Agentimus – AI SEO, llms.txt & MCP for AI Agents

Changelog

1.50.1

• Fixed: the “Put this away” link is gone from the Findings screen. Putting a suggestion away only moved it into a fold at the foot of the same list — a collapsible in disguise — and only one kind of row ever carried the control, so part of the list arrived offering to be dismissed while the rest did not. Every finding now simply stays in its list, anything you had put away comes back on its own, and a connected assistant no longer sees a put-away ledger either.
• Improved: the Dashboard opens with its four small tiles, and the Today card now sits directly above the full-width cards it belongs with. The tile row used to sit between Today and everything else, cutting the run of cards in two.

1.50.0

• New: you can write your own robots.txt rules from the Crawler Policy card. A plain text box is appended to the end of the generated file, exactly as written — appended, never replacing, so an edit cannot break the Content-Signal line or the crawler blocks the plugin writes, and clearing the box is the whole reset. The live robots.txt is shown right above the box, fetched from your site exactly as a crawler receives it — which also honestly shows when a robots.txt file on disk or an SEO plugin owns the job instead (then the extra rules do not apply, and the box says so), and warns when your server wraps the file in an error code, which a strict crawler may ignore. Write your own Sitemap line and the plugin stops adding its own rather than printing two.
• New: the review bell’s cards now say why a visitor is unchecked, instead of one blank “not verified” covering every case. Each card names its own cause — identity checking is turned off, this bot cannot be checked because its operator publishes nothing to check against, its visits gave no clear answer yet, or it carries a real cryptographic signature from an operator this site does not know. Each cause reads differently on the card, so you can tell at a glance which unchecked rows are waiting for an answer and which will never have one — and the small info icon carries the full reason.
• New: each recommendation on a review card is now a small labelled chip that names the kind of action it asks for — nothing left to do, one plugin setting, or a rule at your host or CDN — with the full advice shown on hover. Three cards used to read as three similar sentences; now you can see before reading a word that one asks nothing of you, one is a click in Settings, and one needs your hosting panel. The cards are several lines shorter, and on a proven impostor the genuine engine’s verified traffic is counted right in the stats line so the fake never wears the real crawler’s numbers.
• New: the Client Decisions dialog opens each list with a plain notice saying what your decision actually reaches. A block is enforced at this plugin’s machine files — llms.txt, the markdown pages, discovery — not across your whole site, and it is separate from robots.txt, which is only a request a bot may ignore. Trusting a client sends nothing to the bot and changes nothing in robots.txt. When blocking is switched off in Settings, the Blocked list says so instead of quietly presenting rules that are not being enforced. And a trusted client whose name is also on the robots.txt block list now says so on its row — trusted at the door but publicly asked to stay away can be exactly what you meant, but the two may only disagree out loud.
• Fixed: the review panel could advise turning on “Verify bot identities” when it was already on — the nudge showed for any unchecked row that claimed a checkable bot, whether or not the setting was enabled. It now knows the setting’s live state and only suggests what would actually change something. The same fault was in the forged-signature card, which could recommend “turn on blocking” while blocking was already on; it now says the next attempt will be refused automatically.
• Fixed: a valid signature from an operator this plugin does not recognise displayed as “Passed” in a card’s details, indistinguishable from a recognised operator’s. Proof of identity is not a reason to trust someone: it now reads “Not confirmed”, names the signer, and says the site cannot vouch for them.
• Fixed: an entire explanation paragraph on review cards showed the “help” cursor as if it were hoverable — a style collision, and the question-mark cursor now appears only on the icons that actually carry a bubble.
• Fixed: the readiness card said “Everything looks good” when its advisor simply had no setup step left to suggest — a verdict it could not make, since the Cited figure above it can sit at 38% with every setup box ticked. It now says what it measured: “Nothing left to set up.”
• Fixed: the Report’s Citations card said “checks run on a schedule” to owners whose schedule is switched off. It now reads the schedule’s real state: off means “a check runs only when you start one.” And its “Open Citations” link landed on the Search tab; it now opens the Citations results it names.
• Improved: the Client Decisions rows read cleanly — a known crawler’s name is no longer printed twice, the decision verb no longer repeats the tab’s own name on every row, columns line up, long names shorten instead of pushing the date out of line, and on a phone each client folds to two tidy lines with a proper thumb-sized button.
• Improved: plain words on the undo buttons — “Stop trusting” and “Stop ignoring” replace the made-up “Un-trust” and “Un-ignore” — and the buttons look like buttons before you hover them, without turning red for what is a routine, reversible act.

