convexworks/cookie-banners

Cookie Banner Text and Controls

The wording, controls and markup of 6,738 cookie banners: 2,647 on the homepages of European companies listed in EU-Startups and 4,091 on those of their competitors, in Europe and elsewhere, as they were in October and November 2026. Each row is one banner: its text, every button, link and toggle on its first screen, and the banner's HTML, in the original wording. For 6,480 banners there is also a screenshot of the banner, and for 4,112 a screenshot of the settings window its settings button opens.

Use it to study how consent is worded, or to prototype tools that read banner text and button labels. It does not measure legal compliance.

Prepared by Convex Works from the European Software Data Study.

Quick start

import pandas as pd
from datasets import load_dataset

repo = "convexworks/cookie-banners"
banners = load_dataset(repo, "banners", split="train").to_pandas()
judgments = load_dataset(repo, "model_judgments", split="train").to_pandas()
df = banners.merge(judgments, on="domain", validate="one_to_one")
screenshots = load_dataset(repo, "screenshots", split="train")  # PIL images; .to_pandas() gives the PNG bytes

# One row per control, keeping the domain it belongs to.
controls = pd.DataFrame(
    [{"domain": row.domain, "consent_tool": row.consent_tool, **c} for row in df.itertuples() for c in row.controls]
)
buttons = controls[controls.kind == "button"].dropna(subset=["label"])

# Where the model found no direct reject button, what do the buttons say?
no_reject = df.loc[df.reject_button == "not_detected", "domain"]
print(buttons[buttons.domain.isin(no_reject)].label.str.casefold().value_counts().head(8))

label
accept                262
accept all            245
customize             162
cookie settings       144
allow all             113
accept all cookies    112
ok                    106
cookies settings       78

Many sites use a consent tool's default wording: the 6,738 banners have 4,377 distinct wordings. Use text_id to count each wording once.

Tables

Config	Rows	One row per
banners (default)	6,738	banner
model_judgments	6,738	banner: the model's reading of its accept and reject buttons
screenshots	6,480	banner with a screenshot that shows it as published

The tables join on domain, the company's domain. Each table has one row per domain. source in banners says which list a company comes from.

Of the 5,143 EU-Startups homepages, 2,951 had a cookie banner and 2,647 of those passed every quality check. The model found no banner on 2,018 and could not tell on 121; 53 sites are gone (an expired or suspended site, a parked domain, a page not found or a domain that no longer resolves).

Of the 8,491 competitor homepages, 8,276 could be read: 38 were gone and 177 could not be seen. 5,049 of those had a cookie banner and 4,091 of those are published. 3,027 had none, and 2,018 of these had no cookie or consent wording at all; on 200 the model could not tell.

Versions and segments

Version 2.0 covers every segment of the source directory: what the company sells (vertical, such as software, deeptech or consulting) and who pays (business_model, such as businesses or consumers). Version 1.x covered only B2B software, which is vertical = software with business_model = b2b or b2b2c. Every banner was judged again for 2.0, so rows for the same site can differ from 1.x. Filter on the segment columns in banners to keep a fixed population, and pin a version with load_dataset(repo, "banners", revision="v3.0.0").

Fields

banners

Field	Type	Description
domain	string	Join key: the company's domain as listed
source	string	The list the company comes from: eu-startups, listed in the EU-Startups directory, or eu-startups-competitor, a competitor of a listed company that is not listed itself
page_url	string	URL of the saved homepage after redirects
captured_at	timestamp (UTC)	When the page was saved: 3 November 2026 for the 4,091 competitor banners, 26 October 2026 for 844 banners, 21–22 June 2026 for the other 1,803
checked_at	timestamp (UTC)	When the banner was last seen on the live page with this row's wording: captured_at, or later, when the screenshot run saw it with the same words. For competitor banners it is captured_at. Of the others, 2,587 were seen in October and 60 June banners were not seen again (see How it was made).
vertical	string	Segment: what the company sells, such as software or deeptech_science
business_model	string	Segment: who pays, such as b2b, b2c or marketplace
page_language	string, nullable	The page's lang attribute as written, such as en-GB. Null when missing. The banner itself may use another language; see banner_language.
banner_language	string	Language of the banner's own words (text and control labels), detected by Meta's fastText language identifier lid.176, such as en or de. It differs from page_language on 236 banners, mostly English banners on pages in other languages.
banner_language_score	float	The identifier's probability for banner_language, from 0 to 1. Median 0.89. Short banners and closely related languages, such as Croatian and Serbian, score lower. No cutoff is applied.
consent_tool	string, nullable	Consent tool or website builder that made the banner, recognised from the banner's markup, such as Cookiebot, Cookie Notice or Wix Cookie Banner. For the eight tools listed under How it was made, the first screen was located with the tool's own markup; other banners use the whole banner element. Null means no known tool's markup was found: the banner is custom-built or from a tool not in the list.
text	string	Banner wording, one line per paragraph, starting with the banner's heading when it has one. Buttons and links that stand between sentences are removed; links inside a sentence stay.
controls	list of struct	Every control on the first screen, in page order
html	string	The banner's HTML in the same scope as text and controls: later layers, elements marked hidden and hidden form inputs are removed, as are scripts, styles and SVG; embedded data: images are shortened. Otherwise as saved, including classes, roles and inline styles.
text_id	string	Hash of text. Identical wording gives the same id.

