Privacy

Privacy policy

Semblen LLC · Effective 2026-11-06 · Questions: [email protected]

Who we are

The site at semblen.com ("Semblen") is operated by Semblen LLC, a Texas limited liability company. Semblen LLC is the controller of the personal data described here. Contact: [email protected] · 5900 Balcones Drive STE 100, Austin, TX 78731.

The short version

You do not need an account. If you sign in, we store your email address; we do not know your name unless you tell us.
We keep the text you submit for analysis, with your IP address and browser user-agent, for 180 days, then delete it.
The text is sent to a small number of service providers so the analysis can happen. They are named below.
We do not sell personal data or show advertising. Our pages load one third-party tag, from Google Ads, to measure whether our own ads bring visitors; for visitors in the United States it sets Google’s advertising cookies for that count (not to personalise advertising, and never if your browser sends Global Privacy Control), and elsewhere it sets none. If you use the analyzer without signing in, it sets one cookie, holding a random code, to count free analyses fairly; apart from those, this site sets no cookies unless you sign in, and then only the one that keeps you signed in.
Do not submit anything confidential or personal that you would not want stored for 180 days.

What we collect, why, and on what basis

Each use below names its lawful basis under the EU and UK General Data Protection Regulation, which applies because our servers are in Germany and we serve visitors in Europe.

