A content chain in four links
Automatic content filling of a site on Tilda is not one button, but a chain of four links:
- Source data: a supplier price list, a product table, article drafts, notes.
- AI processing: a neural network turns the data into texts, names, descriptions, meta tags.
- A check: a person compares the result with the source data and the rules of the site.
- Publication: the finished materials go into the Catalog, Feeds, or onto Tilda pages.
Each link can be sped up, but none can be thrown out. If the check is skipped, the neural network’s mistakes go straight onto the site. If you save on the source data, the neural network starts filling in the gaps.
Link 1. Source data
Everything that appears on the site has to come from somewhere. For products that is a table with SKUs, characteristics, prices and stock. For a blog it is the topic, theses, facts, sources, the author’s draft.
The main rule: the source data must contain everything that has to be published as a fact. Price, size, composition, the date of an event, an address — the neural network must not invent them. If some characteristic is missing from the data, it is better to leave the field empty than to get a plausible invention.
It is most convenient to keep the source data in a table, for example in Google Sheets: each product or entry is a row, each field is a column. Such a table is easy to pass to a neural network, and then turn into a file for import.
For a blog the source data looks different, but the rule is the same. A neural network can write an entry on a topic, but the facts, the numbers, the examples from practice, the author’s opinion must come from a person. A convenient format is a short card for each entry: the topic, who it is for, three to five main thoughts, facts with sources, what must not be written. From such a card you get a draft that does not have to be rewritten from scratch.
Separately it is worth agreeing who is responsible for the data being current. If the supplier’s price list was updated and the old prices stayed in the table, the automation will publish the mistake neatly and quickly. The faster the chain, the more important it is that its start is correct.
Link 2. AI processing
On this link the neural network does what it does best: it turns dry data into connected text. For products — names by a template, descriptions, Title and Description. For a blog — drafts of entries, announcements, meta tags.
For the result to be predictable, the neural network needs a clear prompt: which fields to use, how long the text is, in what style, what not to write. A sample works well: one finished card or entry that the neural network follows.
For blog entries the prompt differs from a product one. What matters more is not the template, but the voice: how the author addresses the reader, which words they use, which they avoid, how long the paragraphs are. A sample conveys this best — one or two published entries that the author likes. The draft will still have to be edited, but there will be noticeably fewer edits.
Another task for this link is to bring the data to a format Tilda understands. A neural network can assemble from the table a file with the columns needed for import: rename fields, split or join values, remove what is extra. This is routine work where it is easy to make a mistake by hand.
Processing can be done in a chat, copying rows of the table in portions, or through an AI agent that reads the table itself and writes the result. In the second case it is important that the agent writes the result into new columns and does not overwrite the source data.
Link 3. The check
The check is the link people most often want to skip. The neural network produced neat texts, everything looks right. But this is exactly where the mistakes are caught that would otherwise end up on the site: a characteristic that is not in the data, a wrong price, a promise the business does not make.
What to check:
- the facts in the texts match the source data;
- there are no properties and promises that are not in the data;
- the length of Title and Description is within the needed limits;
- there are no identical texts on different products or entries;
- SKUs and identifiers are in place, nothing got mixed up.
Part of the checks is automated with formulas in the table: length, duplicates, the presence of forbidden words. Comparing facts can be given to a second pass of the neural network. But the result still has to be looked at by eye, selectively.
For blog entries the check is wider than for products. Besides the facts, it is worth reading the text as a whole: whether there are repetitions, whether it contradicts other materials on the site, whether it sounds like the author. A neural network writes smoothly, and that is exactly why a mistake in smooth text is easy to miss. It is useful to read the draft with the question “where did this come from?” for every concrete statement.
Link 4. Publication on Tilda
Tilda has several ways to upload prepared materials. They suit different tasks and protect against mistakes in different ways.
Importing a file into the Catalog
Products can be uploaded into the Tilda Catalog from a CSV or YML file. According to Tilda’s help, on import the columns of the file are matched to the catalog fields, and unused columns are ignored. In CSV the separator is a semicolon, the maximum file size is 50 MB. The Catalog is available only on paid plans.
