Policy
Responsible engagement
Recognly helps companies be described accurately by AI answer engines. That work touches public conversations, so the product enforces many of these rules in code, and every customer agrees to all of them.
- Only real, named people post, from their own accounts, and they say who they work for in the post.
- Recognly never posts anything to a third-party platform. A person does, by hand.
- No fake accounts, personas, bought accounts, reviews, testimonials, upvotes or vote manipulation.
- Everything drafted about a company is based on facts that company approved.
Version 1.0. This policy applies to the Recognly software, which Rivac AI Labs makes and owns, to everyone who works on it, and to every organisation that uses it. Where a community's own rules are stricter than ours, the community's rules win.
1. Named humans only
Contributions on third-party sites come from one of two kinds of account, and no others:
- the customer's official brand accounts (its website, its company page, its own GitHub organisation, its brand account replying where it was mentioned); and
- named employees or founders of the customer who have agreed to take part, posting from their own existing personal accounts.
The product has no feature to do any of the following, and we will not build one:
- create, run, buy, rent, "age" or "warm up" accounts, or let one person use several accounts;
- operate personas, or pose as ordinary users, customers or fans;
- write anything in the voice of a customer, user or reviewer;
- write first-person experience the poster did not have ("I switched to X last year", "we cut our bill by 40%") unless it is a fact that person confirmed;
- log into a customer's or an expert's social accounts, or store their passwords or session cookies.
2. Disclosure in the post
Every contribution on a third-party site that mentions the customer, a competitor or the customer's product category states the poster's connection in the body of the post, in plain words, for example:
Disclosure: I'm the founder of Acme.
A profile bio, a hashtag on its own, or a disclosure hidden behind a click or a hover is not enough. The same applies to outreach emails. Recognly inserts the disclosure sentence into every such draft and re-checks it after every edit. Nobody can edit it out, and a draft that mentions the customer or a competitor without it cannot be approved.
3. People post; software does not
- Nothing is posted, commented, liked, followed or voted on by Recognly on any third-party platform.
- Each contribution is approved by a named person and posted by the named expert, by hand, through the platform's own posting box.
- The product does not use browser automation or bots on social platforms, and it does not auto-reply to keywords. The browser extension never posts, comments or fills in forms on any site.
- Changes to the customer's own website or documentation are delivered as downloadable files, draft pull requests or CMS drafts. A person at the customer reviews and publishes them.
- Connecting GitHub is optional. If connected, the product only adds a draft pull request on the one repository the customer chose. It does not read, copy or store the rest of the repository, and it never writes to the base branch.
- Outreach emails are sent one at a time by a person from their own mailbox.
- Every outreach draft ends with a line inviting the reader to reply "no thanks". A person or a site that says no goes on the customer's do-not-contact list, and the product never drafts another email to them.
- A case file's fix plan only drafts and suggests. The owned-page fix goes to the review queue like every draft; outreach to a third-party page or thread is an opportunity a person decides on (one customer per thread); engine feedback is a checklist a person completes from their own account and session. We never ask a publisher or an engine for a ranking or a mention, only for the fact to be stated correctly.
- Listings. When a customer's own listing on a review site, company database or marketplace (G2, Capterra, Crunchbase, a cloud marketplace and the like) says something that differs from the customer's approved facts, we give the customer a task to correct it in that platform's own vendor portal. Our software never signs in to a platform, never operates an account and never edits a listing. Listing pages, the customer's or anyone else's, are never outreach targets: we do not write to a platform or a reviewer about a listing.
- Replay predictions. A prediction never publishes, edits or approves anything. The draft goes through the same review and compliance checks as every draft, and a person decides whether to publish it.