What you analyze. When you analyze a link, a video or pasted text, we store the text you submit — for a link or video, the paragraphs or transcript segments we analyzed — with the results, the source link, the time, your IP address and your browser's user-agent. Part of every analysis is a short summary (the TL;DR) and a read of how the piece is built (the kind of piece, who is quoted, the headline against the article); for pasted text this is done when it is 150 words or more. So that the same link is not analyzed twice, the analysis of a link (the article, not who asked for it) is cached for a day, or for 30 days when the article is one of those shown on our own /daily pages and 180 days for our /government pages; pasted text and its result are cached for a day under a one-way code of the text, then deleted. If you press Share, we keep the analysis, and for pasted text the text, under a link anyone who has it can open, for 180 days, then delete it. A shared link holds no IP address or account; the page tells you before it makes a link to pasted text. If your browser is set up the way an automated crawler's we have seen is (its language setting and type), the record also carries a mark saying so, which keeps it out of our counts of analyses. Purpose: to perform the analysis you asked for, to find and fix errors in our models, and to prevent abuse (limits are applied per IP address). Basis: performance of the service you requested (contract), and our legitimate interest in running and securing it.
The browser extension. The Semblen extension does nothing until you click its icon. When you do, it reads the article text of the page you are on (or the text you selected) and sends that text to us for analysis; for a whole page it also sends the page's title, which we use to describe the article and do not keep apart from the analysis. It does not send your browsing history or anything from other tabs. For a whole article of 100 words or more (not a selection), it also sends the page's address once the analysis is done, to compare the article with other outlets' coverage of the same story: the address tells us which outlet published it, and the comparison is saved under a one-way code of the address, not the address itself, so that the next reader of the story gets it at once. The saved comparison can quote the article, is not linked to you, and is kept for up to 7 days. If you press "Check the claims", the address also goes with that request, so the article is not used as evidence for its own claims, and we do not store it. We keep the result, the first 200 characters of the text, the time, your IP address and your browser's user-agent; we do not keep the full text in our records and do not use it to train our models. So that the same page is not analyzed twice, the text and its result are held in a cache for 24 hours and deleted within two days. The historical background and any claim checks for the page, including the sentences that were checked, are kept as they are for the website (30 days). The extension also reports anonymous counts of which parts of an analysis are used (for example that the claim checks were run), with a random identifier that lasts until its panel is closed; these never include the page, its address or its text. Basis: performance of the service you requested, and our legitimate interest in preventing abuse (the free daily limit is applied per IP address).
Accounts. You do not need an account. If you sign in, we store your email address, the date the account was created, how it was made (an email link or Google), your plan, the page or link the account was made from (for example a race page, so we can count which pages bring sign-ups), and a one-way code for each sign-in link and each signed-in browser (never the link or the token itself). We use the address to send the sign-in link you ask for. If you choose to sign in with Google instead, Google tells us your email address and that it is verified, and nothing else; we ask for no other part of your Google account, and Google learns that you signed in to Semblen. A new account gets a 14-day trial of Semblen Pro, with no payment details asked for; when it ends the account returns to the free plan unless you subscribe. To give one trial per person, we keep a one-way code of each account’s email address (ignoring dots and anything after a “+” in a Gmail address, and anything after a “+” in others), even after the account is deleted, and a new account whose address matches one starts on the free plan without a trial. During the trial we keep a running count of the analyses the account runs; when the trial ends we keep only an anonymous figure (the day it ended, the count, and whether the account subscribed) with nothing tying it to the account, and delete the count kept with it. A new account gets a welcome email, and a reminder three days before its trial ends. While you are signed in, your analysis allowance (a day's on the free plan and during the trial, a month's on Pro) is counted against your account instead of your network address. Analyses are not linked to your account: the count is a number per account per day or month, and earlier days' and months' counts are deleted. If you set up a daily digest, we store the outlets and topics you chose and when a digest was last sent, and email it to your account's address until you switch it off or unsubscribe; the stories analyzed for it are kept as short summaries (headline, link, the flagged passages), not the articles, with a record of which ones each digest sent so that none is sent twice. Now and then we email Pro and trial accounts about new features; every such email has an unsubscribe link, and unsubscribing stops them. Basis: legitimate interest in telling you about the service you use. If you subscribe, the purchase is made through Stripe, which acts as the seller of record and handles payment and tax: we send Stripe your account's email address and a reference to the account, so the payment is tied to it, and we never receive or store your card details. We store Stripe's identifiers for you and your subscription, the plan you chose, its status and its renewal date; Stripe holds your payment details and billing history under its own privacy policy. Umami (below) also records the amount of each payment, with no name, email address or account attached. So that a first payment is counted with the visit it came from, we keep the IP address and browser user-agent from the request that starts a checkout until the payment arrives; Umami makes its daily visitor code from them, as for a page view, and we delete them then, or by the next day if no payment arrives. Basis: legitimate interest in knowing which of our pages and links bring subscriptions. You can delete the account, and everything in this paragraph with it except that one-way code, from the account page at any time. Basis: performance of the service you requested.
Improving our models. We may use submitted text, results and feedback to evaluate and improve our detection models, including as training examples. Basis: your consent, given when you submit text for analysis — the analyze form says so beside the button. You may withdraw it by asking us to delete what you submitted (below); withdrawal does not affect processing that already took place. Text analyzed through a teacher's class link is never used to train or evaluate our model.
Claim checks. When a claim check runs, we store the claims found, the sources retrieved and the verdicts. Results for the same text, and a verdict for the same claim checked against the same sources, are reused for up to 30 days. Basis: performance of the service.
Feedback. If you press Agree, Disagree, thumbs up or thumbs down, we store that choice, what it refers to, and anything you type. Basis: legitimate interest in correcting our results.