1.49.0

• New: a warning that your CDN is blocking an AI company is now verified before it may keep saying so. Cloudflare counts crawlers by their User-Agent string, and anyone can wear one — on the site this was built for, a vulnerability scanner wore OpenAI’s crawler name through a thousand blocked requests a day, and the warning read “Cloudflare is blocking OpenAI” about a company that never visited. While such a warning is live, the blocked traffic’s source addresses are now checked hourly against what the company publishes. When everything checked proves to be an impostor, the warning stands down to a note that your edge is doing its job — nothing needs allowing. One verified request keeps the warning, and so does an inconclusive check: “could not say” is never read as “not them”.
• Improved: the demoted notice keeps everything the warning had — the date the blocking started, and your fold if you had made one — and it becomes a warning again by itself the moment verified traffic starts being refused.
• Improved: a connected assistant reading the edge summary sees the same verdict: how many of the blocked requests proved to be the company, how many proved to be impostors, and how many could not be determined.

1.48.0

• New: the Request Log records what crawlers read on your pages, not only which AI files they fetched. Until now an AI crawler could read every article on the site and the log showed nothing but its llms.txt fetch. Only machines are recorded — your readers are never logged here, they are counted separately and without a User-Agent — and a switch in Settings turns the whole stream off if you would rather keep a longer history of the files alone. A page served from a CDN or a caching plugin never reaches WordPress, so on a cached site these are a floor rather than a total, and the screen says so.
• New: a request that arrives cryptographically signed is recorded as signed even when the signature belongs to an operator this plugin does not recognise. The check was already running, and a valid signature from anyone other than Google or OpenAI proved who sent it and then left no trace at all — so the first company to sign your site would have been invisible. It is written down now, without granting anyone standing they have not earned: proof of identity is not a reason to trust someone.
• New: a Signature filter on the Request Log, including “signed by anyone” — the question an owner asks first, and one no list of names can express, because you cannot pick the name of a company you have never seen. A connected assistant can ask the same question.
• New: a conflict between your CDN and your own AI policy now appears on the Findings list, dated. A CDN in front of your site can turn an assistant away before the request reaches WordPress, and Agentimus could already see that happening — but only said so on the Request Log, so an owner found out by going and looking. One had been running for weeks unseen. It is a finding now, urgent when the edge is refusing an assistant your policy welcomes, and it carries the date it started rather than only the window it was measured over.
• Improved: a CDN conflict says when it actually began, worked out from the hourly figures already kept rather than from the moment Agentimus first looked. Installing this feature today would otherwise date a problem that started last week to today. Where the figures cannot answer, the wording changes rather than the date being guessed at: “Started the 26th” when it was derived, “Running since at least the 30th” when it has held for as long as anything was kept, and “First seen today” when there is no history to read.
• Fixed: the Findings list said “Nothing is blocking AI assistants” whenever every setup check passed. That is a promise it could not keep: the checks measure this plugin’s own configuration, and an assistant can be turned away by a CDN sitting in front of the site with every one of them green. It now says what it actually checked — “All N setup checks pass” — and the conflict above it carries the rest.
• Improved: the weekly email names any live conflict at your CDN, so the one channel that reaches you when you are not in wp-admin says it too.
• Improved: the conflict notice on the Request Log folds now instead of hiding. Only your CDN changing its behaviour, or your own policy changing, actually ends a conflict — so the notice collapses to one line you can reopen, and the Findings list and the weekly email keep counting it while it lasts. A notice you had hidden before comes back collapsed, not gone.
• Improved: the three small scripts the plugin adds to your pages are compressed when built, cutting them by about two thirds. Their source ships beside them, as it always has.
• Fixed: the address filter above the Request Log offered internal names — “markdown”, “rest:discovery” — while the table beside it showed the plain ones. One screen, one vocabulary now.
• Fixed: a crawler could name your own site as its home page and have the plugin fetch you on its behalf, then record that as a visit. It never proved anything — of course your own front page answers — and it is refused outright now.
• Fixed: where a fake crawler borrows a real engine’s name, the review card blended their numbers — it could say the forgeries were seen five hours ago when that request was the real engine’s verified visit and the last actual forgery was days old. The failing requests now carry their own count and dates, and the real engine’s verified traffic is noted separately.
• Fixed: a connected assistant’s count of clients awaiting review included ones you had already blocked, so it could disagree with the bell in your admin bar; they read one count now. Each client it lists also carries the recorded IP addresses the review advice tells you to block.
• Fixed: on some hosts — macOS, mainly — text in Cyrillic, Arabic, Greek or Devanagari was corrupted on its way into the search-coverage check, logging PHP warnings and letting a page that contains such text be marked as not answering a search it answers.
• Improved: a post’s structured data now carries its featured image. Machines reading the page’s JSON-LD were told the headline, the author and the dates, but not the one picture the author chose for the piece — the field Google lists as recommended for articles. Emitted only when a featured image is set; nothing is borrowed from the body or the site identity to fill the gap.

