An invoice for five kilograms of textured soy protein, a crepe recipe searched in September 2025 and a Sunday evening set aside for anime are among the personal records that Google's Dreambeans app cited to explain the stories it served one user, according to an analysis published today by Search Engine Land.

The article, written by Damien Andell, co-founder of 1492.vision, and co-written by Olivier de Segonzac, examines the stories and metadata that Dreambeans generated from Andell's own Google account - searches, emails, invoices, purchases, calendar events, subscriptions, YouTube videos, Gemini conversations and other activity records - together with the Android version of the app released in early September, which carries the internal codename Lotus. It runs under the headline "Dreambeans: Google turns your private life into personalized content." The evidence behind that framing is narrower than the headline: one account, a selection of the English-language text the app produced for it, and what the authors found in the client code.

Dreambeans is a Google Labs experiment that produces a fixed number of illustrated stories each day. On September 10, Google opened it to all consumer accounts in the United States for users aged 18 and over, on Android and iOS, including accounts without a paid Google AI subscription, and added Gemini to the apps it can connect to. Google describes Dreambeans as a "proactive lifestyle AI product," according to the analysis.

In Short

A Search Engine Land analysis published today shows how Google's experimental Dreambeans app turns Search history, Gmail and Calendar, YouTube, Google Photos and Gemini chats into a daily feed of illustrated stories, and explains each pick by citing dated searches, invoices and subscriptions. It matters to the US adults who can now connect those apps without a paid plan, and to publishers, because each story arrives already summarized inside the app, with the source website reduced to a link or a button. Deleting an email or a search does not remove stories already built from it, and the findings rest on one account and the app's code, so they show what Dreambeans can draw on rather than how it behaves at scale.

Three months from waitlist to every US account

Google Labs introduced Dreambeans on June 3, 2026. The launch post, by Labs product manager Gozde Oznur, made the app available to eligible Google AI Ultra subscribers aged 18 and over in the US on Android and iOS, with a waitlist for everyone else. With the user's permission, it drew on Gmail, Calendar, Photos, YouTube and Search history through Personal Intelligence, and it needed at least one connected app to work, according to Google.

Access widened quickly. 9to5Google reported on August 28 that free Google accounts in the US could test the app, after Google had extended it to AI Pro subscribers earlier that month. Google made the broader availability official on September 10, listing the connected apps as Calendar, Gmail, Photos, Search, YouTube and, for the first time, Gemini. The illustrations come from Google's Nano Banana 2 image model, according to Google's product page.

One detail does not line up. Google's Labs product page, which carries no date, says Dreambeans is available to users aged 18 and over in the United States and India. The September 10 announcement mentions only the US.

Personal Intelligence itself is older. Google launched it in beta on January 14, 2026 for AI Pro and AI Ultra subscribers in the US, connecting Gmail, Google Photos, YouTube and Search to Gemini. It was off by default, and Photos access required Face Groups to be switched on. Dreambeans takes that same pool of connected data and points it at a different job: not answering a question, but deciding what to show before anyone asks.

Who did the digging

Andell is no stranger to Google's feeds. His Search Engine Land author biography describes him as a Python developer, Google Bug Hunter and algorithm specialist who built proprietary technology for collecting and analyzing data from Google Discover, the personalized feed in the Google app. PPC Land has reported his findings before. In April 2026 he spotted a "You asked to see" label that Discover added to cards picked through its Tailor Your Feed chat, and in September 2026 he reported four configuration flags behind Discover's "Dive deeper" button test.

The Dreambeans piece is a contributed article. Search Engine Land's editorial disclosure says contributing authors are chosen for their expertise, work under the oversight of the editorial staff, and express their own opinions. Because the account under study is Andell's own, his work shows up in the results: one of the stories reproduced in the analysis suggests benchmarking Google Discover swings against the 1492.vision Discover Volatility Index. The article does not include a response from Google.

