Skip to main content
November brings an email of the conversation for visitors who ask for one, files in a team reply on the website, WhatsApp, Messenger, and email, a public page that explains how to delete your SiteGPT account, a 6-digit sign-in code for when you read your email on another device, large files that train again, one 25 MB limit for each file, uploads and website links of PDF, Word, PowerPoint, and Excel files up to 1 GB, text recognition for scanned PDFs, also over 25 MB, chatbot transfers that the new owner accepts, the same link check for answers on every channel, answers that cite the page they used, a reason in plain words for more failed sources, CLI hints that fit the error and name the wait after a rate limit, and, in HIPAA workspaces, conversations and leads closed to the API and read-only in the iPhone app. This page covers changes up to November 16, 2026.

New

  • Visitors can now email a conversation to themselves. Turn on Let visitors email the conversation to themselves in Settings > General, or emailConversationToVisitor in API v2, or, from CLI 0.8.0 (November 16), sitegpt settings general update --email-conversation true. The chat widget then shows Email me this conversation under the latest answer. The visitor enters an address, and SiteGPT emails the conversation as it happened, with the source links unless Hide Sources is on, from your chatbot name followed by “via SiteGPT”. The widget says “Sent to” and the address when it went out, and Chat History shows a line for each email. The address is not saved as a lead. The setting is off by default, and is not available in HIPAA workspaces, where API v2 answers 403 HIPAA_CONVERSATION_EMAIL_DISABLED. See Let visitors email a conversation to themselves and Emailed conversations.
  • The public page Delete your account explains how to delete your SiteGPT account, what is deleted and what is kept, and how to ask SiteGPT by email to delete it. You can open it without signing in. See Delete your account.
  • The sign-in email now has a 6-digit code above the Sign in button, and the subject starts with the code, for example 123456 is your SiteGPT sign-in code. Type the code on the Please check your email page to sign in, for example when you read your email on your phone. The code and the link work one time, for 1 hour, and using one ends the other. The page button to send another email is now Send a new email. Invitation emails that your team sends do not change. See Sign in and The sign-in code does not work.
  • Your team can now send files with a reply from Chat History, for example a photo, a PDF, or a price list. Select the paperclip in the reply box, or drag files onto it. A reply can carry up to 10 files, and it can be files only. In your website chat, the files show under the reply: photos in the chat, other files as a card to open or download. On WhatsApp, the text goes first, then each file as its own message, for the file types that WhatsApp takes. Messenger and email followed on November 6 (see the next item). HIPAA workspaces and other channels, such as Slack, have no attach button. See Send files with a reply, Send files on WhatsApp, and Files in a team reply.
  • Your team can now send files with a reply in Messenger and email conversations too. On Messenger, the text goes first, then each file as its own message. Messenger takes any file type: JPEG, PNG, and GIF photos, MP4 videos, and MP3, M4A, and AAC audio play in the chat, and other files arrive as a file. By email, the files go as attachments of the reply email. Email takes photos, PDF, EPUB, Office and OpenDocument files, text, CSV, RTF, ZIP, contact cards, and calendar invites, and the files of one email can be up to 18 MB together. See Send files with a reply, Send files on Messenger, Send files by email, and Files in a team reply.
  • In the dashboard, a PDF, DOCX, PPTX, or XLSX file can now be up to 1 GB. Other files keep the 25 MB limit. The Upload Local Files form now says “PDF, DOCX, PPTX and XLSX files up to 1 GB, other formats (CSV, DOC, etc.) up to 25 MB each.” A larger file is left out with “NAME is over 1 GB. Split it into smaller files.”, and another file over 25 MB with “NAME is over 25 MB. Split it into smaller files, or save it as PDF, DOCX, PPTX or XLSX (these can be up to 1 GB).” SiteGPT reads a file over 25 MB without its images and other media. API v2 and the CLI keep 25 MB for each file. See Upload files and Size of a file that your chatbot learns from.