The segment columns, vertical and business_model, come from gpt-5.4-nano, one pass over each company's homepage, about, pricing and demo text. For competitors they come from gpt-5.6-terra at high effort, with the same categories. They were not reviewed by a person.

Each control:

Field	Description
kind	button, link or toggle, following the HTML. An "Accept" built as a plain link is a link.
label	Visible text. Null when the control shows no text, such as a close icon.
aria_label	The aria-label attribute, if any
element	HTML tag, such as button, a, input or div
role	The role attribute, if any
href	Links only. Absolute target URL; null for in-page or script links.

model_judgments

Field	Description
domain	Join key
accept_button	detected, not_detected or unclear: whether the model found a direct accept button among the captured labels
reject_button	The same for a direct reject button, such as Reject, Deny, Decline or necessary-only. Settings and Manage buttons do not count.

Model gpt-6-sol at medium reasoning effort read each published banner: its text and control labels, with the same prompt as banner detection (privacy_banner_saved_html_core_v7). The answers therefore refer to the published controls. Accept: 6,214 detected, 516 not_detected, 6 unclear. Reject: 5,188 detected, 1,542 not_detected, 6 unclear. not_detected means not found among the published labels, not that the site has no such button. On 2 banners the model read the published text as not a cookie banner and answered not_applicable for both. No person reviewed these judgments.

screenshots

Field	Description
domain	Join key
captured_at	When the screenshots were taken: 24–26 October 2026, or 3 November 2026 for competitor banners and for the 445 banners whose settings window was taken again in 4.0.1
banner_image	PNG of the banner alone, in full, on a transparent background: the rest of the page is hidden and boxes that scroll inside the banner are unrolled
settings_image	PNG of the layer the banner's settings button opened, in the same way. Null when the banner has no settings button, or the screenshot was not taken or not kept, so null does not prove there is no button.

A banner gets a screenshot when the live banner had at least 90% of its row's words and the row at least 90% of the live banner's, so the image shows the banner in the row. 149 EU-Startups banners had changed by more than that between the save and the screenshot, and 107 were not seen live; they have no screenshot. Images that are blank or tiny are left out (2 banners), and so are settings screenshots without any cookie or consent wording, which caught another layer, such as a menu. A competitor banner with no screenshot of its own is left out of the dataset.