Newsletter. The sign-up closed on 24 October 2026. If you subscribed before then, we store your email address, the date and the page you subscribed from, and use it only to send the newsletter. Basis: consent, which you withdraw by unsubscribing.
Weekly story email. On Fridays, when a story page on Semblen changed that week, we email a roundup of those stories: for each, what happened, who covered it, and the ways it is being told. It goes to every account, on any plan, and to every newsletter subscriber, each address once; for a newsletter subscriber it is the newsletter. We use only the email address we already hold, and keep whether each copy was delivered, bounced or marked as spam, without the address. Every copy has an unsubscribe link; unsubscribing stops it, and our emails about new features with it. Basis: for accounts, legitimate interest in telling you about the service you use; for newsletter subscribers, the consent you gave to the newsletter, which you withdraw by unsubscribing.
Ratings of ad flags. If you rate a technique we flagged in a political ad, or tell us we missed one, we store your rating, anything you write, and a one-way code derived from your network address so each reader counts once. The code cannot be turned back into your address. We use the ratings to check and improve the model. Basis: legitimate interest in the accuracy of the analysis.
Feedback on the “Other outlets” comparison. If you tell us an angle in the comparison is right or wrong, that an outlet or angle is missing, or whether a question it suggests is fair, we store your answer, anything you write, the article's address, and the same kind of one-way code derived from your network address so each reader counts once. We use it to check and improve the comparison. Basis: legitimate interest in the accuracy of the analysis.
Race updates. If you follow a race, we store your email address, the race, the date and the page you signed up from, and email you a link to confirm. Once confirmed, we use it only to send updates on that race: until Election Day, a guide to the race when you confirm and a reminder the day before; then the race's new ads, the claims we checked in them, and debate analyses. Our email provider tells us whether each email was delivered, bounced or marked as spam, and we keep that record without your address. Basis: consent, which you withdraw with the unsubscribe link in every email.
Story updates. If you follow a story, we store your email address, the story, the date and the page you signed up from, and, unless you are signed in, email you a link to confirm; confirming it also makes your free account (see Accounts above). Once confirmed, we use it only to send updates on that story when its page changes (a new way of telling it, a new week in how it’s framed, or a US government statement), at most once a day in the first week and then once a week, each with a quote and a link; we keep when we last sent you one and what the page showed then. After two weeks with no change we send one last note and the follow ends. Our email provider tells us whether each email was delivered, bounced or marked as spam, and we keep that record without your address. Basis: consent, which you withdraw with the unfollow link in every email.
Classroom and monitor requests. If you fill in the form on the classroom or monitor page, we store your email address, the organisation or school and anything else you tell us, the date and the page, and we email you about that request only. Basis: consent, which you withdraw by asking us to remove you.
Pro waitlist. The waitlist closed on 2026-10-06, when Pro opened. Everyone on it was told, and the list was deleted on 2026-10-25.
Free limits. To apply the free daily limits to readers who are not signed in, we keep a count of analyses per day for each browser, so that a class, library or office on one network does not share one allowance. A cookie, semblen_browser, tells one browser from another: it holds only a random code, signed by us so it cannot be made up, lasts a year, is not readable by scripts, and is used for nothing but this count. We store the count under a one-way code of it, not the code itself. A request without that cookie is counted by its network address instead, and every network also has a ceiling of 30 free analyses a day across all its browsers, so clearing cookies does not start a new count. A network address is stored as a salted one-way code of the address, which cannot be turned back into it, rather than the address itself (for an IPv6 address, the code is of its network part); the website and the extension share one count. Earlier days' counts are deleted. So that the same article counts once a day, on every plan, we also keep a one-way code of each article counted that day together with that browser or network code or account, never the article's address or text; earlier days' codes are deleted. We also count, for each day, how many free allowances were used up and how many free analyses were turned away, and whether by a browser's limit, a network's or the network ceiling, as numbers with no address or code kept. A teacher with Semblen Teacher Pro has a class link. Opening it sets a cookie, semblen_class, that holds the link's code for a week, is not readable by scripts, and is used only to count that browser's analyses against the class's allowance. We keep those counts under the teacher's account and a one-way code of the browser, until the day or month is over. Basis: legitimate interest in preventing abuse; the cookie is strictly necessary to provide the free allowance you ask for.
Visits. We record page views and a few in-product events with the page path, the referring site, the tag of the link you followed if it carries one (the ref= part of a link we shared), a random identifier that lasts for the browser tab, whether this browser has visited before, and the device type (phone, tablet or computer, worked out from your browser's user-agent, which is not stored). When the browser is set up the way an automated crawler's we have seen is (its language setting and type), the record also carries a mark saying so, so we can keep such visits out of our counts; neither setting is stored. This record holds no IP address. A page view of a ZIP code lookup is recorded with the state, not the ZIP code. Our web server and network provider keep standard access logs, which do. We also count visits with Umami, open-source analytics that runs on our own server, not a third party's: for each page view and the same events it records the page address (with the link's tag), the referring site, the browser, operating system, device type, screen size and language, how quickly the page loaded and responded, and the country, region and city worked out from your IP address. It tells one visitor from another by a code made from the IP address and browser that changes every day, so visits cannot be joined across days, and it does not store the IP address. It sets no cookies. For a sample of visits, Umami also keeps a replay: how the page looked as it changed, and where you scrolled, tapped and clicked, so we can see where the site is hard to use. Anything typed into a form is hidden, and text that may be yours is left out of the recording entirely: text you pasted for analysis and its results, and your email address once you sign in. A replay carries the same daily visitor code as Umami's other records and nothing else about you. Pages of ours embedded on other sites are never recorded. When an account is made, Umami records that as an event in the same visit (how the account was made, the tag of the link that brought you, and whether a trial started), counted the way a page view is; no email address or account goes with it. Basis: legitimate interest in operating and securing the site.

