Whatitmeansforcustomers,marketersandbrands.ByCharlieKayHeadofSEOandDEILeadTOPOSITIVELYIMPACTPEOPLEANDORGANIZATIONS.25+100025+1000CLIENTSEMPLOYEESArtefactisagloballeaderinconsultingservices,specializedindatatransformationanddata&digitalmarketing,fromstrategytothedeploymentofAIsolutions.Weareofferingauniquecombinationofinnovation(Art)anddatascience(Fact).STRATEGY&TRANSFORMATION|AIACCELERATION|DATAFOUNDATIONS&BIIT&DATAPLATFORMS|MARKETINGDATA&DIGITALTABLEOFCONTENTSIt’sallclickablesonavigatetowhereveryouplease!EarlyFoundations(1998–2015):TheMatchingAndRankingEraTheNeuralLeap(2017–2019):TheOneWhereSearchGetsABrainPlatformDiversi?cation(2010–Now):SearchSpreadsBeyondSearchEnginesTheGenerativeShift(2022–2023):SearchStartsToTalkBackTheAI-FirstSERP(2024–2025):SearchBecomesAWorkspaceAIOverviews:ComposedAnswersAtScaleAIMode:ANewSystemOfReasoningAIModevs.AIOverviews:ASide-by-SideComparison04a04aAIsearchTra?cAnalysis:seoClarityResearchStudySearchUsageIn2025:TheDatosxSparkToroReportWhenPeopleAreUsingAIAssistants,WhichAreBeingUsedMost?WhoIsActuallyUsingChatGPT?WhatDoAllTheseStatsMean?TheRetrievalLayer–WhyRankingsStillShapeAIAnswersseoClarityOriginalResearchInsight5.1DeepAudienceResearch&Discovery—It’sAllInTheData5.2AIVisibility&PerformanceMeasurement5.3TailoringYourStrategyToThePlatform–BecauseNotAllAISearchIsBuiltTheSame5.4BrandBuilding&DigitalPR–FuelAwareness,Mentions&AuthorityInAISearch5.5ReputationManagement–ShapingBrandPerceptionInTheAgeOfAI5.6TrustSignals–EarningYourPlaceInTheAIEcosystem5.7SemanticOptimisationandStructuredData-OptimisingStructureForMachines5.8Multi-ModalandMulti-PlatformOptimisation–MakeAllYourContentAccessible5.9UX&ContentArchitecture–DesigningForAIDiscoverabilityAtScale5.10Crawlability&AIAgentAccess–Don’tBlockTheBotsThatMatter5.11FeedbackLoops&ModelMonitoring-Track,Test,LearnAcrossAIEcosystems5.12AgenticInteractionReadiness-OptimiseforDecision-Making(WithoutClicks)There’salotofnoiseabouthowAIisgoingtochangemarketing—foragencies,consumers,andbrands.EvenworkinginSEO,andnavigatingupdatesday-to-day,I?ndtheamountofcontentbeingsharedonthistopicoverwhelming.Thispieceismyattempttodistilthatnoiseintosomethingcohesive.Apieceofresearchtohelppeopleunderstandwhat’sreallyhappening,howwegothere—and,mostimportantly,whattodoaboutit.Idon’tbelievewe’rewitnessingthedeathofSEO.We’reatapointofmajorchangeandenteringanotherchapterintheevolutionofsearch.Theacronymmightchange.Ourfocusmightshift.Wewilllikelyenhancewhatwearedoingandlearnnewskills.Butthefoundationsofwhatwearetryingtohelpbrandsdoandachieveremainintact,withourworkstillrepresentingacorepillarofdigitalmarketing.Ashasalwaysbeenthecase,thebrandsandbusinessesthatwillwinarethosewhohavethebestproducts,dothebestwork,understandtheircustomers,canadapttoachangingworld,aregoodatstorytelling,andprovidegenuinevalue.Sograbacupoftea(ortwo)andsettlein.Thisisn’tashortpiece—butit’slongforareason.There’salottounderstand.AndIhopethisgivesyouaclear,groundedviewofhowwegothere—andwherewegonext.HeadofSEOandDEILead4SearchisaIwaysevoIving—fromhowitIooks,thewayitworks,andthefeaturesmadeavaiIabIe,tothewaypeopIesearchandthepIatformstheyturnto.We’reundoubtedIyatanin?ectionpointbutit’sbeencomingforalongtime.ThemomentweareinispartofamuchIongertimeIineofchange,shapedbynewtechnoIogies,shiftingbehaviour,risingexpectations,andgrowth.SobeforetaIkingabouthowtoadapttothesechanges,Ithoughtit’dbeheIpfuItoIookback.ThissectionexpIoresthebigthingsthathavehappenedinsearchthatbroughtushere—notjustinhowsearchengineshavechanged,butaIsohowsearchbehaviourhasmovedbeyondthem.PayparticuIarattentiontotheitalicisedterms—thesefoundationaIshiftsinsearcharesuperreIevantintheAIworIdwe?ndourseIvesintoday.