Do I Even Need a CMS If I Have AI?
Claude or ChatGPT can build your pages, your components and your whole site in an afternoon. So what is a CMS still for, and at what point do you actually need one?
The short answer
For a five-page portfolio or a ten-page small business site, no: Claude or ChatGPT can build and host static pages faster and better than any CMS. Past a few dozen pages, yes, but as a structured content store the AI works through an MCP server rather than a screen you log into. Sanity and Contentstack are best built for that role; Contentful, Storyblok and Payload CMS are strong alternatives, and SitecoreAI shows the case at DXP scale.
Key takeaways
- Under about ten pages, skip the CMS. Ask Claude or ChatGPT for the pages, review them, deploy them to any static host. The result is faster, more performant and better looking than a WordPress or Wix site, and a person who has never written code can launch it in under an hour.
- Page count is the decider, not technical skill. At a hundred pages a flat folder of files starts to drift; at a thousand, nobody, human or agent, can keep it consistent without structure. The CMS does not disappear. Its job changes from the place you edit pages to the structured database the AI abides by: content types with required fields, one taxonomy, validation, versions and an approval step, followed on the code side and the content side at once.
- Your interaction with the CMS will mostly disappear. You do not need to log in to publish a news item or a blog post. You tell the assistant the topic or hand it the copy, and it fills the type the CMS defines, with the title, body, image and metadata that search and answer engines expect, every time. The query is answered the same way: the ten oldest pages, every page on a topic or every page missing an image comes back from one request against a database, not a read of every file.
- The gateway that makes this possible is the MCP server (Model Context Protocol, the standard that lets Claude, ChatGPT and Copilot operate software). It is how your AI tool reads, writes and queries the CMS under your permissions. When you evaluate a CMS or DXP for AI, the MCP server is the single most important thing to check, ahead of any writing assistant or AI button in the editor. Then check what it reaches: the content only, or the whole suite.
- The case is strongest when the CMS becomes a DXP. A CMS is the core content features. A DXP adds the suite that augments your data: personalization, experimentation, analytics, session tracking, a customer data platform and a digital asset manager. You can still run all of it through Claude or ChatGPT, and that is exactly why one central platform has to manage it consistently. Few suites deliver that today: most were assembled by acquisition, and their MCP servers reach the CMS long before they reach the rest.
Our picks at a glance
SanityModel in code, hosted MCP, agent on every plan.
ContentstackAgents inside workflows, branches and taxonomy rules.
StoryblokSchema as code plus a visual editor on preview.
DatoCMSEvery write waits for approval, schema changes typechecked.
Payload CMSTypeScript model in the same repository as the site.
SitecoreAIThe full suite in one license; MCP reaches pages and tests.
Our analysts’ take
The question is being asked in earnest for the first time. A founder opens Claude, describes a five-page site, and twenty minutes later it is live on Cloudflare or Vercel with a domain attached. A marketing lead asks ChatGPT for a new landing page and gets one that matches the brand. No developer, no template marketplace, no plugin updates, no login. The products that used to make this possible for non-technical people, WordPress and Wix among them, suddenly look like the slow option. If the AI can build the pages, the components and the whole site, what is the CMS for?
The honest answer for a small site is nothing. A portfolio, a restaurant, a ten-page company site or a campaign that lives for a quarter is better served by a folder of static pages than by any platform on this page. There is nothing to patch, nothing to pay for and nothing to learn. The problem starts when the site grows. By the fortieth page there are three slightly different case study layouts, a hero component with two names and a navigation edited in four places. Every new prompt has to be told what consistency means. By the four hundredth page nobody can say which pages exist, which are stale or which share a topic without reading all of them.
This editorial takes a position on where that line sits and what the CMS becomes once you cross it. In short: the CMS survives as the structured content store, the database with a contract that the AI reads and writes under, and the editing screen you used to log into becomes optional. You ask Claude for a new blog post and it fills the fields the CMS defines. You ask for an audit of the oldest pages and the CMS answers in one query. All of it runs through the platform's MCP server, the gateway between your AI tool and your content, which is why that server is now the most important AI feature on any evaluation. And the argument gets stronger, not weaker, at the top of the market: once a CMS becomes a DXP, with personalization, experimentation, analytics, session tracking, a customer data platform and a digital asset manager augmenting the same data, a central platform that keeps all of it consistent is the only thing that lets an AI operate it safely. Here the market has not caught up with the argument. Most enterprise suites were assembled from acquisitions, and their MCP servers reach the CMS first: SitecoreAI's official server works with pages, personalization variants and tests but not with its customer data platform, search or asset manager, while Contentstack's single server does reach its personalization, customer data, hosting and brand tools. Expect to connect several servers to one DXP for a while yet. The platforms rated here are the ones that are built to be used that way. The ratings beside this piece measure what each is like to launch, run and pay for once it exists; the position on which of them an AI should work through rests on the evidence in each verdict. This is where the Principal Analyst Group stands in October 2026.