Cookies and browser storage

If you use the analyzer without signing in, we set one cookie, semblen_browser, which holds a random code used only to count free analyses per browser (see Free limits above); it lasts a year, is not readable by scripts, and is not used to track you. Apart from it and Google’s cookies for visitors in the United States (below), this site sets no cookies unless you sign in. If you open a teacher's class link, we set semblen_class (see Free limits above). If you sign in, it sets one cookie, semblen_session, which holds a random token that keeps you signed in on that browser for 30 days or until you sign out; it is not used to track you and is not readable by scripts. Signing in with Google also sets a cookie for up to 10 minutes, semblen_oauth, which ties Google's reply to your request and is deleted when that reply arrives. Signed in or not, the site keeps in your browser's session storage, erased when you close the tab, one random identifier and the tag or referring site of the link the tab arrived from (so a visit can be counted with the link that brought it), a flag in local storage that says only that this browser has visited before, and, if you sign in, a flag that says this browser is signed in (so the page header can show your account). The flag holds no identifier, so it cannot link one visit to another; clearing your browser's site data removes it. Our network provider, Cloudflare, may set a security cookie to tell people from bots. For visitors in the United States, the Google Ads tag on our pages sets Google’s advertising cookies on semblen.com: _gcl_au, and _gcl_aw, which holds the click identifier, if you came from one of our ads; each lasts 90 days. Google may also set its own cookies on its domains. They are used to count our ads’ results, not to personalise advertising, and a browser that sends Global Privacy Control gets none. Outside the United States the tag runs with consent denied and sets no cookies. Umami, our visit counter, sets no cookies and keeps nothing in your browser.

Who processes it

We share personal data only with processors that need it to run the site, each bound by its own terms:

Anthropic (United States) — text you submit is sent to the Claude API to write technique explanations, the TL;DR and the article read, historical context and what an article leaves out, claim verdicts, the comparison with other outlets' coverage of the same story, and questions to consider; it also gives a second read of passages our model flagged, says whose words a flagged technique is in (the writer's or someone quoted), lists who is quoted, and, for a digest, suggests search terms for the topics you chose. For questions to consider on an opinion piece, Claude may search the web with Anthropic's search tool, using terms from the article. When an article has charts, we load those images from the publisher's site and send them to Claude to check how they are drawn. For our story pages, we load the lead photo of each article a story covers from the publisher's site and send it to Claude to describe what it shows. To check a statistical claim, Claude may look up official figures from FRED, the BLS, the EIA, the SEC, the US Treasury and the UK's ONS, which receive only the series, company or search terms looked up, and may search official sites and press-release wires with Anthropic's search tool. To say who is behind the site of an article you analyze, we send the text of that site's homepage footer and its About and terms pages (not the article, and nothing about you) to Claude, and look the site up on Wikidata, Wikipedia and the public domain registry (RDAP), which receive only the site's address. Under Anthropic's commercial terms it does not train its models on this content.
Exa (United States) — the text of a claim being checked is sent as a search query. Jina AI fetches the source pages found. Google Fact Check Tools may be queried with keywords from a claim.
Supadata — the link of a video you ask us to analyze is sent so its transcript can be retrieved. YouTube is asked for the video's title. Our challenge page shows videos from YouTube's privacy-enhanced player (youtube-nocookie.com), so YouTube receives your IP address and browser details when you open that page, and may store data in your browser if you play one.
X (United States) — our /government pages show the images attached to official accounts' posts on X. Your browser loads them straight from X's servers (pbs.twimg.com), so X receives your IP address and browser details when you open such a page. We send X no referrer, so it is not told which of our pages you were on, and we send X nothing about you ourselves.
News publishers — our story pages show the lead photo of each article a story covers, as its publisher published it. Your browser loads each photo straight from the publisher's site, or the image service it uses, so that site receives your IP address and browser details when you open a story page. We send no referrer, so it is not told which of our pages you were on, and we send publishers nothing about you ourselves.
Webshare — when a news site refuses a direct request for an article you asked us to analyze, we fetch it again through Webshare's proxy network, which then sees the article's address. It never receives your IP address or anything else about you.
Octen (United States and Singapore) — when a news site refuses or bot-checks a direct request for an article you asked us to analyze, or sends back almost none of it, we first ask Octen’s extract service for the page, and it then sees the article’s address. It never receives your IP address or anything else about you. If Octen can’t get the article either, we try Webshare.
Resend (United States) — sends our email, so it receives your address and the message.
Stripe (United States) — if you subscribe or make a payment, Stripe is the seller of record: it collects your card and billing details and any tax; we receive only its identifiers and the state of your subscription.
Google — only if you choose to sign in with Google: it tells us your email address.
Google Ads (United States) — our pages load its tag, which tells Google that a page was viewed, that an analysis was started (not what was analyzed), or that a new account started its trial (with a one-way code for counting it once, never the email address or the account), and, if you came from one of our ads, the click identifier in the link, with your IP address and browser details as any web request carries. Outside the United States it runs with consent denied and sets no cookies. In the United States it sets Google’s advertising cookies (see Cookies above) so Google can tell which ad visit an analysis or a trial came from; your browser can refuse them by sending Global Privacy Control, and then the tag sets none. Everywhere, Google uses this only to count our ads' results, not to personalise advertising. Basis: legitimate interest in measuring our own advertising.
Cloudflare — all traffic to the site passes through its network. Signed-out analyses use Cloudflare Turnstile to check for bots. Cloudflare runs a check in your browser when you open the analyzer, and on our other pages once you start to use something on it that analyzes (you click into a paste box, or a “Check this ad yourself” link comes into view), so an analysis can start at once. A new analysis (not one we already have) runs only after the check has passed, and we send Cloudflare the check’s result and your IP address to confirm it. We keep only a daily count of checks passed and failed.
Sentry (Functional Software, Inc., United States) — when something on the site breaks, our server or your browser sends Sentry a report of the error: what failed and where in our code, the page's address with anything you typed or that could identify you taken out (text you pasted, email addresses, the codes in sign-in and shared links, a ZIP code), and your browser and operating system. A report from your browser reaches Sentry with your IP address, as any web request does; Sentry is set not to store it, but notes the country and city it points to. No replays, cookies, account details or the text you analyze are sent. Basis: legitimate interest in keeping the site working.
Hetzner — our servers, and everything we store, are in Germany.
Modal (United States) — model training runs there and may include examples drawn from submitted text.

We do not sell personal data and do not share it for advertising. We disclose it if the law requires.

