HTML Table to JSON & CSV - Web Table Extractor

For AI agents and data pipelines: give it a web page URL, get back every HTML table on that page as JSON: one object per row, keyed by the column names from the table's own header rows, with merged cells filled in, numbers as numbers, and footnote marks removed.

Page-reading tools often break tables: rows and columns slide out of line when a cell spans two rows, a two-level header turns into two unrelated header rows, and some readers drop the table altogether. This Actor rebuilds each table cell by cell the way a browser lays it out, then gives you rows you can use directly.

Input and output in one look:

{ "urls": ["https://en.wikipedia.org/wiki/List_of_largest_companies_by_revenue"], "tableIndexes": ["1"] }

{
  "status": "ok",
  "url": "https://en.wikipedia.org/wiki/List_of_largest_companies_by_revenue",
  "tablesFound": 6,
  "tablesReturned": 1,
  "tables": [
    {
      "index": 1,
      "title": "List",
      "headerSource": "th-rows",
      "headerRowCount": 2,
      "rowCount": 50,
      "columnCount": 9,
      "columns": [
        { "name": "Ranks", "type": "number" },
        { "name": "Name", "type": "text" },
        { "name": "Revenue › USD (in billions)", "type": "number", "headerParts": ["Revenue", "USD (in billions)"] },
        { "name": "Employees", "type": "number" }
      ],
      "rows": [
        { "Ranks": 1, "Name": "Amazon", "Industry": "Retail Information technology", "Revenue › USD (in billions)": 716, "Profit › USD (in billions)": 79.9, "Employees": 1576000, "Headquarters": "United States", "State-owned": null, "Ref.": null },
        { "Ranks": 2, "Name": "Walmart", "Industry": "Retail", "Revenue › USD (in billions)": 713, "Profit › USD (in billions)": 21.8, "Employees": 2100000, "Headquarters": "United States", "State-owned": null, "Ref.": null }
      ],
      "rowsText": [
        { "Ranks": "1", "Name": "Amazon", "Revenue › USD (in billions)": "716", "Employees": "1,576,000", "Headquarters": "United States" }
      ],
      "spanCopies": [{ "row": 1, "column": "Headquarters", "copiedFrom": "above (rowspan)" }],
      "footnotesRemoved": 52
    }
  ],
  "otherTables": [
    { "index": 2, "title": "Breakdown by country", "reason": "not selected (tableIndexes 1)", "gridRows": 13, "gridColumns": 3 },
    { "index": 3, "title": "External links", "reason": "layout table: navigation box (links to other pages)", "gridRows": 5, "gridColumns": 2 }
  ]
}

(Real values from a run on 2026-09-25, shortened: columns and rowsText show 4 and 1 of their entries. Leave urls empty and the Actor reads the example page https://en.wikipedia.org/wiki/List_of_countries_and_dependencies_by_population.)

What it handles

Problem in the HTML	What you get
Merged cells (rowspan, colspan)	The value is copied into every cell it covers, so every row has every column. Each copy is marked in spanCopies (above (rowspan) or left (colspan); up to 500 per table, spanCopyCount has the total), so you can tell a copied value from a written one. A rowspan stops at the end of its <thead> / <tbody> / <tfoot> and rowspan="0" runs to the end of it, as in a browser.
Two or three header rows	Column names join the levels with ›: Revenue › USD (in billions), 気温(℃) › 平均 › 日最高. The levels are also in columns[].headerParts. A header cell that spans two header rows is used once (Employees, not Employees › Employees).
Numbers written as text	Columns where at least 80% of the cells read as numbers get type: "number" and the numbers in rows: thousands separators (1,234,567 / 1.234.567,8 / 1 234 567 / 1'234), currency signs ($, €, £, ¥, 円, ₹ ...), %, units (km, kg, °C, 人 ...), − minus signs, about / 約 / ~, and the words million / billion / 万 / 億 (multiplied out). The decimal separator is decided per column from the cells themselves, and only when that is impossible from the page language. columns[].unit, currency and decimalSeparator say what was removed.
Cells that are not plain numbers	Never guessed: in a number column a cell like 1,234 (2023) is null in rows and kept as written in rowsText (counted in unparsedNumberCells). The same goes for values with <, >, ±, over, under or up to (a limit is not the value), for a one-letter suffix such as $1.2M, 850K or 5m (million, thousand, metres or minutes cannot be told apart), and for numbers with more than 15 digits (they would be rounded). Values marked about, ~ or 約 are numbers and are counted in columns[].approximateCells. A column with a code that starts with 0 (008, postal codes) stays text, so no zero is lost. Dashes, n/a and empty cells are null, not 0.
Footnote marks	[1], [a], [note 3], [citation needed] and <sup> reference marks are removed from cells and captions (footnotesRemoved). Turn stripFootnotes off to keep them.
Header cells at the start of each row (<th scope="row">)	They stay as the first column of each row.
Group rows inside the body (one cell across the whole table, such as a region name)	Not returned as a row; their text goes into a Section column on the rows under them (marked derived in columns).
Header rows repeated inside a long table, empty rows	Left out and counted (repeatedHeaderRowsDropped, emptyRowsDropped).
Tables inside tables	Each table is returned on its own; the outer table's cell does not swallow the inner table's text (nestedInTable names the outer table).
Text hidden on the page	display:none / hidden elements, Wikipedia's hidden sort keys and empty rows, and text that sites show only on phones (md:hidden, show-mobile, d-none) or only on hover (tooltips) are not read.
Layout tables	Tables used for page layout (one row, one column without header cells, only images, cells holding other tables or long paragraphs, role="presentation", navigation boxes) are not returned; they are listed in otherTables with the reason. includeLayoutTables returns them too.