●●●●1998–GooglesearchislaunchedNotthe?rst,butthemostsuccessfuIofitskind.Abox,abutton,andarankedIistofIinksthatwouIdgoontoshapehowtheworIdaccessesinformation.2001–GoogleaddsmachinelearningtoimprovespellingcorrectionThisisstiIItruetoday.Type“restuarant”andGoogIewiIIsuggest“Didyoumean:restaurant?”—In2001,thisfunctionaIityusedearIymachineIearningmodeIstoIearnfromcommonspeIIingerrorsandimprovetheinterpretationofqueries.ShowingearIysignsofGoogIe’svisiontobestunderstandauserIsqueryandpresenttherightanswers.2003–2009–Majoralgorithmupdates:Florida,Panda,PenguinAtthetime,thesewerebigupdates.TheytargetedIow-quaIitycontent,keywordstu?ng,andsuspectbackIinks-markingashiftfromrewardingmoreeasy-to-manipuIateoptimisationtacticstorelevance-focusedranking.ThisiswhereGoogIebeganprioritisingthequalityofcontentandtheintentoftheusertoprovidethemostreIevantresuIts.2012–GoogleintroducestheKnowledgeGraphTheKnowIedgeGraphisamassivedatabaseofentities-exampIesofentitiesintheknowIedgegraphincIude(Book’,(BookSeries’,’EducationaIOrganization’,(Event’etc.Atthispointsearchshiftedfromstring-matchingtoentityrecognition.ThismeantsearchcouIdbetterunderstandthe(reaI-worId’meaningbehindwords—IikepeopIe,pIaces,andproducts—andrepresentthemvisuaIIyinrichSERP(searchEngineResuItsPage)paneIs.5●●Googlecouldnowunderstandtherealintentbehindaqueryandsurfacecontenttoalignwiththereal-worldconceptorentitytheuserisaskingabout.[TheimagebelowshowsinformationaboutcheesepulledfromGoogle’sKnowledgeGraphinaSERP.]2014–GoogleintroducesE-A-T(Expertise,Authoritativeness,Trustworthiness)IntroducedaspartofGoogle’sSearchQualityEvaluatorGuidelines(guidelinesusedbyrealhumanevaluatorstoratethequalityofsearchresults).E-A-Temphasisestheimportanceofcredible,accurate,andtrustworthycontent—especiallyfortopicslikehealth,?nance,andsafetyknownas“YourMoneyorYourLife”(YMYL).IntheworldofAIsearch,E-A-Tprinciplesarearguablyevenmoreimportant.LLMslikeChatGPT,andfeatureslikeAIOverviewsthatuseLLMsaretrainedtoprioritisecontentthatisfactuallyaccurate,citescrediblesources,anddemonstratesauthortrustsignals—evenifthey’renotspeci?callyusingtheE-A-Tframework.2015–GooglelaunchesRankBrainThiswasGoogle’s?rstsigni?cantuseofmachinelearninginranking.RankBrainhelpedGooglebetterunderstandpreviouslyunseenqueriesbylearningpatternsandrelationshipsbetweenconcepts.Thiswasamajorsteptowardsrealsemanticunderstanding.RankBrainwasthenabletomatchqueriestotheKnowledgeGraphframeworkmoreeffectively,returningevenmoreaccurateandrichresults.●Thesefoundationalshiftslaidthegroundworkforwhatwenowcallsemanticsearch—whereenginesinterpretmeaning,notjustmatchwords.