A jacket for Oslo: one query, two answers depending on the account
Take the query [a warm jacket for my trip to Oslo]. Without an account, AI Mode answers from web pages: an outdoor magazine’s guide, a forum thread, a comparison. For a signed-in user, it starts from the same web cache and also knows when they leave, which brand they have already bought and what they asked it to remember.
Google brought Personal Intelligence to AI Mode on January 22, 2026, a week after the Gemini app, with a similar example: a coat chosen from the user’s preferred brands and their flight confirmation in Gmail, for Chicago in March. Since May 19, 2026, Gmail can be connected with no subscription in nearly 200 countries and territories (Google’s help), outside the European Economic Area, the United Kingdom and Switzerland.
per query, 3 of which decide
on top of the web
including written profiles
What personalizes an answer at Google
The settingsPersonal Intelligence
- the public name of Google’s account-based personalization, in AI Mode, Ask Maps, the Gemini app and Gemini in Chrome
- in Google’s code, it reaches the classic SERP too
- in Search it draws on Search Services History, memory, Gmail, Google Calendar and Google Photos
The decisionmagi
the AI of the results pages, AI Overviews and AI Mode; it decides at every query whether to personalizeThe queryCopley
a Google Search component that links “my hotel” or “my flight” to the user’s bookingMemoryPensieve
the user’s memory service; it summarizes that memory for AI Mode answersThe personal indexMoonshine
a Google index of an account’s personal data, including bookings, orders and bills drawn from GmailRankingP13nRank
Google Search’s personalized ranking (p13n = personalization), from query understanding to twiddlers, the rules that reorder results at the end of the chain
AI Mode keeps a non-personalized web cache; the profile arrives through sixteen separate sources
AI Mode runs its web searches, those of its fan-out, through its tools: the functions the model calls to look something up (web search, Gmail, Google Calendar…). For the web, these tools have a cache of Google Search results that Google itself labels, in its code, as non-personalized.
This cache leaves the account aside. Location, language, device type and current searches still apply, as Google uses them even without personalization (Google’s help): in the same context, two users, signed in or not, start from the same web results.
Two personal elements still come on top: the badge of preferred sources, picked by the user, and, for shopping, the products Shopping selects for the buyer.
What Google knows about the user arrives separately. AI Mode queries it the way it queries the web, through sixteen personal sources in five groups: the account, memory, connected Google apps (Gmail and Google Calendar), the search journey, and shopping and travel.
Gmail and Calendar go through MCP connectors (Model Context Protocol, the protocol that links a language model to tools).
For a brand, the web layer can be measured: without an account, you see the pages cited, their order, the brands they name. The personal layer comes on top for each user: the brand they already bought, the date of their flight, the outlet they chose. Our “Google Shopping, from the SERP to AI Mode” infographic covers what this changes for shopping.
Web layer
The pages the AI fetches
- AI Mode’s tools have a cache of Google Search results labeled as non-personalized
- Without personalization, location, language, device and current searches still count (Google’s help)
- The same cache for a user with no account
AIMToolCacheNonPersonalizedGSRContext
Personal layer
Sixteen separate sources, queried like the web
The account
CeUserFactsResultwhat Google knows about the userPersonalProfileMergingResulttheir profile, merged from several sourcesUserContextResulttheir current situationUserInfotheir personalization info
Memory
PensieveSummaryResulta summary of their memory, by Pensieve, the user’s memory serviceSavesRecallResultwhat they have saved in GoogleSavesListResulttheir saved listsSavesItemResultan item they saved
Connected Google apps
GmailMcpExtensionGmail, through an MCP connectorCalendarMcpExtensionGoogle Calendar, through an MCP connector
The journey
JourneyDetectResultspots a series of searches on the same projectSearchJourneyResultthe search journey under way
Shopping and travel
PurchaseContextResultthe purchase under wayTravelPersonalizationResultthe tool that personalizes travel, in AI Mode and Ask MapsConsumableReorderResultreordering a product they use upAddedInputTravelBookingResulta travel booking attached to the request
The AI Mode answerweb pages, plus what Google decides to add for this query
At every query, Google asks three questions before personalizing
- 1Is the user eligible?18 or older, Search Services History and Personalized Recommendations on
AimPersonalizationUserEligibilitySignalsYesYes - 2Does the question call for it?a lot for a travel jacket, not at all for the date of a battle
AimPersonalizationIsValuableSignalsYesNo - 3Does it pass the safety check?a check specific to personalization, on top of AI Mode’s own
AimP13nSafetySignalsYes - Then the personal sourcesprofile, memory, Gmail, purchases, travel
For every query, Google has ten personalization signals available. For AI Mode, three of them decide; Ask Maps, Google Maps’ conversational search, and Google Finance’s AI make their own call.
