MCP App Store
by KeenEthics
Promvia icon

Promvia MCP Server for ChatGPT

By Atlaven

Overview

Promvia measures whether AI assistants name and cite your website. Connect your existing Promvia account and ask, in plain language, how your site shows up in the AI answers Promvia checks across the major AI assistants and search engines: your AI Visibility Score, which of your tracked questions cite you and which cite other businesses, the sources those answers rely on, your GEO readiness findings and action plan, and the AI-referred visits and revenue Promvia has measured on your site. With write access you can also track or archive questions, queue a new citation check or deep site scan, set the status of an action item, and keep the FAQ answers of your tracked questions in sync. Write access is opt-in: each write permission is unticked on the consent screen until you choose it, and every write runs under the same limits as the Promvia dashboard. Who it is for: marketers, SEO and content teams, and agencies who already use Promvia. Limitations: you need an existing Promvia account with MCP access, and you sign in with OAuth. Promvia reports what it has measured: shares come with their sample size and a 95% confidence interval, and results reflect the most recent completed checks, not a live query of each engine at the moment you ask.

How to connect

  1. ChatGPT: Open the Promvia listing in the ChatGPT app directory and choose Connect. ChatGPT listing Promvia exposes 25 tools here.

Tools

add_prompt

ChatGPT
Track a new prompt on a site. Returns the tracked prompt's queryId and stored text (the id archive_prompt and generate_brief take); a prompt that is already tracked returns that prompt's queryId with alreadyTracked: true, and no second row. Subject to the plan's account-wide tracked-prompt quota — refusals name the reason. When this month's room for automatic checks is used, the prompt is still tracked and automaticChecksFrom says the day its automatic checks start (it can be checked on demand before then). A reworded question is a different measurement, so a prompt is never edited: with replacesQueryId (a tracked prompt's queryId on the same site) the new prompt REPLACES it — in one transaction the new one is tracked under the same quota rules and the old one is archived (its history stays) and linked, and the result carries replacesQueryId. Replacing frees the old prompt's quota slot in the same transaction, so it is accepted at a full quota; but the old prompt keeps its active slot for the rest of the period, so at a full active pool the replacement's automatic checks start next period and the result's automaticChecksFrom says so (it can be checked on demand before then). A replacement is refused, changing nothing, with unknown-query (not a tracked prompt of this site), same-question (the new text is the old prompt's), already-tracked (the new text is already tracked; its queryId is returned) or a quota reason.

archive_prompt

ChatGPT
Stop tracking a prompt (removes it and its future checks; history stays). queryId comes from list_prompts or add_prompt. Returns { ok: true }, or { ok: true, alreadyArchived: true } when the prompt was archived before; a queryId that is not a prompt of this site is refused with { ok: false, reason: "unknown-query" } and archives nothing.

current_site

ChatGPT
Which site this connection is working on, and when it was selected — for "which site am I on?". On a connected app it reads the stored selection (selected: null when none is chosen yet, with the sites to choose from). With an API key nothing is stored, so it always answers selected: null. To pick or change the site, call select_site with one from list_sites.

generate_brief

ChatGPT
Generate an AI content brief for a tracked prompt (Pro plans; shares the account's generation fair-use budget).

get_capabilities

ChatGPT
What this connection can do right now: credential ('oauth' or 'apikey'), the account's plan, your seat (owner, member or viewer), the scopes granted, and every tool on this server with available: true, or available: false and a reason — 'scope' (the grant lacks requiredScope; an OAuth client can ask for it), 'seat' (a viewer seat is read-only whatever the grant says: every tool needing write access reads 'seat' on one), 'disabled' (the feature is not enabled on this deployment) or 'plan' (the account's plan does not include that tool). An API key is read-only by design, so every tool needing write access reads 'scope' on it. Only the caller's own grant and account are described.

get_citation_history

ChatGPT
Citation-check history for a site, newest first, filterable by engine and window (bounded: max 90 days, max 100 rows per page). Pass nextCursor back as cursor for the next page (null on the last), to read the whole window. Each row carries scope: the lane that wrote it (full = a full check, daily = the daily check, sweep = a sweep check), observedQueries: the web searches that engine ran for the answer (null = not exposed by the engine, [] = none ran), and model: the model that answered (null before 2026-10-21 or for engines without one). instrumentChanges lists our own measurement changes (an engine's model switch: engine, date, from, to, notice) dated inside the days the rows span; movement across one may come from the model, not the site.