After a contribution is posted
When an expert posts an approved contribution and records its link, it goes into an append-only ledger with its disclosure, its reviewer and hashes of the approved text (never the text of anyone else's page). The product then checks the live post 1, 7, 30, 90 and 180 days later and every 180 days until two years: through the platform's public API where there is one, otherwise by reading the public page, honouring robots.txt, identifying itself honestly and pacing its requests. It never reads LinkedIn, Facebook or Instagram posts from our servers; a person can check them in their own browser with the extension, which compares the page with the approved text there and sends only the result (whether the disclosure is present, where, and how much of the text matches), never the page itself. If a disclosure has gone missing, the expert is asked to restore it. No text of any live post is stored.
Integrity reviews across customers
To keep customers from coordinating the same wording across several organizations, community replies and outreach emails are compared across customers. At approval a reply is compared only with replies other customers have already posted publicly; a daily review compares the rest and is seen only by our trust and safety staff. A draft that needs a review is held with one fixed message that tells you nothing about any other customer. A review is never an automatic sanction: our staff clear it, dismiss it as a false positive, or confirm it and contact you.
Evidence packs
Customers can export evidence packs that map their ledger to the FTC's endorsement and consumer review rules, Article 50 of the EU AI Act or a platform's rules. The mappings are templates for counsel to review; they are not legal advice, and we never describe anyone's engagement as "compliant". Pack manifests are signed with a key whose public part is published, so anyone can verify a pack was not changed after it was built.
4. No reviews, ratings, testimonials or vote manipulation
Recognly does not write, buy, solicit or incentivise reviews, ratings or testimonials, and does not ask anyone to upvote, like or share. Drafts that ask for votes or reviews fail its checks. It does not add review or rating markup to a site for reviews the customer has not genuinely collected.
When we monitor a customer's listings, a listing's rating is shown as context only: it is never a goal, a finding or a task, and nothing we write suggests asking customers for reviews.
5. Respect for each community's rules
- Hacker News, Stack Overflow and any community that bans AI-generated text get no AI-written text. At most, the expert receives a research brief and writes their own contribution from scratch.
- Wikipedia and Wikidata are never edited by our customers' teams on our advice, or by us. When an article or item differs from a customer's approved fact, we prepare a request brief for a named person, who states their conflict of interest first and writes the edit request themselves, under their own disclosed account, on the talk page (Wikipedia: the
{{Edit COI}}request under its conflict-of-interest and paid-editing policies; Wikidata: the item's talk page under its conflict-of-interest guideline). The brief tells them not to edit the article or item directly and not to ask colleagues or customers to, and suggests independent published sources, with the customer's own site marked as a primary source that is usually not enough alone. At most two such requests a week per project, and every brief goes through human review. - Each community's rules on self-promotion and links are followed. By default a link to the customer appears only when the conversation asks for one.
- Contributions must help the person who asked. If the honest answer does not involve the customer, the expert either gives that answer or does not post.
- Weekly limits per customer and per community keep activity well below anything that could crowd a conversation.
6. One customer per conversation
Recognly never places two of its customers in the same thread. It enforces this across all customers, and it rejects a draft that is too similar to a draft written for another customer. It does not coordinate several people or accounts to push the same message.
7. Grounded in approved facts
- Every factual statement drafted about a customer must match a fact the customer approved. Facts extracted from the customer's own pages are stored with the passage they came from; facts the customer confirms directly are added by hand, and their approval is recorded in the audit log.
- When a customer's own pages contradict each other (two different prices, two different terms), the question goes to the person who owns that decision. Nothing is guessed.
- We never promise citations, positions or placement in AI answers, to customers or in anything the product drafts.
- Structured data the product suggests for a site asserts only what the visible page says.
- The product does not mass-produce pages to "cover more queries". Changes to a customer's site are capped each month, and each one is reviewed by a person.
- Facts service. On a hostname you verify, we serve only the facts a person at your company approved in a reviewed release, verbatim, each with its id and the date it was last confirmed. Nothing is generated; a question no approved fact answers is refused. Facts you retire, reject, edit or flag leave the service automatically; adding facts always needs a new reviewed release. We never post this anywhere else, and publishing does not mean any assistant will use it.