  • A website link to a file now trains as that file, not as a web page: PDF, Word, Excel, PowerPoint, RTF, and plain text files (.txt, .md). This works for links that you add, and for files that a crawl or a sitemap finds. A PDF, DOCX, PPTX, XLSX, or plain text file can be up to 1 GB when the site sends the size of the file, other files up to 25 MB. A large file can stay in Processing for a while, because SiteGPT downloads it in steps. Before, a link to a large PDF or a text file over 3 MB failed. If the site answers with a web page, or the file is over its limit, SiteGPT reads the link as a web page, as before. See Links to files.
  • SiteGPT now reads a PDF of scanned pages with text recognition (OCR), up to 25 MB, from uploads, connected apps, and website links. Before, a scanned PDF failed. A larger scanned PDF shows Failed with “This PDF is scanned pages without a text layer, and SiteGPT cannot read a scan this large yet.” (Later changed: from November 13, a larger scan is read a range of pages at a time. See the Improved section.) See What SiteGPT can read.

Improved

  • API v2 and the MCP settings tools no longer have the chatbot setting cssSelectorsToRemove. It had stopped changing what your pages train on. A request that still sends it gets 400 SETTING_RETIRED, with a message that points to excludeSelectors on the links (PATCH /api/v2/chatbots/{chatbotId}/documents/config). The CLI option --css-selector gets the same error, and CLI 0.7.0 removes it. In the dashboard, use HTML/CSS selectors to exclude on the links. Values that you saved stay, but SiteGPT does not use them. See Crawl options.
  • A chatbot transfer is now a request that the new owner accepts. sitegpt chatbots transfer and POST /api/v2/chatbots/{chatbotId}/transfer no longer move the chatbot at once, and SiteGPT no longer creates an account for the new owner. SiteGPT emails the new owner, who signs in and selects Accept or Decline under the new Pending Transfers on the Chatbots page, or answers with API v2 (GET /api/v2/transfers, then POST /api/v2/transfers/{transferId}/accept or /decline). A request stays open for 7 days, and a new request for the same chatbot replaces it. You can read or cancel the open request with GET and DELETE on the transfer endpoint. CLI 0.7.0 prints the open request. Older CLI versions make the request and then show an error, so update the CLI. See Transfer a chatbot to another owner.
  • A temporary chatbot from agent-first onboarding now stops after 50 AI answers until someone claims it. Before, it kept answering. The chat widget, for example on the onboarding page, then shows the usage notice, and API v2, the CLI, and MCP get 429 TRIAL_MESSAGE_LIMIT_REACHED. See Temporary chatbots.
  • Website, sitemap, link, and YouTube addresses must now be public http or https addresses. SiteGPT refuses localhost, private network addresses, and names such as .internal or .local, and adds nothing from a list that has one. API v2 and the CLI return 400 URL_NOT_PUBLIC with the address. See Addresses that SiteGPT refuses.
  • HTML in answers, custom responses, team replies, and the welcome message now shows only as plain formatting, in the dashboard, the chat widget, and the hosted chat page. Forms and buttons show only their text, form fields, scripts, and frames do not show, and style and class attributes are removed. See Add a custom response.
  • The tooltip above the chat icon now shows only simple formatting and links. Other HTML tags are removed and their text stays. A tooltip over 4,000 characters shows as plain text. See Change the name, welcome message, and tooltip.
  • One file can now be up to 25 MB in the dashboard, API v2, the CLI, and from connected apps, on every plan. The Upload Local Files form now says “up to 25 MB each” (it said 10 MB), and it leaves out a larger file when you choose it, with “NAME is over 25 MB. Split it into smaller files.” Before, a larger file uploaded in full and then failed. API v2 and the CLI now take files up to 25 MB, up from 20 MB, and refuse a larger file with 413 FILE_TOO_LARGE and the same message. Before, a file over 20 MB got a general 400 VALIDATION_FAILED. maxFileSize in GET /api/v2/me and GET /api/v2/account now shows this limit, 26,214,400 bytes, for every account. (Later changed: from November 11, PDF, DOCX, PPTX, and XLSX files can be up to 1 GB in the dashboard. See the New section.) See Size of a file that your chatbot learns from.
  • When Scrape Website or Add from Sitemap cannot add the address, the form now says why, for example that the address is not public, or how many minutes to wait after many imports. Before, these forms showed “Failed to add website. Please try again later or contact support if the issue persists.” (or “Failed to add sitemap. …”) for every refusal. See If adding the links fails.