According to the analysis, Dreambeans can draw on five categories of data. Search covers past searches, saved places and data from Search, Maps, Shopping, Flights, Hotels, Translate, News, AI Mode and Discover. Workspace covers emails, files and events, notably from Gmail, Calendar and Drive. YouTube contributes interests derived from activity on the platform. Gemini supplies chat history, information shared with the assistant and insights already drawn from those conversations. Google Photos contributes photos, videos and information that can be used to infer interests, relationships, places and moments.

The same five categories appear as consent declarations in the app's code: PERSONAL_CONTEXT_LOTUS_USING_SEARCH, WORKSPACE, YOUTUBE, PHOTOS and GEMINI_CHATS. The confirmation screen a user sees after connecting lists Google Workspace, Google Photos, YouTube, Search services and "Gemini App, including chat history."

The Gemini source reaches further than a transcript. According to the analysis, Google states in the app that the shared data may include "Gemini chat history and insights drawn from it." Dreambeans can therefore receive both what was said to Gemini and what Google has already inferred from those conversations. Once the source is switched on, the app says Gemini history will be used in the following day's stories.

Personal Context as plumbing

Google groups the mechanism under the name Personal Context. In the consent libraries, the authors found it expressed as a matrix:

PERSONAL_CONTEXT_USING

The name has a history. In August 2025, "Personal Context" was the label on a Gemini setting that let the assistant learn from past conversations, switched on by default for eligible users outside the EU, UK and Switzerland. In the Dreambeans code it reads less like a single feature and more like an internal layer. The authors infer that Dreambeans, identified as LOTUS, "appears to be one of several consumers of a broader infrastructure," and that its current combination relies on Search, Workspace, YouTube, Photos and Gemini Chats. The analysis names no other consumer.

The design explains how records from different services meet. An invoice in Gmail, a search, a Calendar event and a video can be connected when they describe the same activity, project or interest.

What the stories said about one person

The text Andell collected shows how specific those connections become. Each story carries fields named storyWhat, storyWhy, storyTag, confidence, expireTime, feedRecencyCutoff and storyChat, among others: subject, rationale, category, confidence level and period of relevance are part of the data model, and every story can open its own conversation.

Take food. One rationale reads: "Mon Epicerie Paris invoice documents an August 2026 bulk purchase of 5kg textured soy protein medallions, paired with HappyVore plant-based merguez orders (June 30, 2026) and strict vegan and gluten-free dietary compliance." Another states that it is "Grounded in user profile records confirming a strict 100% vegan, plant-based diet with a preference for gluten-free whole foods and high-protein legume bowls." Products, quantities, retailers and dates end up as a portrait of a diet. The same diet resurfaced in a story about JAPAN Fes NYC, which listed plant-based Japanese street food, gave the event's location and opening hours, linked to the official website and offered a follow-up question: "Any vegan street food?"

Work was handled with similar precision. "User is the co-founder and lead developer of 1492.vision, where he engineers real-time web analytics dashboards using Vue 3 and Vuero admin frameworks," one passage reads, before moving on to "modular composables and shallowRef optimizations for high-frequency data applications."

Searches, with dates attached

Search history appears verbatim. A story about PC performance cited "explicit searches for network adapter driver settings ('transmit buffers ethernet gaming') on August 9, 2026, and past diagnostics for server packet drops." A cooking story went back further: "Grounded in the user's historical query for 'recette crepe' and visit to French recipe portal cuisine.journaldesfemmes.fr on September 3, 2025." That record was roughly a year old when it resurfaced, and the rationale paired it with "culinary science research on gluten relaxation and pan temperature standards."

The wording separates channels. "Search history" and "Search logs" refer to queries, "watch history" to videos and "navigation" to website visits. A single passage cited a video watched on August 19, 2026, searches for "discover" on March 17, 2026, SEO consultant research in April 2025, an explicit preference for Google Search news recorded on August 29, 2026, and "the completed Google Spam Update on August 21, 2026." Google's third spam update of the year began rolling out on August 18. Another rationale reached back to a visit to the cocktail publication PUNCH on December 6, 2025, and to searches about Japanese sake brewing in May and June 2025.

