From a news feed to a decision — what the model proposes, and what you decide

Each row is one outlet you follow, with a health badge built from its recent runs.
1) Read the badge before anything else — a feed that has quietly stopped returning items shows here, not on your dashboard.
2) Use the row menu to run a source on its own, or open it to see every past run.

1) Pick the connector type — the form changes to ask only for what that type needs.
2) Paste the endpoint, choose the languages to keep and how often to check.
3) Press Test connection and wait for the result before saving.
4) Save.

1) Describe your organisation in one sentence — this is read by the model, so write it the way you would explain your remit to a new colleague.
2) Add your sectors, competitors, geographies, key people and keywords, each with a weight.
3) Add exclusions for the subjects you never want scored up (sport, for instance).
4) Set “Alert above” — the score an article has to reach before it is allowed to interrupt anybody.

Group the names you want followed: competitors, customers, suppliers, government bodies, projects, executives or regions. One list per theme keeps alerts readable later.

Check it weekly. Rows are grouped by model with the tokens in and out and an estimated cost, and the whole report exports to Excel and PDF for the same period you see on screen.

Pick a mode when you start one by hand:
• Latest — everything published since the last run. This is what the schedule uses.
• Window — a date range, for backfilling an archive.
• Re-analyse — read articles you already have again, after you changed the interest profile or a prompt.

This is the screen you live in. Filter by source, category, sentiment, relevance, date, review status or story, then act on many rows at once. Save any filter combination you use repeatedly as a view (Stage N) so it is one click tomorrow.

Read the relevance reason, not just the number. A score you cannot interrogate is a score you cannot trust, so the module always shows why it landed where it did.

Under the relevance bar sit the two halves the score was made from: what the model judged, and what your own keyword list matched — with the distance between them named. When those two are far apart it is usually your term list that has drifted, not the model that is wrong, and a wide gap is flagged so you notice instead of averaging past it. An article scored on keywords alone shows no gap at all, because there is no second opinion to disagree with.

1) Open an article the model was unsure about.
2) Approve it, correct a field, reject it, or escalate it to a colleague.
3) Assign it to somebody and leave a comment if it needs a second opinion.

Open a cluster to read the cross-article comparison: how the narrative shifted between outlets, where they contradict each other, what they all agree on, and the detail only one outlet reported.

Check the verdict column before quoting a figure anywhere: Supported, Partially supported, Contradicted, Not found, or Needs human review — each with its reason and the evidence behind it.

Read the recommendation and the quoted rationale, then choose one:
• Create a task — lands in Tasks, assigned and traceable back to this signal.
• Create a lead — lands in CRM as an opportunity to pursue.
• Notify an executive — sends through the normal notification channels.
• Dismiss — with a reason, which is kept.

1) Set the conditions — a watchlist term appearing, a sentiment turning negative, a relevance threshold, a category, a source.
2) Pick who is told and on which channels: in-app, email, SMS, WhatsApp or web push.
3) Set a throttle so one busy story cannot send forty messages.

Recipients keep their own channel preferences, and the message wording is an editable bilingual template — change it once and every channel follows.

Every firing is listed with the rule that caused it and the article behind it — so a complaint of "why was I messaged at midnight?" has an exact answer, and a rule that fires too often is easy to spot and tighten.

Trust breadth over volume. One outlet repeating itself all week is deliberately NOT a trend here — a subject needs more than one source before it is scored at all, which is what keeps this list free of noise.

Ask in Arabic or English — "show me negative coverage this month", "what did this outlet publish about the new regulation". Ask follow-ups; the conversation keeps its thread.

Read the KPI strip, then the coverage timeline, the category and sentiment splits, the relevance spread, your top entities, the source comparison and the geography. Filter the whole page by branch if you run more than one.

Coverage over time, sentiment, source comparison, period-over-period, and AI usage & cost. Set the dates and the branch, then export to Excel or PDF — the export always covers exactly the period on screen.

1) Define a brief: daily, weekly or monthly, with its recipients and formats.
2) Tick the formats you want offered — HTML, PDF, Excel. Whatever you leave unticked has no download button on the issue and cannot be fetched by its link either, so the setting is a real restriction rather than a preference.
3) Let the schedule run it, or generate one on demand to preview.
4) Recipients are notified with a link to the issue, which is stored so you can reopen any past edition.

1) Edit the wording for a task — how it judges relevance, or what counts as a signal.
2) Press Preview to see the exact prompt that would be sent, against a real article from your archive. No provider call, no cost.
3) Press Test run to send it once for real, WITHOUT making it live.
4) Activate when you are happy.

Name any filter combination and keep it — "high relevance, needs review", "negative coverage this month". Keep it to yourself, or share it with the team so everyone works the same queue. One can be your default so the screen opens where you want it.

The module home lists every screen in this manual in one place, split into the daily workspace and the settings you touch rarely. Start here if you are not sure where something is.