Get Plus

Sign in to baba News

One account across the web, iPhone and Android — your subscription follows it.

or use an email code

Search stories

Type at least two characters. Results come from every newsroom baba reads.

↑↓ to move · ↵ to open · esc to close

Sign in to baba News

Sign in to keep asking. News Plus removes the daily limit.

or use an email code

Methodology

Last updated October 6, 2026

How an article on an Israeli news site becomes a short digest here, how reports of the same event become one story, and where each step can go wrong.

1. Collecting the news

baba News reads 44 Israeli newsrooms and 7 official sources through 59 feeds, in Hebrew, Arabic, Russian, English and French. A newsroom counts once however many feeds it has: Ynetnews is Ynet’s English edition, and i24NEWS files in four languages. The full list is on our sources page.

We look for new articles every five minutes through the day and evening, and every fifteen minutes in the small hours (23:00 to 04:00 UTC). We store an article’s text so a digest can be written from it, and we always keep the link to the original. We go back to articles from the last 24 hours in rotation, and when the newsroom has changed the text we write the digest again.

In the seven days from September 29 to October 5, 2026, we collected 22,353 articles, about 3,200 a day.

2. What we leave out

Not everything we collect is published. Before a digest is written, the original text is checked twice: is it an article and not an index page, a ticker or navigation text, and does it contain instructions addressed to an AI model. The digest step also marks foreign wire stories with no Israeli angle, such as another country’s weather or sports league, and we do not publish those. Of the 22,353 articles in that week, 4,927 were left out for having no Israeli angle and 728 for not being articles. 15,942 were published with an English digest.

3. The digest

An AI language model reads each article and writes a digest of it in its own words. It is not a sentence-by-sentence translation, and we never publish the newsroom’s own article text. The model is told to cover every material fact in the source, to keep a neutral tone, and to add no commentary and nothing that is not in the article. Each digest has a body, a summary of a few sentences and up to seven key points. Every article gets two digests, one in English and one for our Hebrew edition.

  • Headlines. The model writes a new headline in natural news English. It is not a word-for-word translation of the original. In the Hebrew edition, a Hebrew newsroom’s headline is shown as the newsroom wrote it.
  • Reports published in English. We show the key points and a link to the newsroom. We do not publish an English body that would compete with the newsroom’s own.
  • Paywalled articles. When a page shows only its opening lines, the digest covers only those lines.
  • Names. The model must list every person it names, with the words of the article that name them. If a digest names someone the article does not, it is written again once. A name that still is not in the article is taken out.

The same step sorts the story into one of ten sections, rates how much it matters from 1 to 5, and records the kind of piece (report, analysis, opinion, live blog, news flash, interview or press release), how the original headline is built, and up to three claims the article attributes to someone.

A digest can still be wrong. It can mistranslate a word, soften or sharpen a claim, or lose nuance. Every article page says so: it carries the line “Translated & summarized from [newsroom] by baba”, or “Summarized from [newsroom] by baba” when the source is in the page’s own language. As of October 6, 2026, digests are written by Google’s Gemini 2.5 Flash-Lite.

4. People, organizations and places

Each digest lists the people, organizations and places the article is about, so coverage can be browsed by name. The many spellings of one person are matched to a single identity. This is automatic, and it sometimes mislabels or misses a name.

5. Grouping reports into one story

When several newsrooms report the same event, we group their reports into one story. Each digest’s headline and summary are turned into a numeric fingerprint of their meaning (an embedding), and a new report is compared with the stories of the last 36 hours. A near-identical copy joins its story at once. For any other close match, a second AI model is asked one question: do these two digests report the same specific event? The report joins only on a clear yes. A clear no starts a new story. An unclear answer waits; we do not guess.

Stories about the same continuing affair are linked over days into an ongoing story. The Hebrew edition uses the same story grouping as the English one. Grouping can still merge two different events or split one event in two.

6. The five groups and the coverage bar

Every outlet we read sits in one of five groups, and the bar on a story counts its newsrooms by group. Each newsroom counts once: a second-language edition or a portal of the same publisher, such as Ynetnews beside Ynet or Mako beside N12, is not a second outlet. “Who covered it” on a story lists those newsrooms by group.