- Buying criteria and competitors. When AI answers credit a competitor on something (price, integrations, where data is stored) or raise an objection against a customer, a fix we draft states only the customer's own approved facts about that criterion. We never write claims about a competitor: a draft sentence whose subject is a competitor fails our check ("State only your own approved facts; claims about a competitor need a public source and counsel review"); a competitor may appear only in a neutral heading or a "compared with" phrase. An objection an approved fact contradicts is corrected with that fact through the normal review; an objection no approved fact answers is reported to the customer as fair and is never "rebutted". A criterion the customer has no approved fact for becomes a question for the customer, never an invented claim.
8. Data minimisation
- Public conversations are read through official APIs, public feeds, licensed data providers, or links a person provides. We do not scrape AI chat apps or LinkedIn ourselves, and we do not buy scraped Reddit data. Consumer AI answers come from a licensed data provider; the rest come from the AI providers' official APIs.
- When the product reads a public web page that is not the customer's own site (for example a page an AI answer cites), it follows the site's robots.txt, sends at most one request per second to a site, and says who it is in every request, with a link to a page that explains what it reads and how to ask us to stop. It does not read LinkedIn, Facebook, Instagram or Threads pages, and one switch turns all such reads off. To explain why a brand is missing from an AI answer, it reads the pages that answer cited, keeps their text for 30 days to compare it with the answer's sentences, and then blanks it (keeping only the title, the status and a fingerprint). Search-result positions come from the search data providers already in use (DataForSEO) or Brave's official search API; the product does not scrape search engines.
- For an outreach draft, the product records the business email address or contact-form link that the cited page, or the contact page of the same site, publishes, so a person can send one email. It keeps that contact for 90 days. It never guesses an address, never buys one or uses an email-finder service, and never sends the email itself.
- To check a customer's own listings, our software reads only the listing addresses the customer gave us, at most once a week each (one page and at most one pricing page), following the site's robots.txt and our usual identity rules, with a pace per site shared by all our customers and a daily cap per site. When a site blocks us or its robots.txt disallows us, we stop and do not try again for a week; we never use a headless browser, another identity or rotating addresses to get around a block. The customer can capture such a listing in their own browser with our extension instead, which sends us only the extracted fields and a fingerprint of the page, never the page. We read Wikidata and Wikipedia only through their official APIs, under their API etiquette (an honest user agent,
maxlag), and never read LinkedIn pages: those listings are entered by hand. We keep only the listing's fields, a fingerprint of the page and short snippets of the element that held each value; we never keep the page, and reviewer names, review texts and other people's details are never read into fields. A listing's value never becomes one of the customer's approved facts. - To warn a brand about suspected lookalike domains that AI answers point people to, the product reads public domain registration data (RDAP) and public DNS records of those domains, and may read the domain's home page once to see whether it is parked or for sale, following the same robots.txt and identity rules. It keeps only the matched phrase (for example "this domain is for sale"), never the page. It never contacts a registrar, an AI engine or the site's owner on the brand's behalf: the evidence bundle is for a person at the brand to send if they choose. Findings say "suspected", never "scam": a registered lookalike may belong to anyone.
- Conversations are stored at thread level. We do not build profiles of the people in them, and we never combine data about a person across customers.
- LinkedIn post text is deleted within 48 hours; only the link to the post is kept.
- Lead reports customers upload 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 are stored only as one-way hashes. When a customer maps its CRM's "How did you hear about us?" field, the answer is read with the contact's first and last name only to remove them from it, and we keep neither. The answers our rules cannot read go to an AI model only when the customer asks, in that redacted form.
- Customers can upload their own sales notes, call transcripts, support tickets or question lists to see which buyer questions their tracked questions miss. A file is kept encrypted only for the minutes before it is read, read once in memory and then deleted. Before anything else, email addresses, phone numbers, IP addresses, card and account numbers, the query part of links, speaker labels and the names in name columns are removed; likely names in free text are removed too, and a name one record gives away is removed from every record of the upload. Only short, generic questions (at most 300 characters) and counts are kept, for 12 months, and a customer can delete an upload's questions at any time. An AI model reads only the redacted text, and what it returns is redacted again; a question that still names a person is dropped. Software cannot spot every name typed in 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. Search Console and Bing data come only from the customer's own verified sites.