What to look for
- 1
A structure the AI must follow
Content types with required fields, allowed values and validation, defined in a file the assistant can read. The point of the CMS now is to say what a news item, a product page or a case study must contain, so that the thousandth one looks like the first.
- 2
An MCP server, before anything else
The official MCP server is the gateway that lets Claude, ChatGPT or Copilot read, write and query the platform under your permissions. It is the most important AI feature to check in any CMS or DXP evaluation. Look for scoped tokens, writes that land as drafts and a log of every action, then ask exactly which products its tools reach. Without it the AI is driving a browser instead of using the system.
- 3
Content you can query, not just store
Ask for the ten oldest pages, every page about one topic or every entry missing an image, and get an answer from one query. Taxonomy, dates and relationships as fields, plus search across the whole store, are what make an audit possible at a thousand pages.
- 4
One source for shared content
Authors, offices, products, legal notices and navigation live once and appear everywhere. Cross-linking, related content and site search all depend on it, and so does the consistency that search and answer engines reward.
- 5
Room to grow into a DXP
If personalization, experimentation, analytics, session tracking, a customer data platform or a digital asset manager are in your future, check whether the platform runs them on the same data and how many MCP servers it takes to reach them all. One AI operating one consistent platform beats one AI stitching six tools together, and today most suites still need a server per product.
The ratings at a glance
Ranked by Fit score for this topic, as rated on October 17, 2026. Dots run from one (Limited) to five (Excellent); each review below spells the ratings out.
| Platform | Fit | AI | Upkeep | Global | Speed | Budget | Self-service |
|---|---|---|---|---|---|---|---|
| 1Contentstack | 7.1/10 | Strong | Good | Fair | Good | Good | Good |
2 Contentful | 6.8/10 | Strong | Fair | Fair | Good | Fair | Good |
| 3SitecoreAI | 6.8/10 | Strong | Fair | Strong | Fair | Limited | Strong |
| 4Storyblok | 6.6/10 | Good | Fair | Good | Strong | Good | Strong |
| 5Sanity | 6.6/10 | Good | Fair | Fair | Good | Good | Good |
6 Webflow | 6.3/10 | Good | Fair | Fair | Good | Fair | Strong |
7 WordPress | 5.5/10 | Fair | Fair | Fair | Good | Strong | Good |
| 8Payload CMS | 4.9/10 | Limited | Fair | Fair | Strong | Strong | Fair |
| 9DatoCMS | 4.8/10 | Limited | Good | Limited | Good | Good | Good |
10 Strapi | 4.5/10 | Limited | Fair | Limited | Strong | Good | Fair |
- 1Contentstack7.1/10
- AI for marketers
- Strong
- Low IT upkeep
- Good
- Global & multi-brand
- Fair
- Speed to launch
- Good
- Budget predictability
- Good
- Marketer self-service
- Good
- 2Contentful6.8/10
- AI for marketers
- Strong
- Low IT upkeep
- Fair
- Global & multi-brand
- Fair
- Speed to launch
- Good
- Budget predictability
- Fair
- Marketer self-service
- Good
- 3SitecoreAI6.8/10
- AI for marketers
- Strong
- Low IT upkeep
- Fair
- Global & multi-brand
- Strong
- Speed to launch
- Fair
- Budget predictability
- Limited
- Marketer self-service
- Strong
- 4Storyblok6.6/10
- AI for marketers
- Good
- Low IT upkeep
- Fair
- Global & multi-brand
- Good
- Speed to launch
- Strong
- Budget predictability
- Good
- Marketer self-service
- Strong
- 5Sanity6.6/10
- AI for marketers
- Good
- Low IT upkeep
- Fair
- Global & multi-brand
- Fair
- Speed to launch
- Good
- Budget predictability
- Good
- Marketer self-service
- Good
- 6Webflow6.3/10
- AI for marketers
- Good
- Low IT upkeep
- Fair
- Global & multi-brand
- Fair
- Speed to launch
- Good
- Budget predictability
- Fair
- Marketer self-service
- Strong
- 7WordPress5.5/10
- AI for marketers
- Fair
- Low IT upkeep
- Fair
- Global & multi-brand
- Fair
- Speed to launch
- Good
- Budget predictability
- Strong
- Marketer self-service
- Good
- 8Payload CMS4.9/10
- AI for marketers
- Limited
- Low IT upkeep
- Fair
- Global & multi-brand
- Fair
- Speed to launch
- Strong
- Budget predictability
- Strong
- Marketer self-service
- Fair
- 9DatoCMS4.8/10
- AI for marketers
- Limited
- Low IT upkeep
- Good
- Global & multi-brand
- Limited
