Convertlyft

Each sale, enquiry or sign-up shared out across the visits the buyer made before it, by where each visit came from and the platform (Paid search ads · Google, Social · Instagram). Any of the six models side by side, from one read: first, last, last_non_direct (default), linear, position (40/20/40), time_decay. Counted from this site's own visits; a path cannot follow one person across devices unless the site identifies them.

Where people got stuck in a window: the pages people spend time on and scroll through, the rage clicks, the dead clicks (the tag’s own, minus the ones a link or a call answered — the Stuck spots page’s count), and the form start-vs-submit funnel. Rows carry the same field names as cvl_dead_clicks, cvl_rage_clicks, cvl_page_attention and cvl_form_friction. Every collection is always present — an empty list means nobody did it, never that this tool does not report it. counts says which lists were capped, how many rows there really are, and the next_cursor that pages_cursor, rage_clicks_cursor or dead_clicks_cursor take. Medians under 30 visits and form rates under 30 starts are null. Instants are UTC.

Run a Convertlyft READ tool that is not in your tool list, by name, with its arguments. It answers exactly as calling that tool directly would, with the same permissions. It refuses every write or spend tool: those are in your list and are called directly. Use cvl_search_tools to find the name and arguments.

Record that the site changed, so the impact can be measured against it. Needs the preview_token from cvl_change_preview of the same note and marked_at: preview, show the person, apply only after they agree. Retry with the same idempotency_key and you get the same single marker back — never a second one. Returns the proposal id you undo with.

Show exactly what cvl_change_apply would write, and the undo it would hand back, without writing anything. Nothing is queued and nothing is stored by this call. Returns the preview_token cvl_change_apply needs for this same change (valid 30 minutes).

Remove a marker you wrote, using the undo computed when it was applied. Takes the proposal_id cvl_change_apply returned. Calling it twice is safe — the second call reports the same result and removes nothing.

Submit the sentence you are about to say plus the evidence ids (receipt_ids) of the data you were given. Returns ok, or the exact span no evidence supports for the figure your sentence names, and a corrected sentence. A numeral is checked against the figure you NAME, so "Visitors were 41" is refused when visitors are 45 and 41 is the visit count — a real number on the wrong figure is the failure this catches, not just an invented one. A rate needs its denominator ("3.2% of 1,200 visits"), and a date or period you state ("last month", "October 10") is checked against the window of the evidence. A refusal names rule_code and rule_text, and a corrected_sentence when one can be built from the evidence. Submit a whole paragraph in one call: each sentence is checked, `sentences` carries every result, and the top level is the first refusal. Deterministic — no model reads your sentence. For page copy (a draft, a page’s text), add against: "site": each sentence is checked against the site’s own crawled pages (home, pricing, product, features, how it works, and any pages you name) and the owner’s site context; receipt_ids are then optional. A sentence passes when the site says it, or when the evidence you submit (receipt_ids) backs it; otherwise rule_code not_on_site, with the reason. source_pages names the pages read. It checks figures, prices, times, guarantees, superlatives, credentials and names; a claim with none of these passes unread (limit says so).

The top 20 sequences of sources buyers came through before buying or getting in touch (e.g. Paid social ads · Meta ×2 > Google & other search · Google), with buyers, revenue per currency, their share of all sales, enquiries or sign-ups, and the ids of the visits they bought in (for recordings), plus how many visits and days it took. Counted from this site's own visits.

Start a crawl of any public website with the Convertlyft crawler and get a crawl_id straight away; the crawl runs for a few minutes. It counts toward the plan’s “Pages scanned”: 1 per crawled page, counted when the crawl ends and never more than limit; the answer says the most it can use and what is left. It obeys the site’s robots.txt, skips sign-in endpoints (/api/auth/*), never fetches private or internal addresses, and runs no AI on the pages. Limits per workspace: one crawl running at a time (an SEO-tab crawl counts), 500 pages per crawl. Past a limit the answer says which and when to retry. mode "links" returns just the map of URLs, never page text; "full" also returns each page’s text. Follow it with cvl_crawl_status, then read pages with cvl_crawl_page. A full audit of a competitor’s site is this tool too, and only on an explicit ask: the problems it finds never enter this workspace’s to-dos, score or alerts. To audit THIS workspace’s own site for the SEO tab, use cvl_run_crawl instead.