●2017–GoogleinventstheTransformerThe‘bigbrains’atGoogleinventedtheTransfomer-a‘neuralnetworkarchitecture’(asystemwhichde?neshowdata?owsandisprocessed)thatchangedAI.ItbecamethefoundationforLargeLanguageModels(LLMs)—AIthatistrainedonmassivedatasetstounderstandandgeneratehumanlanguage.TheseLLMshavebecomethefoundationofAIsearchtoday(moreonthislater).6ThiseventuaIIypavedthewayfor:·Google’sBERT(BidirectionalEncoderRepresentationsfromTransformers)-thisheIpedsearchunderstandIanguageincontext--sowithinthefuIIcontextofasentence,ratherthaninisoIation.·OpenAl’sGPT-thisenabIedChatGPTtogeneratehuman-Iikeresponsestoprompts.·Google’sGemini-amuItimodaImodeIthatunderstandsnotjusttext,butaIsoimages,video,andcode.●2018-GoogleLensandvisualsearchlaunchWiththeIaunchofGoogIeLensyoucouIdsearchusinganimage.YoucannowpointyourphoneatapIantandask,“Whatisthis?”andyou’IIbeprovidedwithananswerbysearch.OryoucantakeapictureofajumperyouIike,addasearch“thisbutingreen”andresuItswouIdbereturnedforthatexactjumperinadifferentcoIour(tryitifyouhaven’t).●2019-GooglelaunchesBERTGoogIebeganusingLLMstobetterunderstandnaturaIIanguagequeries.ForexampIe,insteadofsimpIymatchingthekeywordsinasearchIike“canIbookatrainformymumfromCanterburytoBromleySouth”,BERTheIpsGoogIeunderstandthat“formymum”isessentiaItotheintent—distinguishingitfromsomeonebookingforthemseIvesorneedingadditionaIaccessibiIityinformation.●2021-GooglelaunchespassageindexingPassageranking(aIsoknownas(passageindexing’)meantthatGoogIewouIdnowrankindividuaIpassagesfromawebpageevenifitwasn’tthemaintopicofapageifitbestanswerstheuserIsquery.Thiswasaranking,notanindexing,changeimpactingapproximateIy7%ofaIIsearchqueries.●2022-E-A-TevolvesintoE-E-A-T(adding(Experience’)GoogIeupdateditsguideIinestoincIude(Experience’aIongsideExpertise,Authoritativeness,andTrustworthiness—re?ectingthegrowingvaIuepIacedon?rst-handperspectives,reviews,anduser-generatedcontent.ThisshiftaIignedcIoseIywiththeriseofReddit,YouTube,andTikTokastrustedcontentformatsintheSERPs.Seenextsectionformoreonthis.yBERTallowedGoogletounderstandsearchqueriesbetterthaneverbefore—itwasabreakthroughthatpavedthewayfortheconversationaIAIweseetoday.“Inthepast,aquerylikethiswouldconfuseoursystems--weplacedtoomuchimportanceontheword“curb”andignoredtheword“no”,notunderstandinghowcriticalthatwordwastoappropriatelyrespondingtothisquery.Sowe’dreturnresultsforparkingonahillwithacurb!”Google-Understandingsearchesbetterthaneverbefore7ParalleltotheevolutionofsearchEngines,userbehaviourwasalsoshifting—partiallydrivenbytheriseofnewsocialplatforms.Althoughnotbuiltastraditionalsearchengines,theyful?lmanyofthesameroles:discovery,informationgathering,anddecision-making.Asuserintenthasexpandedwiththeevolutionoftechnology,sohasthede?nitionofwhatconstitutesa‘search’experience.Sonotonlyhasthismeantthatusers,particularlytheyoungergeneration,haveshiftedplatformsbutalsotraditionalsearchengineshaverespondedbypullingmorecontentfromforums,videoplatforms,andsocialfeedsdirectlyintotheirresults.●●●●2010onwards–PinterestlaunchesPinterestlaunchesasavisual-?rstdiscoveryengine,especiallyforinteriors,fashion,DIY,andrecipes.PeopletendtosearchinthemesandbyintentonPinterest(e.g.