Repository navigation
Releases
v1.2.0
Minor Changes
- 1d7ce0f:
rulesanddisableRules, and the CLI's--ruleand--disable, take what a rule id has before its dot as well as a rule id or a category:["robots", "og"]runs everyrobots.*andog.*rule, though their categories areindexingandsocial(issue #187). Before, such a name failed as an unknown rule id. An unknown name inrulesor--rulestill fails, and the error now names the rule ids close to it:response.statusgetsresponse.error-status,noindexgetsrobots.noindexand then the other rules about noindex, each with what it requires, such as(requires crawl).
v1.1.5
Patch Changes
- 2514585: The CLI and the local MCP server send their usage analytics and error reports to
https://g.rivi.software, the reverse proxy PostHog runs for RIVI Software, which forwards them to the same EU project. Before, they went tohttps://eu.i.posthog.com. What an event carries, and the ways to turn it off, are unchanged.
v1.1.4
Patch Changes
-
87ab658:
install-browsersends acli_commandevent like the other commands: the flags given (--with-deps), the exit code and the duration, so a download that fails shows up (issue #173). Turn it off as before, withagentic-seo-check telemetry disable,DO_NOT_TRACK=1orAGENTIC_SEO_CHECK_TELEMETRY=0.On the hosted server, a tool call from a client that names itself neither in
initializenor in the call, such as claude.ai's, now has the product token of its headers as$mcp_client_name(Claude-User), and$mcp_client_name_sourcesays where the name came from. Before, its calls had no client name. The local server is unchanged.
v1.1.3
Patch Changes
-
1a11247:
structured-data.missing-contextreads the host of a JSON-LD@context, so a context such ashttps://example.com/schema.orgnow gets the finding (issue #174). Before, any context withschema.organywhere in it passed.https://schema.org,http://schema.org/,www.schema.org,//schema.organd a bareschema.orgstill pass.JSON-LD wrapped in an HTML comment that ends in
--!>is read as it is with-->, rather than reported as invalid JSON.
v1.1.2
Patch Changes
- 5cb57a0: The hosted MCP server moved to
https://agentic-seo.org/mcp, and the website tohttps://agentic-seo.org, one hostname for both (issue #168). The server's address is also its page: a browser that opens it gets the setup, the limits and the privacy notice. The former addresses,https://mcp.agentic-seo.rivi.software/mcpandhttps://mcp.agentic-seo-check.rivi.software/mcp, keep working, so clients set up with them need no change; the former pages, share links included, redirect to the same address onagentic-seo.org. The hosted server's requests keep their User-Agent and Signature-Agent for now. The MCP servers' instructions link the rules athttps://agentic-seo.org/rules.
v1.1.1
Patch Changes
- a6d28f2: A crawl of a large site now fits in the hosted server's memory, and a local crawl or directory audit uses far less of it (issue #165). Before, every page kept its whole HTML until the audit ended, through the tags its findings quote, and kept its own copy of the menu and footer links every page repeats. A crawl of 100 pages of 600 KB with an 800-link menu needed a 256 MB heap; it now runs in 72 MB. On the website, a crawl of such a site ran out of memory partway and stopped without a report. The findings, their fixes and the audit JSON are unchanged.
v1.1.0
Minor Changes
-
d88ba39: The CLI's text report shows its lists as tables, each with a header row in a rounded frame (issue #154): a crawl's pages, its findings by rule and its field data, and the rules of each category that
agentic-seo-check ruleslists. Before, they had no headers: nothing said that a page's first number was its score, and the rule ids and the field data's values started wherever the text before them ended. The field data now has a column for each metric. The last column of a table wraps, and a path too long for it breaks inside the path; when the other columns leave it too little room, its text goes under each row. Text from a page in Japanese, Chinese or Korean, or with emoji, counts as the two columns a terminal draws it in, so the frames and the wrapped text stay in line. The JSON and the Markdown report are unchanged. -
4052df1: A page can now be compared with one to five competitors' pages, side by side, with what they have that it lacks (issue #129):
agentic-seo-check compare <url> <competitor-url>..., andseo_compare_pageson the local and the hosted MCP server. The tool searches for no competitors: you, or your agent, bring them, from a search for the page's query, from the user, or from the pages an AI answer cited.Each page is requested once and gets both audits, and each site's robots.txt and llms.txt are read once. The answer is a new kind,
comparison, with each page's SEO and AEO score and itssignals: words,<h2>subheadings, question headings and how many are answered right under them, tables, lists, structured data types, the Article's author and dates, the AI search crawlers and answer bots robots.txt keeps out, llms.txt, response time and HTML size. With--keyword(keyword) it adds whether the phrase is in the title, an H1, the description and the text, and with--field-data(fieldData) real visitors' LCP, INP and CLS.gaps: what at least half of the competitors read have and the page lacks, each with the competitors that have it and therulesandfixthat close it on this page, such asaeo.article-author-missing,aeo.question-not-answered-first,aeo.robots-blocks-answer-botorkeyword.not-in-title. A gap no rule covers, such as more text or a table, gives the numbers and no edit.leads: where the page is ahead.
