06Websites and digital assetsReal estate and finance

Nadlan BeIvrit

an encyclopedia of Israeli real estate terms

epedia.co.il

A Hebrew encyclopedia for anyone who runs into a real estate term, from safe room to caveat note, and wants a reliable explanation instead of Wikipedia or a site that is selling something.

350 entries, 376 pages in preview, zero broken links

The challenge

People meet real estate terms in contracts, city letters, at the bank or at a building committee meeting, and Hebrew has no serious encyclopedia for them, only broker glossaries with one line per term. At the same time, when I checked Google Israel, 26 of 27 queries showed an AI summary at the top of the page, so a bare definition no longer earns the click. The site has to give what a summary does not, and be accurate enough to be the source that gets cited.

What I did

Structured the content in the order a regular person meets things: the apartment, construction, the building, the land, what may be built, the money, and then the deal. 16 topics and a fixed entry template: a quotable opening definition, a short summary, headings phrased as real questions, a worked example and common confusions.
Ruled that every entry must still be true in five years. The body explains the concept; any figure that changes (a rate, a ceiling, a fee) lives in a separate data block fed from one central file. When a figure changes, one line is updated and no entry is reopened. Temporary government programs do not get an entry.
Every entry is written fresh and checked against the law. Several sources are gathered for each term and the entry is written fresh. Every batch is then fact checked against the text of the laws and regulations. In the first round 163 entries were checked and 135 were corrected.
Gave the professional reviewer a way to work without learning a system: one Google Doc per entry, and a back office that is simply a list opening those docs. Their edits are pulled back into the site word by word, and anything that cannot be applied automatically is logged for manual handling.
Added a "side by side" section that compares two to four related terms in a table. That is the kind of answer an AI summary does not lay out for the reader.

The decision that made the difference

If it is not certain, it does not go in. The body of an entry carries no notes like "pending verification" or "sources disagree". An unverified figure is removed, and the build prints an error if one comes back. In the same spirit I ran a trimming pass over the entries the reviewer had not yet touched, where deleting was allowed and adding was not: 226 thousand words came down to 159 thousand. The reason: in an encyclopedia, an extra detail that might be wrong does more harm than a short entry that is right throughout.

What AI did, and what stayed with me

AI agents write the entries, and a separate fact checker goes through every batch against the text of the laws and regulations, fixing in place and logging what was verified. Another agent owns every piece of text outside an entry body: the name, headings and microcopy.
I decided what stays out and turned down what did not work: I rejected the first name proposed for the site and the homepage headline the design agent wrote, ruled that unverified data comes out of the entry, and came up with the way the professional reviewer works, one document per entry. The AI writes fast; the reviewer and I decide what is right.

Outcome

350 entries and 376 pages, zero broken links, live in preview (as of October 21, 2026).
42 unverified figures were removed from 34 entries, and every data key added in the latest writing wave was verified.
346 review documents were created in Drive, and the professional reviewer's corrections reach the site through them.
An entry a professional has reviewed carries their name, and the site does not promise that every entry was reviewed that way. A transparency page on the site explains where AI is used.

Link