- The first of the three says whether the user is eligible: users must be 18 or older, with Search Services History and Personalized Recommendations turned on (Google Search Help).
- The second says whether the question calls for it: a lot for the jacket, where the travel date and preferred brands change the answer; not at all for the battle of Hastings, whose date is the same for everyone.
- The third is a safety check specific to personalization, on top of the one AI Mode runs on every query.
The other signals say what to personalize with. Explicit memory keeps what the user asked AI Mode to remember: “Remember I prefer Italian food”, the example in Google’s help.
In Google’s code, a separate setting lets Search draw on what the user tells the Gemini app. For Gmail and Google Calendar, Google Photos, Google Wallet, health data or Chrome, Google works out an effective setting: the user’s choice, once the age, country and language requirements set out in its help are applied.
Where it applies: what is available, region by region
| Feature | United States | Rest of the world, outside the EEA, UK and Switzerland | EEA, United Kingdom, Switzerland |
|---|---|---|---|
| AI Mode personalization from historyfrom 18, with Search Services History and Personalized Recommendations turned on | Yes | Yes | Yes |
| AI Mode memory“remember that…” | Yes, in English | No | No |
| Connecting Gmailno subscription, nearly 200 countries and territories and 98 languages in all | Yes | Yes | No |
| Connecting Google Photos | Yes | In some countries | No |
| Connecting Google Calendar | Yes, in English | No | No |
| Ask Maps with Gmaildirections to the booked hotel or restaurant | Yes, in English | Canada (English, French), India (English, Hindi), Australia, Japan, Indonesia, Mexico, Brazil | No |
| Preferred sourcesthe badge in Top Stories, AI Mode and AI Overviews | Yes | Yes | Yes |
| Personalized SERPwith the “Try without personalization” link | Yes | Yes | Yes |
- Once Gmail or Google Photos is connected, every Search service gets access: Search (incl. AI Mode and Discover), Maps (incl. Ask Maps), Shopping, Flights, Hotels, Translate and News.
- Ask Maps is open in the US, India, Australia, Brazil, Canada, Indonesia, Japan and Mexico, and in English in over 150 countries and territories.
- Google also gives the name Connected Apps to third-party apps such as Instacart, Canva or YouTube Music that users link to AI Mode, in the US, in English (Google’s help).
Sources: Personal Intelligence help, help on connecting Google apps, May 19, 2026 post, August 6, 2026 post, preferred sources documentation, help on personalized results.