A gap is not a finding: it has no severity and changes no score. A competitor that cannot be read is listed with
skippedand why, and left out of the gaps. Headers go to the page's own origin only, never to a competitor. On the hosted server a comparison is one tool call, each competitor's robots.txt is obeyed, and each competitor's site counts toward the sites a caller may audit in a day. The other tools and commands are unchanged.In the library,
comparePages(url, competitors, options)returns aComparison, andrenderComparisonMarkdown,renderComparisonTextandcompactComparisonrender and trim it.
v1.0.0
Major Changes
-
7b67454: Agentic SEO Check is 1.0.0: what agents and CI depend on now follows semantic versioning (issue #150). That is each rule's id, severity, message and recommendation, the MCP tools' names and input schemas on the local and the hosted server, the CLI's flags and exit codes, and the audit JSON (
schemaVersion: 1). Renaming, removing or changing the meaning of any of them now needs a major version. Before, a minor version could.Nothing else changes from 0.22.0: the rules, the tools, the flags and the JSON are the same.
v0.22.0
Minor Changes
-
2f322c3: Audits can now include what Google ranks on: the Core Web Vitals of real visits, from the Chrome UX Report (issue #130).
--field-dataonauditandcrawl, andfieldDataonseo_audit_pageandseo_crawl_site, ask for the URL's record. That is Chrome's 75th percentile of LCP, INP, CLS, FCP and TTFB on phones over the last 28 days, or its origin's when the URL has too few visits. A crawl asks for its first 10 pages. Each page audit getsfieldData, and the text and Markdown reports show it.perf.field-lcp-poor,perf.field-inp-poor,perf.field-cls-poor: LCP over 4 s, INP over 500 ms or CLS over 0.25 at the 75th percentile, as a warning.perf.field-lcp-needs-improvement,perf.field-inp-needs-improvement,perf.field-cls-needs-improvement: LCP over 2.5 s, INP over 200 ms or CLS over 0.1, as info.
Each fix takes what on the page is known to slow that metric: a lazy first image, scripts in
<head>, images without a size, a missing viewport, the page's scripts. When the page shows none of these, the fix says where to look. A page without a record getsfieldData.reason: "no-data"and no finding.The CLI and the local MCP server need a free Google Cloud API key in
CRUX_API_KEY, and fail with how to get one without it. The hosted server uses RIVI Software's own key and keeps each answer for a day. The query goes to Google, never to the audited site. A URL onlocalhost, a private network or an IP address is never sent, and neither is an audit with headers.In the library,
auditUrlandauditSitetakefieldData: true, or{ apiKey, cache }, andPageAudithasfieldData. AFieldDataKeyErroris anExpectedError.
v0.21.0
Minor Changes
-
af3c956: Everything else the CLI prints on a terminal now has the text report's style (issue #144):
--helpstarts with the version, draws a rule after each heading, lists the commands in two groups and each audit's options in four (crawling, checks, output, requests), and dims choices and defaults. Before, it was commander's layout, with one list of 30 options foraudit.- A usage error is a red
✖line, with commander's suggestion under it, then the usage and the command for the help. Before, the command's whole help followed, which pushed the error off the screen. Piped, the whole help still follows, for agents. The CLI's own errors start with✖, where they startedagentic-seo-check:. telemetrysays whether it is on and why, what is sent, where the choice is kept and the command that changes it, andenablesays when the environment keeps it off. The first-run notice is wrapped to the terminal under a heading; piped, it is the one line it was.install-browsershows each download and its percentage on the progress line, and says when the browser is installed. With--with-deps, or piped, Playwright's own lines show as before.mcpandagentic-seo-check-mcp, started by hand in a terminal, say on stderr that they wait for an MCP client and how to add them to one. Started by a client, they print nothing new.rulesnames the categories when--categorymatches none, suggests the nearest rule id for a mistyped--rule, and ends one rule in full with the command that checks only that rule.
The JSON, the Markdown report, the flags and the exit codes are unchanged.