Purchases and hardware were joined up as well. "Transaction history shows PC game key acquisitions (Call of Duty, Ratchet & Clank, Battlefield) via G2A and Gameseal, paired with high-end desktop hardware (Samsung 990 PRO 4TB SSD, Kingston 32GB DDR5 RAM)," one passage reads. A single card, in other words, can rest on months of activity from several services.

Manifold, two prompts and a two-week queue

Generation happens on Google's servers, in a pipeline the code calls Manifold. Dreambeans schedules an update, starts processing and tracks its status through three states: SCHEDULED, ONGOING and FINISHED. The client exposes two prompt versions, generateTopicsPromptVersion and generateStoriesPromptVersion, which the authors read as a two-stage process: topics are first extracted from Personal Context, then stories are written from those topics and the available context. The prompts themselves stay on the server.

Topics have a lifecycle of their own, with two active levels followed by an ARCHIVED state, so an interest can fade without disappearing from the profile. Generation also continues while the app is closed. Dreambeans keeps the credentials it needs for background processing, and after an extended period of inactivity it pauses daily generation until the user returns.

Runtime traces add a calendar to the picture. The authors observed Personal Context keys materialized by date, such as user::pcontext:2026-09-05 through user::pcontext:2026-09-18. In their sample, fourteen dated contexts coexisted, roughly two weeks of queued material. The traces show three processing paths: a dated context prepared in the regular queue, a "sameday" catch-up pass, and a "real" branch fed in parallel with more immediate signals.

The schedule they observed ties each dated key to an overnight run at around 07:00 UTC on the previous day. The production sequence visible in the traces runs context, stories, story_selection, pcontext:dates and images. The feed fills only once the images are ready, which the authors connect to the in-app message that the first stories may take "a few hours." Google's product page says setup can take up to a day.

The clock as a ranking signal

A conventional recommender system predicts what a person is likely to click. Dreambeans, in the authors' reading, tries to work out what deserves attention at a given moment, and its data structures reflect that. Stories carry MORNING, AFTERNOON and EVENING time slots, an expiration time and a feedRecencyCutoff, along with fields named relevanceAndTimeWindowOfInterest and earliestDisplayLocalTime.

The anime story shows the mechanism at work: "Streaming calendars confirm August 30, 2026 episode releases for Mushoku Tensei Season 3, One Piece Elbaf Arc, and Grand Blue Season 3. User subscribes to Rakuten Viki and Netflix. Ideal for Sunday evening, August 30, 2026, offering low-effort entertainment as part of the weekly anime broadcast schedule." Release schedules, subscriptions and a time slot are combined into a single recommendation.

Place and season work the same way. "User is based in a small town on September 1, 2026, immediately adjacent to the Verdon Regional Natural Park," a hiking story begins, before naming the Sentier Blanc-Martel and the Sentier de l'Imbut and describing an "Evergreen interest topic, particularly fertile throughout September and October 2026 due to ideal mild weather conditions and active trail shuttle services." A geographic detail, an interest in walking and early-autumn conditions produced a two-month window of relevance.

The examples Google itself offers during onboarding point in the same direction, according to the analysis: spotting that a passport expires six weeks before a trip planned eight weeks later, finding a dentist appointment buried in an old email, identifying a sourdough workshop after a run of baking videos, or suggesting twenty minutes of Pilates because the user asked for shorter sessions.

Outside facts, inside the story

Several stories mixed the personal profile with recent information: broadcast schedules, game releases, the seasonality of an activity, local conditions. Rationales use phrases such as "Search verification confirms" and "Verified video game release calendars confirm." In the data structures for story-specific chats, the authors found fields named groundingEvidenceAndProvenance and mediaProvenance, along with a webSearch string. Those point to a grounding layer, though the authors say they have not captured enough payloads to reconstruct its precise role in a story's first draft.

Across the rationales, the working text keeps returning to five operations, by the authors' count: linking the story to personal evidence, summarizing what that evidence says about the user, checking external facts, assessing freshness or the window of interest, and explaining why the subject deserves to be shown now. Each story exposes the result through a "Why did I get this?" button that reveals the storyWhy text. That transparency cuts both ways. An unremarkable recommendation, the authors note, can reveal how many separate pieces of information were assembled to produce it.

