Home / What's new
Changelog updated 2026-11-22
What's new
What changed, by day, newest first. EarnCitations hasn't launched yet: everything below runs on your own machine today (quickstart).
- earncitations.com is live: The site now runs on Cloudflare at earncitations.com. Every page answers at its .html address and its short one without a redirect, and an agent that asks for Markdown gets the page's Markdown copy. The free tools that need our server (the audit box and the other live checks) show a static-preview note until it is hosted; each runs locally in one line.
- Unreachable sites: HTTPS not judged: When the audit can't reach a homepage at all (a DNS error, a timeout), the HTTPS check now says it wasn't evaluated, instead of failing with "uses HTTPS". The reachability check already reports the problem.
- Plans from $149 a month: Pricing now has monthly plans by the buyer questions you track and how often we check them: Starter $149 (50 questions, weekly), Growth $319 (400, every other day) and Scale $999 (1,000, daily), each with a 14-day free trial, plus a custom Enterprise plan. Earned, which bills only per verified citation, stays as the alternative. The pricing page's FAQ and its structured data now come from one list, so they can't disagree.
- Pricing copies: a missing summary: The audit's Markdown copy check now also notes a pricing copy that doesn't open with the page's summary (its meta description), the line Mintlify's copies put under the title so an agent knows what the page is before it reads on. A hint only; the score doesn't change.
- More providers in Track: A question like "how to find a dog walker", "how to choose a pet sitter" or "how to find a locksmith at night" now counts as a buying question in Track, as a dentist or plumber question already did. "How to find a walker class in Java" still doesn't.
- Tables in Markdown copies: md-copies (and our own copies) now write an HTML table as a Markdown table, with its header row, instead of loose "Free | 60 |" lines: a pricing table is what an agent quotes, and the header row says which number is which. A table used for layout (headings, code or long text in its cells) keeps the line-by-line reading, and so does any table it can't rebuild without losing text: a nested table, cells whose end tags are left out, or more than 1,000 rows. Captions are kept, and an escaped pipe stays inside its cell.
- Shorter page summaries; review 35 fixes: Every page's meta description is now 50 to 170 characters, the range our own audit checks for (ten of ours ran to 238), and each one is the summary at the top of that page's Markdown copy. A last review pass of the copies' links: an image's alt text keeps a "<" in it, a link holding only a space and an image takes the alt text, brackets written as entities are escaped, more block tags separate a link's words, and "a tutor for exam prep" counts as buying again. In our llms-full.txt, each page's own headings now sit one level under its title, so the outline reads page by page. A final pass: possessive lookups ("a bank's routing number") are flagged, and a backslash in a link's text is escaped.
- Review 34 fixes: A twelfth review pass, of the links and summaries in Markdown copies. Words on separate lines inside a link (a card built from divs, a <br>) no longer run together; a link inside a heading (a blog index's post titles) stays a link; a protocol-relative link keeps https; an entity in a link or a meta description is decoded once, as a browser does. Two slow paths are linear again (many differently named hidden tags; a link holding thousands of line breaks). The question tracker also flags lookups that only mention a provider ("a bank routing number", "a vet tech job").
- Review 33 fixes: An eleventh review pass, of the review 32 fixes. Markdown copies: nested shadow roots (a web component inside another, as server-rendered Lit emits) keep their text instead of emptying the copy; a "<template" inside a script string no longer swallows the page; list items whose end tags are omitted no longer stop later hidden elements being removed; a code block that contains ``` gets a longer fence; a hard-linked llms.txt still gets its index line. Log analysis takes an agent file's latest status by time, so rotated logs read newest first are right. Its error list says "/pricing.md (404, now 200)" for a file that has loaded since, instead of showing the 200 as the error. The question tracker counts "how to find a plumber fast" and "a dentist accepting new patients" as buying again. The audit reports a root llms.txt that errors or refuses us as that, instead of hiding it behind a section's file.
- Links and a summary in Markdown copies: md-copies, and our own copies and llms-full.txt, used to drop every link: an agent reading a copy saw "see the checklist" with nowhere to go. Links now stay Markdown links (Mintlify's copies, which Jam's docs use, keep theirs): absolute when the site is known, root-relative otherwise, so a copy served at another path still points to the right pages. A card link (a heading and a description inside one link) puts the link on the heading. In-page (#…) and script links keep only their text. Each copy also opens with the page's meta description under its title, a one-line summary for an agent deciding whether to read on (Mintlify's copies do the same); md-copies lists the pages that have none. The audit's Markdown-copy check now also says when a site's pricing copy has no links at all though the page has them.