International transfers

Our servers are in Germany. Anthropic, Modal, Resend, Stripe, Cloudflare, Google and the search providers are in the United States; Octen is in the United States and Singapore. Transfers to them rest on the providers' standard contractual clauses or their participation in the EU–US Data Privacy Framework, as each provider's terms state.

Retention

Analysis records (submitted text, results, IP address, user-agent), per-paragraph results and feedback: 180 days, then deleted automatically.
Cached analyses of a link: a day, or 30 days for articles shown on our /daily pages and 180 days for those on /government; they hold the article and its result, not who asked. The text of each /daily article as we analyzed it, kept to tell readers when the outlet later changes the article: the same 30 days. The analysis of the one example article our home and /try pages link to: kept while it is the example. Cached pasted text and its result: a day.
Shared analyses (a link made with Share, and for pasted text the text): 180 days, then deleted automatically. Anyone with the link can open it until then.
What we found about who is behind a site (what it says about itself, what Wikipedia lists, when its address was registered): 30 days.
Passages from articles on our /daily pages chosen for the daily challenge (/challenge), with their analysis: shown only on the challenge, once we approve the pair; kept 30 days from when they are chosen (or until we reject the pair), then deleted.
Cached claim checks and historical-context results: 30 days. Cached comparisons with other outlets: up to 7 days.
Counts for the free limits, and the codes of the articles counted each day: until the day or month is over.
Ratings of ad flags and feedback on the Other outlets comparison: kept with the one-way code, not your address, until we delete them.
Umami replays: 30 days, then deleted automatically.
Error reports at Sentry: 30 days, then deleted by Sentry.
Umami visit records: 180 days, then deleted automatically. Our own visit records, which hold no IP address or lasting identifier: kept, to see how the site is used over time.
Account: until you delete it. Sign-in links: one day. Signed-in sessions: 30 days, or until you sign out.
Newsletter address: until you unsubscribe.
Race-update addresses: until you unsubscribe; we keep a record that you did, so you are not emailed again, and of which race guides and debate emails we sent, so none is sent twice. If you unsubscribe from our update emails, we keep your address on a do-not-email list.
Story-update addresses: until you unfollow or the follow ends; we keep a record that it did, so you are not emailed about that story again unless you follow it again.
Text kept as a training or evaluation example is separated from the record it came from and from your IP address and user-agent; it is retained while the model that learned from it is in use.
Web server access logs: rotated as they fill, within 30 days.
Server logs (what the site's software notes while it runs: its errors and warnings, sometimes with the address of the page or article it was working on): 30 days, then deleted automatically.

Your rights

If you are in the EU, the UK or another place with similar law, you have the right to access the personal data we hold about you, to have it corrected or erased, to restrict or object to processing based on legitimate interest, to portability of data you gave us, and to withdraw consent at any time. Write to [email protected]; if you have an account, tell us its email address; because analyses are not linked to accounts, also tell us roughly when and what you analyzed so we can find it. We answer within one month. You also have the right to complain to a supervisory authority — in Germany, the data protection authority of the state where our servers are, or the authority where you live. Residents of California and other US states with privacy laws have comparable rights of access and deletion, and we do not sell or share personal data as those laws define it.

Children

The site is not directed to children under 13 and we do not knowingly collect their data; if we learn we have, we delete it.

Security

Traffic is encrypted in transit. Stored data is on servers only we administer. No system is perfectly secure; please do not submit sensitive material.

Changes

If this policy changes in a way that matters, we will change the effective date above and say so on this page.

