A mention is a claim
A monitoring report that says you were mentioned 216 times last week, mostly negatively, has made three claims and shown the evidence for none. What each one rested on when we checked it against the Indian press.
A monitoring report lands on a communications director's desk on Monday morning: 216 mentions last week, tone mostly negative. Both numbers are claims, and a third hides under them, that the tool saw the press at all. The report shows the evidence for none of the three. We spent the middle of 2026 checking claims like these against the Indian press, on the pipeline behind Media Monitor, the product OneThought AI runs on this programme's research.
The count
Take a footwear company whose name is also the name of a spyware vendor, a private-equity firm, a shopping mall in Singapore, a dirt-racing speedway, a British bank, an Australian care provider, an American health-policy institute and a poultry group in Bangladesh. A tool that matches the word counts all of them.
The 137 in between were articles about other companies, each one a chance to brief a chief executive on someone else's crisis. How bad this gets depends on the name. Distinctive names were clean. For ambiguous ones, a June audit against Google's own news index found wrong-company articles at about 10% for the footwear brand with no context rule and 2.5% with one; for an electric-vehicle charging company whose name is also a defunct phone brand, 18% without and 13% with. Its context words, charging and battery, describe phones too. No positive word separates the two companies. A negative one does.
So the rule changed in two ways: the context word has to sit within about 90 characters of the mention, and each company can list words that mean "not us". In a dry run over production data, both tuned workspaces went to zero wrong-company matches, with no genuine rejections.
Three days later the count was wrong again. Mentions in article bodies pass through a second check, a language model asked whether this occurrence of the name refers to the company. Its instructions distinguished a company from a common word, apple pie from Apple, and never said which company. A banking group with the same name is unambiguously a company, so the model said yes, at 0.95 confidence, 37 times in one workspace, 36 of them after the exclusion words had been set. The model had answered the question it was asked. The fix was to ask the right one: the company's context words and exclusion words now go into the question.
A count is a claim about a set. The set depends on what was searched, what was excluded and what the model was asked. A report that shows the number and not the set cannot be checked.
The coverage
"216 mentions" assumes the pipeline saw the press.
73%
arrived as a headline and nothing else
Sitemap and search sources carry no summary. A company named in the fourth paragraph was invisible in three of four articles until the full text was fetched.
The story headlined on a metro line getting mobile coverage is, in its body, about one operator's network survey. The story headlined on a vehicle scrappage scheme names, in its body, the truck maker that benefits. Reading the full text found 104 such body-only mentions in production by the middle of June.
Then there are the articles that never arrive. A communications team's hand-kept list of 41 June placements for one company, checked against the pipeline: six found, 35 never fetched. The note on that audit has the breakdown. A tool that finds six of 41 and reports "six mentions" has made a true statement about itself and a false one about the press, and nothing on the report page tells the two apart.
The tone
- 01
Input
The headline and the first 500 characters of the article. The headline alone if no text was fetched.
- 02
Model
A language model asked for one number, with no view of the rest of the article or of the outlet.
- 03
Output
A score from -10 to +10. A question headline is scored on its framing. A bare name scores 0.
- 04
Check
Against human readers: not yet. The first item on the programme's benchmark.
accuracy unmeasured
That is what a tone score is. Language is handled the same way: detected from the headline and summary, accepted only at 85% confidence and only for English or thirteen Indian languages, otherwise recorded as English, so English is over-counted and the regional bar on the dashboard is a floor. Media Monitor shows tone labels as model output. What the report should also say is that a mean of 200 such numbers inherits every one of their errors.
The standard
A mention is a claim, and so is a count, and so is a tone. The standard we are building to is that each one carries its evidence.
- Every mention links to the article it came from, so the reader can open it.
- Every count says what it searched: which outlets, which languages, what time window.
- Every label says what produced it and, once measured, how often it disagrees with a person.
- Every measurement is reported by failure and by language. In the June list, nine of 41 placements were in Hindi outlets. One was found.
The fluent summary that made media monitoring cheap also made it unverifiable, and the person who most needs to verify it, the one about to walk into a room and explain the week's coverage, has the least time to go back to the articles. The report has to do that work for them, in the open. The benchmark that measures how well it does is due in 2027; the audit note is the first of its numbers.