“calmnurseryideas”),andcontentisrankedverydifferently--basedonsaves,engagement,andcurationandnotpagerankorbacklinks.MorerecentlyPinteresthavemadeseveralannouncementsabouttheuseofAIontheplatform--introducingvisualsearchandenablingthecreationofbackgroundsbutalsocarefullylabellingcontentcreatedbyGenAI.2010sonwards–YouTubebecomesthego-toforproductresearchandlearningForMillennialsandGenZYouTubebecame(andstillisinmanycases)the?rststopforhow-tos,productdemos,and,evenlifeadvice--especiallyforcomplexorvisualtopics.YouTubeistheworld’ssecondlargestsearchengine,andasofearly2025itboasts2.53billionmonthlyactiveusersandaround122milliondaily-active-users.Itisareallypowerfulwayforbrandstoconnectwithaudiencesandbuildtrustbyshowcasingthepeople,values,andexpertisebehindthebrand.2021onwards–Redditandforumsriseinin?uenceFoundedin2005,Redditisacommunityplatformmadeupofsmallerinterest-basedforumsknownas‘subreddits’.Today,itseesover365.4millionweeklyactive-users.Sincearound2021,searchengineshaveincreasinglysurfacedforumcontent—withRedditappearingmoreandmoreinsearchresults.(Infact,GooglehavealicensingagreementwithReddittocrawltheircontent).Thisshiftre?ectsagrowingrecognitionbyGoogleofthevalueinsubjective,real-worldquestionslike“isthisworthit?”or“what’sitliketoliveinX?”—andamovetowardprioritisinglivedexperienceovermanagedcontent.2020onwards–TikTokbecomesasearchengineforGenZTikTokwaslaunchedin2016,by2020a700miionusershdll,andisnowestimatedtohave1.8billionmontyhlactveiusers.Itisacrucialplaceforbrandstobe,especiallytoconnectwithyoungeraudiencesbecauseit’swhere40%ofthemnowsearch!Whetherit’s“whattodoinLisbon”or“bestcurlingwands”,TikTokisincreasinglyusedtodiscovercontentacrossallintentsandpurposes.TikTokranksbyrecency,engagement,andauthenticity.鳳Notonlyhavetheseplatformsreframedwhatsearchlookslikeandintroducednewmetricsbywhichtoberankedintraditionalsearchbutsearchhasbecomemorevisual,social,andpersonal.Fromthispointon,SEOoptimisationwasnolongerjustaboutrankinginGoogle,butaboutearningtrustandattentionwherebrand’saudiencesactuallyare—forcingashiftincontentstrategyandthecreationofrolesspeci?ctocontentforsocialmedia.8●Nov2022–OpenAIlaunchesChatGPTChatGPTisalargelanguagemodelchatbotdevelopedbyOpenAI.Itusesnaturallanguageprocessingtohavehuman-likeconversationsandgeneratetext.ItwasinitiallytrainedonGPT-3.5,OpenAI’slargelanguagemodelandgenerativepre-trainedtransformer.ChatGPTwasnotconnectedtothewebat?rst,butwasimmediatelypopular,asdemonstratedbyitsgrowth,andclearlydemonstratedthepowerofGenerativeAI.ChatGPTgrewto100monusersntwomontsilliih(that’sfastergrowththanTikToksawwhenitlaunched).DemonstratingthepopularityandrateofadoptionofGenerativeAI.●Feb2023–MicrosoftlaunchesBingChat(laterCopilot)Bingwasthe?rstmajorsearchenginetoembedOpenAI’stechintotheirexperiencebyintroducingaconversationalassistantdirectlywithinitsresults.Thechatwassearch-adjacentandcouldbeusedtoexpandona‘normal’search.ThismovebyBingwasacatalystforthechangingfaceofsearch--forcingGoogle’shandtorapidlyintroducechangestotheirsearchecosystem.●May2023–GooglelaunchesBard(poweredbyLaMDA),andannouncesPaLM2BardwasGoogle’s?rstattemptatastandaloneconversationalAIassistantandpositionedastheiranswertoChatGPT.Initiallyseparatefromsearch,itlaidthefoundationforintegratinggenerativeAIacrossGoogle’secosystem.Inmid-2023,Bardbecameavailablewithinsearchasanopt-inexperience(viasearchLabs),offeringAI-generatedresponsestoqueriesalongsidetraditionalresults.