- The consented staff panel is off until an owner or admin of the client organization turns it on, and nobody is in it until they join themselves, after reading what it does; nobody can join for someone else. Joining lets a person capture, with one click per answer, an answer they are looking at in their own AI assistant. Only that answer and the question just before it are read, never their history, account or other tabs, and nothing is read in the background. Email addresses, phone numbers, long numbers, the conversation's address, link query strings, the names of the organization's members and the person's own name are removed in the browser and again on the server. A capture is stored under a pseudonymous reference, never with a name; colleagues see the redacted text, never who captured it, and per-role figures appear only when at least three people share the role. A person can delete any capture and leave the panel at any time (deleting their captures, or keeping them with no link to them); captures are deleted after the organization's retention period. We ask people not to capture conversations that contain customer, patient or other personal data.
- To timestamp the answers the product seals as evidence, it sends only a 32-byte fingerprint and a random number to a public timestamp authority; no answer text leaves the product.
- Each customer's data is kept separate from every other customer's.
- Every approval, edit, rejection, posting and publication is written to an append-only audit log. A customer can ask for its log at any time.
9. Honest measurement
Recognly reports how often a brand appears in AI answers as a rate with a confidence interval, compared against a set of questions deliberately left untouched. It never reports a position in an AI answer, never blends different ways of collecting answers into one number, and says plainly what it could not verify. Alerts are statistical tests at a level the customer chooses, not fixed thresholds, and a change it cannot explain is labelled as such: it says "likely platform-driven" or "possible program effect", never that it knows the cause.
When the product connects AI visibility to leads and pipeline, it shows what leads said and what first-touch tracking caught side by side, each with its sample size and interval, and it counts pipeline only in the currency it was recorded in (US dollars), saying how many leads it left out. Program goals are registered and locked (with a published hash) before their results come in; a goal registered after a program started says so forever, and only the estimate at the end of the registered window carries a verdict. Modelled effects are always called estimates.
Buying-committee roles are job contexts (a CFO, a CISO), never people or personas: a role's preamble may not name a person, a company, the brand or a competitor. Each role is measured as its own panel cell with 95% intervals, and a role's difference from the plain questions is called out only when it holds after Holm adjustment across the roles and both sides have at least 30 answers. Staff captures are their own panel, labelled "Consented staff sessions (logged in; people at the brand who chose to share; not a random sample of buyers)", and are never blended into any measured rate, report rate or alert. A committee fix states an approved fact; a concern without one becomes a task for the owner and never an invented claim.
Buyer journeys are scripted conversations of three to five messages that a buyer might send, written as job needs and constraints (never a person, an account or a persona), asked through the providers' official APIs or our licensed data provider. A conversation and a single message with the full brief are two separate panels and are never combined; every rate carries its 95% interval, and lift is measured against holdout journeys we leave alone. A constraint the approved facts support becomes a suggested fix for review; one they do not support becomes a question for the owner, never an invented claim.
Buying criteria are read from the wording of AI answers, by rules and, on some plans, a model check whose agreement with the rules we show. We report them as the share of answers naming each company that credit or object to it on a criterion, with 95% intervals, separately for each way of collecting answers and never combined; cells with fewer than 10 answers are greyed, holdout questions are measured separately and never given work, and a "why they win" explanation is called clear only when it holds after correcting for the number of criteria compared. These figures describe what engines say, not what is true about any company. Pages that carry a competitor's framing may become outreach opportunities only through our usual disclosed flow: one customer per conversation, monthly caps and a person approving every message.
Listing findings are ordered by how often AI answers cite each listing, counted separately for each way of collecting answers with 95% intervals and never combined, times how serious the difference is. The order is a way to choose what to fix first, not a prediction that fixing a listing changes what engines say.