- Review 32 fixes: A tenth review pass. Markdown copies: code blocks keep their line breaks and indentation, named entities outside our short list are decoded, a summary holding a heading is not wrapped in bold, form controls (select, textarea, labels) and plain buttons are left out, and an unclosed element no longer hides the rest of the page. The audit reads a section's llms.txt after a redirect to another host (www to apex, say), keeps WebMCP errors from your HTML when a firewall refuses the manifest, and doesn't count <form-row> custom elements as forms. Log analysis lists an agent file as failing only if its last request failed. The question tracker no longer reads "a cleaner way to write loops" or "pick up Spanish" as buying a service. Site names skip emoji and one-letter title parts. The delivery check keeps text after a literal "<!--" in a Markdown copy.
- What AI sees: images with no size: Lighthouse's Agentic Browsing category measures layout shift, because an agent that clicks by position or screenshot misses a button that moved, and images without dimensions are a common cause. What AI sees now notes images with no width and height in the HTML (two or more). Only a hint: your CSS may reserve their space, and we don't read stylesheets. It also says when inline scripts register WebMCP tools with JavaScript (document.modelContext.registerTool), which a reader of the HTML can't see.
- /pricing.md works too: Our server answers /pricing, so an agent that applies the llms.txt convention (the URL plus .md) asks for /pricing.md. That now returns the page's Markdown copy, as /pricing.html.md does; a page without a copy still answers 404.
- Our Markdown copies point to the rest: Each Markdown copy of our pages now opens with a line naming llms.txt and llms-full.txt, so an agent that lands on one page can find the others. Mintlify's docs copies, which Jam's docs use, open the same way. md-copies does the same when your build has an llms.txt.
- md-copies lists stale copies: A Markdown copy left from an earlier deploy, for a page that is now built by JavaScript (no text in the HTML), still has the old text, and servers keep sending it. md-copies lists such copies and leaves them, since a copy you wrote by hand for a JavaScript page looks the same from the build. (A short-lived --prune option deleted them; it is gone because it deleted hand-written copies too.)
- Cleaner Markdown copies: Markdown copies (ours, and those md-copies writes) leave out form controls and their labels, so a copy no longer has lines like "Check delivery" or "The page you changed" on their own. The text around a form stays.
- Review 31 fixes: An independent review of the Markdown copy work found 21 problems, each reproduced and now fixed with a test. md-copies no longer writes through a symlink or onto a folder at a copy's name, leaves pages that aren't UTF-8 untouched, writes root-relative links (a folder's page served without a trailing slash pointed to the homepage's copy), lists copies left from an earlier run, and refuses unknown options. Its converter runs in linear time (malformed HTML took 30 seconds), removes comments and scripts first (a "<nav" inside a script could empty the page), and an unclosed hidden element no longer drops the rest of the page. Our server finds the right copy for file names with % or #, and the two variants of a URL never share an ETag. The delivery check counts /pricing.html.md in llms.txt for /pricing and /docs/index.html.md for /docs/index.html, and ignores text that is only in a comment of the copy. Also: the question map on keyword domains (helpdeskpro.io), the section llms.txt after a redirect, the next step for a page that timed out, how-tos about a "host" or "x or y", and hidden elements with "<" in an attribute.
- A server error on llms.txt is named as one: When /llms.txt answers with a server error (500, 502), the audit now says so instead of calling the file missing. Chrome's Lighthouse fails its llms.txt audit on a server error and counts a 404 as not applicable. A 503 is still read as our crawler being refused, as before.
- WebMCP form attributes checked: Chrome's Lighthouse has a WebMCP schema validity audit. The audit now checks the part of it that's in the HTML: a form with a toolname but no tooldescription (or the reverse) and a required field without a name in such a form are warnings, with their own fix; an optional field with neither a toolparamdescription nor a label, or more than 40 tools on one page (Lighthouse warns above 40), is worth fixing. Forms that declare neither attribute aren't judged. What AI sees shows the same problems for one page.
