Legal
Privacy policy
What personal data Recognly handles, why, and the choices you have.
1. Who we are
Recognly is made, owned and operated by Rivac AI Labs (rivaclabs.com) ("we", "us"). We are the controller of the personal data described here that we collect for our own purposes, such as account and billing data. For data our customers add to the service about their own business, contacts or leads, we act as a processor on the customer's instructions.
2. What we collect
Account and organisation data
- Name, email address, time zone and optional avatar.
- A password hash (never the password itself), two-factor authentication settings, and session records including IP address and browser details.
- If you sign in with Google or your organisation's single sign-on, the identity details that provider shares.
- Organisation name, members, roles and invitations.
Billing data
Payments are handled by Paddle (Paddle.com Market Limited), our reseller and Merchant of Record. When you start the free trial or buy a plan, you enter your card or other payment details in Paddle's checkout; they go to Paddle, not to us. Paddle charges you, issues the invoices and receipts, and collects and pays sales tax and VAT. We receive from Paddle what we need to run your subscription: a customer id, your billing email and country, the plan, the subscription's status and dates, and each transaction's amount, tax and invoice number. We never see or store your full card number. Paddle's own privacy notice covers what it does with your payment data. For an Enterprise invoice we also keep the billing address and business details you give us, and send them to Paddle for the invoice.
The free trial needs a payment method up front: nothing is charged for 14 days, and the first charge is on day 15 unless you cancel before. We email you three days before the first charge. Refunds are described in our refund policy.
Content customers add
- Websites, competitors, buyer questions, company facts, draft text, comments and files. This can include the names and job titles of the employees a customer names as experts, and their disclosure sentences.
- Lead reports a customer uploads are reduced to date, source, campaign tags, the "How did you hear about us?" answer and a note. Name, email, phone and company columns are never stored. Email addresses, phone numbers and the lead's own name are removed from the free-text answer and note before they are stored. Software cannot reliably spot other people's names typed into free text, so we ask customers to keep them out of the notes they upload. Email addresses from connected CRMs (HubSpot, Salesforce) are stored only as one-way hashes. When a customer maps the CRM field that holds the "How did you hear about us?" answer, we read that answer with the contact's first and last name. The name is used only to remove it from the answer and is not kept; the answer is stored with email addresses, phone numbers and that name removed. If the customer asks for it, the answers our rules cannot read are sent, in that redacted form, to an AI model (Claude) to be labelled by channel; at most 300 a month.
- Question imports: sales notes, call transcripts, support tickets or question lists a customer uploads to see which buyer questions its tracked questions miss. The file is stored encrypted only until it is read, read once and deleted. Email addresses, phone numbers, IP addresses, card and account numbers, link parameters, speaker labels and names from name columns are removed first, and only short, generic questions and counts are kept. Software cannot reliably spot every name typed into free text, so we ask customers to upload only data they may use for this purpose and to leave out recordings of people who did not agree to it.
AI answers and public conversations
Answers that AI engines give to our customers' buyer questions, and public conversations read through official APIs, public feeds, licensed data providers or links a person provides. These are stored at thread level; we do not build profiles of the people in them. LinkedIn post text is deleted within 48 hours and only the link is kept. From those answers we also keep short quoted clauses about the buying criteria they mention (price, where data is stored, integrations and the like) for the customer's brand and its competitors; they are parts of the stored answers, not new data.
Wrong-fact case files
When an AI answer states something wrong about a customer's company, we may read the public pages that answer cited to find where the wrong statement came from. We keep only the short passage that states it, the page address and a fingerprint of the text we read, not the whole page, and we read only pages whose robots.txt allows our fetcher. No new personal data is collected: re-check answers come from the same AI engines we already query, and the people who tick a fix as done are recorded by their account, as for every other review action.
Brand safety
When a customer uses the brand-safety scan, we look for links, email addresses and phone numbers in the AI answers we collect for that customer and compare them with the contacts the customer approved. Phone numbers and email addresses in AI answers may belong to people. We keep only the value, the sentence it appeared in (at most 300 characters) and where it was seen; we do not build profiles of people. Once a finding is resolved or marked a false positive, the sentences are cleared after 12 months. For suspected lookalike domains we read public registration (RDAP) and DNS data, which may include a registrar's abuse contact; this cache holds public data only and is not linked to a customer. We do not contact anyone on the customer's behalf.