get_citations

ChatGPT
Get a site's latest AI-citation results: per tracked prompt, whether AI engines recommend the site, which OTHER domains the answer cited (rivals, but also forums, reference sites and press), and sentiment — with sentimentSource: 'model' (read by Claude Haiku, paid plans), 'keyword' (a keyword estimate — say so when you quote it) or null (no sentiment; an answer that only repeats the brand from the question is not counted). shareOfVoicePct ships with shareOfVoiceSample — its citation counts, the measured discovery questions behind them (questions, which tooThin and earlyReading are decided on), a 95% Wilson interval, tooThin and earlyReading — because 2 of 10 and 200 of 1,000 are the same percentage and not the same finding. Each prompt also has demandBand: how often the question is asked, as a band ('rare', 'occasional' or 'frequent', from search volume on its head term) — never a count, and null when it was not measured or the account is not on a paid plan.

get_deep_scan

ChatGPT
Deep site scan results: latest run lifecycle plus the merged page-level findings (newest full crawl with newer diff overlays) — which pages carry which problems, robots-blocked pages counted but never fetched.

get_evidence

ChatGPT
Raw citation evidence for a site's latest checks: answer excerpts, every cited URL and domain, in-answer position, sentiment, and the observed query fan-out (the web searches the engines ran, per question).

get_geo_actions

ChatGPT
The site's GEO action plan — the same ranked list as the dashboard's Fix list, open items included: each item's taskKey (the key update_action_status takes; that tool needs write access), title, kind (technical / content / offsite), priority, effort and status (open / in progress / done / dismissed). Capped at 50 items; totalActions and omitted say how many exist. The consistency check's findings are in the plan as technical items; each clears when the next site check no longer finds it.

get_geo_audit

ChatGPT
Get a site's latest GEO readiness audit: overall grade/score and the specific findings to fix. The score is a weighted index over the signal checks, not a proportion: scoreBasis states that it carries no confidence interval, and none should be inferred. signals.consistency holds the consistency check, on every plan: six rule checks of the scanned pages against each other (brand name, hreflang links, FAQPage markup against the visible text, offer availability against the page text, the homepage H1's category, page weight), each issue with its URLs and quotes. It is not part of the score.

get_metrics

ChatGPT
AI traffic & revenue: attributed visits, verified revenue by currency and source, unverified conversions kept separate. window states each figure's time window: aiVisits.confirmed, attributedRevenueCents, aiTrafficSharePct and visitorQualityLift are all-time; aiVisits.likely is the last 30 days. Never add or divide figures from different windows. aiVisits.likely is the last 30 days of direct visits landing on a page AI cites (inferred, not counted). aiTrafficSharePct carries aiTrafficShareSample (visit counts + a 95% Wilson interval); visitorQualityLift carries visitorQualityLiftSample — its four IDENTIFIED-visitor/converter counts (visitors with a visitor id; visitsWithoutVisitorId counts visits without one, e.g. cookies not accepted) and a two-proportion z-test — and deliberately NO interval, because a ratio of two rates is not a proportion. Same payload as REST /v1/sites/{siteId}/metrics.

get_opportunities

ChatGPT
Open Prompt Opportunities — scored prompt candidates worth tracking (available when the feature is enabled).

get_prompt_templates

ChatGPT
Prompt-dimension templates and their variants (country/language/persona/funnel/product) with per-variant tracked-prompt ids for filtering results.

get_sources_report

ChatGPT
Which domains AI answers cite or name for this site's prompts — the off-site targets worth earning presence on, classified by kind. count is the number of tracked questions whose latest answers cite the domain or, for a competitor the owner listed, name it in the text: say 'named or cited', not 'cited'. The hosts each answer actually cited are otherCitedDomains in get_citations.