- Agents asking for files you don't have: The log analyzer now lists AI agents' requests for llms.txt, llms-full.txt and .md copies that got an error (a 404 for /pricing.md, say) as a warning of their own, and no longer counts them as fetched. It also counts fetches of agent manifests (/.well-known/ard.json, mcp.json). An agent asking for a file you don't serve is the clearest sign it's worth publishing.
- The log analyzer shows who reads llms.txt: Per AI agent, the log analyzer now counts requests for llms.txt (root or a section's), llms-full.txt and Markdown copies, and says so in plain words, including when no agent fetched any of them. Path lists are capped per agent, so the counts are floors.
- md-copies: Markdown copies at build time: A new command,
earncitations md-copies <build-dir>, writes a Markdown copy next to every built HTML page (pricing.html.md), from the page's main content, offline, and with --link adds the link tag that advertises it. Run it in your deploy so the copies always match the pages. Our own site's copies are built by the same converter. - The audit reads a section's llms.txt: Auditing a section such as acme.com/docs/ when the site has no root llms.txt, the audit now reads /docs/llms.txt (the llms.txt proposal lets a file cover a section). It's reported as a warning that names the file and asks for a root llms.txt linking it, since that is the one agents look for first.
- The drafted llms.txt links Markdown copies: When a page advertises its own Markdown copy (a link tag or a Link header, on the same host), the fix kit's llms.txt draft links the copy instead of the HTML page, as the llms.txt proposal suggests. The audit notes the copy without fetching it.
- Review 30 fixes: The question map again suggests "best CRM software for small business" when the title is "CRM Software | Acme", and on hosts like gocrm.io. The crawler view's hidden-element scan is linear (a crafted 3 MB page went from 45 seconds to 41 ms) and keeps positions on Turkish text. The delivery check resolves relative links in a folder llms.txt, matches HTML, entities and escapes in a Markdown copy, and keeps the earlier fix when the copy is also stale. The audit reports an advertised ARD file that isn't JSON beside a valid one, reads the legacy file behind a catch-all ard.json, and keeps FAQ questions that name different tags apart. The question doctor names a failing page in its findings and comparisons, and the tracker treats "how to choose a dentist" as buying.
- API description fix: openapi.json now describes the delivery check's next as a list (it always was one), and documents checkedAt, expect, since and facts, including the new llmsFile and markdownCopy.
- Markdown for agents that ask for it: Our server now sends a page's Markdown copy to a client whose Accept header prefers text/markdown to text/html, and names the copy in a Link header on the HTML. Browsers, which never ask for Markdown, get HTML as before.
- Five more AI assistants in the log analyzer: Visits referred by Poe (poe.com), Qwen Chat (chat.qwen.ai), Kimi (kimi.com), Duck.ai and NotebookLM now count as AI assistant visits, in the log analyzer, the analytics-export reader and the GA4 channel regex. By referrer host only: their short names as a utm_source could be anyone's campaign.
- Markdown copies of our own pages: Every page in our sitemap now has a Markdown copy at its URL plus .md (pricing.html.md, index.html.md for the homepage), rebuilt with the page and advertised with a
<link rel="alternate" type="text/markdown">, as we suggest to others. Our llms.txt says so. - Delivery check reads the Markdown copy: When a page advertises a Markdown copy (a
<link rel="alternate" type="text/markdown">or a Link header, as the llms.txt proposal suggests), the delivery check looks for your new sentence there too. A deploy that updates the HTML but not the copy leaves agents that read the copy the old text; that, or a copy that doesn't load, is now a warning. - Review 29 fixes: The FAQ markup check now reads questions inside
<details>and<summary>, after a number or badge in its own element, in decomposed accents, and naming a tag ("the <iframe> embed"), and counts a site-wide FAQ once across pages. The crawler view no longer counts controls inside hidden menus or panels, reads image names from long tags and from title, and doesn't take an aria-hidden icon as a name. The audit reads the legacy /.well-known/ai-catalog.json and notes a broken ard.json behind a valid one. The question map skips "Plans & Pricing for …" and the brand on hosts like getacme.com; the question doctor names a page that answers 404 or 410 (gone); and the tracker no longer reads "how to pick a lock" as a buying question. - A Markdown copy in llms.txt counts as the page: The delivery check now treats a link to a page's Markdown copy (/pricing.md, /docs/index.html.md) as a link to the page, and names a folder llms.txt that redirects to the root file once. The research page's count of pricing pages linked from llms.txt was made before this: vercel.com links /pricing.md and was counted as not linking it, so that figure is now marked as a floor.
