
Sources
Sources are where the AI draws ideas: websites, RSS, Telegram, Reddit, forums.
1✨ Auto-discover suitable sources
At the top of the section there's an “Automatically find suitable sources” checkbox. When on, the service picks new sources for your project topic on its own:
- 🎯 favours PRIMARY sources — official sites, original feeds and channels rather than aggregators and reposts;
- ✅ checks each link for reachability and adds only the working ones — within your source limit;
- 🔁 runs periodically, and also when a source is archived for low effectiveness — a replacement fills the freed slot;
- 🔗 hunts for primary sources via links in your posts: it opens the material, finds external links (to a study, a journal, a Telegram channel) and checks whether such a site can itself be a source. For websites it looks for an RSS/Atom feed both on the page and on the home page; Telegram channels and subreddits are added as the matching type.
Every addition is posted to “Notifications” for all project members. Picks rely on the project's “Marketing context” — the more precise it is, the more relevant the sources. The feature makes an AI request (spends tokens); if a source doesn't fit, just disable it.
2🗄️ Source archive (tab)
The “Sources” section has two tabs: “Sources” — active (enabled) ones you work with, and “Archive” — disabled sources. When you turn a source off, or the service auto-disables a silent (irrelevant) one, it isn't deleted but moved to the “Archive” so it doesn't clutter the main list. In the archive each source still shows its stats; expand the card and enable it to bring it back to active. To avoid enabling them one by one, the archive has a “Re-crawl all” button at the top — it returns every disabled source to work at once and restarts their crawl (within the active-source limit; if there's less room, the most recently checked ones come back first). The tabs show source counts.
3Website source: a feed/section page or a single article
A “website” source can point either at a single article or at a feed/section page (a news section, …/category/health/). The service tells these apart:
- 📄 a link to a specific article — taken as one article: title and gist;
- 🗂️ a feed page — the service follows the links into several recent articles (up to 5 per sweep) and takes material from each separately. Dated articles are prioritized (fresher); service links (tags, authors, login pages) and images are skipped.
For a feed page the service also looks once for the site's RSS/Atom feed and, if found, reads the source through it from then on — more precise and cheaper (clean per-article entries).
💡 Where an RSS feed exists, it's easier to add it directly (the “RSS” type). Auto-discovery can find RSS on its own too.
4🔎 Relevance check when adding
As soon as you paste a link into the add form, the service quickly estimates how well the source fits the project topic: it reads a few recent items and compares them with the “Marketing context”. A coloured indicator appears under the field — from “great fit” to “barely on-topic” — with a short note. If the source is unreachable (won't open or blocks requests), that's reported too. This helps weed out clearly irrelevant or non-working feeds up front.
5🖼 Source icons
Each source gets its own icon — the site's favicon or a Telegram/Reddit channel avatar. It's shown next to the name in the list and on the “Your AI content-marketing pipeline” diagram on the “Overview”. Icons are fetched automatically while crawling the source and shared across identical sources in different projects — nothing to set up. If an icon can't be fetched, the source-type emoji stays next to the name.
6What it is and how often
Monitoring is a periodic sweep of your active sources. On a schedule (about every 30 minutes by default) and via the “Generate drafts” button, the service checks whether new material has appeared. It polls the sources rather than receiving instant push notifications: an item is “new” if the service hasn’t processed that link before.
7Smart (adaptive) polling
To avoid overloading either the sources or the models, the service tunes the polling frequency per source. The minimum interval is once every 30 minutes (never more often). A source that regularly yields fresh material stays at that floor; if several consecutive sweeps find nothing new (or the source is unreachable), its interval automatically doubles — 30 min → 1 h → 2 h … up to 6 hours for “sleeping” sources. As soon as material appears again, the frequency returns to 30 minutes. The “Generate drafts” button still runs on demand at any time. So frequently updated feeds are checked more often, rare ones less, and no wasteful model runs happen.
8One source in several projects — no duplicate fetches
If the same source (same link) is added to several of your projects, the service polls it once per sweep and reuses the result for all projects that have it — instead of fetching it separately per project. This avoids duplicate network load on the source and wasted resources. Relevance selection, deduplication and effectiveness are still computed per project — each project takes what fits its own topic from the shared material.