How it was made

1. Pages. Homepages of 5,143 companies listed in the EU-Startups directory, saved as rendered HTML by a headless browser on 21–22 June 2026. A homepage is included when it loaded with at least 500 characters of text, was not a parked, for-sale or unrelated page, had estimated search traffic, and its business could be identified.
2. October refresh. On 24–26 October 2026 the screenshot run (step 9) visited every published banner's homepage. Where the live banner had the same words as its June row in both directions (90% or more), the row was kept and its checked_at set (1,743 banners). Every other homepage of the 5,143 was saved again on 26 October 2026 (3,399 homepages), from Athens, by a browser with an ordinary Chrome's name, a 1365×900 window and the Berlin time zone, 12 seconds after loading and a mouse move. These went through steps 3 to 8 below; the new result replaces the June one. Of those, 53 were gone and 161 could not be seen (a bot check, an error or under 500 characters of text): these 161 keep their June result. Ten June banners the screenshot run saw live, but that did not show in time on the October save, also keep their June row. The October save had no Wappalyzer, so consent tools come from the page's own markup; Wappalyzer had added a tool to 52 of 8,257 June homepages.
3. Candidates. A script collected page fragments with cookie or consent wording, together with their controls.
4. Banner or not. gpt-6-sol at medium reasoning effort, run through the Codex CLI, judged whether each homepage's fragments contained a sitewide cookie banner. Across the June and October saves: 2,951 yes, 2,018 no, 121 unclear. The prompt is the one used for 1.x. Before the full run, gpt-6-sol, gpt-6-luna at low and medium effort, and the 1.x model mistral-medium-2604 classified the same 300 random homepages. On the 36 homepages where any answer differed, Claude (an AI model, not a person) checked each answer against the prompt's rules: gpt-6-sol was wrong on 3, mistral-medium-2604 on 6, and gpt-6-luna on 10 (low) and 13 (medium). The 264 homepages where all agreed were not checked.
5. First screen. For eight consent tools (Cookiebot, CookieYes, HubSpot Cookie Policy Banner, Complianz, Finsweet Cookie Consent, iubenda, OneTrust and CookieScript), the first screen is located with the tool's own markup. Later layers such as settings panels, vendor lists and detail tabs are left out. For Complianz, the hidden parts follow its default stylesheet. Other banners use the whole banner element.
6. Cleaning. Elements the HTML marks as hidden (the hidden attribute, inline display: none or visibility: hidden, or the tool's hide class) and screen-reader-only text are left out. Whitespace is normalised; the wording is otherwise unchanged. Emails and phone numbers are replaced with [email] and [phone] (9 published banners have an email replaced and 13 a phone number; some of the 13 are other long numbers in the markup, not phone numbers).
7. Selection. 2,647 of the 2,951 banners pass every check. The rules were set by inspecting examples. No person checked the released rows.
8. Buttons. gpt-6-sol read each released banner's text and control labels again and judged whether it has a direct accept and a direct reject button (model_judgments).
9. Screenshots. A browser in Athens, with the same window, opened each published banner's homepage again, found the banner (by the published element's id or classes, else by its words), hid the rest of the page and took a screenshot. It then clicked the settings button gpt-6-sol marked among the banner's controls, or a settings label other banners of the same tool use, and took a screenshot of the layer that opened. Nothing was accepted or rejected. The first pass, on 24 October, gave the browser's headless name. The banners it missed, and the new and changed ones, were looked for on 26 October with an ordinary Chrome's name, and again after a mouse move and a scroll when they did not show within 12 seconds. Banners not found from Athens were tried again through a Frankfurt VPN (13 found). On 3 November 2026 the settings windows of the Cookiebot, iubenda and OneTrust banners were taken again from Copenhagen, with the same browser and window, because boxes that scroll inside a settings window had been squeezed back and their last lines hidden under its buttons. Where the live banner still had the words of its row (90% both ways), both screenshots were replaced (445 banners).
10. Competitors. The competitors of the EU-Startups software companies that are not listed themselves, whose business could be identified and that have estimated search traffic: 8,491 homepages. On 3 November 2026 a browser in Copenhagen, with the same window, Chrome name, time zone and waits as the October save, visited each homepage once. It saved the page before clicking anything, took a screenshot of each thing on screen that had a known consent tool's markup, or was a fixed, sticky or dialog box with cookie or consent wording, then clicked a settings button whose whole label is one the model marked on the EU-Startups banners, and took a screenshot of what opened. Nothing was accepted or rejected. Steps 3, 5, 6, 7 and 8 are as above. Step 4 differs: the model did not read every page. For 4,636 homepages, the page fragment closest to what was on screen (at least 70% of its words shared both ways) was taken as the banner, with no model. gpt-6-sol read the other 1,622 with cookie wording: those with nothing close, and those with nothing on screen (413 yes, 1,009 no, 200 unclear). The 2,018 with no cookie or consent wording count as no banner: in the October refresh all 1,296 such homepages came back as no banner. Run both ways on 130 random homepages, 95 of the 100 with something on screen gave the same row, 4 had nothing close, and 1 took a longer fragment. In step 8 the model reads every competitor banner, and one it does not call a banner is left out (30). A competitor banner is published only with a screenshot of its own (356 left out). Settings clicks on labels such as "Learn more", which also lead elsewhere, are not kept.

Exclusion reason	Banners	Meaning
no_screenshot	356	Competitors only: no screenshot shows the banner
repeated_controls	196	The same control appears twice, from a second layer or a duplicate layout
contains_tabs_or_form_fields	115	Tabs, dropdowns or form fields inside the banner
too_many_controls	105	More than 12 controls
text_too_long	82	Over 1,500 characters, with page or settings text mixed in
no_consent_wording	75	No cookie or consent wording
banner_hidden_in_saved_html	75	The banner was marked hidden when the page was saved
contains_page_content	55	The banner element holds navigation, main page content or over 50,000 characters of markup
control_label_too_long	47	A control label over 60 characters
unlabelled_toggle	43	A toggle with no label
duplicate_page	33	The same page as another listed domain
unsupported_tool_layout	31	A Complianz layout the rules do not cover
not_a_banner	30	Competitors only: the model did not read it as a cookie banner, mostly a settings window or a chat prompt
no_notice_text	15	No notice text left after cleaning
multiple_banner_candidates	4	More than one possible banner element