Before a fix to a customer's own site is published we may predict how much it will change what AI answers say, and we say exactly what that is: a controlled replay on one model with the pages the engines retrieved, not a measurement of the live engines and not a promise. Every prediction comes with its 80% interval and its chance of a measurable rise, per panel, never combined. Each prediction is fixed before publishing (it cannot be edited or deleted), scored six weeks later against the questions we deliberately leave alone (the holdout), and counted in the track record whatever the outcome; exclusions are decided by rules fixed in advance and shown with their counts. The public track record shows failures and exclusions as plainly as successes and appears only once enough organizations contribute.
When the product weights the appearance rate by what buyers really ask (Search Console queries, Bing grounding queries and the customer's own uploads), the weighted rate is an additional view shown beside the raw rate, each with its own interval and sample size, and never merged with it; it is hidden when too few questions carry demand. Comparisons with Google's and Bing's own AI reports are descriptive, and the Bing comparison is labelled a proxy because the product does not sample Copilot. Suggested questions go to the measured questions, never silently to the holdout, and adding one is the customer's choice.
Evidence the product provides says exactly what it proves. A sealed, timestamped answer proves that the product captured that text from that engine no later than the timestamp; it does not show what the engine says to other people or at other times, and an evidence bundle is never called legally binding (its README says it is not legal advice). Timestamps from the demonstration authority are always labelled "demo authority, not trusted". Links, addresses and numbers flagged in AI answers are shown defanged so nobody opens a phishing link from the product.
The lost-question diagnosis is an observation with intervals, not a verdict: search-result positions are a proxy for the engine's own search, a stage an engine does not expose is shown as not observable (never as zero), and the suggested next steps are suggestions, never promises that an engine will cite or name the brand.
The deep-research panel asks each provider's deep-research mode, over its official API, a few vendor-research tasks a month and reads whether the report shortlists the brand and states its facts correctly. It is a separate, small panel: every rate is shown as "k of n reports" with a 95% interval, it is never added to the answer panels, the order of rows in a report's table is described as layout and never as a rank, and the provider labels say "(API)" because API modes can differ from the consumer products. The product does not use consumer accounts, logged-in sessions or personas to run it, and shortlist tasks never name the brand or its competitors, so the report's own choice is what is measured. A wrong price or attribute in a report becomes a suggested fix to the brand's own pages or approved facts, reviewed by a person; we never promise that a report will include or recommend the brand.
When an AI answer states a wrong fact about a customer, we open a case file and say what we know and how we know it. A source is "suspected" from what the answers cited and "implicated" only by a controlled replay on model behaviour (the API, never the live engine), and we say which. We call a fact corrected on an engine only when the upper end of the 95% interval of answers that still get it wrong is at most 3 in 10 (with at least 9 answers that address the fact), and we reopen the case if it comes back.
Answers without web search are asked only through the model providers' official APIs (Claude and Perplexity), never through consumer accounts or logged-in sessions. They describe what an engine says without looking anything up; we never claim to know a model's training data. The pages we read to see where a wrong fact comes from are read like every cited page: robots.txt honoured, rate-limited, honestly identified, and switched off with the platform's public-fetch switch.
An experiment is pre-registered before its change ships: the hypothesis, the metric, the engines, the expected effect and the stop rule are locked and hashed (and signed in the product), the questions or pages are randomised into treatment and control, and the result is read out every week per engine with an interval that stays valid however often we look. It stops only by its registered rule; a stop by hand outside that rule is labelled so, and every amendment is shown. Experiments are never pooled across engines and never promise a ranking.
Each country and language a customer sells in is measured as its own market, with its own questions and holdout. We never add markets together into one figure, and we call a difference between two markets only when it holds after adjusting for the number of comparisons shown.