  • Photos and files in Chat History now open through a link for each file that works for 1 hour: SiteGPT’s copies of WhatsApp files, files that your team sent, and email attachments. While the conversation is open, Chat History gets new links by itself. A link that you copy out of Chat History stops working after 1 hour, with “This link has expired. Open the conversation again.” Before, these links had no time limit. See Files in a team reply.
  • The install script now puts the email and signature of a signed-in visitor (set user:email) after # in the widget address, not after ?. Hosted chat links and iframe embeds take the values after # or after ?. If the widget cannot reach SiteGPT, or SiteGPT cannot finish the sign-in, the chat shows an error page with “The widget could not apply the signed-in user. Try again.” and opens no conversation. See Identity verification.
  • In a CSV export of conversations, the rows of one conversation now stay together. Before, the rows of several conversations could mix. See Export conversations to CSV.
  • A temporary chatbot from agent-first onboarding that nobody claims is now deleted after its 7 days to claim end and 30 more days pass, with its content, conversations, and leads, and its temporary account. Before, SiteGPT kept it. Claimed chatbots are not affected. See If nobody claims the chatbot.
  • For chatbots in HIPAA workspaces, API v2 no longer gives out or changes conversations, messages, or leads, for any caller: your own API tokens, the CLI, the SDKs, and MCP. These requests return 403 HIPAA_CONVERSATION_API_DISABLED, or 503 HIPAA_CHECK_UNAVAILABLE when SiteGPT cannot check the HIPAA status. API v0 conversation and message endpoints refuse the same way. Analytics, settings, content, tags, and the other chatbot endpoints still work. The MCP servers, and CLI versions before 0.6.0 that signed in with sitegpt login, mostly keep 403 HIPAA_AI_ASSISTANTS_DISABLED for conversations and leads, and an update of the CLI no longer makes conversation and lead commands work. Use the SiteGPT dashboard for conversations and leads. See HIPAA workspaces in the API, MCP server, and HIPAA_CONVERSATION_API_DISABLED.
  • In the iPhone app, chatbots in HIPAA workspaces still show conversations and leads, but replies, take-overs, and other changes to conversations and leads fail. Reply from Chat History in the dashboard. See HIPAA workspaces.
  • SiteGPT now turns on HIPAA processing for an account only when none of its chatbots has a connected channel, including an email address and Zendesk. See Integrations in HIPAA workspaces.
  • Uploads in the dashboard now keep going on slow and unsteady connections. The form shows how much of each file has gone, for example “48 of 100 MB”. A request stops only when no data reaches SiteGPT for 60 seconds. A file over 8 MB goes in parts, and each part is tried up to 5 times (not in a HIPAA workspace, where a file goes in one request). When the form gives up, it says why, for example “The upload stopped: no data reached SiteGPT for 60 seconds. Check your internet connection, then try again.” Before, a slow upload could fail after 2 minutes. See A file failed to upload.
  • A PDF over 25 MB that you upload or link to is now read a range of pages at a time when it cannot be read in one piece. A scanned PDF over 25 MB is now read with text recognition (OCR), and a PDF with a lot of text no longer fails with “This PDF holds more text than SiteGPT can read in one piece yet.” Such a PDF takes longer to train. It still fails if one page is too large, or if it has too many pages to read in time, with “This PDF has more pages than SiteGPT can read at once. Split it into smaller PDFs and add those.” Another file whose text is too large to read in one piece, of any size, now shows Failed with “This file holds more text than SiteGPT can read in one piece yet.” See Files over 25 MB.
  • When a site is busy, answers with a server error, or does not answer the download of a linked file, for example after many downloads, and reading the link as a web page fails too, SiteGPT now tries the download again a minute later, up to 3 times in all. A scanned PDF cannot be read as a web page, so before, such a link failed. If the download still fails, the link shows Failed with “This PDF is a scan, which SiteGPT reads from the file itself, but the site did not let SiteGPT download the file. Resync it later, or upload the PDF as a file.” or “The site did not let SiteGPT download this file, and reading the link as a web page failed too. Resync it later.” See Links to files.
  • Very large files and links now train in steps, and SiteGPT checks your page usage again before each step. If other content used up your pages in the meantime, the document stops with Limit Reached. See What happens at the page limit.