In the code: the query signals
AimPersonalizationUserEligibilitySignalswhether the user is eligible for personalization (Google’s help: 18 or older, history and recommendations on)AimPersonalizationIsValuableSignalswhether the question calls for personalizationAimP13nSafetySignalsa safety check specific to personalizationExplicitMemoryQuerySignalsexplicit memory: what the user asked AI Mode to rememberP13nSourcesthe sources of personalizationResolvedP13nToolConfigthe configuration chosen for the personalization toolPersonalizationSignala general personalization signal of magi (AI Overviews and AI Mode)FinanceAiPersonalizationEligibilitySignalsthe decision of Google Finance’s AIAskMapsP13nClassifierSignalsthe decision of Ask Maps, Google Maps’ conversational searchP13nEvalGroundTrutha ground truth, to evaluate the personalization produced
In the code: account settings
web_and_app_history_statusthe state of Web & App Activitysearch_services_history_statusthe state of Search Services Historysearch_services_personalization_statusthe state of Personalized RecommendationsEFFECTIVE_PCONTEXT_SEARCH_USING_WORKSPACEthe effective setting for Gmail and Calendar: the user’s choice, once access requirements are appliedEFFECTIVE_PCONTEXT_SEARCH_USING_CHROMEthe effective setting for ChromePERSONAL_CONTEXT_SEARCH_USING_GEMINIthe setting that lets Search draw on what the user has shared with the Gemini app
Google writes a profile of each user for its models
Google writes profiles of its users in natural language, so that a language model can read them as they are. They are among the 392 kinds of profile data a Google service can request for a query. These data flow between Search, Maps, YouTube, Discover and ads: each produces and consumes them.
The general profile comes in versions drawn from search alone, from local search, or from search plus YouTube: a Search answer can rely on a profile drawn from searches alone.
Memory is written too: recent facts, what the user asked Google to remember, a version updated continuously, and a predictive memory, what Google anticipates. Pensieve, the memory service that supplies AI Mode answers with a summary, keeps its own profile.
Google scores, entity by entity, how familiar the user is with a brand, a place or a person, and keeps their areas of interest, topics, creators and places. Affinity for a topic has its own twiddler, a rule that reorders SERP results after ranking.
Other kinds of profile data describe their life: the people who matter to them, signals of a big life decision, their job-search, learner and traveler profiles.
Google also condenses the user’s tastes into embeddings, several per surface: the results page, Gemini, places, hotels, style in shopping. For shopping, it writes a shopping profile for the AI Mode model; the Google Shopping infographic covers it.
Google Search files forums, social networks and creators in a content family of their own, Human Voices (HuVo in its code). For every user, it surfaces more of this first-hand content: the hidden gems, announced in May 2023 as a ranking improvement (Google’s post). In its code, a twiddler is dedicated to them, Natural Pearls.
For each user, the profile also keeps an embedding of their taste for user-generated content, plus their affinity for creators. Autocomplete also offers them personalized Human Voices suggestions.
Profiles written for the models
A general profile, and profiles drawn from search alone, from local search, or from search plus YouTube.
Common prefix: NATURAL_LANGUAGE_USER_PROFILE
NATURAL_LANGUAGE_USER_PROFILEthe general profile…_SEARCH_ONLYdrawn from search only…_SEARCH_LOCALdrawn from local search…_SEARCH_YOUTUBEdrawn from search plus YouTubePENSIEVE_USER_PROFILEthe profile kept by Pensieve, the memory service
A written memory
What happened recently, what the user asked Google to remember, and what Google anticipates.
Common prefix: NATURAL_LANGUAGE_MEMORY
NATURAL_LANGUAGE_MEMORYmemory in natural language…_RECENTrecent…_EXPLICITexplicit: “remember that…”…_ONLINEupdated continuouslyPREDICTIVE_MEMORYa predictive memory: what Google anticipates, also updated continuously (_ONLINE)
Affinities
What the user knows, and the domains, topics, creators and places they care about.
ENTITY_FAMILIARITY · _SCORESthe entities they know, and a score for eachDOMAIN_INTERESTStheir areas of interestP13N_TOPIC_AFFINITYtheir affinity for a topicCREATOR_AFFINITYthe creators they likePLACE_INTERESTSthe places they care about
Their life
The people who matter, big decisions, jobs, learning, travel.
PEOPLE_WHO_MATTERthe people who matter to themBIG_LIFE_DECISION_SIGNALSsignals of a big life decisionJOBS_PROFILEtheir job-search profileLEARNER_PROFILEtheir learner profileUNIVERSAL_PROFILE_TRAVELtheir traveler profile
Their traces in AI Mode
Their notebooks, the state of their conversations with the AI, and Personal Intelligence.