1.47.0

• Fixed: on a site not written in the Latin alphabet, every page carrying a reported search was told its words answer nothing. The part that pulls the meaningful words out of a search reads Latin letters and numbers, so a search in Cyrillic, Chinese, Japanese, Arabic, Greek or Devanagari yielded nothing to look for — and that came out the far end as “none of it is on the page”. The whole site sat on the list permanently and no amount of writing could clear it. A search the check cannot read is now reported as exactly that: not measured, not a failure.
• Fixed: a URL somebody pasted into a search box, and somebody’s prompt pasted in whole, were judged against your pages like real searches. Both make rows no edit can ever close. Web addresses and anything far longer than a person types are now left out, alongside the machine probes that already were.
• New: you can set a search aside, not just a page. When a page does not answer the search it is found for, the page may be lacking — or the search was never a question your site should answer. Until now the only lever was to set the page aside, which takes good writing off the list to quiet a bad query. The row offers both now, and anything you set aside can be put back from the Search screen.
• New: an assistant asked to check one page gets the whole verdict for it. It graded only how readable a page was, so a page flagged for not answering its search came back spotless and an assistant read it as finished. It now answers both halves and says which of the searcher’s words the page never uses.
• New: a connected assistant can set a search aside too, one at a time. It refuses any search no engine has actually reported for your site, and putting one back is always allowed.
• Improved: the list of searches left out of the Search screen now says why each one was left out, and the figure reading “weren’t people” counts machine probes alone rather than everything excluded for any reason.
• Note: your site reads its pages once after updating, so a finding that was true all along may appear on your list.

1.46.0

• New: a connected assistant can read a page and change one passage of it instead of replacing the whole thing. It quotes the passage back exactly and that passage has to appear once — not found and found-twice are both refusals that change nothing, because a fix applied to a guess is how content gets mangled. The previous version is kept as a revision, and the saved page is read back and compared.
• New: an assistant can describe a picture that has no description — the commonest thing wrong with content on most sites, and the one thing it previously could not do.
• New: and it can describe the pictures inside a page, which is a different job. WordPress copies an image’s description into the page when you insert it and never looks at it again, so the words a reader gets from a published page are the ones written into that page — describing the library copy changes nothing a reader sees. There is a tool for each now, and each says which copy it writes.
• New: a way to list the pictures in your media library that carry no description at all. Search matches words, and the images needing description are exactly the ones with no words on them, so that question could not be asked before. A description that is only the file name counts as none, the same way the page checks count it.
• Fixed: the tool for describing a picture refused the exact pages the checks were flagging. Where a description was only the file name, the check asked for it to be replaced and the tool refused it as already described — and told the assistant the checks were satisfied. Those pages could be found and never closed. The checks and the tools that fix them now share one definition of what a description is.
• Fixed: an empty description on a captioned picture is now a finding. An empty description means “this is decoration” and that is still respected, but WordPress leaves the field empty by default, so a captioned screenshot could reach every screen reader undescribed while the row said every image had one. A caption is evidence the picture means something.
• Fixed: that new check would never have run on anything already published — a page is read again only when the set of checks changes, and a new rule inside an existing check did not register as one. Your site re-reads its pages once after updating.
• Fixed: captions come in three shapes and only one was understood — the block editor’s, the classic editor’s, and the way the same classic caption is drawn on any theme that has not opted into HTML5 markup. All three are read now.
• Fixed: an assistant was told “this page has no featured image, add one” even on kinds of content that cannot have one. The two absences are told apart now.
• Fixed: the two counts of your worklist could disagree, and the probe that reads what your theme does with a picture’s description had a gap in its filter.
• Improved: on the dashboard’s Today line, how each figure compares with yesterday is a small arrow at the end of its label rather than a coloured sentence beneath it. Hover it for the number; a figure that did not move says nothing at all.

1.45.0

• Fixed: a page could be told it does not answer a search it is no longer found for. That reading is measured in the background against the search the engines showed the page for at the time, and the engines move that search without anybody touching the page — so the reading beside it became an answer to a different question. Pages are read again when their search changes, and the old reading stays on screen, marked, until the new one replaces it.
• Fixed: a day Bing reported nothing about was drawn as a drop to zero pages in the index. Bing answers for every date in its window whether or not it holds numbers for that date, so a date with nothing in it was stored and drawn as a measurement — a fall to zero and back, on a site that never left the index. Those days now keep their place in the chart as a gap you can hover, and say what they are when opened.
• Fixed: for the same reason the “Pages in Bing’s index” figure above that chart could have read zero, and the warning about Bing’s crawler hitting errors could have gone quiet on a site that was still erroring. Both now read the most recent day Bing actually reported.
• Fixed: fresh numbers from Bing could be overwritten by an empty answer, because Bing re-sends its whole window on every poll. An empty answer can no longer overwrite a reading.
• Fixed: “Report an Issue” in the More menu had no icon, and the Report screen wore the flag drawn for it. Both have their own mark again.
• Improved: where two searches on a page are tied for busiest, the same one wins every time — the order could vary between readings, which meant the search a page was measured against could change on its own.

The most recent releases are listed here, back to 1.45.0. Every earlier one, back to 1.22.0, is kept in the full history in the same detail: https://github.com/heera/agentimus/blob/main/CHANGELOG.md