Why a media monitor misses a story
A communications team's list of 41 placements for one month, of which the pipeline had found six. Where the other 35 were, what we changed, and what one company's coverage does not prove.
Media Monitor, the press-monitoring product OneThought AI runs on this programme's research, reads the Indian press for mentions of an organisation and reports what was said, where and in what tone. Every count it shows assumes that the articles it reads are the articles that exist. In June 2026 we could test that for one company, on the product's production pipeline.
41
placements in the team's own list
One company, June 2026, three stories.
6
found by the pipeline
All 35 misses were articles it never fetched.
13
of the 35 recovered by open-web search
Retrospective test, 6 July 2026.
The list
The company buys and resells used phones. Its communications team keeps a spreadsheet of every placement it knows of: outlet, headline, date. For June 2026 it held 41, from a press release on a data-security survey (31), a comment for World Environment Day (9) and one industry story.
Where the 35 went
Had the matcher seen the articles and failed to recognise the company in them? No. Every one of the 35 was an article that never entered the pipeline. The losses were all at the first stage.
- 01
Discover
A curated list of national and trade feeds, and a Google News search in English.
35 of 41 lost here
- 02
Fetch
The full text of each article. Until June, 73% of articles arrived as a headline only.
measured 11 June, fixed
- 03
Match
Find the company in headline and body, then check it is the right company.
0 of 41 lost here
- 04
Label
Tone, topic, language and the people quoted.
accuracy not measured yet
Each discovery channel misses in its own way. A curated list cannot contain a PR wire site nobody has heard of, and some large business outlets block any reader that is not a browser. Google's search merges the 31 copies of a press release into a few results, keeps the large outlets, and was only ever asked in English.
The test
On 6 July we ran open-web search as a discovery source, after the fact: five queries on the company and its three stories, limited to India, 103 unique results, 24 search credits.
Of the 13 recovered: five wire copies, four trade placements, two national outlets, two Hindi outlets. Of the 103 results, 68 were about the company; 31 of the rest were social media posts, which can be excluded at the source.
The same day we ran a fairer comparison: six brands, one 24-hour window, the same query, region and recency for each. Open-web search returned every one of the 44 articles the existing channel had matched, and 12 more, nine of them regional-language coverage of one retailer's results in Telugu, Hindi, Marathi, Gujarati and Punjabi. It lost two fields: the publisher's name, now derived from the domain, and the byline, which the article page's own metadata gives back about nine times in ten.
One detail nearly gave the wrong answer. The first run capped results at 20 per query and showed the new channel finding 59% of the old one's articles. The whole gap was one brand with 39 articles; a cap of 50 closed it to 100%. A result cap can manufacture a recall gap.
What we changed
Four discovery channels were added in July 2026: a per-company feed of Hindi and regional-language outlets; a Google News feed per company with a Hindi edition for search terms written in Devanagari; a check that fetches the body of any item Google matched that our headline rule had dropped; and a daily open-web search sweep. Each runs behind a switch and writes to the fetch log, so the next audit can say which channel found what.
What this does not show
- One company, one month, three stories. Press-release coverage skews to wires and syndicated copies, so the 15 wire misses may overstate the gap for reported news.
- The retrospective search is a lower bound for a live monitor, since a story is easier to find on the day, and an upper bound in another way, since we searched with the stories in hand.
- Six brands over 24 hours, two of them with no coverage in that window. The comparison reports overlap between two search channels, not accuracy.
- The team's list is what one team found by hand for one month. It is the best list we have.
Next
In this note a placement is one headline in one outlet, and a story counts as found if any placement of it is. The audit is to be repeated on a comparable cluster of coverage once the July channels have run for a month. The attribution benchmark follows in 2027, with missed mentions, wrong tone and invented attribution measured separately and by language, as how we evaluate sets out. The numbers will be published whichever way they come out.