- Delivery check reads folder-level llms.txt: The llms.txt proposal now allows a file at a subpath, such as /docs/llms.txt for pages under /docs/, with the most specific file applying. The delivery check tries the nearest two folder-level files before the root one, and says which file lists the page, or that a nearer file doesn't.
- What AI sees: controls an agent can't name: a note when buttons, links or form fields in a page's HTML have no name (no text, image alt, aria-label, title, label or placeholder). Agents that operate pages read the accessibility tree, where an unnamed control is just "button"; Chrome's Lighthouse checks this under "Accessibility for agents". Not scored, and it reads the HTML as sent: scripts may add names later.
- Audit: legacy ai-catalog manifests: an agent resource manifest linked with rel="ai-catalog", ARD's old name, is now found and checked, with a note to move it to /.well-known/ard.json or rel="ard", as Lighthouse advises.
- Review 28 fixes: an independent review of this afternoon's code found twelve problems, each reproduced and now fixed with a test. The draft JSON-LD in the question doctor and its coding-agent prompt can no longer close its own script tag. data-rel, data-type and data-toolname attributes no longer count as real ones. A valid ARD manifest is used even when another location has a broken one. A Markdown copy needs rel="alternate", so an llms.txt linked with rel="describedby" isn't compared with your pricing page. A wrong price in an answer is caught whichever order the source passages come in. A homepage that answers 404 gets a next step about the address, not downtime.
- Audit: FAQ markup that the page doesn't show: the FAQ check now warns when two or more questions in a page's FAQPage markup (or all of them) never appear in the page's own text. Google's structured data policies say not to mark up what readers can't see, and an assistant reading the markup gets questions people never saw. Our own homepage had this: its markup now matches its visible questions word for word.
- Agent resource manifests (ARD) and WebMCP forms in the audit: Chrome's Lighthouse added an experimental Agentic Browsing category that checks ard.json, a manifest of a site's MCP servers and other agent resources. The audit's MCP check now also finds an ARD manifest (at /.well-known/ard.json, or where robots.txt, a <link rel="ard"> or a Link header points), checks it against the rules Lighthouse documents (urn:air: identifiers, exactly one of url or data, 2 to 5 representative queries), and reports an advertised manifest that doesn't load. It also counts forms that declare a WebMCP tool. ARD and WebMCP are in the glossary.
- llms.txt version 2: llmstxt.org updated the proposal on 2026-08-10. The validator now accepts a byte-order mark before the H1 and notes headings between the summary and the first section, which v2 doesn't allow there. The audit and the crawler view now find a Markdown copy advertised by an HTTP Link header (rel="alternate"; type="text/markdown") as well as by a <link> tag, and the fix kit offers the header as an option. The page's questions now say what v2 changed, and that llms-full.txt is a convention, not part of the proposal.
- Answer log: a question set now points out discovery questions that read like a how-to or a definition ("how to close deals faster", "what is lead scoring") and name nothing a buyer would choose between, after Jam's advice to track the prompts that stand for buying decisions. It's a note; nothing is dropped.
earncitations track setprints the same. - Fix kit: a promotional description ("the highest-rated…") is no longer drafted into llms.txt or the Organization JSON-LD. Each gets a TODO that says why, as facts.json already did.
- Question doctor: a homepage that timed out or errored for our crawler now reads "Couldn't load your homepage", with no crawl verdict about AI crawlers and no delivery check.
- Answer check: plural endorsements ("solid alternatives", "good options") count as recommendations. A page that says "a strong alternative" backs a claim of "a strong open-source alternative".
- Answer log: the pace line's swing is now for a change between two months (√2 wider than for one month: ±11 points at 171 answers a month). The first comparison is 2 to 3 intervals after a change.
- Question map: categories that start like a verb ("Log management", "Contact center software") get their best-for question again.