Checked against a browser: on 48 real pages captured on 2026-09-25 (Wikipedia in English, German, French and Japanese, BLS, IRS, the Federal Reserve, EIA, the Bank of England, the ECB, GOV.UK, the Japan Meteorological Agency, the Bank of Japan, Japan's National Tax Agency, iban.com, x-rates.com, worldometers, TIOBE, ESPN, CBS Sports, Sky Sports, Formula 1, Hacker News, MDN and W3Schools) and one hand-written page of broken and unusual table markup, the rows, columns and cell text of 454 tables (58,621 cells) match what Chromium draws from the same HTML, cell by cell. 12 of the real pages were captured only after the reading rules were finished, as a blind check: 80 of their 82 tables matched on the first try, and the 2 that did not (a table with stray </tr> tags, and a hidden coordinate format on Wikipedia) were fixed. 4 tables with zero-width spacer cells cannot be compared by position and are not in the count.

Choosing tables

By default every data table on the page is returned. To get fewer:

• tableIndexes: the nth <table> on the page, counted from 1 (the same index you see in tables and otherTables).
• titleContains: text in the table's <caption>, aria-label or the nearest heading above it.
• columnContains: text in any column name.
• minRows: skip small tables.

Every table that is not returned is listed in otherTables with its index, title, size and the reason, so an agent can pick the right one in a second call.

What it cannot read (and says so)

• Tables drawn by JavaScript. The Actor reads the HTML the server sends and does not run JavaScript. If that HTML has no <table>, the row says no-tables and that a table you see in a browser is drawn by JavaScript or built from <div> tags. It never says "this page has no table" for such a page. (Example: https://aws.amazon.com/ec2/pricing/on-demand/ sends no <table>.) A table whose HTML has header rows but no data rows (or only a "Loading..." row) gets its rows from JavaScript too: it is listed in otherTables as no data rows in the HTML, never returned with its header as a data row, and the page is marked likelyNeedsJavaScript.
• Pages behind a sign-in or a check page (Cloudflare, DataDome, PerimeterX, AWS WAF, Akamai ...): the row says login-required or blocked. The Actor never signs in, never solves or bypasses check pages and does not switch to proxies to get around a refusal.
• Addresses robots.txt does not allow. The Actor checks each site's robots.txt (RFC 9309, as NeverEmptyTables; if a site has no group for that name, User-agent: * applies) for every URL and every redirect, and does not request disallowed addresses (robots-disallowed).
• Classes that a site hides with its own CSS file (other than the common ones above) cannot be seen without that CSS; such text is read.

Input

Field	Type	Default	What it does
urls	array of strings	(empty: the Wikipedia population list is used)	Pages to read, 1 to 1,000 per run. A URL without https:// is read as https://. The same URL given twice is read and charged once.
tableIndexes	array	(all)	Only the nth <table> elements of the page, counted from 1.
titleContains	string	(none)	Only tables whose caption, aria-label or nearest heading contains this text (case does not matter).
columnContains	string	(none)	Only tables with a column name containing this text.
minRows	integer	1	Skip tables with fewer data rows.
includeLayoutTables	boolean	false	Also return layout tables (with layoutReason).
stripFootnotes	boolean	true	Remove footnote marks.
includeTextValues	boolean	true	Also return rowsText (every cell as written).
includeCsv	boolean	false	Also return each table as RFC 4180 CSV in csv.
includeLinks	boolean	false	Also return links: the first link in each cell, per row and column.
maxRowsPerTable	integer	10000	Upper limit of data rows per table (1 to 100,000); the rest is counted in rowsTruncated.
onlyChanges	boolean	false	Monitor mode: return only pages where a selected table gained or lost rows since the last run with the same watchName.
watchName	string	(none)	Name of the remembered state (letters, digits, ., -, _; up to 40). Setting it fills changeType and changes.
resetMonitoringState	boolean	false	Forget what this watch remembered.
maxConcurrency	integer	4	Pages read in parallel, 1 to 8.
requestTimeoutSecs	integer	20	Time limit for one request, 5 to 60 seconds. HTTP 429, 500, 502, 503 and 504 are asked again up to two more times.