  • Left (3 newsrooms). Outlets on the Israeli left: Haaretz, Davar, Local Call.
  • Centre (11). The national mainstream: the broadcasters and the big portals and dailies. Right-wing readers see several of them as leaning left, and we say so.
  • Right (9). Right-wing outlets, including the religious-Zionist press (Arutz Sheva, Srugim, Kipa, Makor Rishon).
  • Haredi press (3). The ultra-Orthodox community’s news sites. Their framing sits to the right of the average newsroom; they are shown on their own because their agenda is the community’s.
  • Arab press (5). Arabic-language outlets written for Arab citizens of Israel, including Makan, the public broadcaster’s Arabic service.

Outlets outside the groups are counted as other (13): the Russian-language, business and sports outlets, and outlets the evidence cannot place. They select for a language or a beat, so their silence on a story is expected. They have their own grey segment in the bar and are listed in the coverage panel, but a story is never said to be missing from them. Official sources (the Prime Minister’s Office, the health and foreign ministries, the police, the IDF, Home Front Command and Magen David Adom) publish statements, not coverage, and are not counted at all.

The religious-Zionist press is in Right because on the political questions we track, from the government to the use of force, it frames the news the way the right-wing press does. A story the religious-Zionist sites skipped is almost always the same political claim as one the right-wing sites skipped, and together they are several newsrooms with several owners, not two sites with one.

The Haredi and Arab press have segments of their own because each is a community with its own agenda, which shows in the stories it chooses to run, and neither sits honestly on a left-right line. The Israel Democracy Institute builds its left, centre and right camps from Jewish respondents only, because Arab citizens are not part of that division. Counted as right, the Haredi sites’ community news turned ordinary stories into right-leaning coverage.

The line beside the bar (“67% right-leaning”, “Mixed coverage”) is read over the left, the centre and the right only, and only when at least two of the story’s newsrooms are in them. Each outlet also keeps its own lean, shown on its source page: the Jerusalem Post is in the Centre group, and its own rating reads center-right. How groups and lean are assigned is described in our editorial standards.

Not Everywhere marks a story that one part of the press ran and another did not. It is only about the left, the centre, the right and the Haredi press. A story is marked when all of these hold: at least three newsrooms have it; it is rated at least 3 of 5; it is at least an hour old; one of those four groups, with at least two newsrooms publishing that day, has not run it; and another of them has run it in at least half of its newsrooms that published that day. Opinion pieces are not counted. The story then carries a line such as “No coverage observed in the Haredi press”. A story that only the Arab press ran is marked “Only in the Arab press”. The Arab press is listed as missing where it is, but is never the reason for a verdict on its own, and the other outlets never are. “No coverage observed” means we did not find a report in the feeds we read. A newsroom may have covered the story somewhere we do not read, or later.

7. How a story reads: tone and stance

Every digest is also read for three separate things, and we keep them apart on purpose. Valence is what happened, from catastrophe to relief; it belongs to the event, and every outlet reporting a funeral shares it. Tone is how the copy is written: neutral and attributed, charged, or inflammatory throughout. Stance is who the copy is for or against, asked separately about the sitting government, about force versus restraint, and about each tracked person or institution the piece names.

These readings are source-blind by construction. The model that answers is given our English digest and nothing about who published it, so a masthead cannot sway the answer. The questions are written as observable situations (who is quoted, which adjectives, what is left out), never as political positions, so the same rubric judges a hostile column about a left-wing figure and one about a right-wing figure. A test set of mirrored pairs checks that it does. As of October 6, 2026, these readings come from TypeSafe’s Jev judgment model.

A group’s stance toward someone is never a raw score. Each article is first read against the average of every newsroom on the same story. A group that only covers a person’s bad days is not “hostile” when everyone reported the same days, and a group that writes the same event two levels harder than the average is exactly that. A group needs at least five judged pieces in a 7-day or 30-day window before it appears. A single article shows a one-word tone label (Neutral, Charged or Inflammatory) and never a number, because repeated reads of the same text drift slightly.

What this cannot do: it reads the digest, not the original layout, placement or photo; it does not know what an outlet chose not to write about at all (that is the coverage bar’s job); and it is a statistical reading, not a verdict on anyone’s honesty.

