How it works
This page follows one form submission from the visitor's click to the entry in your admin, and lists exactly what leaves your site.
The path of one submission
Formsieve works the same way for each form plugin it protects. Where Gravity Forms and Contact Form 7 differ, both are described. The short version is on the home page, and the Gravity Forms and Contact Form 7 pages on formsieve.com show what each form plugin gets.
- The visitor submits the form. On Gravity Forms, Formsieve runs inside Gravity Forms' own spam check, the
gform_entry_is_spamfilter, at priority 50: after Gravity Forms' honeypot (priority 1) and before Gravity Forms' built-in Akismet check (priority 90). If the honeypot, or another tool that hooks in before Formsieve, has already flagged the submission as spam, Formsieve leaves it alone and makes no API call. Akismet runs after Formsieve and only sees submissions that Formsieve allowed. On Contact Form 7, Formsieve runs inside Contact Form 7's own spam check, thewpcf7_spamfilter, at priority 20, after Contact Form 7's reCAPTCHA and Turnstile (priority 9) and its Akismet and Disallowed Comment Keys checks (priority 10). If one of them has already flagged the submission as spam, Formsieve leaves it alone and makes no API call. - Skip rules. Nothing is checked, not even by the free local pre-filters below, when Formsieve is off for the form or when the
formsieve_should_checkfilter says so. Those submissions are delivered and logged as "not checked"; on Contact Form 7, a form with Formsieve switched off is skipped without a log row. A missing setup is not a skip: without a Formsieve key or before consent is given, the free local pre-filters still run on both form plugins, and a submission they do not decide is delivered and logged as "not checked" (Test mode needs neither key nor consent). - Free local pre-filters, cheapest first, stopping at the first decision: allowlist → honeypot → time-to-submit → per-IP rate limit → blocklists → trivial content → replay cache → heuristics. See Pre-filters. They run with or without a key, and a submission stopped here uses no AI check.
- Guards. If your per-form hourly limit is reached (it counts every request sent, retries included) or the circuit breaker is open after repeated errors, Formsieve does not call the service and applies your fail mode (deliver by default). If your plan has stopped the AI check (trial over, no active plan, this site not on the plan, monthly allowance used), the submission is delivered, or held for review if you chose that, and Formsieve asks again at most every 15 minutes. See Plans, AI checks and limits.
- One HTTPS call to Formsieve's service (
api.formsieve.com, run by Thinking42), which asks TypeSafe's Jev model through OpenRouter with zero-data-retention routing and returns the answers. The request asks one main question ("is this spam?", with structured criteria) and five short supporting questions (category, attempts to manipulate the classifier, relevance to your site, deceptive urgency, coherent text). Timeout: 3 seconds, with at most one quick retry. Only a fresh answer counts as an AI check on your plan. - Decision. The answers become one probability
pbetween 0 and 1, which your thresholds turn into allow, review or block. See Verdict bands and actions. - Outcome and explanation. Gravity Forms: a blocked entry is marked as spam (or discarded, or refused with a validation message, as you choose per form), and the entry gets a note, a Formsieve panel and entry meta; review-band notifications to the site owner get a "[Possible spam NN%]" subject prefix. Contact Form 7: a blocked submission gets Contact Form 7's native spam status and a spam-log line, no mail is sent, and Flamingo (if installed) stores the message with the details; for the review band, the site owner's mail gets a "[Possible spam NN%]" subject prefix.
- Log. A row with numbers and identifiers only (never the message) is written to Formsieve's decision log, one log for every form plugin, which feeds the dashboard.
- Your corrections teach it. Marking an entry "Spam" or "Not Spam" (Gravity Forms) or moving a Flamingo message to Spam or back to the Inbox (Contact Form 7) labels the log row and updates the learned sender lists. See Feedback loop.
The probability is never sent back to the visitor's browser.
What is sent, and what is not
Only after an administrator has entered a Formsieve key and given consent on the consent screen.
Sent to Formsieve's service with each check (as of 2026-10-09):
- your site name, the site description you wrote and the site languages;
- the form name, and the per-form context text if you wrote one;
- the label, type and value of each ordinary field in the submission (for example name, subject, message, company, choices), with phone numbers redacted by default, HTML removed and each value capped (2,000 characters, 4,000 for the message, whose beginning and end are kept, about 12 KB in total; a value longer than 20,000 characters is already cut on your server before any check);
- the domain of the sender's e-mail address (for example
example.com), never the full address (you can switch even the domain off); - with every request, your Formsieve key, your site's address and environment type (production, staging, development or local; used to count the sites on your plan), the plugin, its version and which form plugin handled the submission (for example
gffor Gravity Forms), and your WordPress and PHP versions (in the User-Agent header). Formsieve's service does not pass your key, the site address and environment, the form plugin or these versions on to OpenRouter or TypeSafe; - numeric signals computed on your server: number of links and their domains (only from fields that may be sent), whether a phone number is present, seconds from page load to submit, submissions from the same IP in the last hour (a count, not the IP, and only while the rate limit is on), the writing systems used (script mix), message length (as submitted), and whether the sender was previously marked "spam" or "not spam" by you. These are the only signals sent; the other local hints (repeated values, crypto addresses, prompt-injection patterns and so on) never leave your site.
The key calls (checking your plan and usage, or releasing this site from your key) send your key, this site's address and environment, the plugin and its version and which form plugins are active, and no form data.
Never sent: the visitor's IP address, browser user agent and cookies, the full e-mail address, passwords, payment and card fields, file uploads, hidden fields, consent and CAPTCHA fields, and fields that look like health, national ID, passport, driving-licence, bank or card data (detected by label and type and excluded automatically, with a warning on forms that look medical or legal). You can exclude more fields per form.
Stored on your site by Formsieve: settings, the key (obfuscated at rest, autoload off), a consent record, and the decision log (time, form, verdict, score, category, model, route, request ID, latency, HMAC hashes of the IP and e-mail address, your label). The log never contains what the visitor typed. It is deleted after 90 days by default (Privacy kit). The submission itself is stored by Gravity Forms as an entry, or by Flamingo when it is installed (Contact Form 7 stores nothing itself).
Where it goes: see Data processing: your site → Cloudflare → Formsieve's service (Thinking42, AWS in the United States) → OpenRouter → TypeSafe AI, Inc. Formsieve's service does not store the content of submissions and deletes its request records after 30 days (as of 2026-09-26).
Why a probability, and why three bands
Jev answers with calibrated probabilities rather than a yes/no guess, so a score of 0.97 and a score of 0.55 can be treated differently. Formsieve blocks only confident spam, delivers the uncertain middle with a tag so a borderline lead still reaches you, and lets everything else through. Calibration explains which model version the thresholds were set for and how Formsieve notices when the service answers with a different one.