Limitations

• Saved HTML is not a screen recording, and stylesheets loaded from separate files were not available. The screenshots show how a banner and its settings window looked once, in one desktop window, not what happens after a choice.
• Each banner was seen from one place at one time: Athens, except 13 screenshots taken through Frankfurt, the 445 banners and settings windows taken again for 4.0.1 from Copenhagen, and every competitor banner, seen from Copenhagen. A site may show visitors elsewhere another banner or none. 1,803 rows keep the wording saved in June: 1,743 of them were seen live in October with the same words, 60 were not seen again.
• 425 published banners are on homepages where the June save found none. Whether they were added after June or did not show in time in June cannot be told apart.
• The dataset covers companies listed in EU-Startups and competitors of its software companies whose homepage passed the checks above, not a sample of all websites. It does not say whose competitor a company is.
• The two sources were not selected in the same way: competitor banners need a screenshot of their own and were pre-screened by what was on screen, so compare them with care.
• 1,532 of the 4,091 competitor banners have no settings screenshot. A rough look at their button labels suggests about 440 of them do have a settings button: its label was not one we knew, or the click or screenshot failed.
• kind follows the HTML: accept or reject options built as links appear as link.
• A few banners still have a button label run into the text, where a link and a button sit next to each other with no space, such as cookie policyAccept.
• Banner detection, the accept/reject judgments, the settings buttons and the segment columns come from model passes and were not reviewed by a person. Screenshots were checked by eye on random samples only.
• The dataset is not evidence of legal compliance or non-compliance.

Licence and rights

Our selection, structure and labels are released under CC BY 4.0. The banner wording is quoted from the websites listed; rights in it remain with the site owners and consent-tool vendors, and CC BY 4.0 does not cover it.

Corrections and removal

Site owners can ask for a row to be corrected or removed by writing to hello@convex.works or opening a discussion in this dataset's Community tab. Changes are applied in the next version.

Citation

@misc{convexworks_cookie_banners_2026,
  title  = {Cookie Banner Text and Controls},
  author = {{Convex Works}},
  year   = {2026},
  url    = {https://huggingface.co/datasets/convexworks/cookie-banners},
  note   = {Version 5.1.0}
}

Version

5.1.0 (November 2026): adds 4,091 banners from the competitors of the EU-Startups software companies, for 6,738 in all. banners gains source, eu-startups or eu-startups-competitor. screenshots now has 6,480 banners and 4,112 settings windows. The 2,647 EU-Startups rows are unchanged.

5.0.0 (November 2026): fewer, cleaner columns. banners drops offering_type, which mostly repeated vertical, and title: a banner's heading, when it has one, is now the first line of text for every banner (it was split out for 918 banners from six consent tools). controls drops checked_in_html and disabled_in_html, which showed the saved markup, not what the visitor saw. text_id is now a hash of text, so every id changes; there are 1,782 distinct wordings. Rows, model_judgments and screenshots are unchanged.

4.0.1 (November 2026): fixed settings screenshots. In 4.0.0, a box that scrolls inside a settings window was unrolled but squeezed back by the box around it, so its last lines slid under the window's buttons: Cookiebot's "last updated" line, whole iubenda rows, and some OneTrust rows. The settings windows of 522 Cookiebot, iubenda and OneTrust banners were taken again on 3 November 2026, from Copenhagen; the 445 whose live banner still matched its row replace their 4.0.0 screenshots, the others keep them. screenshots now has 2,389 banners and 1,553 settings windows. banners and model_judgments are unchanged.

4.0.0 (October 2026): the dataset follows the sites as they are now. Every homepage whose banner was not seen unchanged was saved again on 26 October 2026: changed banners are updated, banners that are gone are dropped, and banners found for the first time are added, for 2,647 banners in all. banners gains checked_at. A new screenshots table has a screenshot of 2,387 banners and of 1,543 settings windows.

3.0.0 (October 2026): the sites table is removed, so the dataset lists only homepages with a published banner. Its segment columns vertical, business_model and offering_type move into banners; consent_tools_in_page is dropped. Banners and judgments are otherwise unchanged.

2.1.0 (October 2026): banners gains banner_language and banner_language_score, the language of the banner's own words, detected with fastText lid.176.

2.0.1 (October 2026): consent_tool names any known consent tool or website builder found in the banner's markup, such as Cookie Notice or Wix Cookie Banner, instead of only eight tools; 679 banners (29%) remain unknown. consent_tools_in_page recognises the same tools.

2.0.0 (October 2026): covers every segment of the source directory, not only B2B software: 5,143 homepages and 2,343 banners. Every homepage, including the 1.x ones, was judged again by gpt-6-sol with the same prompt, so banner detection can differ from 1.x for the same site. model_judgments are judged from the published banner, so they match controls, and can be unclear. Buttons and links that stand between sentences no longer appear in text.

1.1.0 (October 2026): consent_tools_in_page counts a tool only from its own script or banner markup, which removes false matches such as tool names in cookie lists. HubSpot's banner is named HubSpot Cookie Policy Banner. Buttons built as links without an address are now controls (10 banners). rechecked is removed from model_judgments.

1.0.0 (October 2026): first release, B2B software companies.