get_usage

ChatGPT
Every allowance on this Promvia account in one read. fullChecks.left is the check budget: how many full checks (one question asked on every engine in fullChecks.engines) are left on demand this month. citationCalls.budgetModelVersion names the ledger; on versions 1 and 2 citationCalls also lists the scheduled and on-demand buckets (with purchased top-ups) in provider calls, on version 3 those internal quantities are omitted. engines.selected lists the engines the account's scheduled checks run on (its own choice, or the plan's default when it made none) and engines.daily the ones among them on the frequent check this month (empty on a period still on the old model), and engines.sites how many days apart that check actually runs for each site (checkedEveryDays: 1 is every day; a site can be checked less often than its plan's frequent check, and 7 means it has only the weekly check); fullChecks.engines is the on-demand lane, a different set. Also expansion packs, sites, tracked prompts and the AI generation budget. These are the budgets start_citation_run, start_deep_scan and generate_brief draw on; those tools need write access, so a read-only connection does not list them.

get_visibility_score

ChatGPT
Get a site's AI Visibility Score (0–100 + letter grade) and competitive rank — how findable the site is across AI search engines. score is null until at least one unbranded (discovery) question has been measured: the site audit alone is readiness (get_geo_audit), not visibility, and is never reported as this score. shareOfVoicePct is a rounded integer, so a share above zero can read 0: shareOfVoiceSample.note carries the bounded form (<1%, >99%) — quote the note, not the integer. shareOfVoicePct comes with shareOfVoiceSample: the answer slots it was counted on (n), the measured discovery questions those slots came from (questions), a 95% Wilson interval, tooThin when the sample cannot carry the share — fewer than 10 questions — and earlyReading from 10 up to 32 questions, where the share is shown but is provisional and firms up as more questions are measured (say 'early reading' when you quote it): one answer names many businesses, so its slots are not independent and the sample size is the questions, not n. Quote the share as over its questions (the note says 'over N questions') with the interval, never the rate alone. The score itself is a weighted index, not a count out of a count, so scoreBasis says it carries no confidence interval; do not invent one. When score.thinBasis is present, fewer than its threshold of the questions have an answer naming the site and one answer can move the score by many points: say so whenever you quote the score. score.modelVersion names the formula; do not compare scores across versions. rivals answers 'who is ahead?' in the same call: the top five businesses by share of voice over the same answer slots (the dashboard's own ranking), plus your own row (you: true) with its real rank when it is not among the five, each with its rank, shareOfVoicePct and shareOfVoiceSample (counts and 95% interval). A competitor known only by a name has domain null and its name in name; your own rank is null while you hold no answer slot. Quote each share with its interval. When two businesses' intervals overlap it is a statistical tie, not a lead: versusYou says 'tied', and only 'ahead'/'behind' when the intervals are clear of each other — decided on interval95PctFine, the same interval to two decimals, because whole-percent bounds of small shares touch after rounding. rivalsBasis states the denominator. rivals is empty until a check has completed.

list_prompts

ChatGPT
A site's tracked prompts, oldest first, 100 per page: each with queryId, query (the tracked text exactly as stored), faqQuestion (the question the FAQPage JSON-LD carries: a language variant's trailing engine instruction removed), language (the variant's language, null for the default variant), templateId, answer (the FAQ answer verbatim, null when there is none), answerHash, answerUpdatedAt (when the answer last changed; null if it has not changed since 2026-10-26), archived, automaticChecksFrom (YYYY-MM-DD when its automatic checks start in a later month, else null), replacesQueryId (the prompt this one replaced, else null) and replacedByQueryId (the newest prompt that replaced this one, else null; a replaced prompt is archived and keeps its history). answerHash is the lowercase hex SHA-256 of the answer's UTF-8 bytes — hash exactly the answer string returned here (answers are stored with LF line endings and trimmed ends, nothing else changed) and compare to check that your page and Promvia hold the same text; it is also the expectedHash the answer writes take. Pass nextCursor back as cursor for the next page (null on the last); archived prompts only with includeArchived: true. Writing answers is set_prompt_answer or sync_prompt_answers, which need write access (prompts:write).

list_sites

ChatGPT
List the sites on this Promvia account, with their ids and domains.

select_site

ChatGPT
Choose the site to work on, e.g. for "promvia select 2" or "promvia select client-b.com". site is a number from list_sites, a site id, a domain or a name (a unique part of one is enough). On a connected app the choice is stored for this connection: every site tool then works on this site, without a siteId, until another select_site. It changes no data in Promvia. With an API key nothing is stored and it returns the siteId to pass. Only sites this connection can use are matched; no match or several matches returns the candidates and changes nothing.