SEARCH_AIM_JOURNEYStheir AI Mode notebooks, where they keep instructions, sources and threadsSEARCH_AIM_JOURNEYS_EXPLICIT_THREADSthe threads they filed there themselvesSEARCH_MAGI_STATEthe state of their exchanges with magi, the AI behind AI Overviews and AI ModePERSONAL_INTELLIGENCEthe state of Personal Intelligence for this account
Embeddings
72kinds of profile data are embeddings, several per surface
ASTRIA_SRP_USER_EMBEDDINGS for the results page · ASTRIA_GEMINI_USER_EMBEDDINGS for Gemini · LOCAL_USER_PLACE_EMBEDDINGS for places · HOTEL_HOBBES_USER_EMBEDDINGS for hotels · ASTRIA_KITE_UGC_USER_EMBEDDINGS for user-generated content · SHOPPING_USER_STYLE_EMBEDDINGS for style, in shopping
Personalization weighs most on shopping, travel and places
Google illustrates Personal Intelligence with shopping and travel, and that is where it has built its personalization tools: a travel tool in AI Mode and in Ask Maps, Google Maps’ conversational search; a hotel twiddler; personalized predictions in Shopping.
| Vertical | What Google reads | What changes on screen |
|---|---|---|
| Travel | the flight confirmation in Gmail, the traveler profile | in the SERP, the user’s flight instead of the Google Flights module; in AI Mode and Ask Maps, answers built around their booked trip |
| Hotels | bookings, searches for hotels and places to visit, hotel pages opened, confirmation emails | hotels tied to their history move up the list |
| Places | bookings, the local profile, the places they care about | Ask Maps gives directions to the booked hotel or restaurant |
| Shopping | purchases read in Gmail, the shopping profile, the purchase under way | brands already bought, products picked for this buyer |
The Google Flights module steps aside for the user’s flight
In the classic SERP, Google removes the Google Flights module when Personal Intelligence, its account-based personalization, answers instead, whatever intent it reads into the query.
Google shows flights and hotels found in Gmail for [my flights] or [my reservations] (Google Search Help). For [my flight to Oslo], the personal answer takes the place of the module and of competing airlines’ prices.
For an airline or an agency, the confirmation email decides what travelers see of their flight in Google.
In the code
SuppressionReason.Reason.PERSONAL_INTELLIGENCEPersonal Intelligence answers from the accountInterpretationIndependentSuppressiona removal reason that applies whatever meaning the query is givenTravelPersonalizationResultthe tool that personalizes travel, in AI Mode and Ask Maps
No personal answer
Google Flights · New York → Oslo
- Airline A · nonstop$612
- Airline B · 1 stop$489
- Airline C · nonstop$705
With Personal Intelligence
Your flight · confirmation in Gmail
JFK → OSLDecember 12 · 6:40 pm · seat 14C
The Google Flights module and its prices are removed.
A hotel personalization twiddler counts past bookings
In hotel results, a personalization twiddler, a rule that reorders results after ranking, takes the user’s history into account: their hotel bookings, their searches for hotels, places to visit and amenities, the hotel pages they opened. Their price preferences and how many guests they usually book for count too.
Google tests confirmation emails in it against a control group. It also records when a hotel was boosted thanks to Happyhour, a part of Moonshine, the index of an account’s personal data. Happyhour holds bookings and orders drawn from Gmail, searches, Maps searches, and saved or visited places.
A hotel tied to the user’s history starts with an edge in their results.
