.

Both: the platforms actually winning

Contentstack lets you choose the model behind each agent and ships an MCP server with about a hundred tools. Contentful runs its AI Actions on five providers with your keys and hosts an MCP server with OAuth. Sanity builds its own agent on managed models and opens two MCP servers beside it. These are the platforms at the top of this editorial, and the lesson is not subtle: the vendors with the most finished AI are also the ones least interested in locking you into their model.

Is anyone winning yet? What the evidence across 41 platforms says

Read across every platform we score and the picture is less flattering than the home pages. Fewer than one platform in eight has built-in AI we would call strong across the board. More than a third sit in the bottom half, with an AI button at most. And the strongest AI belongs, almost without exception, to the platforms that were already the strongest content platforms before AI arrived. AI has produced no upsets. It has widened the gap between the large vendors and the small ones, and between the platforms with structured content and the ones storing pages as HTML.

What is solved: writing and the door

In-editor text generation with some brand control is now the norm, and an official or well-supported MCP server is the single most common AI feature in the market: nearly three quarters of the platforms we score ship one. Several platforms with almost no AI inside the editor, among them Strapi, Agility CMS, DatoCMS and Ghost, have excellent MCP servers. The door is open everywhere. What is behind it varies enormously.

What is thin: search, personalization, audit, cost

Semantic search your own applications can query exists on about five platforms. A real machine-learning personalization engine exists on about eight, nearly all of them enterprise suites. AI usage dashboards with cost tracking are rarer still, and a full audit trail of who invoked AI and what changed is confined to the same handful. These are the four capabilities that decide whether AI pays for itself and whether legal lets you keep it on, and most of the market has not built them.

What is announced: the vaporware list

Acquia AI is early access with no service commitment, one sandbox agent and no production use. Squiz Spark is recruiting beta testers and appears in no release note. Jahia's eighteen agents are a vendor-run migration service, not a product you operate. Every agent in Xperience by Kentico's marketing suite is flagged as a preview feature. Source by Webflow is a limited preview. Payload's AI workflows are a beta. HCL Digital Experience, Concrete CMS and Kentico Xperience 13 have no agents and no MCP server at all. Gartner coined the term agent washing for this in June 2025, predicted that more than 40 percent of agentic AI projects would be cancelled by the end of 2027, and found in October 2025 that 45 percent of marketing technology leaders running vendor agents said they fell short of the promised results.

Who is winning: marketers in the suites, developers at the door

Marketers win where the agent sits inside the editor with brand rules and a visible approval step: SitecoreAI, Optimizely SaaS CMS, Contentstack and HubSpot Content Hub do this today, and Adobe does it for the teams that can fund it. Developers win where the door is open and the pricing is published: Sanity, Contentful, Drupal, Directus and Strapi. Nobody wins where the vendor has wrapped a model, hidden its name, metered it and shipped no approval gate. Several platforms in the middle of the market are exactly there.

How to tell bleeding edge from buzzword before you sign

Ask which model, in writing

Ask for the name of the lab and the cloud behind each AI feature, whether the vendor can change it without notice, and whether the contract bars training on your content. A vendor that will not name the provider is reselling something it does not want you to price. Contentstack, Contentful, HubSpot, Sanity and Kontent.ai publish the answer. Strapi, Prismic, Squiz and Uniform's Scout agent do not.

Ask for generally available, not announced

Agent suites move from keynote to preview to early access to general availability, and the pricing page rarely says which stage you are buying. Ask for the release note and the date. If the demo environment is the vendor's and not yours, the feature is not shipping.

Watch the agent get rejected

Have the agent build a campaign or translate a section, then ask to see the screen where an editor rejects it and the log entry that recorded who asked, what changed and which model ran. If the answer is a chat transcript, the approval gate is a conversation, not a control.

Price a month of real use

Get the credit price, the included allowance, the overage rate and the cap, then ask what a product launch with twelve translations consumes. Sanity publishes five cents a credit; SitecoreAI caps interactions at 50,000 a year; Optimizely sells AI as an add-on that can add 20 to 40 percent. The platform that looks cheapest on the license is often not.

Try the door with your own assistant

Connect Claude, ChatGPT or Copilot to the vendor's MCP server using an ordinary editor's account and ask it to draft, change and publish something. You will learn in ten minutes whether permissions hold, whether the action lands in the audit log, and whether the vendor's door is a product or a press release.

Platforms in this piece

Frequently asked questions

Which CMS is winning at AI in 2026?

Contentstack, SitecoreAI and Contentful lead among the platforms we score, because each pairs working agents with governance and, in Contentstack's and Contentful's case, lets you choose the model. Optimizely SaaS CMS and Adobe Experience Manager follow for enterprises, Sanity leads for teams building their own agents, and HubSpot Content Hub is the strongest choice for marketing teams without developers.

Does any CMS vendor have its own AI model?

No CMS vendor trains a frontier language model. Every platform's writing, translation and agent features run on models from OpenAI, Anthropic or Google, usually through Microsoft Azure, Amazon Bedrock or Google Vertex. Adobe's Firefly is the one home-grown generative model of note, and it generates images. Bloomreach and Adobe also run their own machine-learning models for personalization, which is a different thing from generative AI.

Which AI models do SitecoreAI, Optimizely and Adobe actually use?

SitecoreAI's Stream and Agentic Studio run on Microsoft's Azure OpenAI service. Optimizely's Opal was built on Google's Gemini and since March 2026 also routes to Anthropic's Claude through Google Vertex, with Optimizely choosing the model. Adobe runs its agents and assistants on Adobe-managed models, with Microsoft's Azure OpenAI service named on its sub-processor list, uses its own Firefly model for images, and lets you connect your own Azure OpenAI subscription only for translation.

Should I choose a CMS with built-in AI agents or one with an MCP server?

If your marketing team will use AI daily without a developer, choose a platform with finished agents, brand rules and an approval gate inside the editor, such as SitecoreAI, Contentstack, Optimizely SaaS CMS or HubSpot Content Hub. If you have developers and want AI under your own model contracts, an open door matters more: Contentful, Sanity, Drupal, Directus and Strapi. The best platforms offer both, and an MCP server is worth insisting on either way.

Is AI in a CMS worth paying extra for in 2026?

Yes when the platform has agents that do multi-step work with an approval step and a usage dashboard, because that saves real editorial hours and the cost is visible. No when the AI is a rewrite button metered in credits with no cap. Ask for the credit price and the included allowance before signing. Sanity publishes $0.05 per credit, HubSpot and Webflow include credits with their plans, SitecoreAI caps use at 50,000 interactions a year, and Optimizely's AI add-on can raise a contract by 20 to 40 percent.

About this editorial

This is a position our principal analysts took on October 3, 2026. It is not updated: when the view changes, we publish a new editorial. The scorecards linked above carry the current ratings for every platform named here.

Scorecards and platform inclusion are unsponsored: no vendor pays to be named in an editorial or to change a rating. Disagree with us? Tell us.

Still deciding?

Answer seven questions for a shortlist built around your team, or put two of the platforms named here side by side.