{"id":2755,"date":"2026-08-20T12:42:11","date_gmt":"2026-08-20T10:42:11","guid":{"rendered":"https:\/\/mairateam.com\/?post_type=glossary&p=2755"},"modified":"2026-08-20T15:49:37","modified_gmt":"2026-08-20T13:49:37","slug":"geo-lift-test","status":"publish","type":"glossary","link":"https:\/\/mairateam.com\/cs\/slovnik\/geo-lift-test\/","title":{"rendered":"Geo lift test"},"content":{"rendered":"

Geo lift test<\/strong> je experiment, p\u0159i kter\u00e9m se reklama zapne nebo vypne jen ve vybran\u00fdch regionech a v\u00fdvoj tr\u017eeb se porovn\u00e1 s kontroln\u00edmi regiony, kde se nic nezm\u011bnilo. Rozd\u00edl mezi skupinami je inkrement\u00e1ln\u00ed p\u0159\u00ednos kampan\u011b. Je to nejprakti\u010dt\u011bj\u0161\u00ed zp\u016fsob m\u011b\u0159en\u00ed inkrementality, proto\u017ee nepot\u0159ebuje cookies, souhlasy ani z\u00e1sah do m\u011b\u0159en\u00ed, jen tr\u017eby rozd\u011blen\u00e9 podle region\u016f.<\/p>\n

V kostce<\/h2>\n\n\n\n\n\n\n
Co to je<\/strong><\/td>\nZapnut\u00ed nebo vypnut\u00ed reklamy ve vybran\u00fdch regionech proti kontroln\u00edm<\/td>\n<\/tr>\n
Co pot\u0159ebuje<\/strong><\/td>\nTr\u017eby \u010dleniteln\u00e9 podle region\u016f a mo\u017enost c\u00edlit kampan\u011b podle region\u016f<\/td>\n<\/tr>\n
Co nepot\u0159ebuje<\/strong><\/td>\nCookies, souhlasy, pixely, z\u00e1sahy do webu<\/td>\n<\/tr>\n
Typick\u00e1 d\u00e9lka<\/strong><\/td>\n\u010cty\u0159i a\u017e osm t\u00fddn\u016f, plus obdob\u00ed n\u00e1b\u011bhu a dob\u011bhu<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Jak test prob\u00edh\u00e1<\/h2>\n

F\u00e1ze a jejich d\u00e9lka<\/h3>\n

Test m\u00e1 pevn\u00fd \u010dasov\u00fd rozvrh. Kdo n\u011bkterou f\u00e1zi vynech\u00e1, dostane \u010d\u00edslo, kter\u00e9 nejde obh\u00e1jit.<\/p>\n

    \n
  1. Pretest obdob\u00ed.<\/strong> Alespo\u0148 8 t\u00fddn\u016f t\u00fddenn\u00edch tr\u017eeb po regionech, l\u00e9pe 12 a v\u00edc. Slou\u017e\u00ed k v\u00fdb\u011bru kontroln\u00edch region\u016f a kalibraci modelu, kampan\u011b v n\u011bm b\u011b\u017e\u00ed beze zm\u011bny.<\/li>\n
  2. Minim\u00e1ln\u00ed detekovateln\u00fd efekt.<\/strong> Z rozptylu historick\u00fdch dat se p\u0159ed startem spo\u010d\u00edt\u00e1 nejmen\u0161\u00ed lift, kter\u00fd sestava dok\u00e1\u017ee odli\u0161it od \u0161umu. Kdy\u017e je o\u010dek\u00e1van\u00fd efekt men\u0161\u00ed, zv\u011bt\u0161\u00ed se zm\u011bna rozpo\u010dtu nebo prodlou\u017e\u00ed test.<\/li>\n
  3. N\u00e1b\u011bh.<\/strong> Prvn\u00ed jeden a\u017e dva t\u00fddny po zm\u011bn\u011b se nehodnot\u00ed: kampa\u0148 se u\u010d\u00ed a n\u00e1kupy p\u0159ich\u00e1zej\u00ed se zpo\u017ed\u011bn\u00edm.<\/li>\n
  4. Hodnocen\u00e9 obdob\u00ed.<\/strong> Minim\u00e1ln\u011b 4 t\u00fddny, obvykle 4 a\u017e 8. Krat\u0161\u00ed test nerozli\u0161\u00ed efekt od b\u011b\u017en\u00fdch t\u00fddenn\u00edch v\u00fdkyv\u016f.<\/li>\n
  5. Dob\u011bh.<\/strong> Jeden a\u017e dva t\u00fddny po skon\u010den\u00ed zm\u011bny, aby se zachytily opo\u017ed\u011bn\u00e9 objedn\u00e1vky a n\u00e1vrat region\u016f k norm\u00e1lu.<\/li>\n<\/ol>\n

    Kontrola a v\u00fdpo\u010det liftu<\/h3>\n

    Kontrolou nen\u00ed jeden region, ale v\u00e1\u017een\u00e1 kombinace kontroln\u00edch region\u016f, jej\u00ed\u017e historie kop\u00edruje v\u00fdvoj testovac\u00ed skupiny: syntetick\u00e1 kontrola. V\u00e1hy se najdou v pretest obdob\u00ed, b\u011bhem testu se stejn\u00e9 v\u00e1hy pou\u017eij\u00ed na aktu\u00e1ln\u00ed tr\u017eby kontroln\u00edch region\u016f a v\u00fdsledek je odhad, jak by se testovac\u00ed regiony vyv\u00edjely bez zm\u011bny. Lift = (tr\u017eby testovac\u00edch region\u016f minus odhad) \/ odhad. Inkrement\u00e1ln\u00ed tr\u017eby jsou tento rozd\u00edl v korun\u00e1ch, inkrement\u00e1ln\u00ed ROAS je rozd\u00edl d\u011blen\u00fd \u00fatratou v testovac\u00edch regionech za hodnocen\u00e9 obdob\u00ed.<\/p>\n