Search Convertlyft’s public documentation for agents — https://convertlyft.com/llms.txt and https://convertlyft.com/llms-full.txt (the API, the MCP door, dates, limits, worked examples). Returns the best-matching sections, quoted, each with where it came from. Plain word matching, not a summary. No account needed; same per-address limits as the other public tools.

Tell Convertlyft that the fix for one error is live on the site. Call it only once the change is deployed — not when it is written or merged. It starts the clock on "stayed fixed": once the affected pages have had enough page loads that the old rate would have produced at least 3 of these errors and none came, the error reads as stayed fixed. The earliest report wins; reporting again changes nothing. An error that is set aside, or merged into another, is refused — the refusal names the error to report instead. It records a date; it changes nothing on the site.

The methodology behind one figure on the KPI header: its definition, the population it was computed over (denominator, numerator and row count), the filters in force, the condition that decides when it is withheld rather than shown, and the attribution epoch before which campaign-filtered counts are undercounted. It returns no figure value — call cvl_kpi for that, and present the same window.handle here so the explanation describes the numbers you already have. Repeat any source or campaign you passed to cvl_kpi: the handle carries the window, not the filters, so without them this describes the unfiltered population and says so in `filters`.

Visitors, visits, the share of visits that bought or got in touch (conversion rate), revenue by currency, median engaged time and site health for a window — the same figures, from the same code, as the dashboard header. Every figure is an object, never a bare number: read "state" first. When state is "ok" it carries "value", a rendered "display", and its own "previous" and "delta" so you never divide; when state is anything else there is NO "value" key at all and "reason" is a sentence you can show a person. A non-ok figure means the number does not exist for this workspace or this range — never that it is zero. Pass from/to once, then present the returned window.handle on later calls so the whole chain reads one window. Visits are counted on the day they started, as on Home, so a visit that runs past midnight counts on its first day; the crash-free rate counts visits with an unhandled error. A window too large to count with a source or campaign filter or with part days on a busy site is refused as window_too_large: ask for whole days without the filter, or fewer days. home: Home’s four figures — Visits, the result, its rate and From Google — with the result word for the site’s type, and visits by device.

Queue today’s operator run for this workspace. It does NOT execute here — a worker claims the row and runs it, so this returns a run id and a status, never a report. If the workspace has already spent its daily cap the call is refused and names the cap and the spend; asking twice on one day attaches to the existing row rather than starting a second run.

Queue a fix for the OWNER to approve. This does not apply anything and cannot: approval is a human act on this product, and no credential can approve a proposal, including its own. baseline is what you saw before proposing and revert_plan is how to undo it — both are required, because a proposal that cannot be staleness-checked or undone is one an owner cannot safely accept. Proposing the same fix_identity twice replaces the earlier card rather than showing two.

What the scan found about AI agents on a public site: its agent-readiness level (L0 unreachable to L4 agent-native) with what holds it there, each AI crawler or assistant robots.txt lets in or not and what our request in its name got, whether llms.txt exists, and report_url — the full view on convertlyft.com/agent-readiness. Reads the same crawl as cvl_public_seo_scan, never a second one: pass the scan_id you already have, or a url (which starts or reuses that scan). Waits up to about 45 seconds. No account needed. Without one, the limits per address are 20 calls a minute and 200 a day, and 10,000 calls a day across everyone without an account (past one, HTTP 429 naming it, with Retry-After); and 25 new scans a day per address, or per workspace when signed in (a reused scan does not count). Past the scan limit the answer is an error naming it, with retry_after_seconds.

Scan a public website for SEO problems with the Convertlyft crawler: up to 25 pages, robots.txt obeyed, no JavaScript run, nothing changed on the site. Waits up to about 45 seconds and returns the crawler's own summary: the score out of 100 with its four parts, how many critical issues, warnings and notes, the top 3 issues with up to 3 example pages each, the first fix, what this scan did not check, what AI agents could reach, and report_url — the full result on convertlyft.com/seo-audit. If the scan is still running at 45 seconds it returns {status: "running", scan_id}; call cvl_public_seo_scan_result with that scan_id. A scan_id stays readable for at least 24 hours (a finished scan of the same site is reused for 24 hours); the crawler may delete scans older than 30 days. No account needed. Without one, the limits per address are 20 calls a minute and 200 a day, and 10,000 calls a day across everyone without an account (past one, HTTP 429 naming it, with Retry-After); and 25 new scans a day per address, or per workspace when signed in (a reused scan does not count). Past the scan limit the answer is an error naming it, with retry_after_seconds.