9🧹 Auto-deduplication of sources
Sometimes both a website and its RSS feed end up in your sources — or the same address twice (by mistake or via auto-discovery). They yield identical content and waste sweeps. The service periodically (about every 12 hours) finds such duplicates and disables the less suitable one in a pair:
- 🔁 a duplicate is a source with the same address, or when a website's RSS feed matches a separately added feed;
- 🏆 the most effective/reliable one is kept (by on-topic post count, freshness and lack of failures; ties favour the canonical RSS feed);
- 🗄️ disabled ones go to the “Archive” — you can re-enable them;
- 🔔 each such optimization notifies project members with a list of what was disabled and what was kept.
This works regardless of how the sources were added — manually or by auto-discovery.
10Slowing down and auto-archiving silent sources
The service doesn't spend tokens on sources that consistently bring off-topic material. It tracks how long it's been since a source last produced a RELEVANT post (one that entered the queue) and gradually slows down its checks:
- ⏳ over a day with nothing on-topic — the interval grows to 6 hours;
- 🐌 over 2 days — to 24 hours;
- ⏸️ over 3 days — a “probation”: the source is NOT disabled yet, but rechecked once a day for about another week (in case on-topic material shows up), and a note appears in “Notifications”;
- ⛔ around 10 days with no on-topic posts (3 days + a week of rechecking) — if nothing turned up, the source is finally disabled and moved to the “Archive” with a suggestion to check or re-enable it.
As soon as the source yields a relevant post again, the frequency returns to 30 minutes and the probation resets. When you re-enable it manually, the countdown restarts. New sources get a 3-day warm-up.
11⭐ Authoritative source (reference)
You can mark a source as “Authoritative” (a checkbox when adding, or the “☆ Make authoritative” button in the card). It's a source where moderators put reference content to shape the project's desired format. Special rules apply:
- 🔁 never drops out of rotation — always polled, never slowed down or archived, always scores high;
- 📝 every item becomes a post — the relevance threshold and dedup don't apply to it (ALL new items are processed);
- 🧭 for each item the service adapts the project's “Marketing context” — extends it so the item's topic becomes a natural part of the project, and generates the post in the updated context. This way the project gradually adapts to the reference source. A notification is sent when the context changes.
⚠️ Use deliberately: an authoritative source directly changes the project's “Marketing context”. Put in it only the content you actually want the project to adapt to.
12Hourly activity chart + auto-polling
The expanded source card has a 24-hour activity chart — it shows the hours when new material usually appears in the source (hours are counted in the project timezone from “Project info”). As data accumulates, the service starts polling “dead” hours (when the source consistently has nothing) less often — saving tokens and requests without missing fresh content. To avoid going blind if the source changes its schedule, such hours are still re-checked at least every 6 hours. This is a finer, per-source setting on top of the whole-project “quiet hours”.
13Is the source alive + the “Analyze” button
The source details show whether there was new material at the last check and how much — this indicates whether the source is “alive” (green badge — there's new content, yellow — nothing). If a source's effectiveness declines (it hasn't produced on-topic posts for a while), the service not only polls it less often but also sends a notification at each decline stage, recommending you review the source or refine the project context. The “✨ Analyze” button compares the source's fresh material with your project topic and gives a verdict (good fit / partial / poor) plus a concrete recommendation: keep the source, refine the “Marketing context” for it (what exactly to add), or replace it with a more on-topic one. Under the recommendation there's an “✨ Apply recommendation to project” button — it appends the recommendation to the project's “Marketing context” and IMMEDIATELY optimizes the context with the model (removing duplicates and fluff, compacting tokens) so it doesn't bloat; the saving applies to every subsequent generation. Both features use an AI request (spend tokens).
14🔬 Sweep diagnostics: why no new posts
A source can be “alive” and on-topic yet produce no new posts in the plan — because between “new material appeared” and “a draft in the plan” there are several filters. The expanded source card has a “🔬 Last sweep diagnostics” block — it shows the funnel for the last sweep: how many items were fetched and how many were filtered out at each step:
- duplicate by link — an item with this link was already processed (this also covers items the model previously deemed “off-topic”: their links are remembered so the same material isn't sent to the model again, saving tokens);
- semantic duplicate — the topic was already covered by one of the previous posts (repeat protection);
- below relevance threshold — the specific item weakly matches the project topic (the source is on-topic “on average”, but this item isn't);
- over the daily quota — the item is on-topic, but the daily post limit is already reached;
- off-topic (model) — rejected during text writing.
At the bottom — how many posts were actually created. If there are none, the block hints at the main reason. It's a snapshot of the last sweep only (updated on every source check). Common fixes: refine the “Marketing context” (raises relevance of the items you want), increase “Generation limits”/the weekly budget, or replace the source if it consistently brings duplicates or off-topic material.
