How Inquiry and Tone are read

Attention

Attention has three parts, each read on its own. The headline number is the parts that can be read, in equal shares. It needs two of the three: with one, the reading is withheld and the part that could be read still shows.

Each channel against its own normal

A name is compared with itself, never with another name. This week's value is set against the median of the 52 weeks before it, on a log scale, in units of that history's own spread. Then the platform's own move comes out: if Wikipedia traffic fell everywhere this week, that fall isn't news about anyone. The result runs through a curve from 0 to 100 where 50 is this name's normal week, one spread above normal reads 73, and two read 88. It stops at four spreads, so a long story sits near 98.

A channel counts only when it can say something: enough weeks of history, enough volume and enough distinct values. A channel that can't is left out. It is never read as “no change”, and nothing here has a default score.

No AI company publishes how often a name is asked about, so the AI channel reads how many sources the answers to one fixed question cite, against that name's own normal. It is a sign of how much there is to draw on, not a count of people asking. It needs eight weeks of history, as does X; neither channel is live before late December 2026.

Cities

A city is read on up to four channels, in three parts. Asked is its English Wikipedia article, matched to the city by its Wikidata identifier, and, for the 493 most-read cities, Google Search interest in the city's Google topic (its Wikidata Freebase identifier), which Google keeps apart from other things with the same name. Covered is the headlines that name it: page titles from the GDELT Project and headlines from our own feeds whose rights allow it, at most five a week from any one outlet, like every other name. Discussed is the count of English posts on X that name one of the same 493 cities, read on Mondays.

A headline names a city only by the rules our city pages use. Whole words only. Phrases that carry a city's name without being about it come out first: the Paris Agreement, the Phoenix Suns, Real Madrid, the Lima Group. A name two cities share, or one that is also a common word or a place elsewhere (Birmingham, Portland, Valencia), counts only when the same headline names the city's country, or its state in the United States. Plain “Washington” is not the capital. Only English headlines are read. Some headlines that use a city to mean its government (“Beijing says”) still count, because the rules read names, not meaning.

A count of posts can't read each post the way the headline rules do, so the X query is built from the same rules and kept fixed: the city's names as whole phrases, the false friends taken out (Paris Hilton is not Paris), and a shared name only beside its country or state. A city whose name can't be told apart that way isn't counted on X. Bluesky and AI answers are not read for cities.

A city needs two live parts, like every name. Wikipedia and Search are both Asked, so a city also needs Covered or Discussed. Covered needs 26 weeks of headline history, which cities have had since the week of October 19, 2026; a city with too few headlines in a normal week stays unread. X needs eight weeks of history, so Discussed can't be live for a city before January 2027. Until a city has two live parts, its page shows no reading at all. Tone is not read for places.

Tone

Tone is stance toward the name, not the mood of the text. “Apple stock falls” isn't unfavorable to Apple. An obituary isn't. The victim of a bad event isn't. It is a reading of what was written this week, not a judgement of the person or organization, and not a rating of any kind.

A model labels each headline and each sampled post favorable, descriptive or unfavorable, and says whether it is about this name at all. Anything it is unsure of counts as descriptive. We keep the label and a reference, never the text.

The balanced net. Left alone, Bluesky would outvote the news about nineteen to one, and its audience leans critical. So each channel is read on its own, against that channel's normal: the median of its peers over eight weeks. Each channel needs 8 mentions to vote, and leans only past 0.10 and past its own sampling error. The balanced net is those distances from normal in equal votes.

The call. Favorable or Unfavorable only when news and Bluesky agree. When they pull opposite ways, or one is itself evenly divided, the word is Split, never neutral. Otherwise it is No clear lean, or One channel only when one has the numbers. One channel never makes a call.

What AI answers say

Every week the same question goes to Google AI Mode, Gemini and Bing Copilot, for the 600 names on a fixed panel: “What is the current public reputation of {name}, based on reporting from the past month?” A separate labeller scores each answer for stance from −2 to 2 and for the share of it that is favorable and unfavorable, then the answer is gone. Only scores are kept, with the engine, the week and the source domains the answer cited.

Name pages show one composite: the combined stance, the favorable and unfavorable shares, how many answers and engines, whether the engines agree (they differ when two of them sit two or more points apart), and the domains cited most. It is withheld unless at least two engines read the name that week. We never show an answer, a quote or a paraphrase of one, and never one engine's number.

In Tone these answers are one vote, at most 20% of the balanced net, once two engines have read a name. They can't make a call or stand in for a missing channel, because they read the same news.

The whole panel, week of 2026-10-26: 600 names, the engines agreeing on 425.

When we say nothing

• Attention needs two live parts.
• Tone needs 20 scored mentions in the week, and a channel normal to read against.
• Tone is withheld when over 40% of mentions turn out to be someone or something else with the name.
• These show as “Not enough to read this week”, with the reason.

For the week of 2026-10-26, 1,948 names had an Attention reading and 3,645 were withheld. 379 had a Tone reading.

Who is left out

• Political figures and bodies are not read for Tone: anyone Wikidata calls a politician, party member, candidate or office holder, a political organization, and a short hand list. Anyone not yet checked stays out.
• Places are not read for Tone. Cities have an Attention reading only, and a City page shows it only in a week the city has one.
• Minors are not read at all.
• People removed on request are gone from every page the moment the request is decided. Report an error or ask for removal.

Questions people ask inside Palanor are in no public reading, now or later. There is no list ordered by Tone, no most-disliked list and no leaderboard of people by Tone. Nothing here is available as a file, an API or an embed.

How far to trust the labels

Plainly: single labels are often wrong, and the favorable class is weak. We tested one run of the model on 327 mentions, 267 of them about the entity, on 2026-10-31. The reference labels came from one annotator, which was itself an AI agent, not a person or a panel, so these figures measure agreement, not truth. Overall agreement was 82%, with a kappa of 0.57.

• About one “unfavorable” label in three disagrees with the annotator, and more than half of “favorable” labels do, on only 11 true cases.
• Headlines agreed 89% of the time, Bluesky posts 75%.
• Totals hold up better than single labels: 63 predicted unfavorable against 55 in the reference set.
• Mentions of someone else with the same name were caught 94% of the time, and 8.9% of real mentions were wrongly dropped, mostly posts that name the entity only in a link card.
• Bluesky's audience and the model's weak favorable class both push readings toward unfavorable, and the data can't separate the two. That is why every channel is read against its own normal, and why a call needs two channels to agree.

Read a week's Tone as a signal to look closer, never as a finding about anyone.

Sources and credits

• The GDELT Project. Global Knowledge Graph article counts, page titles read for stance, and page titles counted for the cities they name. Data from the GDELT Project. Headlines belong to their publishers: the title is read in memory and never kept or shown.
• Wikimedia pageviews. Daily views of each English Wikipedia article, CC0, with scrapers filtered out by the mix of phone and desktop traffic.
• Bluesky. Weekly post counts, and a sample of public posts read for stance. A label and the post’s address are kept, never its text.
• X. Weekly counts, retweets excluded. No post, handle or text is read, stored or shown.
• Google Trends. One input to the Asked part. The series itself is not shown.
• News feeds. Headlines from sources whose rights allow it, counted and labelled. We show the shares, not the headlines.
• AI answers. Google AI Mode, Gemini and Bing Copilot, read through SerpAPI. Scores, engine, week and source domains are kept. The answer is read once to score it and is never stored, logged or shown.
• StatCounter Global Stats. Google’s share of search, used in setting the channel weights. CC BY-SA 3.0.

See this week's most asked about