·BardhelpedGoogletesthowusersinteractwithAIinsearch—in?uencingtherolloutofAIOverviewsandGeminiintegration.9AwaveofcontroversyfollowedthelaunchofBardandGoogle’ssearchGenerativeExperience(SGE),raisingearlyconcernsabouttheaccuracyandtransparencyofAI-generatedanswers.OnewidelyreportedincidentinvolvedGoogle’sownpromotionaltweetforBard,inwhichthemodelincorrectycaimetatteJamesWeSpaceTeescopetooterstimagesoapoutsideoursolarsystem—amistaetatetoa$100iionropinApaet’smaretvaTisreayigigteterearissoactuainaccuracyanteneeorresponsidepoymentoAIinsearc.lfhGoogeasoacecriticismorreusinguparagrapsopuisercontentinitsAIwitoutconsistentyproviingcearcitationsorinueingeatearouncontentowners—visiiity,anattriutioninteageoAIsearc●●Dec2023–GoogleannouncesGeminiThesuccessortoLaMDAandPaLM2,GeminiisGoogle’snewfamilyofpowerfulmultimodalLLMs,designedtoreasonacrosstext,code,images,andmore.ItreplacedBard,andmarkedasigni?cantstep-changeinperformanceacrossGoogle’sAIecosystem.Moreimportantly,Geminiiswhatreallypreparedthegroundforthenextevolutionofsearch—poweringtheshiftfromexperimentalAIanswerstosomethingmuchmorenative,integrated,andcapable.Late2023–BingDeepsearchrollsoutBingdescribesDeepsearchasatoolto“provideevenmorerelevantandcomprehensiveanswerstothemostcomplexsearchqueries…anenhancementthatofferstheoptionforadeeperandricherexplorationoftheweb…[It]takesthesearchqueryandexpandsitintoamorecomprehensivedescriptionofwhatanidealsetofresultsshouldinclude.”Inmanycases,forBing,thisisthepointwhereanswersbeginreplacinglinks.●Youcanthinkofallthesethingsassearchbecominginteractive—usersstoppedtypingandscanningandstartedchatting,re?ning,andco-creating.●May2024–GooglelaunchesAIOverviewsPoweredbyGemini,GooglestartedshowingAIgeneratedsummariesabovethestandardsearchresults--pullingdatafrommultiplesources.AIOverviewsincludelinks,citations,andsometimesmulti-stepreasoning.Asofnow,thelaunchofAIOverviewsmarksthesinglebiggestchangeevermadetoGooglesearch--replacingalistoflinkstothird-partycontentwithgeneratedanswers.[TheimageatthetopofthenextpageshowsanexampleofanAIOverviewintheGoogleSERP.Thequeryis‘explainthedifferencebetweenlightningandthunder’.Theansweriscompiledof5paragraphs--anAIGeneratedanswerintheSERPwithnolinks.]10●Late2024–AI-poweredShopping&SearchresultsbecomeincreasinglyshoppableasAIcuratesdynamicproductgridsbasedonuserintent,context,andpastbehaviour.Thesegridsaren’tjustkeyword-triggered—they’reassembledbyinterpretingqueriesholistically(e.g.“eco-friendlytrainersforwinterrunning”),factoringinreal-timeavailability,reviews,specs,andevenindividualpreferences.AInowstartstoactmorelikeapersonalshopper,?lteringandrankingproductsbasedonrelevance,trust,andperformance.·Forretailers,thismeansstructureddata,productfeeds,andhigh-qualityproductcontent(includingUGC--UserGeneratedContent--andreviews)arekey·Brandsthatlackproductschema,missingpricesorratings,orunoptimisedproductpagesriskexclusionfromthesesurfacedresults—eveniftheyrankwellorganically.·Thischangealsoexpandsacrosssurfaces,thevariousplatformsandinterfaceswithinGoogleincludingclassicsearchtoGoogleLens,YouTubeShopping,andAIOverviewsetc.whereproductpanelsaredynamicallygenerated.LatestresearchshowsthatOrganicProductGridsshowfor14%ofallqueriesonmobileand15%ondesktopintheUS.