For updating existing products there is a mode in which the import only updates the data and does not create new products. This is the safest way to change texts or prices in a catalog that is already published: no new items will appear by mistake.
Through an MCP server
The Tilda MCP server lets you connect an AI agent to the site directly. According to the help, the agent can add products to the Catalog, create entries in Feeds and pages. Connection is available on the Personal and Business plans, and MCP works only with agents installed as applications on a computer, without browser versions.
Access is given to the agent by a key, and it can be limited to particular sections and actions. For filling content it is reasonable to give exactly the rights the task needs: if the agent adds products, it does not need access to pages. There is a separate article in the MaraCraft blog about setting up MCP and controlling the agent on the site in detail.
When to choose MCP, and when a file import? Import is convenient for large batches: hundreds of products, prepared and checked in a table in advance. MCP is convenient for small regular tasks: add a few new products, publish an entry, fix a description. In both cases the check stays the same, only the way of delivery to the site changes. A file has an advantage here: it can be opened and checked in full before upload.
Entries in Feeds
Feeds is the Tilda section for a blog and news. Entries can be uploaded as a CSV file: according to the help, it has columns for the heading, the text, the date, visibility, SEO Title, SEO Description, the author’s name and other fields. This is convenient when a series of entries has to be uploaded at once, for example to move a blog from another platform or to publish a prepared package of articles.
Through MCP the agent can also create entries in Feeds, one at a time and within the rights it was given.
| Upload method | What it is for | Limits |
|---|---|---|
| CSV or YML import into the Catalog | Bulk upload and update of products | Paid plans only, CSV up to 50 MB, separator “;” |
| Import in the “update only” mode | Changing texts and prices of existing products | New products are not created |
| CSV import into Feeds | A batch of blog or news entries | The file has to be prepared by the Feeds columns |
| MCP server | Products, entries and pages through an AI agent | Personal and Business plans, application agents only |
How the chain looks in work
In practice the chain is assembled from the data to publication. First the source table, where each row has a unique identifier. Then a task for the neural network with a sample and rules, which is brought to shape on a trial selection of products of different types. Then processing in portions, with the result written into separate columns, a check — with formulas, a second pass of the neural network and by eye — an export of the current catalog, a trial upload of several products, and only then the whole file. After publication the pages are looked at on a phone and on a computer.
Each link needs its own skill: the structure of the data, setting the task for the model, the check, working with import and rights in Tilda. So mistakes most often appear at the joints: the table does not match the catalog fields, the check was done selectively where a full one is needed, or the import was started without an export to return to.
How not to break what is already published
The biggest risk of automation is tied to changes in what is already published. An unsuccessful import can overwrite descriptions, mix up prices or create duplicates of products that are already published.
A few rules that protect against such mistakes:
An export before an upload. Before any bulk change the current state of the catalog or Feeds is saved. If something goes wrong, there will be somewhere to restore the data from.
The “update only” mode for existing products. If the task is to change texts or prices, and not to add something new, an import without creating new products is needed.
Minimal rights for the agent. If the filling goes through MCP, the key should be limited to only the sections and actions that are needed. An agent that adds entries to Feeds must not have access to pages and the catalog.
A person confirms publication. The MCP protocol specification recommends that a person always have the possibility to reject an agent’s action, and that clients show a confirmation before operations and keep a log of calls. For a site this means a simple rule: the agent prepares, a person checks and publishes.
A trial portion. First a few products or entries, then everything else. A mistake in three products is simpler to fix than in three hundred.
A change log. I write down what was uploaded, when, and with which file. If a week later a mistake is found in the descriptions, the log will show which import brought it and which products are affected. A simple table is enough for this: the date, what was uploaded, by whom, where, a link to the file.
If filling a site on Tilda is a regular task you want to hand over entirely, look at Tilda site updates. And if you need ongoing help with the site, there is Tilda site support. The task can be discussed through the quiz on mara-craft.com: I will look at which links of the chain you already have and what can be automated.