    Modelov\u00fd p\u0159\u00edklad<\/h3>\n

    Modelov\u00fd p\u0159\u00edklad: e-shop spou\u0161t\u00ed nov\u00fd kan\u00e1l jen ve t\u0159ech kraj\u00edch. V pretest obdob\u00ed m\u011bly tyto kraje pr\u016fm\u011brn\u011b 500 000 K\u010d tr\u017eeb t\u00fddn\u011b a syntetick\u00e1 kontrola se s nimi shoduje s odchylkou do 2 %. Kampa\u0148 b\u011b\u017e\u00ed 7 t\u00fddn\u016f s \u00fatratou 40 000 K\u010d t\u00fddn\u011b, prvn\u00ed t\u00fdden je n\u00e1b\u011bh, hodnot\u00ed se t\u00fddny 2 a\u017e 7.<\/p>\n\n\n\n\n\n\n\n\n
    Odhad bez kampan\u011b (syntetick\u00e1 kontrola)<\/td>\n520 000 K\u010d t\u00fddn\u011b, kontroln\u00ed regiony sez\u00f3nn\u011b vzrostly o 4 %<\/td>\n<\/tr>\n
    Skute\u010dn\u00e9 tr\u017eby testovac\u00edch kraj\u016f<\/td>\n585 000 K\u010d t\u00fddn\u011b<\/td>\n<\/tr>\n
    Lift<\/td>\n(585 000 minus 520 000) \/ 520 000 = 12,5 %<\/td>\n<\/tr>\n
    Inkrement\u00e1ln\u00ed tr\u017eby za 6 t\u00fddn\u016f<\/td>\n65 000 \u00d7 6 = 390 000 K\u010d<\/td>\n<\/tr>\n
    \u00datrata za 6 hodnocen\u00fdch t\u00fddn\u016f<\/td>\n40 000 \u00d7 6 = 240 000 K\u010d<\/td>\n<\/tr>\n
    Inkrement\u00e1ln\u00ed ROAS<\/td>\n390 000 \/ 240 000 = 1,6<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

    Bez kontroly by srovn\u00e1n\u00ed s pretest obdob\u00edm uk\u00e1zalo r\u016fst 17 % (585 000 proti 500 000) a p\u0159ipsalo by kampani i sez\u00f3nu. Skute\u010dn\u00fd efekt je 12,5 % a s ROAS, kter\u00fd si za stejn\u00e9 regiony a t\u00fddny p\u0159ipsala platforma, se porovn\u00e1v\u00e1 inkrement\u00e1ln\u00ed ROAS 1,6. Zda se kan\u00e1l vyplat\u00ed, rozhodne a\u017e mar\u017ee: inkrement\u00e1ln\u00ed ROAS 1,6 znamen\u00e1 1,60 K\u010d tr\u017eeb za korunu \u00fatraty, ne zisk.<\/p>\n

    Jak se test stav\u00ed<\/h2>\n

    Regiony se rozd\u011bl\u00ed na testovac\u00ed a kontroln\u00ed tak, aby si historicky odpov\u00eddaly v\u00fdvojem tr\u017eeb: p\u00e1ry region\u016f s podobnou sez\u00f3nnost\u00ed a velikost\u00ed. V testovac\u00edch se kampa\u0148 vypne, zapne nebo z\u00e1sadn\u011b nav\u00fd\u0161\u00ed, v kontroln\u00edch b\u011b\u017e\u00ed v\u0161e beze zm\u011bny. Po skon\u010den\u00ed se porovn\u00e1 v\u00fdvoj obou skupin proti spole\u010dn\u00e9 historii. Kontroln\u00ed regiony se obvykle skl\u00e1daj\u00ed do syntetick\u00e9 kontroly, kter\u00e1 kop\u00edruje historick\u00fd v\u00fdvoj testovac\u00edch region\u016f, a lift je odchylka od t\u00e9to p\u0159edpov\u011bdi.<\/p>\n

    Z na\u0161\u00ed praxe<\/h2>\n

    Geo p\u0159\u00edstup je z\u00e1klad na\u0161\u00ed platformy DiagnostIQ, kterou pou\u017e\u00edv\u00e1me na m\u011b\u0159en\u00ed inkrement\u00e1ln\u00edho dopadu kan\u00e1l\u016f. Takto jsme nap\u0159\u00edklad zm\u011b\u0159ili, \u017ee TikTok u e-commerce klienta p\u0159in\u00e1\u0161\u00ed 15 procent tr\u017eeb nav\u00edc, kter\u00e9 by bez n\u011bj nevznikly, p\u0159esto\u017ee atribuce kan\u00e1l podce\u0148ovala. \u010cesk\u00e1 specialita, se kterou po\u010d\u00edt\u00e1me p\u0159i n\u00e1vrhu: men\u0161\u00ed trh znamen\u00e1 m\u00e9n\u011b region\u016f a del\u0161\u00ed testy, aby se efekt odd\u011blil od \u0161umu.<\/p>\n

    \u010cast\u00e9 chyby<\/h2>\n