Location, Photos and the "who, where, when"

Dreambeans declares Android location permissions, but the authors found no calls that use them directly in its logic. Geography arrives by other routes. Google explains that Photos data can contain a location estimated from the content or attached to the file, which may itself come from the device, and Search contributes places saved through Maps. A local recommendation can therefore carry geographic context without the app querying GPS when the story is generated.

The description of Google Photos inside Dreambeans is the most explicit passage the authors found. Google states that Photos can be used to infer the user's interests, relationships with people in the images and places they have visited, and that the user's face can be associated with location data and the corresponding timestamps. Activation relies on Face Groups and on a component called Photos Life Understanding. Google adds that this data can be used to improve its services, including to train generative AI models.

Fourteen ways to say no

The dislike screen, built in a component named dislike_feedback_view.dart, asks "Tell us what went wrong" and offers fourteen answers in four groups. Under Accuracy & Substance: "My personal info is wrong," "Topic facts are wrong" and "Lacked depth." Under Visual: "People or pets off," "Irrelevant people" and "Objects or setting off." Under Relevance: "Not for me," "Repeated story," "Seeing this too much" and "Steering ignored." Under Timing: "Already happened," "Old news," "Plans changed" and "Wrong timing."

The first option is the telling one. Google has built in an explicit route for cases where the picture assembled from an account is simply wrong.

Users can also state preferences directly, through interstitials with a "Type your own" field or in chat. In one exchange, Andell wrote "I want more stories about SEO and algorithms," and the app replied that it had updated the feed. In the data model, a preference has text, a polarity (POSITIVE or NEGATIVE) and a source (CHAT, CHIP, INTERSTITIAL or MANUAL). After a "Not for me" response, the app may confirm: "I've noted your preference to move away from stories like this one." Those requests later reappear in story rationales, as with "The user explicitly requested technical deep dives on Knowledge Graphs, entity resolution, and semantic search architecture on Aug. 30 and 31."

Explicit corrections sit on top of behavioral telemetry. The events the authors list include CARD_IMPRESSED, CARD_SWIPED_PAST, VIEW_STORY_FULL_PAGE, VIEW_STORY_FINISHED, STORY_THUMBS_UP, STORY_THUMBS_DOWN, STORY_SAVE, STORY_SHARE, STORY_FULL_IMAGE_VIEW and INTERACTION_TYPE_LONG_PRESS_FULL_IMAGE. The steering by chat echoes Discover's Tailor Your Feed experiment, where cards chosen through a conversation are marked "You asked to see."

Deleting the source does not delete the story

Google states that a change or deletion in a connected app can take 24 to 48 hours to be reflected in new stories. Cards already generated stay in the feed until the user deletes them. Source data and the content derived from it - topics, stories, images, explanations - therefore have separate lifecycles. As the authors put it, "deleting the original information does not automatically delete all the representations already generated from it."

A global "Delete All Dreambeans Data" command covers topics, stories and "profile information." The confirmation screen says the action is irreversible, does not delete the Google account and does not affect other Google services. The profile built by Dreambeans thus appears as a named dataset of its own. Deleting an email in Gmail, removing a story and wiping Dreambeans data each have a different reach.

Google also states that Dreambeans permissions are managed independently of the corresponding settings in other Google services, and the June launch post said choices made in Dreambeans do not change Personal Intelligence settings elsewhere. In the other direction, the app indicates that data from Dreambeans may be shared with other Google services. The client does not say which ones, or for what.

That gap matters to the advertising market because of what changed in October. Google's privacy policy dated October 16, 2026 added an exception to its July 2021 pledge not to personalize ads using content from Gmail, Drive or Photos, covering cases where users let Search or Gemini Apps access those services. The Search Engine Land analysis does not mention advertising, and nothing in it shows Dreambeans data reaching Google's ad systems.