  • Answers on every channel now get the same link check as the chat widget: WhatsApp, Messenger, the other chat integrations, email, Zendesk, and API v0 and API v2, and so the CLI and MCP. A link or image in an answer stays only when its address is in your content, a custom response, a text snippet, your instructions, or an earlier answer or team reply. An address that the visitor typed no longer counts, also in the widget. The short message that the AI writes when it hands a conversation to your team gets the same check, and a link in your own hand-off message counts there. See What visitors see with the answer.
  • A failed page or file now says why for more reasons, in the dashboard, API v2, the CLI, MCP, and the export of the document list: an upload that was canceled, did not finish within a day, or did not arrive completely, a web page over 3 MB, a plain text file that is not UTF-8, a Google Docs, Sheets, or Slides file over Google’s 10 MB export limit, and an edit that was not saved completely. Before, these showed “Processing failed. Please try syncing again or contact support.”, and an edit that was not saved completely showed that the page could not be found. The not-found message is now “The page could not be found. Check that the link still works.” API v2 and the CLI now give the same reason and code as the dashboard, not the stored error text. See A file shows Failed and A website import failed.
  • A page or file that stopped processing every time SiteGPT tried it again now shows “Processing stopped before it finished several times, so SiteGPT stopped retrying it. The page or file may be too large to train. Resync it to try again.” See Fix failed links.
  • Update Content now checks the 8 MB limit before it saves, with “This content is over 8 MB, the most SiteGPT saves from the editor. Save it as a file and upload the file instead.” Before, it showed “Something went wrong — please try again.” See Text that you edit.
  • Form messages on Scrape Website, Add from Sitemap, Add Multiple Links, Add YouTube Videos, and the GitHub forms are now plain sentences, for example “Add at least one link, such as https://example.com/pricing” instead of “urls: Add at least one link, …”. Add YouTube Videos now says why it could not add links, for example “You added links many times in a short while. Try again in N minutes.” See If adding the links fails.
  • When you paste two or more links into Add Multiple Links and every line ends in a comma or semicolon, for example a column saved as CSV, the form now removes those commas and semicolons. See How the form reads a pasted list.
  • Web pages now train with the web address of images that load only when a visitor scrolls, so answers can show these pictures. Resync pages trained before to get the addresses. See What a crawl does not do.
  • The Source column of the document list, its export, and Auto Sync Jobs now use the dashboard’s names for sources, for example Local files and Multiple links.
  • The Unanswered questions panel in Analytics no longer lists one-word replies. For conversations that SiteGPT classifies from November 14, it lists only questions that the visitor asked, never the chatbot’s own questions. See Fix unanswered questions.
  • The plan message for API tokens, the CLI, and MCP now reads “API tokens, the CLI and MCP need the Growth plan or above. Upgrade your plan, or ask the owner of a chatbot you work on to upgrade theirs.” The MCP approval page now shows this message with an Upgrade your plan button, instead of a “403 Access Forbidden” page, and the CLI sign-in page shows the button too. On the Agents page, Specific chatbots lists only the chatbots that a plan covers. See Which plan applies.
  • The two button labels of Let the Visitor Choose must now differ in more than capital letters, spaces, or end punctuation. See Let the visitor choose the AI or a person.
  • In API v2, the CLI, and MCP, the chat-mode settings now return visitorsAsked, which says whether a new conversation starts by asking the visitor. rateLimits.thread.enabled now reads false when the chatbot owner’s plan no longer includes the message limit. The API reference now lists question.forwarded, which is true for a WhatsApp message that the visitor forwarded. See Set up human support.
  • The Messenger row under Webhook verification now checks a new app secret with Facebook before it saves it, so a wrong secret can no longer stop the chatbot from answering. The Messenger dialog now says Webhook verification is in Log only mode in that mode, not that verification is off. See The Webhook verification section.
  • The sign-in email now says to enter the code on the “Please check your email” page, and that page keeps the code box for the full hour that the code works, not 10 minutes. See Sign in.
  • An account on a custom plan can now have a retention period in its contract. SiteGPT then deletes conversations and leads by itself after that period. See Custom retention period.