2026-11-06: Every account, on any plan, and every newsletter subscriber now gets a weekly email of the story pages that changed that week, with an unsubscribe link that stops it. For newsletter subscribers it is the newsletter.
2026-11-06: The daily challenge (/challenge) can now pair two news outlets’ coverage of the same story: we keep the paragraph shown from each article, with its analysis, for 30 days from when it is chosen, then delete it.
2026-11-05: If you are signed out, Cloudflare’s check for bots (Turnstile) now also runs on our pages when you show you are about to analyze something: you click into a paste box, or a “Check this ad yourself” link comes into view. So when you press Analyze, the analysis can start at once instead of waiting for the check. For visitors in the United States, the Google Ads tag now sets Google’s advertising cookies (_gcl_au, and _gcl_aw if you came from one of our ads), so Google can count which of our ad visits lead to an analysis or a trial. They are not used to personalise advertising, and a browser that sends Global Privacy Control gets none. Outside the United States the tag still sets no cookies.
2026-11-04: Teachers can now subscribe to Semblen Teacher Pro, which gives them a class link. A student who opens it gets a cookie, semblen_class, holding the link's code for a week, so their analyses count against the class's allowance instead of the free one; no student account is made.
2026-11-02: You can now follow a story from its page on Semblen: we store your email address and the story, and email you updates on it until you unfollow or two weeks pass with no change. Our /government pages now show the images attached to official accounts’ posts on X. Your browser loads them from X’s servers (pbs.twimg.com), so X receives your IP address and browser details when you open such a page. We send X no referrer, so it is not told which of our pages you were on. Story pages now show the lead photo of each article they cover: your browser loads it from the publisher’s site, which receives your IP address and browser details, and we send each photo to Claude to describe it.
2026-10-30: Umami now keeps a replay of a sample of visits: how the page looked and where you scrolled and tapped, to see where the site is hard to use. Anything typed into a form is hidden, and text you pasted for analysis, its results and your email address are left out. Replays are deleted after 30 days.
2026-10-30: When something on the site breaks, an error report goes to Sentry (United States), with text you pasted, email addresses, sign-in and shared-link codes and ZIP codes taken out; Sentry keeps it 30 days.
2026-10-28: Share now makes a link to the analysis itself (semblen.com/a/…), and opening a shared link shows that analysis as it was and runs nothing. If you press Share, we keep the analysis, and for pasted text the text, under a link anyone who has it can open, for 180 days, then delete it. A shared link holds no IP address or account. Before a link to pasted text is made, the page says that anyone with it can read the text. Every analysis you ask for now counts toward your allowance on every plan, and the same article counts once a day: to do that we keep, for that day only, a one-way code made from the article and your network code or account, never the article’s address or text. Without an account, the extension’s analyses now count toward the same free daily count as the website’s, instead of a count of their own. Without an account, the free daily count is now kept per browser instead of per network, so a class or office on one network no longer shares two: a cookie, semblen_browser, holds a random code for that count and lasts a year, and each network still has a ceiling of 30 free analyses a day. Signed-out analyses use Cloudflare Turnstile to check for bots: before a new analysis runs (not one we already have), Cloudflare runs a check in your browser, and we send Cloudflare the check’s result and your IP address to confirm it. We keep only a daily count of checks passed and failed. To give one trial per person, we keep a one-way code of each account’s email address (ignoring dots and anything after a “+” in a Gmail address, and anything after a “+” in others), even after the account is deleted, and a new account whose address matches one starts on the free plan without a trial.
2026-10-27: An account now records how it was made (an email link or Google), and sign-ups are counted in Umami with the visit they came from. During a trial we count its analyses, and keep only an anonymous figure once the trial ends. We also count, per day, how many free daily allowances were used up.