Outreach contacts
When we draft an outreach email to the owner of a page that AI answers cite, we record the business email address or contact-form link that the page itself, or the contact page of the same site, publishes, and the author's name when the page names one, so that a person at the customer can send one email from their own mailbox. We read only pages the site's robots.txt allows. We keep this contact for 90 days, then delete it. We never guess an address, never buy addresses or use email-finder services, and the product never sends the email. Every draft invites the reader to reply "no thanks"; anyone who does goes on the customer's do-not-contact list, which keeps only the address or domain so we never draft for them again.
Listings
For a customer's brand we read the listing pages the customer gave us on review sites, company databases and marketplaces (at most once a week each, following each site's robots.txt), and Wikidata and Wikipedia through their official APIs. We keep only the facts the listing shows about the company (name, website, prices, founding year, headquarters, company size, integrations and the like), short snippets of the page element that held each value, a fingerprint of the page and when it was read. We never keep the page, and we never read reviewers' names, reviews or photos into these facts. LinkedIn pages are never read; a customer types those facts in.
Replay predictions and the public track record
Before a fix to your own site is published, we may replay the questions it targets on one AI model with the public pages the AI engines retrieved for them, once with your page as it is and once with the draft. We read those pages through the same polite page cache as everywhere else (robots.txt honoured, an honest user agent, page texts kept for 30 days at most) and keep only the documents' addresses, positions and the model's short answers. No personal data is collected.
Each prediction is kept as a fixed record so it can be checked later. If your organization shares its results (on by default; turn it off under Organization settings, "Prediction track record"), the scores of your resolved predictions are pooled with other organizations' into the public track record as numbers only: no brand, organization, question, page or answer text, and only once at least 30 predictions from at least 5 organizations have resolved, so no single customer's results can be read from it. Sample and demo data are never included.
Source watch
For each brand on a plan with the watch, the product stores copies of public third-party web pages that AI answers cite (the page's HTML, compressed, and its visible text), the public domain registration facts of those sites (creation date, registrar and, only when the record discloses it, the registrant organisation), the signals computed from them and the team's reviews (label, note, reviewer and time). The pages are read only when robots.txt allows it; sites that hold member data (such as LinkedIn) are never read. Page copies are deleted after 90 days unless they belong to an open campaign or an evidence bundle; evidence bundles are kept one year; all of it is deleted with the brand. Statements about the brand found on these pages are sent to the AI model provider configured for analysis (as for every analysis step), with instructions to treat the page text as data. Review notes are never shown to client guests. No personal data of page authors is collected beyond what the page itself publishes.
Server and CDN logs
When a customer uploads or ships access logs of its own sites, we process them only to find requests from AI agents and to verify them.
- Every line that is not from a known AI agent is discarded while the file is read. For visits that came from an AI assistant (for example a link in ChatGPT) we keep only a daily count per assistant and landing page (the page's address without its query string); no address and no browser string is kept.
- For AI-agent requests we keep per-minute counts per agent, page and verification result, with at most five network prefixes (IPv4 /24, IPv6 /48) per minute and page, never a full address.
- To check a claimed crawler we look up its address in reverse DNS; the result is cached for 7 days under a salted hash of the address, not the address. The host name found is kept only when it belongs to a domain an AI operator publishes for its crawlers; any other name (which can spell out the address) is never stored.
- Uploaded files and shipped batches are stored encrypted (Fernet) only until the import has read them (normally seconds to minutes) and deleted; any left behind are deleted after 24 hours. Files imported with the command line tool are read in place and never copied.
- Retention: per-minute rows 90 days, daily rows 400 days, reverse-DNS cache 30 days, batch records 14 days.
- The customer must have the right to share its logs with us; logs of sites the customer does not operate must not be uploaded.
Assistant directories
- Assistant directory results are public listing data, read once a week and shared by every customer; no personal data.