  • When a chatbot is transferred to an account with a custom retention period, and the chatbot has conversations or leads older than that period, the new owner now confirms their deletion before the chatbot moves. The Accept dialog says, for example, “Your account deletes conversations older than 30 days. This chatbot has records that old, and they will be deleted once you own it.” In API v2, GET /api/v2/transfers returns retention, and the accept needs { "confirmRetention": true }, or it fails with 409 RETENTION_CONFIRMATION_REQUIRED. Before, these records were deleted after the move without a warning. See If your account has a custom retention period.
  • When the chatbot owner has no active plan, the legacy API (v0 and v1), API v2, the CLI, MCP, and the content forms of the dashboard now say “The chatbot owner’s account has no active plan. The owner can choose a plan on the Billing page.” Before, they said “This account has no active plan”, in three versions that did not say what to do. See Billing.
  • A setting that the chatbot owner’s plan or add-ons do not include now says why under the setting, with a link to the Billing page (no link on a white-label dashboard). On Settings > General, the note under Limit Messages Per Conversation says “The chatbot owner’s plan does not include a message limit per conversation. The chatbot owner can upgrade the plan on the Billing page.” On Appearance, one note above Watermark Text says “Watermark settings need the Remove ‘Powered by SiteGPT’ add-on on the chatbot owner’s account. The chatbot owner can add it on the Billing page.” Before, a lock button opened a browser alert with other words. API v2 and the CLI give the same text. See Limit messages per conversation and Remove the “Powered by SiteGPT” branding.
  • Links, source links, follow-up suggestions, the active tab, and focus outlines in the chat widget now read on any brand color. When your Primary Color or Link Color is too light for the chat background, or too dark for dark mode, the widget makes it only as much darker or lighter as it needs, and keeps its hue. Before, a white or very light brand color hid them. Buttons and message bubbles keep your exact color. The lead form button and the unread count now use your Text Color, like the other buttons. Before, their text was always white. A hex code without # now works in the chat window too. If you set --chat-color with the widget SDK, the active tab, focus outlines, and the text cursor now follow --chat-accent-color. See Change colors.
  • The Appearance page and API v2 now show the size and place that the widget uses. A distance, font size, or height saved as 0 uses the default, as the widget always did. The distance fields and GET /api/v2/chatbots/{chatbotId}/settings/appearance now show that default, not 0. A mobile distance that follows the desktop distance reads null in API v2. The Font Size (in px) placeholder now shows the size that the widget uses, and the preview uses the default height. See Move the chat icon away from other buttons.
  • When your role does not allow a request in API v2, the CLI, or MCP, the message now says what your role cannot do in the words of the dashboard, for example “remove Super Admins”, and names someone who can change your role. For actions that only Super Admins can do, such as a chatbot transfer, it says “Only a Super Admin can. The chatbot owner or another Super Admin can change your role.” The transfer message now says “transfer this chatbot, or see or cancel its transfer”, not “change chatbot transfer”. See Authentication.
  • Smart follow-up suggestions in API v2, the chat integrations, and Zendesk now take as much of the conversation into account as in the chat widget.
  • The API v2 reference now lists the fields of every settings section, with their types, allowed values, and ranges: what a PATCH accepts and what GET returns. The MCP search tool finds them too. The settings object of a chatbot now says that it is a summary, and where to read every setting. See Settings updates.
  • SiteGPT now asks answers to cite the page they came from, and Sources lists only the pages that the answer used. The AI is asked to link the words of a fact to its web page, for every chatbot, unless Hide Sources is on. Under the answer, the chat widget, Chat History, the chat integrations, API v0 and API v2 (sources), webhooks, and the email of a conversation list the pages that the answer links, and a page that holds a link that the answer gives. Before, the list showed every page that the search found, also under “I don’t know” answers. An answer with your fallback message now lists none. See What visitors see with the answer and Message sources.
  • Chat History now shows the files that matched the question under Sources, as Files that matched the question (only your team sees these), with their names and no link. API v2 and webhooks have them in sourceFiles. The chat widget never shows file names to visitors, and SiteGPT does not give the AI the names of files. See Read a conversation and Sources of an answer.
