Product · OneThought AI
How Media Monitor reads the press
Media Monitor reads the Indian press for mentions of the organisations you track and reports what was said, where and in what tone. It works in stages, and each stage can be wrong in its own way, so this page says what each one does and where it stops.
Built by OneThought AI on research from Scaling Truth Research Foundation.
Four stages run on a schedule all day, and what they find reaches you as a dashboard, alerts and summary emails. Nothing in them is secret, and the limits are as important as the mechanics.
- 01
Discover
Feeds and sitemaps from a curated list of outlets, polled every few minutes, plus a news search per organisation.
cannot see print, broadcast or paywalled sites
- 02
Fetch
The full text of each article, through a cheap-first chain of fetchers, so a mention in the body counts.
- 03
Match
Your organisation's name, aliases, products and spokespeople, with a context check and a model's second opinion for ambiguous names.
- 04
Label
Tone, topic, language and the people quoted. All model output, shown as such.
accuracy not yet measured
What it reads
A curated list of national, business and trade outlets, read through their feeds and sitemaps every few minutes, and a news search for each organisation you track. Four channels were added in July 2026, each behind a switch: a feed of Hindi and regional-language outlets per organisation, a Google News feed with a Hindi edition for search terms written in Devanagari, a check that fetches the body of any item the search matched that the headline rule had dropped, and a daily open-web sweep.
What it cannot read: print editions, broadcast, paywalled sites, and outlets that block any reader that is not a browser (a free reader service gets through some of these). A count on the dashboard is therefore a count of what was found. How far that can fall short of what was published, and what was done about it, is in Why a media monitor misses a story.
Once an article is found, its full text is fetched: a direct request first, a reader service if the site blocks it, and a paid fetcher as a last resort. Until June 2026, 73% of articles arrived as a headline only, so a company named in the fourth paragraph was invisible. That is why fetching comes before matching.
How it decides a mention is yours
You give it four things: the organisation's name, which is always matched; aliases, such as a stock ticker or a short form; product names; and the people who speak for it. For a name that other organisations share, you add context words that describe your organisation's world, and "not us" words that mark the others.
Meridian Foods said on Tuesday that its dairy plant near Pune would double capacity by March, chief executive Asha Rao told reporters. Shares of MFL rose 2% in early trade.
Separately, Meridian Bank reported a fall in quarterly profit.
Matching runs twice. When an article arrives, its headline and summary are matched at once, so the dashboard updates within minutes. When the full text has been fetched, the whole body is matched again, and that is where ambiguous names are resolved:
| The mention | What happens | Confidence recorded |
|---|---|---|
| A name no other organisation shares | Accepted | 1.0 |
| A shared name with one of your context words nearby | Accepted | 0.8 |
| A shared name with a "not us" word nearby | Rejected | |
| A shared name with neither | A model is asked whether this mention is your organisation, told your context and "not us" words, and told to reject a different organisation that shares the name | The model's own confidence |
Every accepted mention links to the article it came from, so the decision can be checked by opening it. How this went wrong for ambiguous names before the context and exclusion rules existed, and how it was measured, is in A mention is a claim.
What it labels
- Tone. A model reads the headline and the first 500 characters of the article, or the headline alone if the text could not be fetched, and returns a number from -10 to +10. A question headline is scored on its framing; a bare name scores 0. The dashboard's tone filters and averages are built from these numbers.
- Topic. One to three labels from four: financial, technology, interview, general.
- Language. Detected from the headline and summary. Accepted only when the detector is at least 85% sure and the language is English or one of thirteen Indian languages; otherwise recorded as English. English is therefore over-counted, and the language bar is a floor for regional coverage.
- People. The people named in an article and their role: chief executive, executive, spokesperson, analyst, reporter, politician, third party, or unclear.
- Bylines. The journalist's name is read from the article page's own metadata, and only the journalist's published work is kept. The directory built on it is not yet switched on for customers.
None of these labels has been measured against human readers yet. They are model output and are shown as model output. Measuring them is the first item on the research programme's benchmark.
What arrives in your inbox
- An alert within about five minutes when an article about your organisation scores strongly negative, with a 30-minute cooldown per rule so a bad morning arrives as a pattern rather than thirty emails.
- A summary at 10:00 and 17:00 in your timezone: new mentions since the last one, per organisation, with the tone split and the top headlines. Quiet organisations are skipped; if everything is quiet, nothing is sent.
- Anomaly alerts when the pattern changes: a volume spike against the organisation's own baseline, a run of negative tone, a pickup by a national outlet, or a cluster of outlets carrying the same story.
What to hold it to
Every mention should open to its article. Every count should say what was searched. Every label should say what produced it. Where it falls short of that today is written above, and where it fails in production is published under Media intelligence as it is measured.