- Answer log: "What to ask next" now says what your routine gives you, like Jam's pricing table counts answers checked a month: about how many answers a month, how soon the before-and-after compare can read a question, and how far a rate can move between two months by chance (±11 points at 171 answers a month). The same line ends
earncitations track due. - Answer check: a wrong price inside a recommendation ("a good fit for teams at $29 a month" against a page saying $49) is now marked contradicted, not just "the page doesn't recommend this". "A fail-safe restore option" no longer counts as an endorsement.
- Question doctor: the coding-agent prompt never lets the agent write a price answer from figures on the page (those can be an add-on, structured data or a quoted rival). It leaves a TODO for you instead. No prompt when the homepage didn't load at all.
- Question doctor: a homepage that refused our crawler now reads "Not read". It no longer shows a failed crawler step, 0% coverage or a delivery check for a page we never saw.
- Audit: a bare HTTP 429 or 503 (rate-limited or briefly down, no challenge page) is a warning to try again in a few minutes. It is no longer reported as bot protection.
- Promotional descriptions: "best practices", "no 1-year contract", "#10 on G2" and the brand's own name ("Best Buy") are no longer flagged. "World’s" with a curly apostrophe and "the market leader" are now caught. The check is English only. A flagged description is also left out of the drafted facts.json.
- Question map: no "best … for …" questions made from a call to action ("Sign up for free"), a thank-you or the brand itself.
- Answer log for agents: answer_log_report, and definitions: a new MCP tool reads an exported answer log and returns the report, optionally for the last N days. Its output, and
track report --json, now end with plain definitions of every number (what "named", "recommended", "named you first", citation share and the scopes mean), for an agent to quote rather than infer. After Jam's metricDefinitions in its API. - Answer check: more ways to recommend: "a strong open-source alternative", "a solid choice", "a good fit" and "worth considering" now count as recommending the name they're about, as "best" and "top pick" did. A bare adjective ("strong security", "good support") still recommends nothing, and "not a good fit" recommends no one.
- Fix kit: no marketing claims in the Fact Ledger: a site description like "the highest-rated scheduling system" stays your description, but the draft facts.json no longer lists it as a fact (nobody can check "highest-rated", "leading" or "#1"). The kit asks you to write a factual summary instead: what it is and who it's for.
- Question map: a discovery question from any part of the title: a title like "Brand | Scheduling Software for Online Bookings" now gives "best scheduling software for online bookings". Only titles that started with the claim did before, so a site with its name first got no question that leaves its name out.
- Audit: a partial check says so: when a homepage refuses or challenges our crawler, most checks can't run. The result now says "Partial check", how many checks weren't evaluated, and asks you to check whether your bot protection lets AI crawlers through, instead of showing a bare grade for a site we never saw. Same on the command line.
- Docs check: quickstarts an agent can follow: a new note (not scored) for a docs page that lists quickstarts as cards ("Node.js Quickstart", "Python Quickstart") whose links exist only after JavaScript runs. Crawlers and agents that read the HTML see the names and nothing to follow. Found on a real docs site whose introduction names 14 quickstarts and links none in its HTML.
- Question doctor: a site that refuses our crawler: a homepage that answers 401 or 403 to us is no longer reported as "answer engines can't read it". Bot protection can refuse crawlers it doesn't know and lets known AI crawlers through, so the result now says "Couldn't read your site" and asks you to check whether your bot protection allows AI search crawlers and user-triggered fetchers.
- Question doctor: a prompt for your coding agent: the diagnosis now has a "Copy a prompt for your coding agent" button, like the audit, delivery and docs checks: the question, the page, what we found, the draft fix from your own words (or a TODO where we found none) and, with a competitor, what their page has that yours doesn't. It ends with the check to run once it's live. Nothing for a question your site already answers: what's left is off-site. After Jam's copy-prompt on every action.
- Answer log: competitor rows leave out questions that name them: a competitor's named and recommended rates, and your "Named you first" rate, no longer count answers to a question that names that competitor ("HubSpot alternatives"), which name it by construction; nor does its citation share, and its own pages are never listed as sources to pitch. The log says how many were left out, even for a competitor with no other answers. After Jam's rule that such questions are tracked but not scored.
- Answer log: last 7, 30 or 90 days: the rates can cover a window instead of every saved answer, so a change since last month isn't buried in old answers (
track report --days 30on the command line). The summary says how many older lines were left out; exports always carry the whole log. After Jam's 7, 30 and 90-day windows. - Review 23: one rule for short pages: a page under 80 words that runs scripts is treated the same way in the audit, the question doctor and the Answer check: what it says counts, and a claim we can't find there is "unknown", not absent. Under 10 words with a script, it's an app shell everywhere. The question head-to-head no longer calls you "ahead" when their page is the short one, and the llms.txt generator no longer says a file is well-formed when the server refused to show it to us.