  • SiteGPT CLI 0.8.0 is available. After a refusal whose message says why, for example your plan, your role, an account with no active plan, a message limit, or a HIPAA workspace (403), the CLI no longer adds a hint that does not fit, such as sitegpt login after a plan refusal. After a rate limit, the hint now names the wait that SiteGPT asks for, for example “Rate limited. Wait 30 seconds and retry.”, or a minute when SiteGPT sends no wait. Before, it said to wait a few seconds. The new option sitegpt settings general update --email-conversation turns Let visitors email the conversation to themselves on or off. See Hints after an API error, Settings: general, and Upgrade the CLI.

Fixed

  • Large files train again. A file up to 25 MB that you upload, or that comes from Google Drive, Dropbox, OneDrive, or Box, now trains. Google Docs, Sheets, and Slides have a smaller limit that Google sets. Before, some large files failed on every try and every resync. A file over 25 MB now fails within seconds. (Later changed: from November 11, PDF, DOCX, PPTX, and XLSX uploads and links to these files can be larger. See the New section.) See Size of a file that your chatbot learns from.
  • From November 6 to November 7, files from Google Drive could fail to sync, and a resync failed again. For a few hours on November 6, files from other connected apps, and uploaded text or CSV files over 8 MB, could also fail to sync. On November 7, SiteGPT changed how it downloads files from Google Drive to fix this. If a file shows Failed, resync it. See Keep connected files up to date.
  • When the connection to Google Drive, Dropbox, OneDrive, Box, or SharePoint drops before the app answers, SiteGPT now tries the download again by itself. Before, the file showed Failed at once. See A file shows Failed.
  • Team members can use API v2, the CLI, MCP, and the SDKs again on the chatbots they work on, when the chatbot owner’s plan is Growth or above. Since September, SiteGPT checked only the plan of the person who created the token, so a team member without their own plan got 403 PLAN_UPGRADE_REQUIRED. Now your own plan or the chatbot owner’s plan counts. On an owner’s plan, the chatbot list shows only the chatbots that the plan covers, and creating a chatbot still needs your own plan. See Which plan applies.
  • In Chat History, the chat button of the dashboard preview no longer covers the send button of the reply box. It does not show while a conversation is open. It still shows on the other dashboard pages, including the conversation list.
  • Requests that arrive at the same moment no longer go past a limit. Invites and direct adds of team members stop at the member limit, and a direct add that gets no seat creates no account. Several take-overs of one conversation, or several switches back to the AI, add one notice for the visitor. Claims of one temporary onboarding chatbot give one checkout. Through API v2, taking over a conversation that your team already holds, or switching one that the AI already answers, adds no notice: tookOver or switched is then false. See Conversation state conflicts.
  • A widget color saved as a hex code without #, for example 1D4ED8, now shows on the chat icon and the tooltip. Before, the browser ignored it there, so the icon and the tooltip lost that color. See Change colors.
  • When API v2 refuses to add a website, sitemap, or links, it now answers with the real status and code, for example 400 URL_NOT_PUBLIC. Before, these refusals came back as 502 KNOWLEDGE_INGEST_FAILED and “try again”. See Errors from content operations.
  • A file over 25 MB now shows Failed with “This file is larger than 25 MB, the most SiteGPT can read from one file. Split it into smaller files and add them again.” Before, the badge said only “Processing failed. Please try syncing again or contact support.”, and sometimes “The page could not be found (404).” when the file size had the digits 404 in it. Google Docs, Sheets, and Slides over Google’s 10 MB export limit still show the general message. See A file shows Failed.
  • When no pages were left in your quota during an upload, the Upload Local Files form now lists the file as failed with “Quota exceeded. You can only add N more pages.” Before, it counted the file as uploaded, but SiteGPT had not added it. See Upload files.
  • Editing a document’s text with Update Content, or with API v2, the CLI, or MCP, no longer fails when the new text is over about 500 KB. In the dashboard, the text can be up to 8 MB. With API v2, the CLI, and MCP, it can be up to 1,000,000 characters. Before, a large edit failed with “Something went wrong”. See Text that you edit.
  • A sitemap whose addresses have & in them now adds the real addresses, with &. Before, SiteGPT added the addresses with &, which the site often answered with “page not found”. If your Links List has such links, delete them.
  • A document whose text contains <|endoftext|> now trains. Before, it showed Failed.