- Speed to launch
- Good
- Budget predictability
- Good
- Marketer self-service
- Good
- 10Strapi4.5/10
- AI for marketers
- Limited
- Low IT upkeep
- Fair
- Global & multi-brand
- Limited
- Speed to launch
- Strong
- Budget predictability
- Good
- Marketer self-service
- Fair
The shortlist, reviewed
Contentstack
The enterprise version of the same idea, and the clearest case of one MCP server reaching the whole suite: its single server covers content management and delivery, Launch hosting, Brand Kit, Personalize, Lytics customer data and analytics. Agent OS adds a no-code agent builder and the Polaris assistant on top of twenty-stage workflows, taxonomy permissions and branches, every agent run is logged with the model and tokens it used, and about 21 published agent skills teach the assistant the conventions.
- Best for
- Large content teams that need the AI to work inside approval, branching and localization rules from day one, and to reach personalization and customer data through the same door.
- Watch out for
- The MCP server is installed locally rather than hosted, each service behind it needs its own credentials and one Contentstack doc still calls it experimental, a stack is capped at five branches and five environments, and forms, email and commerce come from marketplace apps or other products.
- What it costs
- Free plan, Build at $29 per month, Growth at $299 per month plus $25 per extra user; Enterprise quote-only. AI runs on metered credits.
Contentful
The most established structured store, with the strongest governance around AI writes. The remote MCP server reached general availability in July 2026 with per-space scoping, AI Actions run inside workflows with approval gates, and audit logs separate human edits from AI actions and stream to your security tools. Schema changes travel through a migration CLI the assistant can run.
- Best for
- Enterprises with a security review to pass that want the AI on their own model contracts and content delivered to many channels.
- Watch out for
- Content types are capped at 50 fields with no nested object types, most governance features sit on quote-only tiers, and the Premium plan is not on the pricing page.
- What it costs
- Free plan, Lite at $300 per month; Premium and Enterprise quote-only. AI Actions are metered.
SitecoreAI
The example of what happens when the CMS becomes a DXP, and of how far the MCP server lags the license. The customer data platform, Personalize, Search, a digital asset tier and Agentic Studio ship together, and the official Marketer MCP server lets Claude or ChatGPT work with pages, components, personalization variants, A/B/n tests, briefs and brand kits. Twenty prebuilt agents and a no-code agent builder come with it, and AI governance is certified to ISO 42001.
- Best for
- Enterprises whose site is one part of a personalization, experimentation and customer data operation, and who want the AI to run the content and testing side of it under one governed platform.
- Watch out for
- The MCP server stops at the CMS side: the customer data platform, Search and the Content Hub asset manager have no official server of their own, so agents reach them through bring-your-own connectors in Agentic Studio or not at all. Add quote-only pricing, a fair-use cap of 50,000 AI interactions a year, and implementations that depend on certified Sitecore specialists.
- What it costs
- Quote-only. No published tiers. AI usage is subject to a 50,000 interactions per year fair-use threshold.
Storyblok
The structured store that keeps a visual editor for the people who still want one. The content model became first-party TypeScript in July 2026 with a dry-run push the assistant can diff, the hosted MCP server has a Claude Code setup guide, and marketers compose pages from the same components on a live preview of the real site.
- Best for
- Marketing teams that want the AI to define the structure and write most of the content, while some editors keep working visually.
- Watch out for
- The MCP server is still labelled an experimental Labs project, most 2026 features sit on the quote-only Premium and Elite plans, and search, forms and personalization come from other products.
- What it costs
- Free starter plan, Growth at $99 per month, Growth Plus at $349 per month; Premium and Elite quote-only.