- Directory captures: when a team member clicks "Capture directory results" in the extension, the visible result list of that search (names, publishers, links without query strings) and the member's display name are stored for the brand. Nothing else on the page is read.
- MCP server checks: the tool list of the brand's own server is stored with the brand and deleted with it. A bearer token for the server is stored encrypted as a brand secret, sent only to that server, and never shown again.
- Assistant task answers: kept like other probe answers (answer text blanked after 13 months); the extracted suggestions stay.
Imports from other tools
When a team imports an export from another AI-visibility tool, the person confirms they may share that data with us. The file is encrypted while it waits to be read and deleted when the import ends (a file waiting for its columns to be mapped is kept at most 7 days). Only the columns mapped to a known field are read; columns about people (email, user, owner, author, assignee, IP address, phone) are never read. Prompts, topics and answer texts are cleaned of email addresses, phone numbers and the names of the organisation's members before they are stored; job messages and logs hold counts, never file contents. We keep the prompts, the old tool's results (flags, positions, rates and cited links) and Recognly's own reading of the exported answers; answer texts are blanked after the brand's retention setting (26 weeks after the newest import by default, 4 to 104, measured against today, so a brand that stops importing keeps no answer text past it), except for the weeks of a side-by-side run. Imported answers are never sent to an AI model. Everything imported can be exported and deleted by the team (an import, a source, or the whole brand), and is deleted with the brand. When a pull connector is turned on, the product uses the customer's own key to read only that customer's workspace at the vendor.
Agent buying-task audits
Screenshots of the customer's own public pages and test flows taken by the browser agent, the visible text the fetch agent read, and a request log (URL, method, status and type; no bodies), kept for 90 days (AGENT_AUDIT_RETENTION_DAYS) and deleted with the brand's data; run summaries, facts and findings stay with the brand's audit history. A test identity the customer provides (email, name, company and optionally a phone number and password) for testing their own sign-up form is stored encrypted, set only in a signed-in session, never shown again, used only by the sandboxed browser on the customer's verified domain, never sent to an AI model (a page that echoes it back is masked before the agent reads it) and masked in every record; the customer confirms it is a test identity, not a real person, and can remove it at any time. The brand's audit token goes only to the brand's own hosts, never to third parties its pages load content from.
Answer archive
Every AI answer we capture for a customer is sealed: we store a fingerprint (SHA-256 hash) of the question, the answer and the provider's response next to the answer we already keep, and, unless the customer turns it off, the provider's original response file. No personal data is added. To timestamp the seals we send only a fingerprint and a random number to an independent timestamp authority; no text leaves the service. When a customer deletes a brand's data, the sealed answers, fingerprints and files go with it; only weekly fingerprints (hashes and counts, no content) stay with the audit record. Evidence bundles are built only by members of the customer's organisation and are deleted after 30 days.
Compliance evidence
- Engagement ledger. For every contribution a customer's team records as posted or sent: its link, community, the named expert and their role, the disclosure sentence, the reviewer and approval time, and SHA-256 hashes of the approved and generated text. Live checks record only what they found (whether the disclosure is present, its position, a text-match score, the HTTP status); no text of any live page is stored. Entries are kept for three years, also after a brand is deleted unless the customer asks us to delete the audit trail.
- Evidence packs a customer exports contain its ledger, its approved texts, its experts' names, roles, disclosures and consent records (no email addresses). They are stored in the brand's data folder until the brand is deleted.
- Integrity reviews. To prevent coordinated wording across customers we keep a numeric fingerprint (MinHash) of community replies and outreach emails, never their text, and compare them across customers. Only our trust and safety staff see review cases; no customer ever learns anything about another.
Browser extension
When you click Verify disclosure in the extension on a post your team recorded, the extension reads the visible text of that tab in your browser, compares it with the approved text there, and sends us only the result: the post's address, whether the disclosure was found, its position, a match score between 0 and 1, a SHA-256 hash of the page text and the matcher's version. The page text itself never leaves your browser.
When you click Read this listing in the extension on one of your brand's listings, the extension reads that tab's structured data, meta tags and visible text in your browser (with review sections removed first), extracts the listing's facts there, and shows them to you. Only after you click Send these fields does it send us the listing's address, those facts, a SHA-256 hash of the page text and the extraction rules' version. The page text never leaves your browser. The extension asks our API whether a page is one of your listings only on known listing sites.