  • A file upload that stopped before all of the file arrived now shows Failed in Files List. Before, it could stay queued with nothing stored.
  • A message that is only a link, for example a link that your team sends alone in a reply, now looks like a link in Chat History and in the chat widget. Before, it looked like plain text, although you could click it.
  • An image in an answer whose address SiteGPT did not give the AI now shows as its caption in plain text. Before, the chat widget and the Appearance preview showed ”!” and the caption, for example when a visitor pasted an image into a question. Images from your content, such as a custom response, show as pictures. If nothing is left of an answer, the visitor sees your fallback message. See What visitors see with the answer.
  • Creating, renaming, and deleting tags with API v2, the CLI, and MCP now works for every team role. Before, these requests failed with 404 CHATBOT_NOT_FOUND. Deleting a tag now needs the conversations:write scope, not conversations:delete. See Scopes.
  • Rotating an API token now fails with 403 PLAN_UPGRADE_REQUIRED and keeps the token when no plan covers you, or when no plan covers one of the token’s chatbots anymore. Before, the first case failed with a server error, and in the second case the token rotated. The Agents page shows the plan message, not “Server Error”.
  • A file link that the site refused now shows its own message, also when reading it as a web page hit your site’s error page. Before, it showed your website’s error. See Links to files.
  • Chat History and the chat widget no longer show an empty Sources item, for example under answers from API v2 and chat integrations.
  • When connecting Messenger, Freshdesk Messaging, Google Chat, or Zoho SalesIQ fails, or disconnecting one of these or Slack or Crisp, SiteGPT now says “Could not connect. Check what you entered and try again.” or “Could not disconnect. Try again.” (Later changed: from November 14, a refusal with a reason shows that reason. See the last items of this section.) See Where to find integrations.
  • When Facebook refuses the welcome message after Get Started, Chat History now shows “Not delivered to Messenger”. See Messages that Facebook refuses.
  • Email history now opens a file that your team sent with a reply, and names it as the email did. An email that may have gone out now stays in Email history as Not confirmed after you resolve the conversation or discard the draft. The Email card shows Removal stopped when removing email stops with an error, not Removing. See Send files by email.
  • WhatsApp voice notes and other media whose type has extra details now show in Chat History, and SiteGPT keeps a copy.
  • The Leads page now shows the number of leads and Export CSV for chatbots with more than 2,000 leads. Above 10,000, the count shows “10,000+”. See Find a lead.
  • The v0 API reference now lists the 429 answers of the send-message endpoint, MESSAGE_LIMIT_REACHED and TRIAL_MESSAGE_LIMIT_REACHED. See Errors.
  • SiteGPT’s session cookie no longer keeps the email address of an email-link sign-in from before July 2026. The next page load removes it.
  • After an invitation button that no longer works, the Can’t find the email? help on the Please check your email page now shows the invitation email that SiteGPT sent, not the sign-in email.
  • API v2 now accepts a settings section that you send back as you read it, with one field changed. Before, these sections refused it: general (a null message limit, or the limit on a plan without it), appearance (the watermark and call to action fields on an account without them), user-data, lead-form, and webhooks (fields that you can only read), localization (an empty object), and chat-mode (button labels saved before November 14). A request that changes nothing now saves nothing and gets the warning “No setting changed. Every field sent has the value that is saved.” updatedSections lists only the sections that changed. This also applies to the CLI and MCP. See Settings updates.
  • When API v2 refuses a change to the message limit, the watermark, or the call to action button, the message now says why and who can act, for example “The chatbot owner’s plan does not include a message limit per conversation. The chatbot owner can upgrade the plan on the Billing page.” See Settings updates.
  • API v2 now shows gpt-4.1 in the general settings of a chatbot with no saved model, the model that it answers with. Before, it showed gpt-4.1-mini. See Choose a model.
  • A failed run of a sync job now gives its reason in the same words and code as the Failed badge of a source: the Failed status of the run in the job details, and latestIngestJobRun.errorMessage in API v2. The setup checklist of agent-first onboarding in API v2 and MCP does the same for failed sources. Before, these showed the stored error text. See Check that it works.
