Tone of presidential coverage across major US broadcast networks and digital news outlets, 2025–present.
% positive minus % negative each week, indexed across outlets. Faint lines are weekly values; bold lines are LOESS-smoothed (span = 0.5). The shaded band is a 95% interval that widens in weeks with fewer contributing outlets.
Share of segments/headlines classified as negative each week.
Total segments or headlines mentioning the president each week.
Mean net score across the selected date range, by network.
Track coverage volume for specific topics over time. Currently being refined — keyword definitions are being reviewed to reduce noise before we publish the live trends.
The weekly aggregated CSVs that power this page are released under CC BY 4.0. See About for how the index is built and what it does not capture.
The same index, computed from newspaper headlines back to 1947. Each line is one presidency aligned to its own inauguration, so month 12 of one administration sits directly above month 12 of another.
Bold lines are a seven-month local trend; the faint line behind each is the raw monthly PCI it is fitted to. Shaded bands are 95% intervals and widen when fewer papers contribute. The window is fixed at seven months for every president, so a one-term and a two-term administration are smoothed by the same amount.
Mean PCI across each presidency, ordered chronologically.
The Presidential Coverage Tracker is a research project from Yale political scientists that measures the tone of presidential coverage across major US broadcast and cable news networks and digital news outlets. We collect closed-caption transcripts from the Internet Archive's TV News Archive and news headlines from Media Cloud, classify each segment and headline mentioning the president using a fine-tuned natural language inference model, and update this dashboard every few days. A third corpus of newspaper headlines, described below, extends the same measure back to 1947 so the current administration can be read against every president since Truman rather than on its own.
The classification model, training data, and analysis code for the net coverage score are open source. The paper's code repository is at github.com/kevin-deluca-polisci/presidential_headlines; the code for this dashboard lives at github.com/kevin-deluca-polisci/coverage-tracker. The full paper is forthcoming.
Traditional methods of measuring media tone — dictionary-based word counts and document-level transformer sentiment classifiers — assign an overall positive, negative, or neutral score to a document rather than to the candidate it discusses. This document-level approach cannot identify to whom sentiment is directed. A headline like "Biden struggles to contain inflation" conveys a different signal about the president than "Inflation eases under Biden," yet both might be flagged as negative by a dictionary model that treats "inflation" as a negative term. We adapt methods from natural language inference (NLI) to measure candidate-specific media coverage. Rather than inferring tone from surface-level word usage, our approach uses stance detection to evaluate whether a headline or transcript segment implies that a given candidate is performing well or poorly.
Each broadcast transcript is split into 3-sentence windows. Windows mentioning "Trump" are classified using the Political DEBATE model developed by Michael Burnham. We further train the model on a task that evaluates whether the text conveys a positive, negative, or neutral performance cue about Trump; full training procedure, validation, and reliability tests are documented in our code repository and accompanying paper. The same method applies to digital news coverage using the headline as the unit of analysis.
A segment is classified as positive if the model scores a "yes" on the hypothesis "the author of this text believes that Trump is performing/performed/will perform well" and negative if the model scores a "yes" on the parallel "poorly" hypothesis.
The Performance Cues Index (PCI) is the share of coverage carrying a positive performance cue minus the share carrying a negative one. A PCI of −15 means 15% more of the week's coverage implied the president was performing poorly than implied he was performing well. Zero means the two are balanced. It is not a measure of how favourable coverage is in general, and it is not a sentiment score — see below.
Two problems make a simple average across outlets misleading, and the index corrects for both.
Thin outlet-weeks. Because PCI is a percentage, an outlet publishing a single article in a week scores +100 or −100. Under a simple average that outlet would count as much as one publishing five hundred. Outlet-weeks below 30 headlines are therefore excluded from the headline number. In the current corpus every outlet-week above this threshold falls within ±50; several below it sit at exactly ±100.
The same threshold governs the charts. A week an outlet did not clear it is drawn as a break in that outlet's line rather than being joined across, so the several multi-week gaps left by source outages — Reuters, ABC News and Bloomberg each have one — read as missing data instead of as measurement. An outlet also needs at least eight qualifying weeks before it appears at all; Washington Post and Politico entered the corpus in August 2026 and are collected but not yet plotted.
Changing corpus composition. Which outlets are available varies over time, so a raw average confounds movement in coverage with movement in the sample. Rather than averaging, the index estimates outlet effects and period effects jointly and reports the period effect, so an outlet entering or leaving the corpus no longer moves the series by itself. The shaded band is a 95% interval from that estimate, and it widens in weeks resting on fewer outlets — thin periods are drawn as thin rather than as precise.
The headline is unweighted across outlets, so it answers what did the average outlet publish rather than what did the average reader see. Those are different quantities and the second is not currently reported. No adjustment is made for outlet credibility or prominence; the index requires no judgment about which outlets deserve more weight.