- Answer log: rates by topic: put
## Pricing(or any topic) above questions in your own set and the log shows named and recommended rates per topic, and the competitor named ahead of you in each, so you can see which buying decision you're missing from and who has it. Sets from the question map get topics from its question types (pricing, comparisons, integrations…). The CSV export has a topic column. After Jam's prompt topics. - Schema generator: start from your site: a "Start from my site" button fills the empty fields from what your own pages already say (name, website, description, profiles, and for a local business the phone and street address). A logo is filled only when your markup calls it a logo, and nothing you typed is replaced.
- llms.txt generator: when you already have one: instead of only "paste it into the validator", it now shows what we found: whether the file is well-formed, whether its links resolve, and whether there is an llms-full.txt.
- Answer log: named first, and questions that name a competitor: a "Named you first" rate: of the answers naming you and a competitor you track, how many named you before all of them (after Jam's position metric). And a question that names a competitor ("HubSpot alternatives") no longer makes that competitor "the answer" in the list of questions where a competitor beats you, since those answers name it by construction (Jam scores such questions apart).
- Question map and question doctor: short pages: a short page that isn't an app shell is now a content problem ("weak"), not a crawl problem. We no longer claim its answer "only appears after JavaScript runs" when we can't know that; when the page has scripts, the diagnosis says so as a possibility.
- Answer check: short pages are read: a cited page under 50 words that isn't an empty app shell (example.com) is now read and marked "short", instead of being reported as unreadable. What it states counts; if it runs scripts, a claim not found there is "unknown" rather than "unaddressed", and its title cannot back a claim on its own. An app shell is still "unknown".
- How to use it, in order: a short guide to the free tools in the order to use them: the first hour (can AI read you?), the first week (what buyers ask), then every week and every month.
- Audit: dates that disagree: the freshness check now warns when a page's visible "Updated" date and its JSON-LD dateModified are more than a day and a half apart, naming the page and both dates. It uses the same rule as What AI sees, which reads only unambiguous dates.
- Data and security: what each check fetches, what the server keeps (in memory, minutes to hours) and what stays in your browser, with the promises a test can check checked on every build. Every page now has a content security policy.
- Since your last check: the audit, the docs check, the delivery check and the question map say what changed since you last ran them in this browser: the score, what you fixed, what newly fails (for the map, which questions are now answered or no longer are). A cached audit says so, with a button to check again now.
- Docs check: could a coding agent set your product up from your public docs? Twelve scored steps, plus unscored notes: one name per credential, sections that stand alone, errors an agent can act on, install commands that match npm, and where keys go and how they change. Compare with a competitor, and see where you sit in our study of 123 software sites.
- Delivery check: after you ship a fix, did it reach AI crawlers? The page, your new text in the HTML, robots.txt for each agent, noindex, the sitemap, links and llms.txt, each reported separately, with IndexNow one step away.
- Answer check: each claim in an AI answer against the pages it cites. Whether the answer is about the right company, the four kinds of presence (named, recommended, cited, described by others), competitors, correction drafts and a correction-ledger CSV.
- Answer log: record what assistants answer, by hand, in your browser: what to ask next, which sites get cited, citation share, before and after a change.
- Question map and question doctor: the buyer questions your own pages suggest, each diagnosed (answered, weak, missing or a crawl problem), with drafted fixes in your site's own words and a done-when delivery check.
- What AI sees: one page as a crawler that doesn't run JavaScript gets it, including hidden instructions to AI and Markdown copies for agents.
- AI log analyzer: which AI agents read your site, impersonators caught with the operators' IP ranges, people who clicked through from an AI answer and reached a signup page, and a Google Analytics "AI assistants" channel.
- For coding agents: a "copy a prompt for your coding agent" button on the audit's fix kit, the docs check, the delivery check and the map. The OpenAPI contract for the HTTP API is now listed in an RFC 9727 catalog at /.well-known/api-catalog.
- Research: 150 independent hotels and restaurants, whether software companies answer their own buyers' questions, Markdown copies for agents and the docs study.