  • Where the dashboard shows a source’s status outside Links List and Files List, for example when you pick files from a connected app or GitHub, and in the onboarding preview, Limit Reached now starts with “Your plan’s page limit was reached while processing this source.” Before, it could read “Knowledge page limit reached. Processing failed. Please try syncing again or contact support.” (Later changed: from November 15, Links List and Files List show the same text. See the last items of this section.)
  • When the server refuses a save on a settings page, the page now shows the reason when SiteGPT gives one, for example “The two button labels must differ in more than capital letters, spaces or end punctuation.” on Chat Modes when two people change the labels at the same time. Before, it showed a general error, and the label check showed “Server Error”. Other failures across the dashboard now say “Something went wrong. Please try again.”
  • When connecting or disconnecting a chat integration is refused for a reason, the integration page now shows the reason, for example “The chatbot owner’s plan does not include this integration. The owner can upgrade on the Billing page.”, the HIPAA workspace, or “This integration is already connected. Reload the page to see it.” Before, the integration pages, Messenger included, said “Could not connect. Check what you entered and try again.” or “Could not disconnect. Try again.” for every reason. See Where to find integrations.
  • Saving a row under Webhook verification on Enforce without its credential now says, for every tool except Messenger, “Enter the verification credential before you choose Enforce. Without it, Enforce refuses every message.” Before, it said “Something went wrong. Please try again.” A saved credential now shows (saved) after its label. See Check webhook calls.
  • Messenger: a save in the Messenger row that stays on Enforce and keeps the saved app secret, no longer checks the secret with Facebook again, so it no longer fails when Facebook is slow. Turn on verification in the Messenger dialog now uses an app secret that is already saved, so you can leave App secret blank. See Turn on verification.
  • API v2 now answers 403 ROLE_PERMISSION_REQUIRED when you are a team member of a chatbot and your role does not allow the request, with a message such as “Your role on this chatbot cannot start conversations. The chatbot owner or an admin can change your role.” Before, it answered 404 CHATBOT_NOT_FOUND, for example when an Agent’s token started a conversation. This also applies to the CLI and MCP. A chatbot that you cannot reach still answers 404. See Authentication.
  • Saving Settings > General of a chatbot with no saved model now keeps GPT-4.1, the model that the chatbot answers with. Before, GPT Model showed GPT-4.1 Mini for such a chatbot, and saving the page for another reason moved it to GPT-4.1 Mini. A v0 API call to the general settings without gptModel now keeps the saved model too. Before, it saved GPT-4.1 Mini. See Choose a model.
  • The API v2 reference now describes the preview of each setup checklist item of agent-first onboarding, as SiteGPT returns it. Before, the reference did not allow the field, so clients that check answers against it refused them. See Agent-first onboarding.
  • The Let the Visitor Choose button labels are now checked on Settings > Chat Modes only when you change one. Before, a chatbot whose labels were saved before November 14 and differ only in capital letters, spaces, or end punctuation could not save other changes on this page. API v2 already worked this way. See Let the visitor choose the AI or a person.
  • In Links List and Files List, Limit Reached now shows “Your plan’s page limit was reached while processing this source.”, the same text as everywhere else. Before, it showed “Knowledge page limit reached. Upgrade your plan to add more pages or delete some.” See Fix failed links.
  • On WhatsApp, Freshdesk Messaging, Zoho SalesIQ, and Zendesk, a link in an answer now shows as its words followed by the address in brackets. Before, these channels showed the link as raw Markdown, such as [form](https://example.com/form). On Messenger, such a link now shows its words too, not only the address. See What visitors see with the answer.
  • On Messenger, an AI answer over 2,000 characters, with its sources, now goes as several messages. Before, Facebook refused it. On Messenger and WhatsApp, when someone on your team takes over or resolves the conversation while a long answer goes out in parts, the rest is not sent. See Messenger and Chat message length.
  • In the chat widget and Chat History, an underscore inside a word, as in snake_case, no longer turns text into italics, and a link whose address has _ or * in it opens the full address.
  • In API v0, a team member’s change to a chatbot that their role allows now saves, for example a change to the settings, the appearance, conversation starters, or follow-up suggestions. Before, it failed with a server error (500) for everyone except the chatbot owner. A request that the role does not allow still fails with 403. See Legacy API v0.