15⭐ Primary-source score (stars)
The service builds a “link graph” among your sources: if OTHER of your sources link to a source (in their material), it's likely a primary source with original information — its score grows and is shown as stars (★ up to ★★★) next to the name and in the card. Conversely, a source that links out a lot but isn't linked to is flagged as a possible aggregator/repost. Primary sources rank higher, reposts lower, so quality channels stand out. Scores are recomputed periodically from recent posts' material.
16🧭 Feed quality: no repeats, no trivia
So the feed doesn't loop on the same thing or fill with empty hooks, the service (configurable in “Project”):
- 🧠 extracts the CORE MESSAGE of each item (not the headline, but the gist “who asserts what”) and won't post it if the same message already went out — even in different words or from another source. The message memory is a window in days (default 14; configurable);
- 🎯 judges the item's SIGNIFICANCE and drops the ephemeral, keeping the important. The key question: does the reader learn something they can act on, or something that changes their understanding? News about the LIFE OF AN ORGANISATION — appointments and departures, awards, anniversaries, event announcements, partnerships, grants — counts as trivia even when the organisation is respected and the person appointed works on an interesting field: a hint at a topic is not a finding. What matters: discoveries, research results, launches, changes of rules or practice. Strictness — the “Significance filter”: off / moderate (drop trivia) / strict (important only).
Both show in the source funnel diagnostics (“message duplicate”, “dropped by significance”). Authoritative sources BYPASS the filters (every item → a post). Assessment is one cheap AI call per sweep.
17Why a source produced no posts
The source details show a “material → post” funnel: how many items were found and at which step they were dropped. Steps run from cheap to expensive — the service doesn't pay to write a post for material that won't be published anyway:
- “duplicate link” — this item was already processed;
- “semantic duplicate” and “duplicate message” — the same thing has already been posted, even if worded differently;
- “below topic relevance” — the item is far from the project's topic;
- “off the project's topic (quick check)” — a fast AI check made before writing the post;
- “dropped by the significance filter” — the item carries no useful information (see “Feed quality”);
- “didn't fit the daily post quota” — the item is fine but the day's limit is used up;
- “the model judged it off-topic” — the final decision while writing;
- “the model didn't answer — no post” — the model was unavailable. The item is NOT marked as processed: we'll try again on the next pass. The service never publishes boilerplate filler instead of a post.
If there are no posts but “found” is high, look at which step loses the most: that tells you whether to change the source or refine the project's topic.
18A disabled source and adding it again
Disabling a source means “take it out of service”, not “delete it”: the record stays in the archive together with its statistics (how many posts it produced, how many tokens it spent). So:
- if you add the same address again, the source RETURNS from the archive with its history instead of appearing twice. Its failure counter and polling interval are reset — the source starts fresh;
- returning it takes an active slot, so when the source limit is used up the service will ask you to disable something first;
- addresses are compared ignoring “http/https”, “www” and a trailing slash — the same page is recognised in any spelling.
Duplicates accumulated earlier are cleaned from the archive automatically (once a day). Only disabled records are removed: active and reference sources are never touched, and published posts don't depend on the source and stay where they are.
19🔬 Scientific sources (for researchers and students)
If the topic is scientific or educational, “Recommend sources” and auto-discovery suggest specialized primary sources with feeds: arXiv (per-section RSS), PubMed/PMC, bioRxiv and medRxiv, PLOS, Nature, journal RSS (ScienceDirect, etc.), ScienceDaily, EurekAlert!, DOAJ and subject-specific scholarly societies for your topic. This gives first-hand access to current research. Note: some scientific sites expose only abstracts or require sign-in — such feeds may read partially; prefer resources with open RSS/preprints.
20© Copyright and risks
The service does not copy others' content verbatim: the model rewrites material IN ITS OWN WORDS, conveying facts and meaning (facts and news as such aren't copyrightable — the expression is). Long verbatim fragments are avoided; short quotes are allowed with attribution. Images are generated anew by AI, not taken from the source. How to reduce risk on your side: don't configure sources for verbatim reposting; add a link to the primary source where appropriate; don't republish paid/closed materials or licensed photos in full; remember that you approve and publish the final post — ultimate responsibility for content lies with the project owner.