Output

One row per URL. For each returned table:

Field	Meaning
index	Position of this <table> on the page, counted from 1.
title, caption, heading	The <caption> (or aria-label), and the nearest heading above the table. title is the first of these that exists.
headerSource, headerRowCount	Where the column names came from: thead, th-rows (leading rows made of <th> cells), first-row (a table without <th>: the first row is used only when every cell in it is filled, not a number and unique, and a column below it is numeric), or none (names are column_1, column_2, ...).
columns	name, type (number, text or empty), unit, currency, decimalSeparator, approximateCells, nonEmptyCells, numberCells, headerParts.
rows	One object per data row, keyed by column name. Numbers in number columns; null for empty cells.
rowsText	The same rows with every cell as written on the page.
spanCopies, spanCopyCount	Cells filled from a merged cell (up to 500 listed).
sectionRows, footerRowCount	Group rows turned into the Section column; rows that came from <tfoot> (they are the last rows).
emptyRowsDropped, repeatedHeaderRowsDropped, unparsedNumberCells, footnotesRemoved, rowsTruncated	What was left out or changed, counted.
gridRows, gridColumns	Size of the table as laid out, header rows included.
csv, links	Only when includeCsv / includeLinks is on.

Column order: read it from columns. Apify's dataset may list object keys that contain a dot (such as Pos. or Fert. Rate) after the other keys in rows; the values are unchanged.

Row fields: tablesFound (all <table> elements on the page), tablesReturned, otherTables (tables not returned, with the reason), likelyNeedsJavaScript / jsEvidence, pageTitle, htmlBytesRead, htmlTruncated, and in monitor mode changeType and changes.

Rows that are not charged

status	Meaning
no-tables	The HTML has no <table> element. The note explains that tables drawn by JavaScript cannot be read by this Actor (this is not "the page has no table").
no-data-tables	The page has <table> elements but none with data rows to return: layout tables, tables hidden on the page, or tables with header rows only (their rows are usually added by JavaScript). Each is listed in otherTables with the reason.
no-matching-tables	The page has data tables, but none matches your selection; all of them are listed in otherTables.
robots-disallowed / robots-unreachable	The site's robots.txt does not allow this address, or could not be read (RFC 9309: treated as disallowed). The page was not requested.
login-required	The page needs a sign-in (401 or a redirect to a sign-in page).
blocked	The site refused the request or showed a check page, including check pages sent with HTTP 200 or 202.
not-found / http-error / unreachable / unreadable / not-html / bad-input	As named; the note says why.
too-large	The page was read, but even one of its tables is too large for one dataset row (9 MB), for example a table with tens of thousands of columns. Take it with tableIndexes and a lower maxRowsPerTable.
budget-reached / no-change	The run's maximum charge was reached / monitor mode found no change.

Monitor mode

Set watchName (and onlyChanges to get only changed pages). The first run returns every page as first-check. Later runs compare each selected table row by row with what the watch remembered. Tables are recognised by their title, heading and column names (not by position), so a new layout table elsewhere on the page does not shift anything. A changed page returns only the tables that changed, and changes: [{ tableIndex, title, change, rowsAdded, rowsRemoved, addedRows }] where change is rows-changed, table-added or table-removed (up to 200 added rows are listed). A changed value in an existing row counts as one row removed and one added.

A change is re-read 1.5 seconds later and reported only if the table is the same in both reads; a table that differs between two reads a moment apart (for example a live price ticker) is not reported that time and is compared again in the next run. A page read only up to the 5 MB limit never reports tables as removed. If you change the table selection, each page starts over as new.

Pricing

Pay per event:

• Page with tables returned: charged once per URL that comes back with status: "ok" (at least one table returned, however many tables it has). Rows with any other status are free.
• Run start: charged once per run that returns at least one page, or, in monitor mode, once per run that read and compared pages (even if nothing changed). A run that returns nothing is not charged at all.

If a run's maximum total charge has no room for the start fee plus one page, the Actor requests nothing and charges nothing.

Calling it from code

curl -X POST "https://api.apify.com/v2/acts/neverempty~html-table-to-json/run-sync-get-dataset-items?token=<YOUR_TOKEN>" \
  -H "Content-Type: application/json" \
  -d '{"urls":["https://en.wikipedia.org/wiki/List_of_countries_by_GDP_(nominal)"],"columnContains":"IMF","includeTextValues":false}'

AI agents can call it through the Apify MCP server (https://mcp.apify.com) as neverempty/html-table-to-json.

Limits

• Runs use 1 GB of memory (also the minimum), so that a 5 MB page of small cells fits; pages with large tables are turned into rows one at a time.
• Up to 1,000 URLs per run; up to 5 MB of HTML and 400,000 table cells per page; up to 100,000 rows per table; one dataset row per page is kept under Apify's 9 MB item limit (CSV, then rowsText, then the last rows of the largest tables are left out first, and the note says so).
• Tables built from <div> elements (role="grid") are not read; no-tables rows mention them when the HTML has them.
• In monitor mode, avoid two overlapping schedules with the same watchName.

Support

Something missing or wrong? Open an issue in the Issues tab with the URL, the table and what you expected. This Actor is not affiliated with Wikipedia or any site named on this page.