Bold lines are LOESS smooths (span = 0.5) fit to the weekly index. Faint lines underneath are the weekly values the smoother is fit to.
The Historical tab uses a window fixed at seven months rather than a span proportion. A proportion would smooth a two-term presidency over roughly four years and a partial first term over a few months, and that difference in wiggliness would read as a difference in coverage rather than as an artefact of how long each line is. Fixing the window in months damps month-to-month variation by about three quarters for every administration alike.
It measures cues about performance, not whether coverage is accurate, fair, or deserved — the object of measurement is the coverage, not the conduct being covered. It says nothing about volume: a week of little but harsh coverage and a week of heavy but balanced coverage can share a PCI. It treats each qualifying outlet equally regardless of audience size. And it depends on the corpus actually available, which has gaps; see the outlets listed below and the coverage counts in the volume chart.
Broadcast and cable: CBS, CNN, Fox News, ABC, NBC, and MSNBC/MSNow. Data runs from January 2025 to the present. We include national broadcast and cable news programs only — no local affiliates or regional newscasts.
Digital news: Reuters, Fox News, CBS News, Bloomberg, CNN, ABC News, USA Today, New York Times, NBC News, Los Angeles Times, NPR, Washington Post, and Politico. Headlines are collected from four complementary sources and deduplicated by article URL:
The Historical tab is built from a different corpus and a different collection method, and the two should not be pooled. It comes directly from the headlines corpus assembled for the paper: newspaper headlines only — no broadcast transcripts and no digital outlets — drawn from 44 US daily papers running from November 1947 to the end of 2025.
These headlines are not scraped. They are extracted from digitised newspaper archives through ProQuest TDM Studio, which provides text and data mining access to full-run newspaper databases under licence. That is why the historical corpus reaches back seven decades where the live corpus does not, and also why it is fixed rather than updating: the daily collection pipeline feeds the Current tab only. The historical series changes when the paper's corpus is rebuilt, not on a schedule.
Data availability. The monthly series charted here is in the tracker repository already. The underlying headline-level corpus — every classified headline behind these monthly figures — will be released when the performance cues paper is published. It is not in the repository in the meantime because of its size and because the paper it belongs to is still under review, not because access is restricted; if you need it before then, write to us.
Classification is identical to the current series — the same fine-tuned performance cues model, the same positive and negative hypotheses, applied to the headline as the unit of analysis. A 1952 headline and a 2026 headline are scored by the same classifier asking the same question, which is what makes the comparison across presidencies meaningful in the first place.
Monthly rather than weekly. Mid-century papers ran a handful of headlines a week about a president where a wire service now runs hundreds, so the aggregation period is a month and an outlet needs 5 headlines in a month to contribute. We tested a weekly version: it is technically feasible, but it roughly doubles the width of each interval while leaving the smoothed trend almost unchanged (the two correlate at 0.97), and for the earliest presidencies essentially none of the week-to-week movement is distinguishable from sampling error. Monthly is the finer measurement despite being the coarser period.
Two further rules. A month resting on a single paper is dropped: with one outlet the outlet effect and the month effect cannot be separated, so the cross-outlet index is just that paper. This excludes one month in 953. Outlet and period effects are then estimated jointly across the whole 1947–present panel, so a paper entering or leaving the archive does not move the series on its own. The seven-paper basket restricts to titles publishing across all fifteen presidencies, so movement in the index can be checked against movement in the sample. The two orderings of the presidencies are near-identical (rank correlation 0.99, term averages correlating at 0.98), but they are not interchangeable: the basket tracks the full corpus at 0.92 month to month, and the seven legacy papers run about 2 points less negative on average — most visibly for Obama, where the gap is 5.7 points. Use the basket to check whether a change in the line reflects a change in the sample, not as a substitute series.
What the historical series cannot tell you. Which papers were digitised varies a great deal across eras, and presidencies covered by more papers show more months that register as significantly unusual — roughly two and a half times as many as thin-corpus presidencies. Some of that is real and some of it is simply that a denser corpus resolves more. Read the level and the trend with more confidence than the spikes, particularly before 1980.
Kevin DeLuca is an Assistant Professor of Political Science at Yale University, where he studies American politics with a focus on elections and media. His research examines how news coverage shapes political information and accountability, including work on newspaper endorsements and the effects of local news. He is a Resident Fellow at Yale's Institution for Social and Policy Studies and a Faculty Affiliate at the Center for the Study of American Politics. Further work is at kevinmdeluca.com.
This tracker was developed with Zoe Kava, who built and validated the classification model the index rests on.
Questions about the data or the measure: kevin.deluca@yale.edu