# Custom extraction and search

> Rules that read something from every page's HTML — a price, a tag, a count of links, whether a script is there — and put it in a column of its own. Each rule is tested on one page before you save it.

Describes GLLmolfar 0.9.35, checked 2026-11-08. Canonical page: https://www.getlooploop.com/gllmolfar/docs/custom-extraction

- **Where:** Crawl › Custom extraction & search… (Ctrl+Shift+K); the values in the Custom tab
- **Mode:** The dialog in both modes; the Custom tab in Pro
- **Built from:** Each page's HTML, after JavaScript when the page was rendered
- **Changes:** On the next crawl, or on rows you re-crawl

## On the screen

The dialog holds the rules, one per row, and a test on one page. Here four rules read getlooploop.com/vectorscope.

![The Custom extraction and source search dialog of GLLmolfar with four rules tested on getlooploop.com/vectorscope](https://www.getlooploop.com/images/gllmolfar/docs/custom-extraction.webp)

Numbered parts of the screenshot:

1. **What the rules read.** Up to 10 rules; each is one column in the Custom tab, read from the page HTML (after JavaScript when the page is rendered).
2. **XPath (subset).** The line lists the XPath the app understands. Anything else is shown as an error under the rule.
3. **Column name.** The name of the column the rule fills: here **H2 headings**, **og:type**, **Links in main** and **JSON-LD type**.
4. **Rule type.** **XPath (subset)** twice, a **CSS selector** and a **Regular expression**; **HTML contains** and **does not contain** are the other two.
5. **The expression.** The path, selector or pattern. The regular expression keeps its first group: the JSON-LD types.
6. **Take.** What to keep from the matches: **Count**, **Text**, **First group or match** and the others the rule type allows.
7. **Remove.** Drops the rule. **Add a rule** adds a row, up to 10.
8. **Test on.** A page address and **Test**: each rule's result on that page, before anything is saved.
9. **Result.** One row per rule: 15 H2 headings, og:type **website**, 0 links inside main, and every JSON-LD type on the page.
10. **Save.** Keeps the rules for the next crawl and for re-crawled pages.

## Rule types

| Rule type | What it matches | Take |
| --- | --- | --- |
| CSS selector | Elements, e.g. `.price` or `meta[property="og:type"]`. | Text, Attribute value, Inner HTML, Count, Yes / no |
| XPath (subset) | Nodes, e.g. `//div[@class='price']/span[1]` or `//link[@rel='canonical']/@href`. A path that ends in an attribute or `text()` gives that value. | Text, Attribute value, Inner HTML, Count, Yes / no |
| Regular expression | A pattern over the HTML, case-insensitive, e.g. `gtag\('config', '([^']+)'`. | First group or match, Count, Yes / no |
| HTML contains | Plain text in the HTML, case-insensitive, e.g. `googletagmanager.com`. | Count, Yes / no |
| HTML does not contain | Text that should be on every page. | Yes / no |

Text and attribute values keep up to 20 matches, joined, and up to 1,000 characters.

## The XPath subset

Molfar evaluates XPath with its own small reader, not a full XPath 1.0 engine. What it reads, and what it refuses with a message under the rule instead of an empty column:

| Supported | Refused |
| --- | --- |
| Paths with `/` and `//`, `.`, `..`; element names, `*`, `@name`, `@*`, `text()`, `node()` | Axes written out (`child::`, `ancestor::`…) |
| Predicates `[n]`, `[last()]`, `[position() < n]` (=, !=, <, <=, >, >=) | Unions (`\|`), arithmetic, variables |
| `[@a]`, `[@a='v']`, `[@a!='v']`, `[contains(@a,'v')]`, `[contains(text(),'v')]`, `[contains(.,'v')]`, `[starts-with(@a,'v')]` | Other functions: `count()`, `normalize-space()`, `translate()`… |
| A relative path inside a predicate (`[a/@href]`), `not(...)`, `and`, `or`, brackets, several predicates in a row | A function or a literal as the whole expression |

> **One difference from XPath 1.0.** `contains()`, `starts-with()` and `=` over a path are true when **any** selected node matches; XPath 1.0 looks at the first node only for the two functions. It changes nothing when the path selects one node.

## The Custom tab

Each saved rule is a column of the Custom tab, one value per page; columns from [Prompt on pages](https://www.getlooploop.com/gllmolfar/docs/prompt-column) appear there too, marked **· AI**. Each column has two filters: **{name}: found** and **{name}: not found**.

![The Custom tab of GLLmolfar with the four extraction columns for 500 pages of getlooploop.com](https://www.getlooploop.com/images/gllmolfar/docs/custom-columns.webp)

Numbered parts of the screenshot:

1. **Custom.** The column set the rules fill, in the row of tabs.
2. **Filter.** **All**, or for each column **{name}: found** and **{name}: not found**.
3. **H2 headings, og:type.** The two XPath rules: a count and an attribute's value, per page.
4. **Links in main.** The CSS rule: links inside the main element, counted on each crawled page.
5. **JSON-LD type.** The regular expression's first group from every match, joined.

## States and messages

- **Rule problem** — the app says: "Rule problem: {why}" Shown under the row while you type: a selector or pattern that does not parse, or XPath outside the subset. A rule like this is saved as typed, and the toast names it.
- **Saved** — the app says: "Rules saved. Crawl again, or select rows and Re-crawl, to fill the Custom tab." Rules read pages when they are crawled; the open crawl is not read again by itself.
- **Test page** — the app says: "Test on" The test reads the page from the open crawl's saved HTML when it has it, else asks the site once.

## Related pages

- [Prompt on pages](https://www.getlooploop.com/gllmolfar/docs/prompt-column.md): Agent · test › Prompt on pages: one question to your own model about many pages, each answer saved as a column in the Custom tab.
- [All pages (Internal)](https://www.getlooploop.com/gllmolfar/docs/all-pages.md): The All pages tab, called Internal in Pro mode: every crawled URL in one table, with filters, search, column sets, an overview panel and a page inspector.
- [Settings › Crawling](https://www.getlooploop.com/gllmolfar/docs/settings-crawling.md): Settings › Crawling: request speed, connections, the user agent, the timeout, robots.txt, a pause at a CAPTCHA, JavaScript rendering and service pages.
- [JavaScript](https://www.getlooploop.com/gllmolfar/docs/javascript.md): The JavaScript column set: what differs between served HTML and the rendered page (text, title, canonical, links) and images seen only after scrolling.