Coding-agent panel
- The panel runs third-party coding agents (Claude Code, Codex CLI, Gemini CLI, Aider) on our own starter repositories, in isolated sandboxes, on the platform's or the organization's API keys. It operates no account, no persona and no public repository, and it never posts or publishes anything: no package publish, no push to a base branch.
- Fixes it drafts (AGENTS.md, an Agent Skills SKILL.md, framework docs pages, the SDK quick reference of llms-full.txt) are built only from owner-approved claims and an owner-approved API surface, reviewed by a person, and delivered as draft pull requests to the brand's own repository, or as files to download. Copy says they help coding agents use the SDK correctly; it never promises that agents will choose the brand.
- Results are shown to the customer only and never published as a public comparison of agents or vendors.
- Automated evaluation use of each agent CLI and its API is checked against the provider's current terms before live runs (an owner decision records it); the panel is on by default and can be turned off per brand (the panel's switch) and installation-wide by the operator.
Source watch
When AI answers cite pages that say something negative or false about a brand, the product reads those public pages (within robots.txt, with an honest user agent, at a polite pace) and looks for signs that they were made in coordination: newly registered domains, near-identical text across sites, one tracking id or template, a burst of new pages, text hidden from people that addresses AI assistants, pages linking to each other. A page or a group of pages labelled "suspected" shows signs of coordination; it is not proof, and the label is never an accusation. A group is labelled suspected only when it also shows shared infrastructure (newly registered domains, one tracking id or template, hidden text, pages linking to each other or one registrant). Copied text that appeared at once and nothing else, such as a news story syndicated by established sites, is shown as "copied content, likely syndication": it raises no alert, and its publishers can still be asked to correct a false claim. Labels stay inside the team: they are not in shared or scheduled reports unless an owner or admin turns that on, and client guests see them only when an owner or admin shares the watch with them.
What the product never does: it never posts, comments or reports on any platform; it never contacts a site, a host, a registrar or an AI engine; it never asks anyone to remove a page. The response plan lists factual steps a person takes: report through each engine's and host's own published channel (the person attaches the evidence bundle), state the approved facts on the brand's own site (drafts that never mention the pages or accuse anyone, reviewed by a person before anything is published), ask legitimate publishers that repeat a false claim to correct it (a disclosed draft a person reviews and sends; suspected sites are never contacted), and consider legal advice when pages make accusations. Nothing promises that engines will stop citing a page.
The brand's own content: every draft is checked for text addressed to AI assistants, hidden markup and invisible characters, and fails until it is removed; the brand's own site is checked for the same after each crawl. The files the product composes for coding agents (AGENTS.md and SKILL.md) may give coding agents operational instructions, such as how to install, test or call the product, but never steer AI answers about any brand; they are not held to the keyword-stuffing warning.
Agent buying-task audits on the brand's own site
- Own site only. Audits run only on the brand's own site and domains (and their subdomains). Browser runs and any form submission need a DNS-verified
sitedomain: the owner proved they control it. - Honest identity. Every request names Recognly and our contact page in its user agent (
Recognly/<version> (+<info page>; agent audit)) and carries the brand'sX-PM-Auditheader, so the owner can allowlist audits or leave them out of analytics. On a verified domain the owner may choose a standard browser user agent to see what agents with a plain browser string see; the audit header is always sent. - robots.txt. The fetch agent honours robots.txt for our product token and reports a disallowed page as a finding (
robots_disallowed); it also reports, without acting on it, which AI agents' tokens the file blocks. - No evasion. CAPTCHAs and bot challenges end the run as findings; nothing is solved or bypassed.
- No accounts elsewhere, no personas. Off-site navigation is refused. A form is submitted only on a verified domain, for a task the owner allowed, with the brand's own test identity, which the owner confirms is a test identity and not a real person; it is typed by the runner, never shown to the AI model and masked in every record. Email confirmation links are never followed.
- No payments, no posting. A payment form stops the run; no review, comment or message is ever submitted.