- The method: how a question goes from diagnosis to a fix to a measured answer, step by step.
- Answer log: pages cited: which of your pages the answers cite, and which other pages they lean on most, each counted once per answer, with how often you were named. Links are grouped by address with utm_… and #sections removed; ?version=2 stays a different page (also in
earncitations track report). - Answer log: sources to pitch: for each other page the answers cite, how many of those answers named a competitor and not you. Those are the pages worth asking to include you, with a reason that's true.
- Answer log: write your own question set: no question map needed: type your buyers' questions one a line (or
earncitations track set questions.txt) and they become a frozen, versioned set (discovery or diagnostic by whether they name you), with the schedule of what to ask next and a download to keep. Wordings that look like one decision are pointed out straight away. - Answer log: where a competitor is the answer: per question, the competitor named in a larger share of the answers than you, widest gap first, with a link to diagnose that question. Shares, not counts, so a competitor you added later isn't overrated; fewer than three answers is marked thin.
- Delivery check: several pages after one deploy:
earncitations delivery --from pages.txt(a page a line, a tab, the sentence it should now contain) or several URLs. The pages are checked one after another, the site-level files are fetched once for the run, and you get a one-line-a-page summary. It exits 2 if any page isn't delivered. - Log analyzer: read your analytics export: no server logs? Paste or choose a GA4, Plausible, Fathom or Matomo CSV with one row per source. It counts sessions from the same AI assistants the log analyzer knows, and sets their key events and revenue against everyone else's, with a warning when there are too few to compare. With a landing-page column it also lists the pages assistants send people to. Read in your browser.
- Answer log: questions that may be one decision: two wordings of one buying decision count as two questions and weigh it twice. The question set and the report now point out wordings that share almost all their words once "best", "top", "tool" and the like are set aside, so you can keep one (also in
earncitations track report). - Audit: download as a backlog: every failing check on the audit result as a CSV, most important first: who usually does each (developer, content editor, marketing), what to ship, effort, and how you'll know it's done (the check passes, and for page changes the delivery check confirms crawlers got it). Or copy the same list as a Markdown task list to paste into a GitHub, GitLab or Linear issue (the site's own text goes in code spans, and @handles are broken so nobody gets notified).
- Docs check: a test for your coding agent: each result can be copied as a test to run yourself: a fixed start, the task, a success check and what to record, with the named and unnamed versions kept apart. We run nothing.
- What AI sees: dates that disagree: a note when a page's visible "Updated" date and its structured data's dateModified name different days. Only unambiguous dates are read (a month name or YYYY-MM-DD).
- Answer log: CSV export: one row per attempt (question, engine, time, what was counted, the sites cited and the exact answer), for a spreadsheet a colleague can check. Cells a spreadsheet would run as formulas are escaped. Also
earncitations track csv. - Answer log: attempts with no answer: when an engine refuses, errors or asks a question back, log the attempt. The report shows how many attempts got an answer next to every rate, and the question stays due (also
earncitations track add … --no-answer refused). - Docs check: old pages on the current path: a new unscored note when the quickstart, the API reference or the keys page it links says at the top that it's deprecated, legacy or replaced. It reads no extra pages.
- Docs check in CI:
docs --min 10fails a build when fewer than ten of the twelve checks pass, and the GitHub Action takesdocsanddocs-min. - Who does which part of the work: for the four jobs any AI-visibility setup needs an owner for, what EarnCitations does today and what stays with you.
- Accessibility: each tool announces its result to screen readers (the score, the verdict, the counts), not just that a result arrived, and the audit field shows keyboard focus.
- AI-readiness audit and fix kit: can 26 AI agents read your site, and drafted files to fix what they can't: llms.txt, robots.txt rules, Organization, FAQPage and Product markup from your own words and printed prices, and a Fact Ledger draft.
- AI crawler directory: every agent sourced to its operator's own docs, with a robots.txt tester and builder.
- Free generators: the llms.txt generator and validator, the schema generator and the glossary.
- Research: 124 software companies: how AI-ready they are, from our own audit of each homepage.
- CLI, MCP server and API: every check as a command, an MCP tool and an HTTP endpoint, plus
watchalerts, a robots.txt guard for CI, a GitHub Action and a WordPress plugin.