Facts service question log
When an AI agent or a person asks a customer's facts service a question (/ask or an MCP tool), we store the question with email addresses, phone numbers, long numbers and web addresses removed (a web address is reduced to its host name), cut to 300 characters, together with the answer's fact ids, the time, the kind of agent the request's user agent claims (not verified) and a salted hash of the user agent. We never store the IP address: rate limits use a daily-salted hash of the network prefix that expires with the day. The question log is kept 180 days and then deleted; daily request counts are kept. A customer can turn the question log off (only daily counts are kept then). Facts hostnames set and read no cookies.
Coding-agent panel
Coding agents run scripted tasks on our own starter repositories in isolated sandboxes. No customer code and no personal data enter a sandbox. We keep each run's transcript, code changes, verification results and network log for 90 days, then delete them; the derived results (which package was used, whether it worked, the API-surface findings) stay with the brand's data until the brand is deleted. The vendor test key a customer may store for its own test environment is encrypted at rest, added to requests by the sandbox's egress proxy only, and never written to the run's files.
Public bundle verification
When anyone uploads an evidence bundle on our verification page, we check it in memory and return the result; we do not store the file or its contents, and we log only counts (how many answers, whether it passed). We keep a short-lived counter per network address to limit uploads to 10 an hour.
Integrations
When a customer connects Google, HubSpot, Salesforce, GitHub, WordPress or Slack, we store the access tokens (encrypted) and the data needed for the feature the customer uses.
GitHub is used only to open draft pull requests. Through the connection we read the names of the repositories you can choose from, the latest commit of the branch you chose and, when we need to know whether a file we are about to add already exists, that one file. We then create a new branch and add the files for the fix. We do not list, download, read or store the other files in your repository, and we never write to your main branch. You review and merge the pull request yourself, and you can remove our access in GitHub at any time.
Free AI visibility check
The website, email address, optional brand name, questions and competitors you enter, whether you agreed to marketing email, and your IP address for rate limiting.
Without an account each device gets one free check, and a shared network address a few. To keep to that, we keep two kinds of keyed fingerprints for 30 days: one of a random device identifier kept in a cookie on your browser (pm_fc, see Cookies and storage) and one of your network address. The fingerprint records hold only one-way, keyed hashes, never the identifier or the address, are used for nothing else and are deleted after 30 days. The check itself keeps your IP address as described above. With a free account you can run more checks.
Help chat
When you use the help chat on our website or in the app, we keep:
- your questions and our answers, after we remove email addresses (also written out, such as "name (at) company dot com"), phone numbers, card and bank account numbers, passwords and keys, postal addresses and the name you give ("my name is ...") from your questions. We remove them before anything is stored or sent to our AI provider. We cannot recognise every personal detail you type, such as a name you mention in passing, so please leave out what you would not want kept;
- the page you were on (inside the app without your organisation's address or your brand's identifier), the language you wrote in, your thumbs up or down with the reason and comment you give, and when the conversation happened;
- a random identifier stored in a cookie (
pm_chat_vid, 30 days) so the chat can continue your conversation, which we keep only in a scrambled form. It is replaced whenever you sign in or out, and the chat of a browser tab is cleared when you sign out, so the next person on a shared computer does not see your conversation. If you are signed in, the conversation is linked to your account (and only you can open it) and, inside the app, to the organisation whose pages you were on.
We use this to answer you, to keep the chat within its limits, and to improve our help pages and answers. People at Recognly who look at chat statistics see the conversations without your name, email address, organisation or brand: organisation and brand names and the addresses of your pages are hidden from them.
When you are signed in, the chat's answers are written by an AI model (Anthropic's Claude). Your question, the parts of our public pages that match it and, inside the app, a short summary of your own organisation's plan, limits and setup progress (for a client guest of an agency, only the brands shared with you) are sent to Anthropic to write the answer. Anthropic processes them as our service provider and does not use them to train its models. When you are not signed in, the chat only shows the parts of our help pages that match your question, and nothing you type is sent to Anthropic.