set_prompt_answer

ChatGPT
Set the FAQ answer of one tracked prompt (the same answer the dashboard's FAQ editor saves and the site's FAQPage JSON-LD carries). The answer is stored as sent, except that CRLF/CR line endings become LF and the ends are trimmed; an empty answer clears it. More than 600 characters (Unicode code points) is refused — { ok: false, reason: "invalid", field: "answer", max: 600 } — and never cut. source is your own label for the write (≤100 characters, e.g. a release tag), kept in the answer's change history. expectedHash (optional) is the answerHash you last read, or null for "no answer": when the stored answer's hash differs, nothing is saved and the result is { ok: false, reason: "conflict", currentHash } — so a stale copy cannot overwrite a newer answer; without expectedHash the write replaces whatever is stored. Returns { ok: true, queryId, answerHash, updatedAt, changed }: answerHash is the lowercase hex SHA-256 of the stored answer's UTF-8 bytes (null when cleared); sending the text already stored returns the same answerHash with changed: false and records nothing. queryId comes from list_prompts; an unknown or archived prompt is refused with reason unknown-query. Many answers at once is sync_prompt_answers.

start_citation_run

ChatGPT
Queue a full citation check-run for a site. Runs the SAME gate ladder as the dashboard button (active-run convergence, verification, quotas, monthly COGS ceiling) — refusals name the gate.

start_deep_scan

ChatGPT
Queue a deep site scan (sitemap-first crawl up to the plan's page ladder). Runs the SAME gate ladder as the dashboard button (feature flag, plan allowance, ownership verification, burst cap) — refusals name the gate. No provider cost; polite crawl of the site's own pages.

sync_prompt_answers

ChatGPT
Set the FAQ answers of up to 100 tracked prompts in one transaction, e.g. from a release that keeps its answers in a repository. items is [{ queryId or query, answer, expectedHash? }]. query matches a tracked prompt by its text with case and spacing ignored — the query or the faqQuestion list_prompts returns, a dimension variant included; a text that matches several tracked prompts is refused for that item with reason ambiguous-query and their candidateQueryIds (send one as queryId). A prompt that is not tracked is tracked first, under add_prompt's quota; an archived one with that text is revived (revived: true, its history continues), except one that was replaced by a reworded prompt (refused, reason replaced, with replacedByQueryId) or one tracked only as an archived variant (refused, reason archived-variant). A created prompt whose automatic checks start in a later month carries automaticChecksFrom, as add_prompt does. Every item is checked with set_prompt_answer's rule before anything is written: one invalid item — or the same prompt twice — refuses the whole batch with { ok: false, reason: "invalid", index, field, max }. dryRun: true returns exactly what the run would do and writes nothing. A real run needs idempotencyKey (≤200 characters) — one idempotencyKey per batch, e.g. <release>-<batch> such as "v4.2-1", since a release of more than 100 answers is several batches. Only a run in which every item applied (unchanged, changed or created) is kept under its key: the same key with the same request again returns that first result and writes nothing, and the same key with a different request is refused with reason idempotency-mismatch. A run with a conflict, a refusal or a missing prompt is not kept, so after fixing the cause the same key runs again. A kept key lives 7 to 14 days (records older than 7 days are pruned weekly). Returns items [{ index, queryId, status, answerHash, previousHash, reason? }] and counts; status is unchanged, changed, created (would-create in a dry run), missing (no such tracked prompt), conflict (the stored answer's hash is not the item's expectedHash; nothing written for it) or refused (with its reason: add_prompt's, e.g. over-quota, or one above). answerHash is the item's answer after the run, previousHash before it. An unchanged item records nothing; each change is recorded with source in the answer's history, as set_prompt_answer records it.

update_action_status

ChatGPT
Set a GEO action item's status: open, in_progress, done or dismissed. taskKey is an item's key from get_geo_actions; a key that is not in the site's action plan is refused with reason unknown-task.

Capabilities

Writes

Example Prompts

Click any prompt to copy it.

Similar Apps

Listed on MCP App Store

Listed on MCP App Store

Add this badge to your site or README. It links to this page.

HTML

<a href="https://mcp-app-store.keenethics-labs.com/app/promvia"><img src="https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg" alt="Listed on MCP App Store" width="220" height="44"></a>

Markdown

[![Listed on MCP App Store](https://mcp-app-store.keenethics-labs.com/badges/listed-on-mcp-app-store.svg)](https://mcp-app-store.keenethics-labs.com/app/promvia)