The result of a scan cvl_public_seo_scan started, by its scan_id. Same shape as that tool's answer; waits up to about 45 seconds if the scan is still running, then returns {status: "running", scan_id} again. Does not start a scan and does not count against the 25 scans a day. A scan_id stays readable for at least 24 hours (a finished scan of the same site is reused for 24 hours); the crawler may delete scans older than 30 days.

Append one row to this workspace’s ledger so the next conversation does not derive it again. A claim about the data (kind "finding") MUST carry the receipt_id (the evidence id) of the call it came from and the window it covers — a claim with no evidence behind it is refused, by design. A decision, preference or question needs no evidence. You cannot delete anything here; only the owner can, from their settings.

Remove one of the owner’s page type rules (its id from cvl_page_type_rules). A PREVIEW unless the call says dry_run:false: without it nothing changes, and the answer says how many pages it holds and what each would go back to — the type another of your rules gives it, or the status it had without your rule — with up to 10 examples. Show the owner and apply only on their yes in the chat, with dry_run:false. Applied, those pages go back, and undo is the exact call that brings back your rule and its pages. It changes how pages are grouped in reports; it changes nothing on the site.

Start a crawl of this workspace’s own site with the Convertlyft spider: every page’s status, title, description, headings, words and technical issues. A first-party crawler, not a paid provider; it counts toward the plan’s “Pages scanned”, at most max_pages (dry_run: true says how many pages are left and checks nothing). It starts immediately and runs for a few minutes; the SEO tab follows it live, and cvl_site_pages / cvl_page_content read it once it has finished. Use it when the person asks for a crawl or an audit, or when they ask about their pages and no crawl exists yet.

Find a Convertlyft tool that is not in your tool list, and the skills (step-by-step playbooks) that fit. Your list holds a core set; behaviour, SEO, errors, site and workspace reads are here. Answers name, title and one line per tool; schema:true adds the input schema. Run a read it finds with cvl_call_read. Writes are always in your list already: call them directly.

Add searches to the keyword list, optionally for one page. Idempotent.

Start tracking this site in another market (country, language and device). The new market copies the primary market’s tracked searches and is checked weekly (Mondays, 02:00 UTC); its checks are included in the plan and count toward keywords tracked; the primary market does not change. Nothing is bought when it is added. dry_run: true answers what would be added and when the first check runs, without adding anything. A market already tracked is refused.

Whether AI assistants name a site: answers from the provider’s tracked panel of prompts, each with the platform, the question, whether the site was mentioned and an excerpt. A sample of tracked answers, not every answer an assistant gives. Research is included in the plan, inside this site’s daily limit for paid search data, with no per-call confirmation. The answer is stored, so asking the same question again is instant (refresh: true buys again). cvl_usage shows what was used. Included in the plan, inside this site’s daily limit. dry_run: true gives, without buying, how much of today’s limit is used (rounded) and whether this call fits (fits_today).

The latest full crawl of this site as a summary (a newer list re-check, a 0-page run or a crawl still running never replaces it; full_run names it then, and a re-check clears what it no longer finds): the run (run_id, status, pages crawled), the issues it found in lanes (you_fix, developer, not_a_problem), counted by severity and listed by id with how many pages each touches (the first 10 with up to 3 sample paths), the ten pages people land on most over four weeks of measured visits, the crawl-over-crawl changes and past runs. One issue in full — its fix text and sample pages — with issue: "<id>". Page rows a page at a time with limit and cursor; a crawl whose rows stay on the crawler (run.rows_on_crawler) is read with cvl_site_pages.

Any site’s backlink headline counts from the provider’s link index: referring domains, backlinks, broken backlinks, referring pages and when it was first seen. Research is included in the plan, inside this site’s daily limit for paid search data, with no per-call confirmation. The answer is stored, so asking the same question again is instant (refresh: true buys again). cvl_usage shows what was used. Included in the plan, inside this site’s daily limit. dry_run: true gives, without buying, how much of today’s limit is used (rounded) and whether this call fits (fits_today).