When you are signed in, you can also press the "i" button next to the question box to let the assistant read the page you are on. Only while it is switched on (the chat shows "Reading this page"), each question also sends the page's title and address and up to about 6,000 characters of the text you can see in the page's main area, including what you typed into visible form fields. It never reads the chat itself, password, hidden or payment fields, email and phone fields, or anything that looks like a key or token. We remove the same personal details from this text as from your questions before it is stored with your question or sent to Anthropic, we keep it and delete it together with the conversation, and it is switched off again when you go to another page.
If you choose "Talk to a person", we email the conversation and the email address you enter to our support team so we can reply. We do this only after you tick the consent box.
We delete chat conversations 90 days after the last message. You can delete a conversation yourself at any time with the bin icon in the chat. When you delete your account, your chat conversations are deleted with it, and when an organisation is deleted, the conversations held in its pages are deleted too. Of a deleted conversation we keep only the price of its answers (for our daily spending limits), never its text. Your other rights (access, deletion, objection) are described below.
Consented staff panel
Only when a customer organization turns it on, and only for people in that organization who join it themselves after reading what it does: the redacted question and answer they capture from their own AI assistant with one click in the browser extension, the links in that answer (without query strings and without the conversation address), the assistant and model label shown, the time, and an optional job title they type. Email addresses, phone numbers, long numbers, the names of people in the organization and the person's own name and email are removed in the browser and again on our servers before anything is stored. A capture is stored under a pseudonymous reference, never with the person's name or account; we keep a record of each person's consent (who joined, when, for which brands and which consent text) so they can see and delete their own captures. Nothing is read from the assistant page except the answer the person chose to capture and the question just before it: never their history, account, settings or other tabs.
Affiliate program
For affiliates: payout email, optional website and notes, and referral activity. For visitors who follow a referral link: a referral cookie and a click count, with the IP address used for rate limiting.
Usage and security logs
An audit log of actions taken in the product (for example approvals, edits and publications), request logs and error reports, which can include IP addresses and browser details.
Security and abuse prevention
To keep the service available and to stop automated sign-ups, password guessing and other abuse, we process the network address your requests come from and count requests per address and per network for short periods (from one minute to one day). When we refuse a request for this reason we keep a short record (the time, the kind of refusal, the page or action, and your network address, shortened for IPv6 to its /64 network) for seven days, and daily totals for eight days. For failed sign-ins we keep a count per account for one hour after the last failure, stored under a one-way hash of the email address, not the address itself. When you sign in successfully we place a cookie (pm_device, one year, sent only to our sign-in endpoints) that tells us this browser signed in to your account before, so attacks on your account from elsewhere do not slow you down. It holds a signed hash, not your email. The legal basis is our legitimate interest in protecting the service and its users (GDPR Article 6(1)(f)).
When you create an account we check that the email domain is not a disposable mailbox service; if you try one, we keep only the domain in the refusal record. To protect the waitlist from abuse we apply the same limits and checks as for sign-up (rate limits per network address, a refusal of disposable mailboxes and, when switched on, a human check).
We also keep daily counters of paid provider usage per user, per organisation and for the whole service, to keep costs within limits; they hold amounts and request counts, not content, and are deleted after 40 days.
Human verification (CAPTCHA)
When human verification is switched on, we may ask you to complete a check that you are human on sign-up, password reset requests and after repeated failed sign-ins. It is provided by Cloudflare Turnstile (Cloudflare, Inc.) or by hCaptcha (Intuition Machines, Inc.), whichever is in use. The check runs in your browser and the provider receives information about your browser and device to tell people from bots; we send the provider the check's token and your network address to confirm it. See Cloudflare's privacy policy (hCaptcha: hcaptcha.com/privacy).