8. Comparing the reports on a story

  • How it broke orders the newsrooms by the time each first published on the story. Where a newsroom’s page gives a date but no time, we use the time we first saw the article.
  • How the headlines differ is written by an AI model from the newsrooms’ first headlines, in the original and in our translation: one neutral sentence and the wordings the headlines share. It describes wording, not which report is right. It is made for a few stories at a time, so many stories do not have it.
  • Who reports what takes the claims the lead report attributes to someone, and an AI model checks which of the other newsrooms report each one. It is made for some stories with at least three reports.
  • How it reads shows the average tone across the newsrooms and, on some stories covered by at least three groups, the tone of each newsroom’s original headline.

9. Ranking, Trending and Most read

Stories are ranked by scores, not picked by an editor. A story’s score rises with its 1 to 5 rating, with how much of each group covered it and with how fast new newsrooms are picking it up, and it falls as the story ages. A story is marked Breaking when it is under two hours old and, as a rule, at least three newsrooms filed on it within 90 minutes. We can pin or hide a story by hand; such an override always carries an expiry time.

Trending blends coverage with attention: how many newsrooms picked a story up in the last three hours, how many have it in all, and readers, shares, questions to Duki, follows, likes and clicks through to the newsroom. Each input is damped so one large number cannot swamp the rest, and the score fades as the story’s latest report ages.

Readers are counted as people, not page loads. On the website a reader is one browser on one network on one day; in the app it is one installation. Since October 6, 2026, a website open counts as a reader only when the page recorded at least five interactions (a scroll, a tap, a pointer move, a key press), because automated traffic was passing the earlier, lower bar. Opens in the app are held to a daily limit per installation. Most read ranks stories by readers in the last 24 hours and hides itself until at least three stories have five readers each.

10. Pictures, clips and portraits

  • Illustrations. Story pictures are generated by an AI image model and are labeled “Illustration by baba”. They are symbolic: the model is told to draw no people, no text and no logos. They are not photographs of the event.
  • Newsroom clips and photos. These are the newsrooms’ own posts on YouTube, Instagram, Telegram, X, TikTok and Facebook. We show them through the platform’s own embed and link to the post. One exception: Instagram’s embed does not load for some readers, so we keep a copy of an Instagram Reel attached to a story for up to 30 days and play that. We do not show the photos a newsroom publishes inside its articles.
  • Official photos. Photos released by the IDF Spokesperson and by Magen David Adom are copied and shown with their credit.
  • Portraits. Portraits of people are AI watercolours made from freely licensed photos. The photos, photographers and licences are listed on our portrait credits page.

11. What baba News writes and says itself

The daily brief, The Desk (an hourly spoken bulletin from 06:00 to 23:00 Jerusalem time; a quiet hour gets none), the Good Morning, Israel podcast and our posts on social media are written by AI models from the stories above. The Desk and the podcast are read by AI voices; the podcast’s two hosts, Maya and Daniel, are not real people.

Ask Duki is an AI assistant. What it says about the reporting is taken from the coverage we have collected and cited. A question can also ask for context the articles do not contain; that part is labeled Background (רקע in Hebrew) and is the model’s general knowledge, not a citation and not a live lookup. After each answer, a second model checks the cited sentences against those reports, and the answer shows how many were supported. The background section is not part of that count. Duki can still be wrong.

The Desk, the podcast, our social posts, Duki’s answers and the summary at the top of a full coverage page, which baba writes from the newsrooms’ reports, are baba News speaking in its own voice. They follow a house style that the digest of a single article does not. Our editorial standards set it out.

12. Indexes and forecasts

  • News Mood is the valence reading turned into a daily number from −1 to +1. Reports of one event are averaged into one story, widely covered stories weigh more, and sport, weather and culture are left out. The page explains the measure.
  • The Mamad forecast estimates the chance of a wide siren wave from Home Front Command alert history and a reading of the news. Its method page gives the detail and the limits.
  • The election desk lists the polls the Israeli press publishes. An AI model reads each poll out of the article, and a poll is kept only if its figures pass checks. The average is the plain mean of each outlet’s latest complete poll from the last ten days.

13. Linking out

Every digest links to the original article at its source. baba News is a way into the Israeli press, not a replacement for it.

What this is and is not

This pipeline collects and digests other people’s journalism. It does no reporting of its own and it does not check the newsrooms’ facts. Its AI steps can err. For anything that matters, read the linked original. If you spot a mistake at any stage, email news@itsbaba.com. Our corrections page explains what happens next.