Sanity
The clearest example of a CMS as a structured store the AI works through. The content model is TypeScript, two hosted MCP servers let Claude or ChatGPT read and write under your permissions, the Content Agent runs multi-step jobs across hundreds of documents as reviewable drafts, and semantic queries are included on every plan.
- Best for
- Teams whose content will run into the hundreds or thousands of entries and who want the AI, not a person, doing the publishing.
- Watch out for
- The front end is yours to build and deploy, there is no personalization engine, the content backend runs in a single EU region, and GROQ and Portable Text are Sanity's own concepts.
- What it costs
- Free plan, Growth at $15 per seat per month (viewers free), Enterprise quote-only. Studio hosting is free on every plan; AI credits are published at $0.05 each.
Webflow
The pick for teams that want structure and a finished, hosted site in one product. MCP 2.0 lets the AI build pages, collections and components against the real design system on isolated branches with every action logged, and an AI audit panel fixes metadata, schema markup and answer engine signals site-wide. Nothing is left to deploy.
- Best for
- Design-led marketing teams who want the assistant to build inside the product a person will still manage visually afterwards.
- Watch out for
- Collections are a flat field model capped at 60 fields on standard plans and 100 on Enterprise, with no taxonomy and no schema as code, so the structure this editorial argues for stays shallow. Leaving means static HTML only.
- What it costs
- Published. Site plans from $15 per month, Premium at $25 per month (annual), workspace seats from $15 to $39 per month. Agent-driven builds consume metered credits.
WordPress
The baseline, and the platform the models have read most about. WordPress 7.x adds an Abilities API that tells agents what a site can do and a Connectors vault for your own model keys, and the core AI team maintains the official MCP Adapter plugin. Custom post types and fields give structure, if a plugin or the assistant defines them.
- Best for
- Teams that already run WordPress, or need its plugin ecosystem, and want the assistant to do the development and publishing a freelancer used to.
- Watch out for
- Structure is not the default: a stock install stores pages as HTML blobs and configuration lives in the database with no configuration as code. You still choose a host, patch plugins and own security, and the MCP Adapter is a pre-1.0 plugin.
- What it costs
- Free software. Budget for managed hosting from roughly $25 to $35 per month, premium plugins and someone, or some agent, to update them.
Payload CMS
The CMS that lives inside the code the AI is already writing. The content model is TypeScript in the same Next.js repository as the site, so one prompt changes a field, the component that renders it and the migration together. The scaffold command installs a coding skill for Claude Code, and there is no license fee and no feature gating.
- Best for
- Teams that want one repository for site and content, full ownership of their data and no monthly license, and accept that the assistant is also their operations team.
- Watch out for
- Payload Cloud is closed to new sign-ups, so you host it and its database yourself. There are no compliance certifications, renaming a field still produces a drop-and-add migration, and the admin interface is built for developers.
- What it costs
- Free MIT-licensed software with no paid feature tiers. Budget for Vercel or a container host and a managed database, from free tiers to a few hundred dollars a month.
DatoCMS
The most careful store for AI writes. The hosted MCP server lets reads run freely while every write waits for your approval, a scripting mode built for agents typechecks every schema change before it touches the API, and the 2026 pricing restructure published overage rates for records, bandwidth and API calls.
- Best for
- Content-heavy marketing and media sites that want a conservative, reviewable path for everything the assistant changes.
- Watch out for
- Native AI inside the editor is deliberately minimal, hosting is a single EU region, compliance stops at ISO 27001, and multi-site architecture is weak.
- What it costs
- Free Developer plan, Professional at €99 per month, Scale at €199 per month, each with published limits and overage rates. Enterprise is quote-only.
Strapi
The most widely used open-source structured store, with the agent door in the core product. Strapi's built-in MCP server reached general availability in June 2026, re-checks the user's permissions on every call and lands writes in the audit log, content types live as files in the repository, and generated REST and GraphQL APIs make every query the assistant needs available in an hour.
- Best for
- JavaScript-comfortable teams that want a self-hosted or Strapi Cloud backend behind a front end the AI writes, with no vendor deciding the model.
- Watch out for
- Renaming a field still orphans existing data, the free Cloud plan was retired in July 2026, SSO is a paid add-on, and self-hosting puts upgrades and security on you or the assistant.
- What it costs
- Free open-source core. Strapi Cloud from $35 per project per month, Pro at $90 per month; the Growth license for AI features is $45 per month. Enterprise is quote-only.