Error monitoring
When error monitoring is switched on, to find and fix faults, we record technical error reports from our servers and from your browser. A report holds the error message, the place in our code where it happened, the page or route that was open (with personal parts such as email addresses, tokens and one-time links removed), the browser and operating system, and the time. Before a report leaves our systems we remove email addresses, network addresses, passwords, tokens, cookies and authorization headers, and we never include the content of sign-in, sign-up, invitation, waitlist, billing or secret-key requests. We do not attach your identity to these reports: no name, no email address, no account identifier and no IP address. Reports are kept by our error monitoring provider for 90 days. The legal basis is our legitimate interest in keeping the service working and secure (GDPR Article 6(1)(f)).
The waitlist
When sign-ups are limited, you can leave your email address on a waitlist. The sign-up page asks you to tick a box that says: "I agree that Recognly may email me about my place on the waitlist and when my invitation is ready. I can ask to be removed at any time." We store the address (lower case), the time you agreed and the wording's version, the time you confirmed the address, and, if we invite you, the time of each invitation and when you created an account. We send you a confirmation email, one email when your invitation is ready (more if an administrator invites you again), and nothing else. The legal basis is your consent (GDPR Article 6(1)(a)), which you give by ticking the box and confirming the address, and which you can withdraw at any time: every waitlist email has a link that deletes your address at once (/waitlist/leave), and you can write to us. We delete an address that is not confirmed within 72 hours, and a confirmed one 12 months after you joined the waitlist or after your invitation ended; the entry is also deleted when you ask. We do not share the waitlist with anyone, and we do not use it for any other purpose.
3. Cookies and storage
- pm_session: keeps you signed in. Strictly necessary, HTTP-only.
- pm_ref: remembers an affiliate referral code for 60 days. Set only when you arrive through a referral link.
- pm_oauth: set for 10 minutes while you sign in with Google or single sign-on, so the sign-in can only be completed in the browser that started it. Strictly necessary, HTTP-only.
- pm_device: remembers that this browser signed in to an account before, so sign-in protections against attacks from elsewhere do not hold you back. Strictly necessary for security, HTTP-only, one year, sent only to our sign-in endpoints.
- pm_fc: a random device identifier, signed, that keeps the free AI visibility check to one per device. Set when you open the check, kept 30 days, HTTP-only and sent only to the check's address. Used for fraud prevention only.
- pm_chat_vid: a random identifier that lets the help chat continue your conversation, kept 30 days and replaced whenever you sign in or out (see Help chat). Functional.
- pm_last_org: remembers the last organisation you opened in the app, for one year. Functional.
- Paddle checkout: when you pay, Paddle's checkout (loaded from Paddle) may set its own cookies to process and protect the payment, under Paddle's privacy notice.
- sidebar_state: remembers whether the app's sidebar is open or collapsed, for seven days. Functional.
- Theme preference: your light or dark choice, kept in your browser's local storage.
- We do not use analytics or advertising cookies.
4. How we use data
- To provide the service: measurement, drafts, reviews, reports, alerts and integrations.
- Staff captures are shown to the customer as their own panel, labelled as consented, logged-in sessions that are not a random sample, and are never added to the measured answer rates.
- To secure the service: authentication, rate limits, abuse prevention and audit logs.
- To bill customers and run the affiliate program.
- To send service emails, such as verification, invitations, reports and alerts you have enabled.
- To send operational alert emails about the health of the service (for example a stalled job worker) to our own administrators; they contain no customer data.
- To send marketing email only where you agreed to it. You can unsubscribe at any time.
- To comply with the law and enforce our terms.
5. Legal bases
Where the GDPR or UK GDPR applies, we rely on: performance of a contract (providing the service), legitimate interests (security and abuse prevention, error monitoring, product improvement and business communications, balanced against your rights), consent (marketing email and the waitlist) and legal obligations (tax and accounting records).
6. Sharing and processors
We do not sell personal data. We share it only with:
- service providers who process it for us under contract: our hosting and email delivery providers, AI model providers used to generate and analyse answers (including Anthropic, which writes the help chat's answers, and, on plans with the model check, short clauses of stored AI answers sent to Anthropic to read their buying criteria) and, when a customer turns on the deep-research panel, to write research reports (OpenAI, Google Gemini, Perplexity, Anthropic), and licensed answer and search data providers;
- Paddle (Paddle.com Market Limited), our reseller and Merchant of Record, which receives your payment and billing details to process orders, invoices, taxes and refunds;
- integrations a customer chooses to connect;
- other members of your organisation, according to their roles;
- authorities, where the law requires it; or a successor in a merger or acquisition, with notice.
