A mention that nobody retrieves is a mention that did not happen.

It is the attempt to say what coverage you did not pay for is worth. A press mention, an interview, a review, a citation in someone else's article: none of it was bought, so none of it has a price, and the whole field exists to put a defensible number next to it.

Outputs first: mentions, reach, share of voice, page visits, content downloads. Then sentiment as a tone label, and message delivery as the question whether the key statements appeared. Conversions are attributed where the tooling allows it. On top sits Earned Media Value, a money figure built from advertising prices.

Because it is a proxy. Earned Media Value prices a mention as though it were an advertisement of the same size, which is a category error: an article is not an advertisement, its audience did not choose to see an advertisement, and the price of an advertisement is not the value of being mentioned. The measurement profession has argued against the practice for years, and the figure survives because it is easy.

Outputs are what you produced: pieces, mentions, reach. Outtakes are what the audience took from it: awareness, recall, understanding, favourability. Outcomes are what changed afterwards: attitude, behaviour, revenue, risk. Most reporting stops at outputs and then speaks about outcomes as though the two were the same.

Because campaigns are reported as absolute numbers. Without a measurement taken before the campaign, an increase cannot be separated from a season, a news cycle or a competitor's mistake. A baseline costs one measurement and decides whether every later number means anything.

It is the connection between communication and what the organisation actually achieves: sales, recruitment, cost of capital, licence to operate. It is rarely shown because it requires data the communications department does not own, and because a weak connection is more visible than no connection at all.

Precision at k asks how many of the first k results are relevant. Applied to coverage it asks a question no media report asks: when somebody searches your brand and the topic of the article, does the article appear in the first ten results, or on page four? A placement that cannot be retrieved is an archive entry.

Because reach counts an opportunity and a ranked list decides an encounter. Graded measures such as normalized discounted cumulative gain exist precisely because a result at the top is worth more than the same result further down. Coverage reports value both the same.

Modern systems do not retrieve whole articles, they retrieve passages. The unit of value is therefore the paragraph that carries your statement, not the article that contains it. A long feature with your quote buried in the middle can retrieve worse than a short piece that states the same thing cleanly.

A system has to decide that the string in the text refers to your company and not to a similar name, a product, or a person. Where that fails, the mention exists for a human reader and does not exist for any machine that answers questions about you.

They fall out of most counts. A paragraph that says the company announced, without repeating the name, is invisible to keyword monitoring and hard for systems that do not resolve references. The effect is systematic: the more naturally an article is written, the fewer of its mentions are counted.

Because a match no longer requires the same words. Coverage that describes your product in its own language can be retrieved for a query that uses none of your terms. That cuts both ways: coverage you never wrote can answer questions about you, and coverage that repeats your slogans may retrieve for nothing at all.

“retrieval can be practically implemented using dense representations alone”

Study: Karpukhin et al., Dense Passage Retrieval for Open-Domain Question Answering, arXiv 2004.04906, 2020. https://arxiv.org/abs/2004.04906

That common words carry little. Weighting schemes such as term frequency with inverse document frequency exist because a word everyone uses cannot distinguish anything. A press release built from category vocabulary is expensive to produce and cheap to retrieve.

Different from the tag list. Statistical topic models group coverage by what the texts actually share, not by the categories a monitoring service offers. The clusters that appear without being asked for are the ones worth reading.

Because different systems rank differently, and a placement visible in one can be absent in another. Combining several ranked lists, for example through reciprocal rank fusion, is standard practice in retrieval and unknown in media reporting, where a single screenshot still passes as evidence.

It is the attempt to divide a conversion among the channels that touched it. For earned media it fails twice: the touch often happens without a click, and the click that does happen carries no campaign parameter. The honest version states which part of the journey is unobserved rather than hiding it in a model.

“The inference about the influence of each channel plays an important role in budget allocation and inventory pricing decisions”

Study: Li et al., Deep Neural Net with Attention for Multi-channel Multi-touch Attribution, arXiv 1809.02230, 2018. https://arxiv.org/abs/1809.02230

By citing each other and the same origins. Co-citation and bibliographic coupling, borrowed from bibliometrics, show which outlets move together and which one starts a story that the others then carry. That map explains where to place one piece instead of ten.

Outputs with their date, outtakes where they were actually surveyed, outcomes where the data exists, a baseline, and the retrieval part: whether the coverage is indexed, at which position it appears for the queries that matter, and whether a passage from it can stand alone. Anything else is a screenshot with a number under it.

It usually means the coverage is being read on the result page and in generated answers rather than on the publisher's page, and that the publisher's page is where the link sits. Check whether the articles are indexed, at which position they appear for the relevant queries, and whether an answer engine quotes them without linking. The decline is a retrieval question, not a public relations failure.

That outputs are known exactly, that outtakes require a survey, that outcomes can be connected where the data exists, and that the retrieval part is measurable today at no cost. A number without that structure is a guess in a suit.