2026-10-25: Analyses of the articles on our /daily pages are now kept 30 days instead of 180. We also keep the text of each /daily article as we analyzed it for those 30 days, to compare with the current article and tell readers when the outlet has changed it. Corrected and completed after a review: the free-limit code for a network address is now salted, so it cannot be reversed; cached claim checks and historical-context results, and the record of cached analyses you were shown, are now deleted from our disk on time, not only no longer used; deleting an account now also deletes the record of which stories its digests sent; the closed newsletter sign-up stores nothing; the Pro waitlist has been deleted. Added: the short-lived cookie used when you sign in with Google, what the browser tab keeps about the link you arrived from, the device type in our visit records and that we keep those records, the YouTube videos on the challenge page, the other analyses Claude writes, the web search used to check statistics, that Stripe receives your email address at checkout, and the records kept with a race-update address. New: to say who is behind the site of an article you analyze, we send its homepage footer and About and terms pages to Claude and look the site up on Wikidata, Wikipedia and the domain registry; the result is kept 30 days. Visit and analysis records from a browser set up the way an automated crawler’s is now carry a mark saying so, to keep them out of our counts. A new account now also records the page or link it was made from (for example a race page), to count which pages bring sign-ups.
2026-10-24: When an article has charts, we load those images from the publisher’s site and send them to Claude to check how they are drawn. To check a statistical claim, Claude may look up official figures from FRED, the BLS, the EIA, the SEC, the US Treasury and the UK’s ONS; those services receive only the series or company looked up. The newsletter sign-up and the feedback form are gone; addresses already on the list still get our update emails, each with an unsubscribe link.
2026-10-22: When a news site blocks a direct request for an article you asked us to analyze, we now first ask Octen, an extract service based in the United States and Singapore, for the page, before Webshare’s proxy. Octen sees only the article’s address. The 14-day Pro trial now has the free plan’s 5 analyses a day, with Pro’s features. A new account now gets a welcome email, and a reminder three days before its trial ends. The Google Ads tag now also tells Google when a new account starts its trial, with a one-way code and never the address. Claim checks outside the debate pages are now part of Pro.
2026-10-21: The Google Ads tag now also tells Google when an analysis was started on the site, so we can count which ad visits led to one. It sends no text, link or result, and still runs with consent denied and sets no cookies. Pro and trial accounts may now and then get an email about new features, with an unsubscribe link that stops them.
2026-10-20: Our visit counter, Umami, now records the amount of each payment, with no name, email address or account attached. For a first payment it is counted with the visit it came from: we keep the IP address and browser user-agent from the request that starts a checkout until the payment arrives, and they are deleted by the next day.
2026-10-19: Corrected the short version, which said there are no accounts: you do not need one, but you can sign in. Added what the TL;DR and the article read send to Anthropic (pasted text of 150 words or more included), the Other outlets comparison and its suggested questions, the 14-day Pro trial, how the free limits are counted, how long analyses of a link are cached, Webshare (a proxy used when a news site blocks a direct request for an article) and YouTube (a video’s title). The Pro waitlist closed when Pro opened. The browser extension now sends the page’s address for a whole article, to compare it with other outlets’ coverage of the same story.
2026-10-13: We now also count visits with Umami, open-source analytics that runs on our own server. It sets no cookies, stores no IP address, and its visitor identifier changes every day. It also records how quickly each page loaded. Pro readers can now tell us whether an angle in the Other outlets comparison is right; we store the answer with a one-way code, not who gave it.
2026-10-12: Claims are now searched with Exa instead of Tavily; Tavily no longer receives anything you send us. Jina AI now only fetches the source pages found. Our pages now load the Google Ads tag, to measure whether our own ads bring visitors; it runs with consent denied, so it sets no cookies.
2026-10-08: The browser extension now also sends the title of the page you analyze (not of a selection), used to describe the article.
2026-10-02: You can now sign in, with a link sent to your email address or with Google. If you do, we store the address and set one cookie that keeps you signed in; the site still sets no cookies for anyone who does not sign in. A signed-in reader can also set up a daily digest; we store the outlets and topics chosen. Paid subscriptions are handled by Stripe; we store its identifiers and the state of the subscription, never card details.
2026-10-01: We added what the Semblen browser extension sends and keeps, and what we store when you fill in the classroom or monitor form.
2026-09-28: We now record whether a browser has visited before, as a flag in local storage with no identifier; record ZIP code lookups by state instead of ZIP code; and keep whether each race-update email was delivered, bounced or marked as spam, without your address.

Earlier version effective 2026-09-22.