While they are switched on, we also use Cloudflare, Inc. (Turnstile) or Intuition Machines, Inc. (hCaptcha) for human verification on sign-up, password reset and repeated sign-in failures, and Functional Software, Inc. (Sentry) for scrubbed technical error reports of the web app and the servers, in the United States or, when the EU region is chosen, the European Union. Waitlist emails go through the email provider already listed. Write to us through our contact page for the current list of service providers.
7. International transfers
Data may be processed in countries other than your own, where we and our service providers operate. Where data leaves the EEA or the UK, we rely on a lawful transfer mechanism, such as the European Commission's Standard Contractual Clauses.
8. Retention
- Account and organisation data: while the account is active, then deleted after closure, except where the law requires us to keep it.
- LinkedIn post text: 48 hours.
- Sentences around brand-safety findings: 12 months after the finding is resolved or marked a false positive.
- Answer archive evidence bundles: 30 days.
- Text of public pages that AI answers cite (read to explain why a brand is missing or where a wrong fact comes from): 30 days, then blanked.
- Answers AI engines give without web search (memory and search probes): 13 months, then blanked; the counts and flags stay. They contain no personal data.
- Deep-research report texts: 13 months, then removed (the parsed results, such as which vendors a report named, are kept).
- Uploaded question files (notes, transcripts, tickets): deleted once read, and within 24 hours at most.
- Questions kept from those uploads: 12 months after they were last seen; the import records with counts: 24 months.
- Engagement ledger entries (compliance evidence): 3 years.
- Staff panel captures: the retention the customer organization sets (30 to 730 days, 12 months by default); deleted at once when the person who captured them deletes them or leaves the panel with "withdraw and delete my captures", when they leave the organization or delete their account.
- Staff panel consent records: as long as the organization account exists, for accountability.
- Facts service question log (questions agents asked a customer's facts service, redacted): 180 days; the daily request counts are kept.
- Coding-agent run files (transcripts, code changes, verification results, network logs): 90 days.
- Listing readings and findings: as long as the customer's brand exists; deleted with its data.
- Source watch page copies: 90 days, unless the page belongs to an open campaign or an evidence bundle; source-watch evidence bundles: one year.
- Replay predictions: kept as fixed records as long as the customer's brand exists; deleted with its data.
- Refused-request records (security): 7 days; their daily totals: 8 days; failed sign-in counters: 1 hour after the last failure.
- Daily usage-limit counters and their alert records (per user and organisation): 40 days.
- Error reports at our error monitoring provider: 90 days.
- Waitlist addresses: 72 hours if not confirmed; 12 months after joining or after the invitation ended; at once when you ask.
- Free check results and the email entered: until you ask us to delete them.
- Help chat conversations: 90 days after the last message, or earlier when you delete them or your account or organisation is deleted; only the price of its answers is kept longer (for our daily spending limits), never its text.
- Billing records: as long as tax and accounting law requires.
- Security and request logs: only as long as they are needed to keep the service secure.
9. Security
Passwords are hashed with Argon2. Provider keys and integration tokens are encrypted at rest. Traffic is encrypted in transit, two-factor authentication and single sign-on are available, each customer's data is kept separate, and sensitive actions are written to an audit log. No system is perfectly secure; write to us through our contact page if you find a vulnerability.
10. Your rights
Depending on where you live, you may have the right to access, correct, delete or export your personal data, to object to or restrict some processing, and to withdraw consent. California residents have rights under the CCPA, including to know and to delete; we do not sell or share personal data for cross-context behavioural advertising. You can delete your account from your account settings, or email us. If your data was added by one of our customers, we will refer your request to them. You can also complain to your data protection authority. People in a customer organization who joined its staff panel can leave it at any time from their account page or the browser extension, and can delete any capture they made.
11. Children
Recognly is a business service and is not directed at children under 16.
12. Changes
We will post changes here and update the date above. For material changes we will notify account owners by email.
13. Contact
Privacy questions and requests: our contact page.