In the code
HAS_NO_HOTEL_BOOKINGSno hotel bookingsHAS_2_PLUS_BOOKINGStwo or more bookingsHAS_1_BOOKING_1_OR_MORE_SEARCHESone booking and at least one searchHAS_6_OR_MORE_SEARCHES_ONLYsix or more hotel searches, no bookingHAS_6_OR_MORE_POI_SEARCHESsix or more searches for places to visitHAS_6_OR_MORE_HOTEL_PLACESHEETSsix or more hotel pages openedHAS_USER_PRICE_PREFERENCEa known price preferenceIS_P13N_TWIDDLER_TRIGGEREDthe personalization twiddler firedHPS_EMAIL_CONFIRMATION_IN_P13N_TWIDDLER_EXPERIMENTthe test of confirmation emails in this twiddler, against a control group (_CONTROL)P13N_BOOSTED_BY_HAPPYHOURa hotel boosted thanks to Happyhour, the part of Moonshine that holds bookings, orders, searches and placesRESULTS_WERE_PERSONALIZEDthe results were personalized
In Shopping, candidate products differ from one buyer to the next
In Shopping, four prediction models have personalized versions: the ranking of free listings, predicted CTR, buyer-to-query matching and retrieval, which picks candidate products before ranking. So candidate products differ from one buyer to the next before they are even ranked.
In the code
ORGANIC_P13Nfree listings, personalizedPCTR_P13N_ADJUSTMENTa personalized adjustment of the predicted CTRUSER_TO_QUERY_P13Nmatching the user to the queryORGANIC_SCUBA_RETRIEVAL_P13Npersonalized product retrieval
With the confirmation email, Google records the brand in the customer’s profile
- The confirmation emailmarked up by the sender, or read by Google’s extractors
- The booking or order in Moonshinethe account’s personal index: hotel, flight, order, with price, dates, merchant
- The customer’s profilepurchases, brands, merchants, bookings
- The answerAI Mode, Ask Maps, hotel results
An order or a booking leaves an email, and Google files it. Moonshine, the Google index of an account’s personal data, stores the booking or order with its details.
For a hotel, Google knows who took the booking, the hotel itself or an online travel agency, with the total price and the cancellation policy. For an order, it keeps the merchant, the seller, the amount, the products, and whether it is a gift.
Copley, a Google Search component that understands “my hotel” or “my flight” in a query, links those words to the right booking.
For each booking or order, Google notes where its details came from: the email’s schema.org markup, an attached .pkpass file (Apple Wallet), an XML file or one of Google’s own extractors. A merchant that marks up its emails supplies these details itself; without markup, Google rebuilds them.
Purchases read in Gmail feed four kinds of profile data: purchases, affinity for brands and merchants, package tracking and shopping events. Google’s help says so for Search services such as AI Mode, which “might look into past purchases from Workspace apps, like Gmail” (Google Search Help).
Google documents the markup and states: “Gmail, Google Calendar, Google Search, and the Google app all use this structured data” (markup documentation). It covers flights, trains, buses, hotels, restaurants, car rentals, event tickets, invoices, orders and parcels.
Google asks senders to register before sending marked-up emails: emails authenticated with DKIM or SPF, a fixed sender address, a volume on the order of hundreds of emails a day to Gmail for at least a few weeks, very few spam complaints (registration).
In the code: a hotel booking
hotelthe propertycheckin_time · checkout_timethe stay datestotal_pricethe total pricebooking_agentwho took the booking: the hotel or an agencycancellation_policythe cancellation policyemail_referencethe original emailparser_sourcewhere its details came from
In the code: an order
merchant · seller_namethe merchant and the sellertotalthe amountproduct_groupthe productspurchase_location_typeonline or in storeintended_as_giftbought as a giftparser_sourcewhere its details came from
In the code: where the details come from
MICRODATAthe email’s schema.org markupPKPASSan attached .pkpass file (Apple Wallet)XMLan XML fileGENERIC · CRAWZALL · JUICER · STRAINER · KRUPS · PINPOINTGoogle’s own extractors
In the code: purchases read in Gmail
SHOPPING_GMAIL_PURCHASESpurchasesSHOPPING_GMAIL_BRAND_AND_MERCHANT_AFFINITYaffinity for brands and merchantsSHOPPING_GMAIL_PACKAGE_TRACKINGpackage trackingSHOPPING_GMAIL_EVENTSshopping events
In the code: what “my…” points to
copley.PersonalDataTypePERSONAL_HOTEL, PERSONAL_FLIGHT, PERSONAL_TRAIN, PERSONAL_CAR_RENTAL, PERSONAL_RESTAURANT, PERSONAL_BILL…: the personal data “my…” points to
A preferred source moves up in Top Stories and carries a badge into AI Mode
In the code
SearchSourcePreferenceMetadatathe user’s choice, stored in their account settingsP13nRankDomainPreferenceTwiddlerParamsthe domain preference twiddler of personalized rankingPrefTwiddlerParamsthe settings of the Pref twiddlerPrefResultSignalsthe Pref twiddler’s signal on each result
Preferred sources are the personalization a publisher can get for itself. The user picks sites; Google surfaces them more in Top Stories with a “Preferred” badge, and adds that badge to their content in AI Mode and AI Overviews.