Three sizes of site, three different answers
Whether you need a CMS used to depend on whether you could code. Now it depends on how many pages you have, because the AI removed the coding problem and left the consistency problem. Here is where the line sits in our view.
Up to about ten pages: no CMS
A portfolio, a local business, a product launch. Ask Claude or ChatGPT for the pages, review them, deploy the folder to Cloudflare, Vercel, Netlify or any static host. The result loads faster than any CMS-rendered page, costs nothing to license and has nothing to patch. A non-technical person can do the whole thing in under an hour, which is the job WordPress and Wix used to have.
Tens to a few hundred pages: the drift point
The site still works, but every new page is a chance for the assistant to invent a slightly different layout, a new field name or a second copy of the navigation. You spend prompts describing what consistency means instead of describing content. This is where a structured store starts paying for itself: the AI reads the content type and fills it, and the page cannot come out wrong.
Thousands of pages: the CMS is not optional
News, documentation, product catalogues, multi-brand or multi-language sites. Nobody can keep a folder of thousands of files consistent, searchable and cross-linked, and neither can an assistant that has to read all of them to answer a question. A structured store with types, taxonomy, dates and relationships answers in one query, shares global content once and feeds internal search, external search and the answer engines from the same source.
What the CMS becomes when you stop logging into it
The CMS of the last twenty years was an editing screen with a database behind it. The one that survives the AI is the database with the screen made optional. Four things change.
The structure is the product
A news item needs a title, a summary, a body, a hero image, an author, a date and a topic. A product page needs different fields. The CMS defines those types once and the assistant obeys them on every page, on the code side (the component that renders the type) and the content side (the entry that fills it). That is what keeps page one thousand consistent with page one, and it is what the search and answer engine metadata depends on.
The assistant is the interface
Build me a news page, here is the content. Write a blog post on this topic. Translate these twelve product pages. The request goes to Claude or ChatGPT, which writes into the CMS through its MCP server under your permissions, and the entry lands as a draft for approval or publishes, depending on the rules you set. The editing screen is still there for the people who prefer it. It is one door among two.
The CMS answers questions, not just requests
Give me the ten oldest pages. Find every page about pricing. Which case studies have no image? In a structured store these are queries, and the assistant runs them in seconds. In a folder of files, even one pushed to git, the assistant has to open and read everything, and the answer gets slower and less reliable with every page you add.
Shared content lives once
Authors, locations, legal notices, product data and navigation are entries that many pages reference, so a change happens in one place and appears everywhere. Cross-linking and related content follow from the relationships, internal search indexes the fields rather than the HTML, and external search and answer engines get the same clean structure. None of that exists in a flat folder without someone rebuilding it by hand.
When the CMS becomes a DXP, the case is made
Our definition: a CMS is the core content features. A DXP is the mature, full suite of marketing tools around them, personalization, experimentation, analytics, session tracking, a customer data platform and a digital asset manager, all augmenting the same data. You can still ask Claude or ChatGPT to build the segment, launch the test or pull the report, and that is the point. Six tools operated by an AI with no shared data model is six places for it to be wrong. One platform that manages all of it consistently, and exposes all of it through as few MCP servers as possible, is what makes the AI safe to let loose on an enterprise. Check the reach before you believe the slide: SitecoreAI's official server works with pages, personalization variants and tests but not its customer data platform, search or asset manager; Contentstack's one server does reach Personalize, Lytics, Launch and Brand Kit, with separate credentials for each; Webflow's covers design, content and publishing, with analytics behind an add-on. Most acquired suites will need several servers for a while yet.
Before you pick a CMS for the AI to run: five checks
Count the pages you will have, not the ones you have
If the honest answer is under ten and will stay there, build static pages with the assistant and stop reading. If it is a blog, a news section, a catalogue or more than one brand or language, plan for the structured store now. Migrating a few hundred drifted pages into types later costs more than starting with them.
Make the structure a file
On Sanity, Payload CMS, Storyblok, Strapi and Contentful the content model lives in code or a migration the assistant can read, change and revert. On Webflow and a stock WordPress install it lives behind a screen or in a database. If the AI cannot read the structure, it cannot be made to follow it.