Publishers get a button

For publishers, the comparison is with Google Discover. A Discover card links to a publisher's page. A Dreambeans story, once opened, already presents expanded text and an illustration inside the app, with links to external sites and the option to keep talking. In the stories shown, the Discover benchmarking card linked to 1492.vision and to Search Engine Roundtable and carried a "Read on SERoundtable" button, while the JAPAN Fes NYC card offered a "Visit Official Site" button.

Discover is no side channel for news sites. NewzDash data covering more than 400 publishers showed Discover's share of Google mobile traffic to news rising from 37.03% in 2023 to 67.51% in the fourth quarter of 2025, while web search fell from 51.10% to 27.42%. The feed has also been filling with Google's own material: Marfeel data reported in December 2025 found that 51% of Discover feed positions in the US, Brazil and Mexico were AI Summaries, and that 77% of US exits from those summaries went to inline YouTube plays. Dreambeans goes a step further, building the summary around the reader's own life rather than around a single article.

The authors are careful about what their evidence can carry. A mention in a story, an in-text link and an outbound button do not give a source the same visibility, they write, adding that the formats they observed do not allow them to measure the traffic those formats generate. They close the section with a question: "When a story already provides a personalized answer, what still makes someone want to open the source website?"

The question is not abstract for marketers either. Dreambeans is a working example of hyper-personalization applied to content selection rather than to ad delivery, using signals such as invoices, subscriptions, saved places and photos that sit inside Google's own services.

What the analysis does not establish

The findings come from one account, the authors' own, and from a selection of English-language text. They describe what Dreambeans can draw on and how it explains itself, not how often each source is used or how the feed behaves for other people. Several conclusions are inferences from code: that Personal Context serves other products, that generation runs in two prompt stages, and how the grounding layer works. The authors say plainly that they found no direct use of location permissions and could not reconstruct the grounding layer's role, and that the scope of data sharing from Dreambeans to other Google services is not detailed in the client. The stories place the account in a town in south-eastern France, while Google describes availability as limited to the US, or to the US and India on its product page; the article does not explain how the account was enrolled. No traffic figures are offered, and Google did not comment in the piece.

What the analysis does document is the machinery. Behind each illustrated card sits a rationale that can name a retailer, a date, a search term or a subscription, and a pipeline that prepares those cards up to two weeks ahead. As the authors frame it, the question has shifted from "What is this person likely to click on?" to "Given what we know about their life, their interests and what is happening now, what deserves their attention today?"

Timeline

Summary

Who: Google, through its Google Labs app Dreambeans; Damien Andell, co-founder of 1492.vision, and co-author Olivier de Segonzac, whose analysis Search Engine Land published as a contributed article.

What: An analysis of the stories, metadata and Android code of Dreambeans, which builds a daily feed of illustrated stories from five connected Google sources - Search, Workspace, YouTube, Photos and Gemini chats, including "insights drawn from" Gemini conversations. It documents rationales that cite dated searches, invoices, subscriptions and hardware purchases; a server-side pipeline called Manifold with separate topic and story prompts; around two weeks of dated Personal Context queued in advance; time slots and windows of relevance; a fourteen-option dislike taxonomy; and deletion rules under which stories already generated stay in the feed until removed by hand.

When: The analysis was published today, November 13, 2026. Dreambeans launched on June 3, 2026, and Google opened it to all US accounts aged 18 and over on September 10, 2026. The Android version examined was released in early September 2026.

Where: Dreambeans is available on Android and iOS in the US, according to Google's announcement, and in the US and India, according to its undated product page. The account analyzed is placed by its own stories near the Verdon Regional Natural Park in France.

Why: The analysis shows how much personal data from separate Google services can be combined into one recommendation, and that deleting a source does not remove what was generated from it. For publishers, Dreambeans presents a finished, personalized summary with the source site as an optional link or button, extending a shift already visible in Google Discover. For the advertising market, it arrives weeks after Google's privacy policy began allowing connected Gmail, Drive and Photos data to shape ads, although the analysis does not show Dreambeans data being used for advertising.