The feature is open wherever Top Stories, AI Mode and AI Overviews exist (Search Central documentation), and users had picked more than 600,000 distinct sources by August 2026 (Google’s post).
The number of people using the feature has grown more than sevenfold since the start of 2026, and the news carousels in Search’s AI answers highlight preferred sources too (October 15, 2026 post).
In Google’s code, the choice is stored in the account’s settings, then acts in personalized ranking: a twiddler, a rule that reorders results, moves the domain up. A second twiddler, Pref (for preferences), keeps a preference signal for the query and for each result.
The documentation sets two conditions. To appear in AI Mode and AI Overviews, the site must stay included in Search generative AI features, a Search Console setting. Only a domain or a subdomain is eligible.
Our map of preferred sources by country shows which outlets Google offers users in each country.
In the classic SERP, Google can personalize every step
From query understanding on, KScorer, the Google Search component that picks a query’s interpretation, weighs each meaning with a “Personalization Score”. P13nRank, Google Search’s personalized ranking, steps in at that stage, then in retrieval, the step that picks candidate documents.
At the end of the chain, twiddlers reorder the results for the user: For You, domain preference, affinity for a topic.
According to Google’s help, personalization changes the order of some results or of blocks, with videos on top for someone who often watches them (Google Search Help).
Google says so itself at the bottom of a signed-in user’s SERP: under the pagination, the line “Results are personalized” is followed by a “Try without personalization” link, which shows the page without personalization.
Even signed out, autocomplete takes previous searches into account (Google Search Help); a signed-in account adds history and personalized predictions.
Three cases of personalized SERP are covered above: the Google Flights module, which gives way to the user’s flight, hotel results, reordered by their history, and preferred sources, which move up in Top Stories. The “Try without personalization” link is available in every country (“Where it applies” table).
In the code
P13nRankOutputthe output of personalized ranking, passed along with the queryForYouResultSignalsthe For You twiddler, measured on each resultTopicAffinityTwiddlerResultSignalsthe twiddler of the user’s affinity for a topicP13nBadgingState.ELIGIBLE_FOR_BADGINGa block can show the personalization labelWhyThisResult.Reason.SIGNED_OUT_PQ_BASEDa suggestion drawn from previous queries, signed out
For SEO and AI Search: getting into the customer’s profile
For customers in France, as in the rest of the European Economic Area, the United Kingdom and Switzerland, two kinds of personalization apply today: preferred sources and search history, in the SERP and in AI Mode (“Where it applies” table). Email markup and loyalty pay off in the markets where Gmail connects to AI Mode: nearly 200 countries and territories, the United States among them, outside those three areas.
Mark up confirmation emails
schema.org markup in order, shipping and booking emails gives Google the details it keeps of the order or booking in the customer’s account. With markup, the brand sets what Google keeps: merchant, products, amount, dates, property.
Registering with Google takes DKIM or SPF, a fixed sender address and a steady volume to Gmail.
Treat loyalty as a factor in AI visibility
Each order or booking stays on file under the brand’s name in the customer’s account, and AI Mode reads it once the customer has connected Gmail.
For that customer, a brand already bought and a hotel already booked start with an edge: the first sale sets up the brand’s place in their next answers.
Get picked as a preferred source
Google provides an “Add to preferred sources” button that takes two lines of HTML; a direct link also works in a newsletter or a post. The badge follows the reader into AI Mode and AI Overviews, wherever they exist.