[Theimagetotherightshows8productsinanOrganicProductGridintheSERPfor(bootsforwinterinLondonintheUK,]11●●●May2025–GoogleI/O:Over100AI-relatedannouncements,including:·AIMode–AnAIAssistantbuiltintoGooglesearchasaseparatetabToggIeabIe,interactiveresearchexperience.AIModehasturnedsearchintoadynamicdiaIogue—userscanaskfoIIow-ups,compareinformation,andgetsummarisedinsightsfromacrosstheweb,aIIwithinoneinterface.ThinkofitasChatGPTbuiItintoGoogIe,withaccesstoreaI-timedata,citations,andabIetoassistwitharangeoftasksIikepIanning,coding,brainstormingetc.[TheimagebeIowshowsascreenshotofGoogIe’sAIModereIeasedinMaythisyearintheUS.Thepromptis(WhatisGoogIe’sAIModeandhowdoesitchangesearch?TheanswerbeIowismadeupofaparagraphandabuIIetedIist.TotherightisaseparateboxwithIinksto11differentsitesseIectedoncetheresponsetothequeryisbuiIt].SearchLive–Exploretopicsinreal-timeandacrossmediatypesSearchLiveaIIowsusersexpIoretopicsthroughspokenreaI-timeconversationswithsearch.ItismuItimodaI,andcontextuaI—itdoesn’tjustshowtraditionaItextresuIts,butaIsopuIIsfromnews,videos,sociaIposts,podcasts,forums,andmore.ThismeanssurfacinginsearchnoIongerdependsonrankingapage—yourbrand’spresenceacrossformatsandpIatformsnowmattersmorethanever.Gemini2.5–Google’smostadvanced,fastestLLMyetGemini2.5,reIeasedinmid-2024,ismultimodal,memory-enabled,andcapabIeofdeeperreasoning.ItpowersGoogIeIsWorkspacefeatures,AIOverviews,AIModeandsearchenhancementswithbetterfactualgroundingandfasterretrieval.Asitcontinuestointegrateacrossproducts,Geminireshapeshowinformationisaccessed—notjustviaIinks,butviasynthesisedinsightsdrawnfromtrustedsources.12●●AgenticCheckout–AIcannowguideandcompletetransactions.WithAgenticCheckout,userscannowaskAIassistantstohelpthemresearch,compare,andcompleteapurchase,allwithinaconversationalinterface.Forexample,ChatGPTorGeminicanguideusersfromdiscovery(“Whichhikingbootsarebestforwetterrain?”)topurchase(“Buythetop-ratedwaterproofpairinmysize”).●We’renowintheageofagenticsearch—whereAIdoesn’tjusthelpyousearch,ithelpsyoudecide,plan,andact.AskPhotos–UsenaturallanguagetosearchyourownGooglePhotos.WithAskphotos,Googleletsuserssearchtheirphotolibrarywithquerieslike“Showmeourbeachtripin2021whenAnnawaswearingareddress.”ThisshowcasesLLM-poweredpersonalmemoryrecall—combiningvisualunderstandingwithcontextualprompts.So,asIhopethissectionhasshown,AIhasbeenshapingsearchforyears—butthetype,scaleandspeedofchangewe’reseeingnowisdifferent.It’snotjustaboutnewtechnology;it’sarede?nitionofhow,where,andwhypeoplesearch,andhowsearchoperates.Today’ssearchecosystemisnolongerlimitedtoGoogleorBing.ItspansTikTok,YouTube,Reddit,Pinterest,andafast-growingwaveofAI-poweredassistants.Eachoftheseplatformshasitsowndiscoverypatterns,trustsignals,anduserexpectations—andsuccessnowmeansunderstandinghowtobevisible,credible,andusefulacrossallofthem.13We’veIookedbackathowsearchhaschangedandevoIvedandwhatthathasmeantformarketing,usersandbrands.ArmedwiththatknowIedge一tounderstandhochangingsearchnow一you?rstneedtoundersmodeIs,orLLMs.ThesemodeIsarethebrainsbehindChatGPT,Gemini,CIaude,andeverychatbotorAI-poweredsummaryyouIveseenappearinsearchenginesrecentIy.ALargeLanguageModeI(LLM)isatypeofAItrainedonhugeamountsoftext—books,websites,articIes,forumposts,documentation,andmore.TheyIearnpatternsinIanguagefromthisdocumentationandusethatknowIedgeto:·UnderstandnaturaIIanguagequeries,·Generate?uent,human-Iikeresponses,·Summariseorrephrasecontent,·InferreIationshipsbetweenwords,concepts,andideas.Whenyou’reusingChatGPTit’seasytoforgetthatLLMsdon’tactuaIIy'know'thingsIikeaperson,orIreadIcontentthewayhumansdo.And,queryanindexandreturntheresuItsmostreIevanttoaquery.Instead,theygenerateresponsesbasedonpatternstheyIveIearnedfromtrainingdata.It’simportanttounderstandhowLLMsactuallyworktoknowhowsearchischangingbasedon1.TokenisationBeforetrainingLLMsonIargedatasets,aIIofthetrainingtexthastobebrokendownintosmaIIunitscaIIed‘tokens’—thesecanbewhoIewords,partsofwords,orevenpunctuationmarks.ThesetokensarehowthemodeIseestheinput—IiketransIatingmachine-readabIechunks.“searchischanging”mightbecome→[search][is][chang][##ing]This(tokenisation’heIpsthemodeIprocessIanguagemathematicaIIy.ThemodeIseessequencesoftokensandIearnstopredictwhatcomesnext.Thesetokensarethenmappedtovectors(arraysofnumbers)throughsomethingcaIIedanembeddinglayer.ThisaIIowsthemodeItorepresentthemeaningofwordsnumericaIIy—soitcanIearn,forexampIe,that“puppy”and“dog”aremorecIoseIyreIatedthan“puppy”and“car.”O(jiān)ncethedatasetsaretokenisedthemodeIscanbetrainedonthem.ThecoremodeIistrainedonbillions(sometimestrillions)oftokens!ExampIesofthetypesofdatasetsthatwouIdbetokenisedincIudepubIicIyavaiIabIeorIicensedcontentsuchasbooks,Wikis,newssites,documentation,forumsIikeReddit,andmore.ThesesourcesarereferredtoasthemodeI’strainingdata.ThemodeIIearnsbypredictingthenexttokeninasequence,miIIionsoftimes—essentiaIIycompIetingsentencesatscaIe.ThisprocessheIpsthemodeIIearn:·Howgrammarandsentencestructurework,·WhatfactsarecommonIyrepeated(e.g.“ParisisthecapitaIofFrance”),·HowtoneandstyIevaryacrosscontexts(e.g.newsvs.bIogsvs.conversations),·Howdiscourse?owsIogicaIIy(e.g.howtoargue,Iist,expIain),·WhatconceptsfrequentIyappeartogether(e.g.“booking”and“hoteI”),·WhatquestionspeopIecommonIyask—andhowthey’reanswered.3.Fine-TuningOncethemodeIistrained,deveIoperswiIIoften?ne-tuneittoimprovethewayitrespondstoreaI-worIdprompts.Thisisdoneindifferentways:·ReinforcementLearningfromHumanFeedback(RLHF):WherehumanreviewersratemodeIoutputsforheIpfuInessandsafety.·Instructiontuning:ThemodeIistrainedtofoIIowuserdirectionsbettertoprovideabetterexperience.·Custom?ne-tuning:Sometimes,modeIsaretaiIoredforaspeci?cbrand,tone,orpurpose.WeII-knownmodeIsthatuse?ne-tuning,incIudeChatGPT(GPT-3.5/GPT-4),Claude(Anthropic),Gemini(Google),Perplexity,andBingCopilot(basedonGPT-4Turbo)whichareadaptedfortone,safety,productintegration,orusecases.Fine-tuningimprovessafety,accuracy,andaIignmentwithuserneeds—heIpingturnrawcapabiIityintousefuI,trustworthyinteraction.(AIthough,ashasbeenprovenmanytimes,accuracyissomethingaIIthesemodeIsarestiIIimproving).154.GeneratingaResponse(Inference)ThisiswhathappenswhenyouaskaquestionortypeapromptintoanAIAssistantorLLM-poweredinterfaceIikeChatGPT,CIaude,orGoogIe’snewAIMode.ThemodeIdoesn’tIookupananswerfromadatabaseorperformasearch(unIessretrievaIisenabIed—seenextsection).Instead,ituseseverythingitIearnedduringtrainingtogeneratearesponse,onetoken(i.e.wordfragment)atatime.ThisprocessiscaIIed(inference’.Prompt→“BestfamiIyhoteIsinCornwaII?”Response→ThemodeIdrawsonpatternsit’sseeninsimiIarqueriestogenerateapIausibIe,?uentanswer—notbyretrieving,butbypredictingwhatco