21⏱ Parse time and auto-disabling unreachable sources
On every check the service measures how long parsing the source took (shown in the card: “⏱ Parse: N s”). If parsing takes longer than 20 seconds, the source is considered slow: its effectiveness is lowered (it sinks in the sort order) and its polling frequency is reduced — so a heavy source doesn't slow down the whole sweep. If the source doesn't respond (timeout) or returns an error (HTTP/network), the request counts as failed; after 3 consecutive failed requests the source is auto-disabled and a notification with the reason (timeout/response error) lands in the feed. Any successful request resets the failure counter; so does manually re-enabling. Check the link and the source's availability, then enable it again.
22Token efficiency (posts / 1k tokens)
The source details show how many on-topic posts it produced and how many tokens were spent on them (text rewrite + image generation), plus a “posts per 1000 tokens” metric. This helps compare sources not just by post count but by “cost”: a source that yields many relevant posts at low token spend is more cost-effective. Low efficiency (few posts for a lot of tokens) is a reason to replace the source with a more on-topic one.
23Source effectiveness (three dots)
Each source shows three dots — how often it actually produces posts ON the project's TOPIC (not just any new material). Three green — effective (an on-topic post almost daily); two green — medium (every couple of days); one green — low (about once every 3 days); three red — ineffective or disabled. Grey dots mean the source hasn't produced a single on-topic post yet (new or irrelevant) — the rating appears after the first real post. Important: green dots only appear when posts were actually created from the source; if there are none in “Post queue”, the source will show grey or red dots. Hover for a tooltip. The service uses the same signal to save tokens: effective sources are polled often, weak ones less, silent ones get disabled. Persistently red/grey is a reason to replace the source with a more relevant one or remove it and add a new one.
24Recommend sources (model picks them)
The “💡 Recommend sources” → “Pick” button asks the model to suggest up to 10 sources for your project topic — a mix of popular/well-known and niche/non-obvious ones, of different types (website, RSS, Telegram channel, subreddit). The service fills in the type, link and name. Note: the links are AI-suggested and the model may hallucinate, so we immediately check each recommendation for reachability — we actually try to open it and tag it: “✓ works” (the source opened, with the item count in brackets) or “⚠️ didn't open” (better verify the link manually). Working ones are listed first. You add the ones you want with a single “Add” button (and enable “Analyze on add” to assess topic fit). The picks rely on the project's “Marketing context” — the more precise it is, the more relevant the sources. There's an optional “Refine search” field: describe the exact source you want to add (type, topic, region/language, specific outlets) and the model will prioritise that refinement without breaking the project topic.
25Analyze before adding
The add form has a “🔍 Analyze fit before adding” checkbox. With it on, the button first becomes “Analyze”: on click the service evaluates the source (reads recent items and compares with the project topic) and shows the result above the field — and only then does the button turn into “Add source”. So you see the assessment first and then decide whether to add the source. The button is disabled while analyzing. If you simply moved focus out of the link field, the quick estimate is usually already computed — then the button is “Add source” right away.
26Relevance check when adding
When you enter a new source link and leave the field, the service immediately tries to read it and estimates how well its material matches the project topic — showing dots (like the effectiveness ones) and a short comment. If the match is low or zero, you have two options: refine the project's “Marketing context” (so selection understands your topic better) or pick a more on-topic source. If the source can't be read (unreachable, empty, or blocking server requests — common for Reddit without keys), it will say so. It's a light one-time check on adding — it helps avoid accumulating “empty” sources.
27Move a source to another project
Expand a source and click “⧉ To project ▾” — pick a target project to copy the source there. The “Move” checkbox transfers it (removing from the current one). Bots, channels and templates can be copied between your projects the same way.
28Reddit: which link to use
A subreddit link like https://www.reddit.com/r/medicine/ is enough — the service fetches that community's “hot” feed and takes up to 15 posts (title, text/link). You don't have to pick the source type: Reddit is detected by domain automatically. Important: access to Reddit can be unreliable, so such a source may occasionally return nothing — if Reddit isn't essential for you, RSS, sites and Telegram channels are more dependable. You can also point to a specific feed, e.g. …/r/medicine/top/ or a thread link (/comments/…).
29YouTube: channel and video transcripts
You can add YouTube as a source. The logic:
- 🔗 paste ANY link — a video, Shorts,
/@handleor a channel; the service resolves the CHANNEL that hosts the video and follows that channel (not a single video); - 📝 the video itself isn't used — the TEXT is: for the channel's recent videos it fetches the transcript (captions) and writes the post from it;
- 💬 if a video has no captions — the video description is used.