Ask for the MCP server first, then what it reaches, then for a key you can take back
If there is no official MCP server, the AI cannot use the platform and the rest of the evaluation does not matter. Every platform here ships one. Then ask which products its tools actually reach: on a single-product CMS the answer is everything, on a DXP it is often the content side only, with the customer data platform, search and asset manager needing servers of their own or none at all. Then ask for a dedicated token or user with the least access it needs, writes that land as drafts or need approval, and a log that attributes every action to the agent. DatoCMS requires approval on every write, Strapi re-checks permissions on every call, Contentful streams AI actions to your security tools. Test it by asking the assistant to delete something.
Run the audit query on the demo
Before you sign, ask the vendor, or the assistant through the MCP server, for the ten oldest entries, every entry missing a field and every entry on one topic. If the answer takes a developer, the platform is a page store, not a structured one, and it will not help you at a thousand pages.
Price the store, not the editor
You are buying a database with a contract and an agent door, not seats for people who will rarely log in. Favour platforms that publish prices and meter the things the AI will actually consume: entries, API calls and AI credits. Sanity, DatoCMS, Storyblok and Strapi publish those rates; the enterprise tiers everywhere are quote-only.
Frequently asked questions
Do I still need a CMS if I can build my website with AI?
Not for a small site. Claude or ChatGPT can build a five or ten-page site as static pages that are faster, cheaper and better looking than a CMS-driven one, and a non-technical person can launch it in under an hour. Once the site grows past a few dozen pages, and certainly into the hundreds or thousands, you need a CMS as the structured content store the AI writes into, so every page follows the same types and the whole site stays searchable and consistent.
Can ChatGPT or Claude replace WordPress?
For the sites WordPress was often used for, a small business or a portfolio, yes. An assistant can produce static pages with no hosting cost, no plugins to patch and no login. For a content-heavy site WordPress still has a role as a structured store if custom post types and fields define the structure and the assistant works through its official MCP Adapter. Platforms built around structure in code, such as Sanity, Storyblok and Payload CMS, do that job more cleanly.
What is a CMS actually for in the age of AI?
It is the structured database your content lives in and the contract the AI follows. It defines what a news item, a product page or a case study must contain, holds shared content such as authors and products once, keeps versions and approvals, and answers questions such as which pages are oldest or which cover a topic. The editing screen becomes optional; the assistant becomes the main way you publish.
Will I still have to log in to my CMS to publish content?
Increasingly, no. With a CMS that ships an official MCP server, you ask Claude or ChatGPT to create a news page, write a blog post on a topic or update a product entry, and it writes the fields the CMS defines, with the title, body, image and metadata that search and answer engines expect. The entry lands as a draft or publishes according to your rules. The screen stays available for people who prefer it.
Do I need a DXP or just a CMS if I am using AI?
A CMS covers the core content features. A DXP adds personalization, experimentation, analytics, session tracking, a customer data platform and a digital asset manager on the same data. If you need those, AI makes the central platform more important, not less: Claude or ChatGPT can operate every one of them, and doing so safely needs one consistent system reached through as few MCP servers as possible rather than six tools stitched together. Check the reach before you buy: SitecoreAI's official server works with its pages, personalization and tests but not its customer data platform, search or asset manager, while Contentstack's single server does reach its personalization, customer data, hosting and brand tools.
What is the most important AI feature to look for in a CMS?
An official MCP server. The Model Context Protocol is the gateway that lets your AI tool, whether Claude, ChatGPT or Copilot, read, write and query the CMS or DXP under your permissions. Writing assistants and AI buttons in the editor are secondary; without the MCP server the AI cannot work with your content at all. Check that it ships scoped tokens, draft or approval writes and an action log, and ask which products it reaches: on most DXPs today it is the content only, and the personalization, customer data and asset tools need their own servers.
How many pages before a CMS makes sense?
In our view a CMS adds nothing under about ten pages and becomes hard to do without somewhere in the tens to low hundreds, when layouts, field names and navigation start to drift between pages the assistant built separately. Anything with a blog, a news section, a catalogue or several brands or languages should start with a structured store, because moving drifted pages into types later costs more than starting with them.
How we rated these platforms
Each rating combines the relevant criteria from our independent scorecards into a five-point scale. The Fit score blends the ratings below, weighted for this topic, into one number out of 10. The ratings were taken from the scorecards on October 17, 2026 and are fixed with this editorial; follow the scorecard links above for the current ratings.
Scorecards and platform inclusion are unsponsored: no vendor pays to appear in this guide or to change a rating. Full criteria, weights and evidence are in the methodology and on each platform’s scorecard. Spotted something out of date? Tell us.
Still deciding?
Answer seven questions for a shortlist built around your team, or put the top two side by side.