- Rate limits. One navigation per second, at most 60 page loads per run, one browser run per brand at a time, one audit per brand per hour on request and one scheduled audit a month (or a week).
- Instructions in pages. Text addressed to AI agents is reported (
suspicious_instructions) and never followed; the AI model acts only through a few restricted tools on the brand's own pages and must quote page text verbatim. - Human review. Fixes are owned-fix drafts grounded in approved claims and go through the review queue; the audit never invents a fact.
Assistant directories
- We never submit, edit, rate, review, upvote or install anything in an assistant app, plugin, connector or MCP directory, and never create accounts there. Listing copy we draft is pasted into the directory's own form by a person.
- Directory search pages are read once a week per keyword for every brand together, only after a person checked the directory's robots.txt and terms, with our honest bot identity and a platform-wide pace per host; sign-in-only directories are only captured in a team member's own browser after a click.
- When we check a brand's MCP server we send only
initialize, the initialized notification,tools/list,prompts/listandresources/list; we never call a tool. - Clearer tool descriptions use approved facts only, carry no instructions aimed at AI models, and promise no placement.
Server logs and retrieval proof
A customer's own server and CDN logs are read only to find requests from AI agents and to verify them; every other line is discarded while the file is read and no full IP address is kept. The product reads only the customer's own robots.txt (once a day per site, switchable in Settings, Server logs), the Web Bot Auth key directories of the AI operators in its registry and the IP range files they publish. A fetch of a page by an AI agent is shown as a read, never as proof that an answer used the page.
Comparisons with other tools
The switching kit imports a customer's own exports from another AI-visibility tool and compares them with Recognly's measurements of the same questions. The comparison is factual and sourced: it reports two measurements, how each was made and why they differ (sampling noise, how each tool decides the brand was named, which answers each tool captured, aggregation, surface, model, location and language, prompt wording, timing, branded prompts, the competitors each tracks), with 95% intervals and a statement of where every statement about a tool comes from (the export itself, the vendor's published documentation, or what the customer told us). It never says which tool is right, correct, better or more accurate, and never disparages a vendor; vendor names are used only to name the data source. Imported history is shown as its own panel, labelled with its source, and is never re-labelled as Recognly data or combined with Recognly's numbers. The kit never signs in to another tool, never reads its pages and never operates an account there: data comes only from files the customer exports or, when a connector is turned on, from the vendor's API with the customer's own key and within the vendor's API terms. Nothing in a switching report promises rankings or what switching will change.
10. What we ask of our customers
- Name the people who will post, confirm they agree to take part, and keep their disclosures in place.
- Name one person who approves drafts and who owns decisions about business facts.
- Pause community work promptly if a contribution is removed by moderators or an account receives a warning, until you understand why.
- Do not use Recognly, or ask anyone else, to post undisclosed promotion on your behalf. We may suspend the account of an organisation that does, under our terms.
11. Reporting misuse
If you see a contribution you believe was drafted with Recognly and breaks this policy (a missing disclosure, an account that is not who it claims to be, a request for votes or reviews, or anything else here), write to us through our contact page. Include the link to the post, a screenshot if you can, and what you think is wrong.
- We acknowledge your report.
- We check it against the audit log, which records who approved and who posted each contribution.
- If it breaks this policy, we ask the customer to correct or remove it and fix the cause. Where the breach is serious or repeated, we may suspend the organisation's account under our terms while we review how it happened.
- Where we can, we tell you what we found and what we did.
You do not need to be a customer, or to give your name, to make a report. We will not retaliate against anyone for reporting a concern in good faith.
Why these rules
They follow the US FTC Endorsement Guides (16 CFR Part 255), which require people with a connection to a company to disclose it when they talk about it; the FTC's rule on consumer reviews and testimonials (16 CFR Part 465); comparable rules in the EU, the UK and India; the rules of the platforms where experts post; and the usage policies of the AI providers whose models the product uses. This policy describes how the product works. It is not legal advice.
Changes to this policy
We will publish any change here with a new version number and effective date.