You don't have to pick the type: YouTube is detected from the link automatically. Works without an API key (the channel's public feed + the video page); captions are occasionally unavailable (disabled by the author, or region/consent) — then the description is used. To stay within time and token limits, a few of the latest videos are taken per sweep.
30Source update frequency
Each source card shows how often new material appears in it — e.g. “updates roughly every 30 min” or “every 2 hours”. This is the current adaptive polling interval: the service tunes it to the source's real rhythm. Until a source has been polled at least once there's no estimate yet — it shows “update frequency is still being calculated”; it appears after the first sweeps.
31Topic-based sources (Google News, arXiv, PubMed)
Three source types work by QUERY rather than by link — you enter keywords and the service builds a fresh feed from them:
- 🗞️ Google News — fresh news on the topic (across the web).
- 🔬 arXiv — recent scientific preprints on the topic (for science channels).
- ⚕️ PubMed — recent medical studies on the topic (evidence base).
How to add: in the add form click “change type”, pick the type and type keywords into the link field (e.g. “B2B acquiring 2026” or “metformin longevity”). This effectively turns your project's topic into a live stream of hooks and facts. Works without an API key and without blocks. You can also paste a ready link (a Google News query, a PubMed search) — it's accepted too.
32Social without blocks (Mastodon, Bluesky)
Among social networks, Mastodon and Bluesky can be read reliably (no anti-bot, no sign-in) via their public RSS:
- 🐘 Mastodon — paste a link to the account
https://<server>/@name(or the ready feed…/@name.rss). - 🦋 Bluesky — paste a profile link
https://bsky.app/profile/<handle>(the service appends…/rssitself).
X (Twitter), Instagram*, TikTok and VK can't be scanned reliably — they use anti-bot/sign-in, so they aren't in the list.
33Source types and what is collected
Supported: RSS/Atom feeds (title, link, description), websites and forums (title and a short description or the first paragraphs), public Telegram channels (latest posts via web preview), YouTube (the video's channel → transcripts of recent videos), Teletype.in and Telegraph (articles/blogs — article title and text), Google News (a fresh feed by keywords), Mastodon and Bluesky (public social accounts via RSS, no blocks), arXiv and PubMed (recent preprints/studies by topic — for science and medical channels), Reddit (hot posts), LinkedIn (a public page/post — best-effort; such pages often require sign-in and may not parse), and private Telegram channels if a separate reader is connected. Only the title and a short summary are taken, not the full article. Sources can be toggled on and off.
You don't have to pick the type manually: when you paste a link the service detects the type from the address (Telegram, Reddit, YouTube, Teletype, Telegraph, LinkedIn, RSS). The type select is tucked under an “change type” link below the URL — you usually don't need it, but you can expand it and adjust manually. For “strong” domains (t.me, reddit, youtube, linkedin) the link wins over the chosen type — the source is processed correctly even if the type is set inaccurately.
34Duplicate protection (deduplication)
The same material won’t become a draft twice. First, already-seen links are filtered out. If model keys with embedding support are set, items that are semantically similar to previously created posts are also filtered out — even if their link differs. The memory of processed items persists across runs.
35Relevance scoring and selection
Each item is scored against the project’s “Marketing context” and the day’s topic. With keys, scoring is by meaning (semantic similarity); without keys, by keyword overlap. The few most relevant items move on. The more precise your marketing context, the better the selection.
36Turning into a draft
A selected item is rewritten by the chosen model in your tone of voice, language and audience region; an image is generated in your social network’s format. The draft lands in the queue with the status “Draft” and waits for your approval — nothing is published without confirmation.
37Schedule and manual run
In “Plan”, each weekday gets a topic and a mode: “from sources” or “from the model’s knowledge”. On schedule the service sweeps in the background; the “Generate drafts” button runs the same thing right now. Each run’s outcome (how much was scanned and created) is visible in the logs.
38If a source is unavailable
If a site won’t load, a feed is empty, or a platform blocks server access (this can happen with Reddit), the service doesn’t fail: the problematic source is skipped, and if there is no new material at all, it prepares an “idea” around the project topic so you have something to work with.
39Good to know (limitations)
It’s a lightweight parser: it takes the title and a short description, not the full text; pages rendered entirely by scripts may not be read. Reddit works more reliably with access keys; private Telegram channels need a connected reader; semantic relevance and smart deduplication work only when model keys are set.
* Деятельность компании Meta Platforms Inc., а также принадлежащих ей социальных сетей Facebook и Instagram запрещена на территории Российской Федерации
Didn't find an answer? Write to us — we'll help.
Contact support